Private Inference Beat

Route assessments

Dated explanations of what is established, what is missing, and which operator powers remain relevant. These are preliminary source/evidence reviews, not completed end-to-end devproof audits.

Assessment obligations · version 3

A route is a named client mode, recipient and model path. Plaintext inside approved attested code does not imply operator access. The question is whether an operator can obtain it or alter the computation without the client refusing.

Each cell names a concrete finding or task. “Our … missing/pending” identifies unfinished verification or evidence in this repository. It is not evidence of a provider defect. Select a cell for what was checked, what remains to prove, and a linked issue with acceptance tests.

Work in our verifier
A required check is absent or incomplete in our implementation, such as AMD report-signature verification or GPU appraisal. The linked task states what an implementation and rejection tests must establish.
Evidence we still need
A route-specific test, build reproduction or current review has not been retained. This is unfinished assessment work, not a finding that the provider lacks the protection.
Finding in the reviewed system
A dated source review or experiment identifies a specific protection boundary, such as unmeasured serving code or a substitutable reply key. Read the exact client/version and evidence scope before applying it to another deployment.
Scope of this route
A standard or retention-only path makes a different claim from client-verified confidential inference. It cannot inherit the guarantees of a separate encrypted path.
Routes differ
The provider has distinct client paths or unresolved obligations. Expand its row or follow the cell; no best-route selection or averaging is performed.

Collection is independent: “returned” includes responses an adapter rejected or could not fully check; “unavailable” means no usable response; “stale / mixed” includes samples more than 48 hours older than the snapshot. Exact observation dates and every raw check are on Measurements.

Assessments remain dated source findings as references change. Provider-wide reference differences from the retained review baseline flag re-review; they do not automatically rewrite the assessment or expand accepted policy. A changed document may be unrelated to a particular route. Follow its history and trace the dependency before assigning impact.

Promotion requires a recorded reviewer, exact client and workload versions, linked proof records and reproducible negative controls. No overall security stages are assigned. Existing internal states are retained without promotion; visible labels describe the actual work or finding. Dates refer to source review, not fresh deployment certification.

Hardware

Is platform evidence authenticated under the relevant hardware and TCB policy?

Authenticate the vendor-root chain, quote signature, TCB/debug policy and any GPU evidence required by the route. Release-signature verification alone is insufficient. Changed signatures and unacceptable TCB must reject.

Recipient key

Is the key the prompt is encrypted to bound to the attested workload and the live channel before the prompt is sent?

Bind the key the prompt is encrypted to, or the TLS key that carries it, to the attested workload and the live channel before the prompt leaves the client. Test a substituted recipient key, a substituted TLS endpoint and a key from an earlier quote.

Reply key + freshness

Is the reply key bound to the same attested workload, and is the evidence fresh for this exchange?

Bind the reply key to the same attested workload, and tie the evidence to this exchange with a nonce or an authenticated session. Test reply-key substitution, replay and stale evidence. A receipt checked after disclosure authenticates the reply but does not protect the prompt.

Code

Does measured workload identity resolve to approved, reproducible source?

Trace source, build, artifact and measurement to the approved workload, including serving code and model identity where claimed. Distinguish reproducibility from signed provenance and matching digests.

Operator access

Are operator-controlled inputs and access paths constrained by the approved computation?

Trace root/debug access, unmeasured configuration, secrets, control-plane inputs and mutable routing to actual effects on prompts and replies. A configuration finding is specific to the reviewed code and deployment.

Downstream

Are required worker hops and the requested model bound to the verified request?

Close necessary gateway-to-worker trust links and requested-model mapping. Test substituted workers, models and route fallback. A direct route can eliminate a gateway hop but still needs its model proof.

Updates

Can the client reject changes to accepted code until its approval policy authorizes them?

Identify who changes accepted values and when the client adopts them. Test an unapproved release, rollback and manifest update. Automatic adoption and independently reviewed pins imply different authority.

ACI gateway

3 paths · 3/3 returned · Reference history

5 provider reference surfaces differ from the review baseline. Determine route relevance before relying on the dated assessment.

Hardware · Our live TCB run to do

TEE RedPill: Our live TCB run to do — Evidence we still need

Evidence we still need

Evidence reviewed
verifiers/aci.py and verifiers/dstack_boot_registers.py (merge a565dee). tests/test_aci_boot_registers.py replays data/observations/evidence/sha256/2d27859e....json with tee_boot_config and tee_hardware_config True, and rejects each tampered register and a genuine quote for an unpinned image; Intel collateral is mocked in that replay. teep 6548996 on tee.redpill.ai 2026-11-15 (data/evidence/teep-2026-11-15/runs/phalacloud_qwen_qwen3.8-27b_20261115T170151Z/report.txt): gateway quote signature, Intel chain, debug bit 0 and 4-RTMR event-log replay pass; aci_session_verified pinned 8 backend sessions as genuine TDX and GPU with allowed MRTD/MRSEAM, workload policy not checked (teep's check, not ours).
Next proof / success condition
Run the merged adapter live and retain the gateway Intel PCS TCB status; appraise backend (inference host) CPU/GPU evidence in our adapter, where backend_evidence_verified is still false.

Tracked work and acceptance tests: #27

Phala hostname: Our live TCB run to do — Evidence we still need

Evidence we still need

Evidence reviewed
verifiers/aci.py and verifiers/dstack_boot_registers.py (merge a565dee). tests/test_aci_boot_registers.py replays data/observations/evidence/sha256/2d27859e....json with tee_boot_config and tee_hardware_config True, and rejects each tampered register and a genuine quote for an unpinned image; Intel collateral is mocked in that replay. The 2d27859e quote was captured via tee.redpill.ai; it applies to this hostname through the shared gateway workload (verifiers/aci.py docstring: hostnames share the quote and measured Compose).
Next proof / success condition
Run the merged adapter live and retain the gateway Intel PCS TCB status; appraise backend (inference host) CPU/GPU evidence in our adapter, where backend_evidence_verified is still false.

Tracked work and acceptance tests: #27

Ordinary RedPill: Our live TCB run to do — Evidence we still need

Evidence we still need

Evidence reviewed
verifiers/aci.py and verifiers/dstack_boot_registers.py (merge a565dee). tests/test_aci_boot_registers.py shows tee_boot_config and tee_hardware_config True on data/observations/evidence/sha256/2d27859e....json, and rejects each tampered register and a genuine quote for an unpinned image; Intel collateral is mocked in that replay. That evidence applies here only through the shared gateway workload; the full replay rejects api.redpill.ai as outside the measured tee_only_domains (Operator and Downstream columns).
Next proof / success condition
Run the merged adapter live and retain the gateway Intel PCS TCB status; appraise backend (inference host) CPU/GPU evidence in our adapter, where backend_evidence_verified is still false.

Tracked work and acceptance tests: #27

Recipient key · Routes differ

TEE RedPill: Gateway key tested by teep only — Evidence we still need

Evidence we still need

Evidence reviewed
teep 6548996, phalacloud qwen/qwen3.8-27b (data/evidence/teep-2026-11-15/runs/phalacloud_qwen_qwen3.8-27b_20261115T170151Z/report.txt): aci_key_custody, tls_key_binding and gateway_tee_reportdata_binding pass; e2ee_usable decrypted 17 fields. teep exited 1 with "capture self-check: reverify report differs from capture" (data/evidence/teep-2026-11-15/SUMMARY.md). For z-ai/glm-5.3-flash teep refused before sending (data/evidence/teep-2026-11-15/runs/phalacloud_z-ai_glm-5.3-flash_20261115T170116Z/report.txt): a session uses channel binding type proxy_image_sha256, which the ACI/1 draft (Dstack-TEE/private-ai-gateway spec/aci.md, extension points; revision not pinned here) lists as a published channel-binding extension and teep 6548996 does not implement. teep@6548996 docs/attestation_gaps/aci_gateway.md residual 1: the inference host is unattested to the client and the client nonce does not reach it.
Next proof / success condition
Run the recipient test in our adapter and reject a substituted recipient key before send; bind the backend that receives the forwarded prompt to an approved workload policy; explain the teep capture self-check failure.

Tracked work and acceptance tests: #30

Phala hostname: Our backend key test to do — Evidence we still need

Evidence we still need

Evidence reviewed
Gateway keyset binding is implemented; no complete retained request test binds the selected backend.
Next proof / success condition
Verify a harmless prompt is encrypted to a key bound to the selected backend and the live channel; reject a substituted recipient key before release. Include keyset continuity.

Tracked work and acceptance tests: #30

Ordinary RedPill: Our backend key test to do — Evidence we still need

Evidence we still need

Evidence reviewed
Gateway keyset binding is implemented; no complete retained request test binds the selected backend.
Next proof / success condition
Verify a harmless prompt is encrypted to a key bound to the selected backend and the live channel; reject a substituted recipient key before release. Include keyset continuity.

Tracked work and acceptance tests: #30

Reply key + freshness · Routes differ

TEE RedPill: Receipt only after stream — Several open obligations

Several open obligations

Evidence reviewed
data/evidence/teep-2026-11-15/runs/phalacloud_qwen_qwen3.8-27b_20261115T170151Z/report.txt and manifest.json: aci_receipt_verified "receipt verified, including the request hash". teep@6548996 docs/attestation_gaps/aci_gateway.md residual 10 (E2EE v2 responses unauthenticated except through the receipt) and 12 (stream delivered before any signed statement; on 2026-11-13 a cut stream's receipt body_hash matched no received prefix). nonce_match fails: freshness rests on the gateway REPORTDATA nonce. Our adapter does not verify receipts. teep exited 1 on this run: "capture self-check: reverify report differs from capture".
Next proof / success condition
Verify receipts in our adapter and reject a response whose receipt is missing or does not match before relying on it. Authenticating chunks before delivery, or a receipt for a cancelled stream, needs a gateway change.

Tracked work and acceptance tests: #30

Phala hostname: Our backend reply test to do — Evidence we still need

Evidence we still need

Evidence reviewed
Gateway nonce binding is implemented; no complete retained request test covers the reply from the selected backend.
Next proof / success condition
Verify a harmless prompt/reply exchange through the selected backend; reject substituted reply keys, stale evidence and replay.

Tracked work and acceptance tests: #30

Ordinary RedPill: Our backend reply test to do — Evidence we still need

Evidence we still need

Evidence reviewed
Gateway nonce binding is implemented; no complete retained request test covers the reply from the selected backend.
Next proof / success condition
Verify a harmless prompt/reply exchange through the selected backend; reject substituted reply keys, stale evidence and replay.

Tracked work and acceptance tests: #30

Code · Our rebuild proof to do

TEE RedPill: Our rebuild proof to do — Evidence we still need

Evidence we still need

Evidence reviewed
Compose preimage and event bindings are checked by the adapter; a complete independently reproduced deployment is not retained.
Next proof / success condition
Trace the deployed measured artifact to reviewed source/build and model identities, including mutable configuration; test an unapproved artifact.

Tracked work and acceptance tests: #30

Phala hostname: Our rebuild proof to do — Evidence we still need

Evidence we still need

Evidence reviewed
Compose preimage and event bindings are checked by the adapter; a complete independently reproduced deployment is not retained.
Next proof / success condition
Trace the deployed measured artifact to reviewed source/build and model identities, including mutable configuration; test an unapproved artifact.

Tracked work and acceptance tests: #30

Ordinary RedPill: Our rebuild proof to do — Evidence we still need

Evidence we still need

Evidence reviewed
Compose preimage and event bindings are checked by the adapter; a complete independently reproduced deployment is not retained.
Next proof / success condition
Trace the deployed measured artifact to reviewed source/build and model identities, including mutable configuration; test an unapproved artifact.

Tracked work and acceptance tests: #30

Operator access · Routes differ

TEE RedPill: Routes outside measurement — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
private-ai-gateway e9b6ae8a; live compose 66e91ac6 and sessions 2026-11-15T17:40Z. research/operator-review-2026-11-15/aci.md
Next proof / success condition
A measured or receipt-cited routing table with code pins; authenticated config secrets.

Tracked work and acceptance tests: #30

Phala hostname: Routes outside measurement — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
private-ai-gateway e9b6ae8a; live compose 66e91ac6 2026-11-15. research/operator-review-2026-11-15/aci.md
Next proof / success condition
A measured or receipt-cited routing table with code pins; authenticated config secrets.

Tracked work and acceptance tests: #30

Ordinary RedPill: Non-TEE backend allowed — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
Gateway compose line 273; handlers.rs:1023 at e9b6ae8a. research/operator-review-2026-11-15/aci.md
Next proof / success condition
TEE-only enforcement on this host, plus a pinned routing table.

Tracked work and acceptance tests: #30

Downstream · Routes differ

TEE RedPill: Sessions pinned; workload unchecked — Several open obligations

Several open obligations

Evidence reviewed
data/evidence/teep-2026-11-15/runs/phalacloud_qwen_qwen3.8-27b_20261115T170151Z/report.txt: aci_session_verified "8 attested sessions pinned: genuine TDX and GPU with allowed MRTD/MRSEAM, workload policy not checked"; aci_backend_named fails ("the attestation binds no backend for the requested model"); compose_binding fails ("ACI/1 publishes no compose manifest for the inference host"); measured_model_weights fails. teep exited 1 on this run: "capture self-check: reverify report differs from capture". teep@6548996 docs/attestation_gaps/aci_gateway.md residual 1.
Next proof / success condition
Apply an approved workload policy to the backend model CVM and bind the requested model to the request in our adapter; test worker and model substitution.

Tracked work and acceptance tests: #30

Phala hostname: Our worker test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Ordinary RedPill: Non-TEE backend allowed — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
This hostname does not enforce TEE-only serving. No request-bound backend proof closes that gap.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback. Use a route that enforces attested serving.

Tracked work and acceptance tests: #30

Updates · Approval not established

TEE RedPill: Approval not established — Several open obligations

Several open obligations

Evidence reviewed
Compose/keyset identities are captured. The reviewed deployed profile lacks a continuity endorsement and independently established client approval.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback. Authenticate workload/keyset continuity.

Tracked work and acceptance tests: #30

Phala hostname: Approval not established — Several open obligations

Several open obligations

Evidence reviewed
Compose/keyset identities are captured. The reviewed deployed profile lacks a continuity endorsement and independently established client approval.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback. Authenticate workload/keyset continuity.

Tracked work and acceptance tests: #30

Ordinary RedPill: Approval not established — Several open obligations

Several open obligations

Evidence reviewed
Compose/keyset identities are captured. The reviewed deployed profile lacks a continuity endorsement and independently established client approval.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback. Authenticate workload/keyset continuity.

Tracked work and acceptance tests: #30

TEE RedPill Intel TDX

Reviewed 2026-11-15 by claude-code dated re-review 2026-11-15 of the cells whose text cites 2026-11-15 evidence; other cells carry the 2026-10-14 codex-source-review (neither is independent audit approval) · source revision 1eca00d00f17 · Route inventory and raw checks

Our live TCB run to do Hardware

Evidence we still need

Evidence reviewed
verifiers/aci.py and verifiers/dstack_boot_registers.py (merge a565dee). tests/test_aci_boot_registers.py replays data/observations/evidence/sha256/2d27859e....json with tee_boot_config and tee_hardware_config True, and rejects each tampered register and a genuine quote for an unpinned image; Intel collateral is mocked in that replay. teep 6548996 on tee.redpill.ai 2026-11-15 (data/evidence/teep-2026-11-15/runs/phalacloud_qwen_qwen3.8-27b_20261115T170151Z/report.txt): gateway quote signature, Intel chain, debug bit 0 and 4-RTMR event-log replay pass; aci_session_verified pinned 8 backend sessions as genuine TDX and GPU with allowed MRTD/MRSEAM, workload policy not checked (teep's check, not ours).
Next proof / success condition
Run the merged adapter live and retain the gateway Intel PCS TCB status; appraise backend (inference host) CPU/GPU evidence in our adapter, where backend_evidence_verified is still false.

Tracked work and acceptance tests: #27

Evidence scope: dated repository source/review; no new live route proof is implied.

data/reviews/decisions/chg-b69f2fb6d3774d63dbf5612b560ea77d.jsonverifiers/aci.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdtests/test_aci_boot_registers.pyresearch/operator-review-2026-11-15/aci.md

Gateway key tested by teep only Recipient key

Evidence we still need

Evidence reviewed
teep 6548996, phalacloud qwen/qwen3.8-27b (data/evidence/teep-2026-11-15/runs/phalacloud_qwen_qwen3.8-27b_20261115T170151Z/report.txt): aci_key_custody, tls_key_binding and gateway_tee_reportdata_binding pass; e2ee_usable decrypted 17 fields. teep exited 1 with "capture self-check: reverify report differs from capture" (data/evidence/teep-2026-11-15/SUMMARY.md). For z-ai/glm-5.3-flash teep refused before sending (data/evidence/teep-2026-11-15/runs/phalacloud_z-ai_glm-5.3-flash_20261115T170116Z/report.txt): a session uses channel binding type proxy_image_sha256, which the ACI/1 draft (Dstack-TEE/private-ai-gateway spec/aci.md, extension points; revision not pinned here) lists as a published channel-binding extension and teep 6548996 does not implement. teep@6548996 docs/attestation_gaps/aci_gateway.md residual 1: the inference host is unattested to the client and the client nonce does not reach it.
Next proof / success condition
Run the recipient test in our adapter and reject a substituted recipient key before send; bind the backend that receives the forwarded prompt to an approved workload policy; explain the teep capture self-check failure.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

data/reviews/decisions/chg-b69f2fb6d3774d63dbf5612b560ea77d.jsonverifiers/aci.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdtests/test_aci_boot_registers.pyresearch/operator-review-2026-11-15/aci.md

Receipt only after stream Reply key + freshness

Several open obligations

Evidence reviewed
data/evidence/teep-2026-11-15/runs/phalacloud_qwen_qwen3.8-27b_20261115T170151Z/report.txt and manifest.json: aci_receipt_verified "receipt verified, including the request hash". teep@6548996 docs/attestation_gaps/aci_gateway.md residual 10 (E2EE v2 responses unauthenticated except through the receipt) and 12 (stream delivered before any signed statement; on 2026-11-13 a cut stream's receipt body_hash matched no received prefix). nonce_match fails: freshness rests on the gateway REPORTDATA nonce. Our adapter does not verify receipts. teep exited 1 on this run: "capture self-check: reverify report differs from capture".
Next proof / success condition
Verify receipts in our adapter and reject a response whose receipt is missing or does not match before relying on it. Authenticating chunks before delivery, or a receipt for a cancelled stream, needs a gateway change.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

data/reviews/decisions/chg-b69f2fb6d3774d63dbf5612b560ea77d.jsonverifiers/aci.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdtests/test_aci_boot_registers.pyresearch/operator-review-2026-11-15/aci.md

Our rebuild proof to do Code

Evidence we still need

Evidence reviewed
Compose preimage and event bindings are checked by the adapter; a complete independently reproduced deployment is not retained.
Next proof / success condition
Trace the deployed measured artifact to reviewed source/build and model identities, including mutable configuration; test an unapproved artifact.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

data/reviews/decisions/chg-b69f2fb6d3774d63dbf5612b560ea77d.jsonverifiers/aci.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdtests/test_aci_boot_registers.pyresearch/operator-review-2026-11-15/aci.md

Routes outside measurement Operator access

Finding in the reviewed system

Evidence reviewed
private-ai-gateway e9b6ae8a; live compose 66e91ac6 and sessions 2026-11-15T17:40Z. research/operator-review-2026-11-15/aci.md
Next proof / success condition
A measured or receipt-cited routing table with code pins; authenticated config secrets.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

data/reviews/decisions/chg-b69f2fb6d3774d63dbf5612b560ea77d.jsonverifiers/aci.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdtests/test_aci_boot_registers.pyresearch/operator-review-2026-11-15/aci.md

Sessions pinned; workload unchecked Downstream

Several open obligations

Evidence reviewed
data/evidence/teep-2026-11-15/runs/phalacloud_qwen_qwen3.8-27b_20261115T170151Z/report.txt: aci_session_verified "8 attested sessions pinned: genuine TDX and GPU with allowed MRTD/MRSEAM, workload policy not checked"; aci_backend_named fails ("the attestation binds no backend for the requested model"); compose_binding fails ("ACI/1 publishes no compose manifest for the inference host"); measured_model_weights fails. teep exited 1 on this run: "capture self-check: reverify report differs from capture". teep@6548996 docs/attestation_gaps/aci_gateway.md residual 1.
Next proof / success condition
Apply an approved workload policy to the backend model CVM and bind the requested model to the request in our adapter; test worker and model substitution.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

data/reviews/decisions/chg-b69f2fb6d3774d63dbf5612b560ea77d.jsonverifiers/aci.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdtests/test_aci_boot_registers.pyresearch/operator-review-2026-11-15/aci.md

Approval not established Updates

Several open obligations

Evidence reviewed
Compose/keyset identities are captured. The reviewed deployed profile lacks a continuity endorsement and independently established client approval.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback. Authenticate workload/keyset continuity.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

data/reviews/decisions/chg-b69f2fb6d3774d63dbf5612b560ea77d.jsonverifiers/aci.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdtests/test_aci_boot_registers.pyresearch/operator-review-2026-11-15/aci.md

Phala hostname Intel TDX

Reviewed 2026-11-15 by claude-code dated re-review 2026-11-15 of the cells whose text cites 2026-11-15 evidence; other cells carry the 2026-10-14 codex-source-review (neither is independent audit approval) · source revision 1eca00d00f17 · Route inventory and raw checks

Our live TCB run to do Hardware

Evidence we still need

Evidence reviewed
verifiers/aci.py and verifiers/dstack_boot_registers.py (merge a565dee). tests/test_aci_boot_registers.py replays data/observations/evidence/sha256/2d27859e....json with tee_boot_config and tee_hardware_config True, and rejects each tampered register and a genuine quote for an unpinned image; Intel collateral is mocked in that replay. The 2d27859e quote was captured via tee.redpill.ai; it applies to this hostname through the shared gateway workload (verifiers/aci.py docstring: hostnames share the quote and measured Compose).
Next proof / success condition
Run the merged adapter live and retain the gateway Intel PCS TCB status; appraise backend (inference host) CPU/GPU evidence in our adapter, where backend_evidence_verified is still false.

Tracked work and acceptance tests: #27

Evidence scope: dated repository source/review; no new live route proof is implied.

data/reviews/decisions/chg-b69f2fb6d3774d63dbf5612b560ea77d.jsonverifiers/aci.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdtests/test_aci_boot_registers.pyresearch/operator-review-2026-11-15/aci.md

Our backend key test to do Recipient key

Evidence we still need

Evidence reviewed
Gateway keyset binding is implemented; no complete retained request test binds the selected backend.
Next proof / success condition
Verify a harmless prompt is encrypted to a key bound to the selected backend and the live channel; reject a substituted recipient key before release. Include keyset continuity.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

data/reviews/decisions/chg-b69f2fb6d3774d63dbf5612b560ea77d.jsonverifiers/aci.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdtests/test_aci_boot_registers.pyresearch/operator-review-2026-11-15/aci.md

Our backend reply test to do Reply key + freshness

Evidence we still need

Evidence reviewed
Gateway nonce binding is implemented; no complete retained request test covers the reply from the selected backend.
Next proof / success condition
Verify a harmless prompt/reply exchange through the selected backend; reject substituted reply keys, stale evidence and replay.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

data/reviews/decisions/chg-b69f2fb6d3774d63dbf5612b560ea77d.jsonverifiers/aci.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdtests/test_aci_boot_registers.pyresearch/operator-review-2026-11-15/aci.md

Our rebuild proof to do Code

Evidence we still need

Evidence reviewed
Compose preimage and event bindings are checked by the adapter; a complete independently reproduced deployment is not retained.
Next proof / success condition
Trace the deployed measured artifact to reviewed source/build and model identities, including mutable configuration; test an unapproved artifact.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

data/reviews/decisions/chg-b69f2fb6d3774d63dbf5612b560ea77d.jsonverifiers/aci.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdtests/test_aci_boot_registers.pyresearch/operator-review-2026-11-15/aci.md

Routes outside measurement Operator access

Finding in the reviewed system

Evidence reviewed
private-ai-gateway e9b6ae8a; live compose 66e91ac6 2026-11-15. research/operator-review-2026-11-15/aci.md
Next proof / success condition
A measured or receipt-cited routing table with code pins; authenticated config secrets.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

data/reviews/decisions/chg-b69f2fb6d3774d63dbf5612b560ea77d.jsonverifiers/aci.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdtests/test_aci_boot_registers.pyresearch/operator-review-2026-11-15/aci.md

Our worker test to do Downstream

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

data/reviews/decisions/chg-b69f2fb6d3774d63dbf5612b560ea77d.jsonverifiers/aci.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdtests/test_aci_boot_registers.pyresearch/operator-review-2026-11-15/aci.md

Approval not established Updates

Several open obligations

Evidence reviewed
Compose/keyset identities are captured. The reviewed deployed profile lacks a continuity endorsement and independently established client approval.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback. Authenticate workload/keyset continuity.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

data/reviews/decisions/chg-b69f2fb6d3774d63dbf5612b560ea77d.jsonverifiers/aci.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdtests/test_aci_boot_registers.pyresearch/operator-review-2026-11-15/aci.md

Ordinary RedPill Intel TDX

Reviewed 2026-11-15 by claude-code dated re-review 2026-11-15 of the cells whose text cites 2026-11-15 evidence; other cells carry the 2026-10-14 codex-source-review (neither is independent audit approval) · source revision 1eca00d00f17 · Route inventory and raw checks

Our live TCB run to do Hardware

Evidence we still need

Evidence reviewed
verifiers/aci.py and verifiers/dstack_boot_registers.py (merge a565dee). tests/test_aci_boot_registers.py shows tee_boot_config and tee_hardware_config True on data/observations/evidence/sha256/2d27859e....json, and rejects each tampered register and a genuine quote for an unpinned image; Intel collateral is mocked in that replay. That evidence applies here only through the shared gateway workload; the full replay rejects api.redpill.ai as outside the measured tee_only_domains (Operator and Downstream columns).
Next proof / success condition
Run the merged adapter live and retain the gateway Intel PCS TCB status; appraise backend (inference host) CPU/GPU evidence in our adapter, where backend_evidence_verified is still false.

Tracked work and acceptance tests: #27

Evidence scope: dated repository source/review; no new live route proof is implied.

data/reviews/decisions/chg-b69f2fb6d3774d63dbf5612b560ea77d.jsonverifiers/aci.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdtests/test_aci_boot_registers.pyresearch/operator-review-2026-11-15/aci.md

Our backend key test to do Recipient key

Evidence we still need

Evidence reviewed
Gateway keyset binding is implemented; no complete retained request test binds the selected backend.
Next proof / success condition
Verify a harmless prompt is encrypted to a key bound to the selected backend and the live channel; reject a substituted recipient key before release. Include keyset continuity.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

data/reviews/decisions/chg-b69f2fb6d3774d63dbf5612b560ea77d.jsonverifiers/aci.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdtests/test_aci_boot_registers.pyresearch/operator-review-2026-11-15/aci.md

Our backend reply test to do Reply key + freshness

Evidence we still need

Evidence reviewed
Gateway nonce binding is implemented; no complete retained request test covers the reply from the selected backend.
Next proof / success condition
Verify a harmless prompt/reply exchange through the selected backend; reject substituted reply keys, stale evidence and replay.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

data/reviews/decisions/chg-b69f2fb6d3774d63dbf5612b560ea77d.jsonverifiers/aci.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdtests/test_aci_boot_registers.pyresearch/operator-review-2026-11-15/aci.md

Our rebuild proof to do Code

Evidence we still need

Evidence reviewed
Compose preimage and event bindings are checked by the adapter; a complete independently reproduced deployment is not retained.
Next proof / success condition
Trace the deployed measured artifact to reviewed source/build and model identities, including mutable configuration; test an unapproved artifact.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

data/reviews/decisions/chg-b69f2fb6d3774d63dbf5612b560ea77d.jsonverifiers/aci.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdtests/test_aci_boot_registers.pyresearch/operator-review-2026-11-15/aci.md

Non-TEE backend allowed Operator access

Finding in the reviewed system

Evidence reviewed
Gateway compose line 273; handlers.rs:1023 at e9b6ae8a. research/operator-review-2026-11-15/aci.md
Next proof / success condition
TEE-only enforcement on this host, plus a pinned routing table.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

data/reviews/decisions/chg-b69f2fb6d3774d63dbf5612b560ea77d.jsonverifiers/aci.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdtests/test_aci_boot_registers.pyresearch/operator-review-2026-11-15/aci.md

Non-TEE backend allowed Downstream

Finding in the reviewed system

Evidence reviewed
This hostname does not enforce TEE-only serving. No request-bound backend proof closes that gap.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback. Use a route that enforces attested serving.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

data/reviews/decisions/chg-b69f2fb6d3774d63dbf5612b560ea77d.jsonverifiers/aci.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdtests/test_aci_boot_registers.pyresearch/operator-review-2026-11-15/aci.md

Approval not established Updates

Several open obligations

Evidence reviewed
Compose/keyset identities are captured. The reviewed deployed profile lacks a continuity endorsement and independently established client approval.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback. Authenticate workload/keyset continuity.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

data/reviews/decisions/chg-b69f2fb6d3774d63dbf5612b560ea77d.jsonverifiers/aci.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdtests/test_aci_boot_registers.pyresearch/operator-review-2026-11-15/aci.md

NEAR AI

3 paths · Stale / mixed · Reference history

6 provider reference surfaces differ from the review baseline. Determine route relevance before relying on the dated assessment.

Hardware · Routes differ

Gateway HTTPS: RTMR0 not derivable — Several open obligations

Several open obligations

Evidence reviewed
data/evidence/near-2026-11-15/glm-5.3-flash-verify.json: tcb_current, tee_debug_disabled, event_log_replayed, tee_boot_config and gpu_attested true; tee_hardware_config null; tcb_statuses gateway and 3 models UpToDate. Per-instance checks: commit ef56f97, fixture tests/fixtures/near_glm53_three_instances.json. teep 6548996 nearcloud runs agree (data/evidence/teep-2026-11-15/runs/nearcloud_*/report.txt).
Next proof / success condition
NEAR to publish the VM configuration (or we derive RTMR0 another way) so RTMR0 can be pinned; verifiers/near_ai.py sets tee_hardware_config to None.

Tracked work and acceptance tests: #29

E2EE through gateway: RTMR0 not derivable — Several open obligations

Several open obligations

Evidence reviewed
data/evidence/near-2026-11-15/glm-5.3-flash-verify.json: tcb_current, tee_debug_disabled, event_log_replayed, tee_boot_config and gpu_attested true; tee_hardware_config null; tcb_statuses gateway and 3 models UpToDate. Per-instance checks: commit ef56f97, fixture tests/fixtures/near_glm53_three_instances.json. teep 6548996 nearcloud runs agree (data/evidence/teep-2026-11-15/runs/nearcloud_*/report.txt).
Next proof / success condition
NEAR to publish the VM configuration (or we derive RTMR0 another way) so RTMR0 can be pinned; verifiers/near_ai.py sets tee_hardware_config to None.

Tracked work and acceptance tests: #29

Direct model: Third-party check only — Evidence we still need

Evidence we still need

Evidence reviewed
data/evidence/teep-2026-11-15/runs/neardirect_z-ai_glm-5.3-flash_20261115T165919Z/report.txt, data/evidence/teep-2026-11-15/runs/neardirect_Qwen_Qwen3.8-27B_20261115T165927Z/report.txt; tee_boot_config is an allowed failure in teep.
Next proof / success condition
Run our verifier on the direct endpoints and retain the result, including boot and hardware register pins.

Tracked work and acceptance tests: #29

Recipient key · Routes differ

Gateway HTTPS: Our request test to do — Evidence we still need

Evidence we still need

Evidence reviewed
Some quote/key checks exist, but the actual TLS/request path and selected backend are not fully bound. teep 6548996 on 2026-11-15 found the gateway TLS SPKI bound to its attestation (tls_key_binding pass, data/evidence/teep-2026-11-15/runs/nearcloud_*/report.txt); that run used E2EE, not this plain path.
Next proof / success condition
Bind the actual TLS/request path and selected backend to the attested key; reject a substituted recipient key before the prompt is sent.

Tracked work and acceptance tests: #29

E2EE through gateway: Third-party E2EE test only — Evidence we still need

Evidence we still need

Evidence reviewed
teep 6548996 nearcloud z-ai/glm-5.3-flash (Instances: 3) and Qwen/Qwen3.8-27B: nonce_match, tls_key_binding, e2ee_capable, e2ee_usable pass (data/evidence/teep-2026-11-15/runs/nearcloud_*/report.txt). On 2026-11-15 NEAR listed z-ai/glm-5.3-flash on three instances (19aea169..., a8a25faa..., cc796618...) with three different compose hashes and one signing key 0x3b753b44... (commit ef56f97, fixture tests/fixtures/near_glm53_three_instances.json). teep 74834db: on /v1/completions only prompt is encrypted; stop and user pass in plaintext.
Next proof / success condition
Run a harmless encrypted request in our verifier and reject a substituted recipient key before send; record which request fields stay plaintext per endpoint.

Tracked work and acceptance tests: #29

Direct model: Third-party E2EE test only — Evidence we still need

Evidence we still need

Evidence reviewed
data/evidence/teep-2026-11-15/runs/neardirect_*/report.txt: nonce_match, tls_key_binding, e2ee_capable, e2ee_usable pass for z-ai/glm-5.3-flash and Qwen/Qwen3.8-27B.
Next proof / success condition
Run a harmless encrypted request in our verifier and reject a substituted recipient key before send; record which request fields stay plaintext per endpoint.

Tracked work and acceptance tests: #29

Reply key + freshness · Routes differ

Gateway HTTPS: Our reply test to do — Evidence we still need

Evidence we still need

Evidence reviewed
Some nonce checks exist, but no actual prompt/reply exchange through the selected backend was verified.
Next proof / success condition
Verify a harmless prompt/reply exchange through the selected backend; reject substituted reply keys, stale evidence and replay.

Tracked work and acceptance tests: #29

E2EE through gateway: v3 fix in source only — Evidence we still need

Evidence we still need

Evidence reviewed
Read 2026-11-15 at nearai/inference-proxy 3f7e979, src/encryption.rs:62-75: request_aad() appends canonical_reply_key() under the label nearai-e2ee-request-v3, so a relay that swaps the reply key breaks decryption of the request. Issue 1061 is still open with no comments. The deployed proxy digest 60d08c20… matches no release digest (research/operator-review-2026-11-15/near.md). teep decrypted reply fields on nearcloud through an honest gateway (data/evidence/teep-2026-11-15/runs/nearcloud_*/report.txt), which does not exercise substitution.
Next proof / success condition
Tie the deployed proxy to a commit containing #321, run a v3 exchange with a substituted reply key and observe the rejection, and check whether a v3 client can be downgraded to v1/v2.

Tracked work and acceptance tests: #29

Direct model: Our reply test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that the reply key is bound to the same attested workload; reject a substituted reply key, stale evidence and replay before the reply is disclosed.

Tracked work and acceptance tests: #29

Code · Routes differ

Gateway HTTPS: Our anchor lacks glm-5.3-flash — Work in our verifier

Work in our verifier

Evidence reviewed
data/evidence/near-2026-11-15/glm-5.3-flash-verify.json: model_composes for three instances, compose hashes c900dc73..., e8416a02..., a7de0bba..., each quote_compose_matches and chain_authorized true, anchored false; verification_status rejected. gateway_reference_audited is also false (no audit ledger match for the gateway image); it does not gate the row. Same observation in commit ef56f97. teep sigstore_verification fails for image digest sha256:079a14b41deb11e0 (HTTP 404) on nearcloud and neardirect (data/evidence/teep-2026-11-15/SUMMARY.md).
Next proof / success condition
Review and anchor each compose that serves a model (three distinct composes serve glm-5.3-flash); authenticate compose-manager actions through the quote/event chain; map the requested model to approved reproducible artifacts.

Tracked work and acceptance tests: #29

E2EE through gateway: Our anchor lacks glm-5.3-flash — Work in our verifier

Work in our verifier

Evidence reviewed
data/evidence/near-2026-11-15/glm-5.3-flash-verify.json: model_composes for three instances, compose hashes c900dc73..., e8416a02..., a7de0bba..., each quote_compose_matches and chain_authorized true, anchored false; verification_status rejected. gateway_reference_audited is also false (no audit ledger match for the gateway image); it does not gate the row. Same observation in commit ef56f97. teep sigstore_verification fails for image digest sha256:079a14b41deb11e0 (HTTP 404) on nearcloud and neardirect (data/evidence/teep-2026-11-15/SUMMARY.md).
Next proof / success condition
Review and anchor each compose that serves a model (three distinct composes serve glm-5.3-flash); authenticate compose-manager actions through the quote/event chain; map the requested model to approved reproducible artifacts.

Tracked work and acceptance tests: #29

Direct model: Our direct-path test — Evidence we still need

Evidence we still need

Evidence reviewed
The experimental direct path needs its own client, workload and accepted-policy evidence. Gateway adapter findings cannot establish this route.
Next proof / success condition
Trace the deployed measured artifact to reviewed source/build and model identities, including mutable configuration; test an unapproved artifact.

Tracked work and acceptance tests: #29

Operator access · Routes differ

Gateway HTTPS: Operator picks the backend — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
cloud-api b26eb029 verification.rs:282,326, measurement.rs:93-110, model.rs:1626-1645; gateway compose 2643bc7a (2026-11-14). research/operator-review-2026-11-15/near.md F1
Next proof / success condition
SSH in the non-dev OS images; gateway log and telemetry egress; a live substitution test.

Tracked work and acceptance tests: #29

E2EE through gateway: App-wide key, unlogged env — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
inference-proxy signing.rs:149-167; compose-manager main.rs:404-425,638-645; v0.0.585 action log; Base owner() reads 2026-11-15. research/operator-review-2026-11-15/near.md F2-F4
Next proof / success condition
The commit behind proxy 60d08c20; fusion and redis code; DstackKms owner powers; the full authorized-compose set.

Tracked work and acceptance tests: #29

Direct model: App-wide key, unlogged env — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
teep neardirect capture 2026-11-14; inference-proxy signing.rs:149-167; compose-manager main.rs:404-425. research/operator-review-2026-11-15/near.md
Next proof / success condition
An encrypted request through our client; other models' direct hosts.

Tracked work and acceptance tests: #29

Downstream · Routes differ

Gateway HTTPS: Gateway picks among 3 composes — Several open obligations

Several open obligations

Evidence reviewed
On 2026-11-15 NEAR listed z-ai/glm-5.3-flash on three instances (19aea169..., a8a25faa..., cc796618...) with three different compose hashes and one signing key 0x3b753b44... (commit ef56f97, fixture tests/fixtures/near_glm53_three_instances.json). teep 4ecddb6 accepts an attested model name that differs from the requested one; our verifier records model_names_match_request, true on 2026-11-15 (data/evidence/near-2026-11-15/glm-5.3-flash-verify.json).
Next proof / success condition
Bind the instance and model that served each request to that request; reject instance or model substitution and insecure fallback.

Tracked work and acceptance tests: #29

E2EE through gateway: Gateway picks among 3 composes — Several open obligations

Several open obligations

Evidence reviewed
On 2026-11-15 NEAR listed z-ai/glm-5.3-flash on three instances (19aea169..., a8a25faa..., cc796618...) with three different compose hashes and one signing key 0x3b753b44... (commit ef56f97, fixture tests/fixtures/near_glm53_three_instances.json). teep 4ecddb6 accepts an attested model name that differs from the requested one; our verifier records model_names_match_request, true on 2026-11-15 (data/evidence/near-2026-11-15/glm-5.3-flash-verify.json).
Next proof / success condition
Bind the instance and model that served each request to that request; reject instance or model substitution and insecure fallback.

Tracked work and acceptance tests: #29

Direct model: Our worker test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #29

Updates · Routes differ

Gateway HTTPS: Our refusal test to do — Evidence we still need

Evidence we still need

Evidence reviewed
Repository per-model pins are distinct from on-chain provider authorization; the sending client policy was not validated end to end. On 2026-11-15 NEAR listed z-ai/glm-5.3-flash on three instances (19aea169..., a8a25faa..., cc796618...) with three different compose hashes and one signing key 0x3b753b44... (commit ef56f97, fixture tests/fixtures/near_glm53_three_instances.json). Each compose was chain-authorized, so a client pinned to the key cannot tell which of the three composes served a request.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback.

Tracked work and acceptance tests: #29

E2EE through gateway: Our refusal test to do — Evidence we still need

Evidence we still need

Evidence reviewed
Repository per-model pins are distinct from on-chain provider authorization; the sending client policy was not validated end to end. On 2026-11-15 NEAR listed z-ai/glm-5.3-flash on three instances (19aea169..., a8a25faa..., cc796618...) with three different compose hashes and one signing key 0x3b753b44... (commit ef56f97, fixture tests/fixtures/near_glm53_three_instances.json). Each compose was chain-authorized, so a client pinned to the key cannot tell which of the three composes served a request.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback.

Tracked work and acceptance tests: #29

Direct model: Our direct-path test — Evidence we still need

Evidence we still need

Evidence reviewed
The experimental direct path needs its own client, workload and accepted-policy evidence. Gateway adapter findings cannot establish this route.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback.

Tracked work and acceptance tests: #29

Gateway HTTPS Intel TDX + GPU

Reviewed 2026-11-15 by claude-code dated re-review 2026-11-15 of the cells whose text cites 2026-11-15 evidence; other cells carry the 2026-10-14 codex-source-review (neither is independent audit approval) · source revision 1eca00d00f17 · Route inventory and raw checks

RTMR0 not derivable Hardware

Several open obligations

Evidence reviewed
data/evidence/near-2026-11-15/glm-5.3-flash-verify.json: tcb_current, tee_debug_disabled, event_log_replayed, tee_boot_config and gpu_attested true; tee_hardware_config null; tcb_statuses gateway and 3 models UpToDate. Per-instance checks: commit ef56f97, fixture tests/fixtures/near_glm53_three_instances.json. teep 6548996 nearcloud runs agree (data/evidence/teep-2026-11-15/runs/nearcloud_*/report.txt).
Next proof / success condition
NEAR to publish the VM configuration (or we derive RTMR0 another way) so RTMR0 can be pinned; verifiers/near_ai.py sets tee_hardware_config to None.

Tracked work and acceptance tests: #29

Evidence scope: dated repository source/review; no new live route proof is implied.

providers/near-ai.mdverifiers/near_ai.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mddata/evidence/near-2026-11-15/glm-5.3-flash-verify.jsonresearch/operator-review-2026-11-15/near.md

Our request test to do Recipient key

Evidence we still need

Evidence reviewed
Some quote/key checks exist, but the actual TLS/request path and selected backend are not fully bound. teep 6548996 on 2026-11-15 found the gateway TLS SPKI bound to its attestation (tls_key_binding pass, data/evidence/teep-2026-11-15/runs/nearcloud_*/report.txt); that run used E2EE, not this plain path.
Next proof / success condition
Bind the actual TLS/request path and selected backend to the attested key; reject a substituted recipient key before the prompt is sent.

Tracked work and acceptance tests: #29

Evidence scope: dated repository source/review; no new live route proof is implied.

providers/near-ai.mdverifiers/near_ai.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mddata/evidence/near-2026-11-15/glm-5.3-flash-verify.jsonresearch/operator-review-2026-11-15/near.md

Our reply test to do Reply key + freshness

Evidence we still need

Evidence reviewed
Some nonce checks exist, but no actual prompt/reply exchange through the selected backend was verified.
Next proof / success condition
Verify a harmless prompt/reply exchange through the selected backend; reject substituted reply keys, stale evidence and replay.

Tracked work and acceptance tests: #29

Evidence scope: dated repository source/review; no new live route proof is implied.

providers/near-ai.mdverifiers/near_ai.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mddata/evidence/near-2026-11-15/glm-5.3-flash-verify.jsonresearch/operator-review-2026-11-15/near.md

Our anchor lacks glm-5.3-flash Code

Work in our verifier

Evidence reviewed
data/evidence/near-2026-11-15/glm-5.3-flash-verify.json: model_composes for three instances, compose hashes c900dc73..., e8416a02..., a7de0bba..., each quote_compose_matches and chain_authorized true, anchored false; verification_status rejected. gateway_reference_audited is also false (no audit ledger match for the gateway image); it does not gate the row. Same observation in commit ef56f97. teep sigstore_verification fails for image digest sha256:079a14b41deb11e0 (HTTP 404) on nearcloud and neardirect (data/evidence/teep-2026-11-15/SUMMARY.md).
Next proof / success condition
Review and anchor each compose that serves a model (three distinct composes serve glm-5.3-flash); authenticate compose-manager actions through the quote/event chain; map the requested model to approved reproducible artifacts.

Tracked work and acceptance tests: #29

Evidence scope: dated repository source/review; no new live route proof is implied.

providers/near-ai.mdverifiers/near_ai.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mddata/evidence/near-2026-11-15/glm-5.3-flash-verify.jsonresearch/operator-review-2026-11-15/near.md

Operator picks the backend Operator access

Finding in the reviewed system

Evidence reviewed
cloud-api b26eb029 verification.rs:282,326, measurement.rs:93-110, model.rs:1626-1645; gateway compose 2643bc7a (2026-11-14). research/operator-review-2026-11-15/near.md F1
Next proof / success condition
SSH in the non-dev OS images; gateway log and telemetry egress; a live substitution test.

Tracked work and acceptance tests: #29

Evidence scope: dated repository source/review; no new live route proof is implied.

providers/near-ai.mdverifiers/near_ai.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mddata/evidence/near-2026-11-15/glm-5.3-flash-verify.jsonresearch/operator-review-2026-11-15/near.md

Gateway picks among 3 composes Downstream

Several open obligations

Evidence reviewed
On 2026-11-15 NEAR listed z-ai/glm-5.3-flash on three instances (19aea169..., a8a25faa..., cc796618...) with three different compose hashes and one signing key 0x3b753b44... (commit ef56f97, fixture tests/fixtures/near_glm53_three_instances.json). teep 4ecddb6 accepts an attested model name that differs from the requested one; our verifier records model_names_match_request, true on 2026-11-15 (data/evidence/near-2026-11-15/glm-5.3-flash-verify.json).
Next proof / success condition
Bind the instance and model that served each request to that request; reject instance or model substitution and insecure fallback.

Tracked work and acceptance tests: #29

Evidence scope: dated repository source/review; no new live route proof is implied.

providers/near-ai.mdverifiers/near_ai.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mddata/evidence/near-2026-11-15/glm-5.3-flash-verify.jsonresearch/operator-review-2026-11-15/near.md

Our refusal test to do Updates

Evidence we still need

Evidence reviewed
Repository per-model pins are distinct from on-chain provider authorization; the sending client policy was not validated end to end. On 2026-11-15 NEAR listed z-ai/glm-5.3-flash on three instances (19aea169..., a8a25faa..., cc796618...) with three different compose hashes and one signing key 0x3b753b44... (commit ef56f97, fixture tests/fixtures/near_glm53_three_instances.json). Each compose was chain-authorized, so a client pinned to the key cannot tell which of the three composes served a request.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback.

Tracked work and acceptance tests: #29

Evidence scope: dated repository source/review; no new live route proof is implied.

providers/near-ai.mdverifiers/near_ai.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mddata/evidence/near-2026-11-15/glm-5.3-flash-verify.jsonresearch/operator-review-2026-11-15/near.md

E2EE through gateway Intel TDX + GPU

Reviewed 2026-11-15 by claude-code dated re-review 2026-11-15 of the cells whose text cites 2026-11-15 evidence; other cells carry the 2026-10-14 codex-source-review (neither is independent audit approval) · source revision 1eca00d00f17 · Route inventory and raw checks

RTMR0 not derivable Hardware

Several open obligations

Evidence reviewed
data/evidence/near-2026-11-15/glm-5.3-flash-verify.json: tcb_current, tee_debug_disabled, event_log_replayed, tee_boot_config and gpu_attested true; tee_hardware_config null; tcb_statuses gateway and 3 models UpToDate. Per-instance checks: commit ef56f97, fixture tests/fixtures/near_glm53_three_instances.json. teep 6548996 nearcloud runs agree (data/evidence/teep-2026-11-15/runs/nearcloud_*/report.txt).
Next proof / success condition
NEAR to publish the VM configuration (or we derive RTMR0 another way) so RTMR0 can be pinned; verifiers/near_ai.py sets tee_hardware_config to None.

Tracked work and acceptance tests: #29

Evidence scope: dated repository source/review; no new live route proof is implied.

providers/near-ai.mdverifiers/near_ai.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mddata/evidence/near-2026-11-15/glm-5.3-flash-verify.jsonresearch/operator-review-2026-11-15/near.md

Third-party E2EE test only Recipient key

Evidence we still need

Evidence reviewed
teep 6548996 nearcloud z-ai/glm-5.3-flash (Instances: 3) and Qwen/Qwen3.8-27B: nonce_match, tls_key_binding, e2ee_capable, e2ee_usable pass (data/evidence/teep-2026-11-15/runs/nearcloud_*/report.txt). On 2026-11-15 NEAR listed z-ai/glm-5.3-flash on three instances (19aea169..., a8a25faa..., cc796618...) with three different compose hashes and one signing key 0x3b753b44... (commit ef56f97, fixture tests/fixtures/near_glm53_three_instances.json). teep 74834db: on /v1/completions only prompt is encrypted; stop and user pass in plaintext.
Next proof / success condition
Run a harmless encrypted request in our verifier and reject a substituted recipient key before send; record which request fields stay plaintext per endpoint.

Tracked work and acceptance tests: #29

Evidence scope: dated repository source/review; no new live route proof is implied.

providers/near-ai.mdverifiers/near_ai.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mddata/evidence/near-2026-11-15/glm-5.3-flash-verify.jsonresearch/operator-review-2026-11-15/near.md

v3 fix in source only Reply key + freshness

Evidence we still need

Evidence reviewed
Read 2026-11-15 at nearai/inference-proxy 3f7e979, src/encryption.rs:62-75: request_aad() appends canonical_reply_key() under the label nearai-e2ee-request-v3, so a relay that swaps the reply key breaks decryption of the request. Issue 1061 is still open with no comments. The deployed proxy digest 60d08c20… matches no release digest (research/operator-review-2026-11-15/near.md). teep decrypted reply fields on nearcloud through an honest gateway (data/evidence/teep-2026-11-15/runs/nearcloud_*/report.txt), which does not exercise substitution.
Next proof / success condition
Tie the deployed proxy to a commit containing #321, run a v3 exchange with a substituted reply key and observe the rejection, and check whether a v3 client can be downgraded to v1/v2.

Tracked work and acceptance tests: #29

Evidence scope: dated repository source/review; no new live route proof is implied.

providers/near-ai.mdverifiers/near_ai.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mddata/evidence/near-2026-11-15/glm-5.3-flash-verify.jsonresearch/operator-review-2026-11-15/near.md

Our anchor lacks glm-5.3-flash Code

Work in our verifier

Evidence reviewed
data/evidence/near-2026-11-15/glm-5.3-flash-verify.json: model_composes for three instances, compose hashes c900dc73..., e8416a02..., a7de0bba..., each quote_compose_matches and chain_authorized true, anchored false; verification_status rejected. gateway_reference_audited is also false (no audit ledger match for the gateway image); it does not gate the row. Same observation in commit ef56f97. teep sigstore_verification fails for image digest sha256:079a14b41deb11e0 (HTTP 404) on nearcloud and neardirect (data/evidence/teep-2026-11-15/SUMMARY.md).
Next proof / success condition
Review and anchor each compose that serves a model (three distinct composes serve glm-5.3-flash); authenticate compose-manager actions through the quote/event chain; map the requested model to approved reproducible artifacts.

Tracked work and acceptance tests: #29

Evidence scope: dated repository source/review; no new live route proof is implied.

providers/near-ai.mdverifiers/near_ai.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mddata/evidence/near-2026-11-15/glm-5.3-flash-verify.jsonresearch/operator-review-2026-11-15/near.md

App-wide key, unlogged env Operator access

Finding in the reviewed system

Evidence reviewed
inference-proxy signing.rs:149-167; compose-manager main.rs:404-425,638-645; v0.0.585 action log; Base owner() reads 2026-11-15. research/operator-review-2026-11-15/near.md F2-F4
Next proof / success condition
The commit behind proxy 60d08c20; fusion and redis code; DstackKms owner powers; the full authorized-compose set.

Tracked work and acceptance tests: #29

Evidence scope: dated repository source/review; no new live route proof is implied.

providers/near-ai.mdverifiers/near_ai.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mddata/evidence/near-2026-11-15/glm-5.3-flash-verify.jsonresearch/operator-review-2026-11-15/near.md

Gateway picks among 3 composes Downstream

Several open obligations

Evidence reviewed
On 2026-11-15 NEAR listed z-ai/glm-5.3-flash on three instances (19aea169..., a8a25faa..., cc796618...) with three different compose hashes and one signing key 0x3b753b44... (commit ef56f97, fixture tests/fixtures/near_glm53_three_instances.json). teep 4ecddb6 accepts an attested model name that differs from the requested one; our verifier records model_names_match_request, true on 2026-11-15 (data/evidence/near-2026-11-15/glm-5.3-flash-verify.json).
Next proof / success condition
Bind the instance and model that served each request to that request; reject instance or model substitution and insecure fallback.

Tracked work and acceptance tests: #29

Evidence scope: dated repository source/review; no new live route proof is implied.

providers/near-ai.mdverifiers/near_ai.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mddata/evidence/near-2026-11-15/glm-5.3-flash-verify.jsonresearch/operator-review-2026-11-15/near.md

Our refusal test to do Updates

Evidence we still need

Evidence reviewed
Repository per-model pins are distinct from on-chain provider authorization; the sending client policy was not validated end to end. On 2026-11-15 NEAR listed z-ai/glm-5.3-flash on three instances (19aea169..., a8a25faa..., cc796618...) with three different compose hashes and one signing key 0x3b753b44... (commit ef56f97, fixture tests/fixtures/near_glm53_three_instances.json). Each compose was chain-authorized, so a client pinned to the key cannot tell which of the three composes served a request.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback.

Tracked work and acceptance tests: #29

Evidence scope: dated repository source/review; no new live route proof is implied.

providers/near-ai.mdverifiers/near_ai.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mddata/evidence/near-2026-11-15/glm-5.3-flash-verify.jsonresearch/operator-review-2026-11-15/near.md

Direct model Intel TDX + GPU

Reviewed 2026-11-15 by claude-code dated re-review 2026-11-15 of the cells whose text cites 2026-11-15 evidence; other cells carry the 2026-10-14 codex-source-review (neither is independent audit approval) · source revision 1eca00d00f17 · Route inventory and raw checks

Third-party check only Hardware

Evidence we still need

Evidence reviewed
data/evidence/teep-2026-11-15/runs/neardirect_z-ai_glm-5.3-flash_20261115T165919Z/report.txt, data/evidence/teep-2026-11-15/runs/neardirect_Qwen_Qwen3.8-27B_20261115T165927Z/report.txt; tee_boot_config is an allowed failure in teep.
Next proof / success condition
Run our verifier on the direct endpoints and retain the result, including boot and hardware register pins.

Tracked work and acceptance tests: #29

Evidence scope: dated repository source/review; no new live route proof is implied.

providers/near-ai.mdverifiers/near_ai.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/near.md

Third-party E2EE test only Recipient key

Evidence we still need

Evidence reviewed
data/evidence/teep-2026-11-15/runs/neardirect_*/report.txt: nonce_match, tls_key_binding, e2ee_capable, e2ee_usable pass for z-ai/glm-5.3-flash and Qwen/Qwen3.8-27B.
Next proof / success condition
Run a harmless encrypted request in our verifier and reject a substituted recipient key before send; record which request fields stay plaintext per endpoint.

Tracked work and acceptance tests: #29

Evidence scope: dated repository source/review; no new live route proof is implied.

providers/near-ai.mdverifiers/near_ai.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/near.md

Our reply test to do Reply key + freshness

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that the reply key is bound to the same attested workload; reject a substituted reply key, stale evidence and replay before the reply is disclosed.

Tracked work and acceptance tests: #29

Evidence scope: dated repository source/review; no new live route proof is implied.

providers/near-ai.mdverifiers/near_ai.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/near.md

Our direct-path test Code

Evidence we still need

Evidence reviewed
The experimental direct path needs its own client, workload and accepted-policy evidence. Gateway adapter findings cannot establish this route.
Next proof / success condition
Trace the deployed measured artifact to reviewed source/build and model identities, including mutable configuration; test an unapproved artifact.

Tracked work and acceptance tests: #29

Evidence scope: dated repository source/review; no new live route proof is implied.

providers/near-ai.mdverifiers/near_ai.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/near.md

App-wide key, unlogged env Operator access

Finding in the reviewed system

Evidence reviewed
teep neardirect capture 2026-11-14; inference-proxy signing.rs:149-167; compose-manager main.rs:404-425. research/operator-review-2026-11-15/near.md
Next proof / success condition
An encrypted request through our client; other models' direct hosts.

Tracked work and acceptance tests: #29

Evidence scope: dated repository source/review; no new live route proof is implied.

providers/near-ai.mdverifiers/near_ai.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/near.md

Our worker test to do Downstream

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #29

Evidence scope: dated repository source/review; no new live route proof is implied.

providers/near-ai.mdverifiers/near_ai.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/near.md

Our direct-path test Updates

Evidence we still need

Evidence reviewed
The experimental direct path needs its own client, workload and accepted-policy evidence. Gateway adapter findings cannot establish this route.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback.

Tracked work and acceptance tests: #29

Evidence scope: dated repository source/review; no new live route proof is implied.

providers/near-ai.mdverifiers/near_ai.pydata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/near.md

Tinfoil

1 paths · 1/1 returned · Reference history

2 provider reference surfaces differ from the review baseline. Determine route relevance before relying on the dated assessment.

Hardware · TCB below AMD fix level

Router → worker: TCB below AMD fix level — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
verifiers/tinfoil_amd.py and verifiers/certs/amd_genoa_tcb_minimum.json (merge 9b4d03a). Live reads 2026-11-15: router-0.tinfoil.sh bl=0x0a tee=0 snp=0x1B ucode=0x58; inference.tinfoil.sh (17:15Z) bl=0x0a tee=0 snp=0x17 ucode=0x54, also below the AMD-SB-3027 microcode minimum 0x56. providers/tinfoil.md records that the chain verifies and the verifier rejects the row on TCB.
Next proof / success condition
Tinfoil hosts reporting TCB[SNP] >= 0x1D and the stepping microcode minimum, retained as an authenticated report. The bl and tee minimums are pinned from observation; no AMD bulletin states them.

Tracked work and acceptance tests: #26

Recipient key · Our recipient proof to do

Router → worker: Our recipient proof to do — Evidence we still need

Evidence we still need

Evidence reviewed
Live TLS-key pin and HPKE/report-data consistency are checked. Since 2026-11-15 the AMD chain and report signature are authenticated (merge 9b4d03a); the platform TCB is below AMD's minimum (Hardware column). No encrypted request is verified.
Next proof / success condition
Verify a harmless prompt is encrypted to the HPKE key bound to the attested workload and the pinned TLS channel; reject a substituted recipient key. AMD report authentication passes as of 2026-11-15; the platform TCB is below AMD's minimum (Hardware).

Tracked work and acceptance tests: #26#30

Reply key + freshness · Our freshness proof to do

Router → worker: Our freshness proof to do — Evidence we still need

Evidence we still need

Evidence reviewed
Authenticated session freshness and the encrypted reply remain unverified.
Next proof / success condition
Verify authenticated freshness for the exchange and a harmless encrypted reply; reject substituted reply keys, stale evidence and replay. AMD report authentication passes as of 2026-11-15; the platform TCB is below AMD's minimum (Hardware).

Tracked work and acceptance tests: #26#30

Code · Our rebuild proof to do

Router → worker: Our rebuild proof to do — Evidence we still need

Evidence we still need

Evidence reviewed
Sigstore release provenance and measurement agreement are checked. The verifier explicitly does not independently reproduce the build.
Next proof / success condition
Independently reproduce the exact released artifact and bind it to the authenticated measured deployment.

Tracked work and acceptance tests: #30

Operator access · Control-plane prompt rewriting

Router → worker: Control-plane prompt rewriting — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
Read at tinfoilsh/confidential-model-router ef64384: manager/model_metadata.go:147-183 (reasoning_params from the control plane), autoroute/autoroute.go:271-304,343-359 (merge without a key allow-list), manager/manager.go:743 (router-config fetched every minute, no hash). Live TCB 2026-11-15. research/operator-review-2026-11-15/tinfoil.md
Next proof / success condition
Production reasoning_params and router-config contents; the live minimum_tcb Tinfoil publishes; the admin-SSH implementation in tinfoil-config.

Tracked work and acceptance tests: #30

Downstream · Our worker test to do

Router → worker: Our worker test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Updates · Our refusal test to do

Router → worker: Our refusal test to do — Evidence we still need

Evidence we still need

Evidence reviewed
The measured router configuration can pin worker repositories. Automatic release authority, client approval and a live worker request still require a scoped assessment.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback. Verify measured worker-repository policy on a live route.

Tracked work and acceptance tests: #30

Router → worker AMD SEV-SNP

Reviewed 2026-11-15 by claude-code dated re-review 2026-11-15 of the cells whose text cites 2026-11-15 evidence; other cells carry the 2026-10-14 codex-source-review (neither is independent audit approval) · source revision 1eca00d00f17 · Route inventory and raw checks

TCB below AMD fix level Hardware

Finding in the reviewed system

Evidence reviewed
verifiers/tinfoil_amd.py and verifiers/certs/amd_genoa_tcb_minimum.json (merge 9b4d03a). Live reads 2026-11-15: router-0.tinfoil.sh bl=0x0a tee=0 snp=0x1B ucode=0x58; inference.tinfoil.sh (17:15Z) bl=0x0a tee=0 snp=0x17 ucode=0x54, also below the AMD-SB-3027 microcode minimum 0x56. providers/tinfoil.md records that the chain verifies and the verifier rejects the row on TCB.
Next proof / success condition
Tinfoil hosts reporting TCB[SNP] >= 0x1D and the stepping microcode minimum, retained as an authenticated report. The bl and tee minimums are pinned from observation; no AMD bulletin states them.

Tracked work and acceptance tests: #26

Evidence scope: dated repository source/review; no new live route proof is implied.

verifiers/tinfoil.pyresearch/review-2026-10-13/review.mddata/routes.jsonproviders/tinfoil.mdverifiers/certs/amd_genoa_tcb_minimum.jsonresearch/operator-review-2026-11-15/tinfoil.md

Our recipient proof to do Recipient key

Evidence we still need

Evidence reviewed
Live TLS-key pin and HPKE/report-data consistency are checked. Since 2026-11-15 the AMD chain and report signature are authenticated (merge 9b4d03a); the platform TCB is below AMD's minimum (Hardware column). No encrypted request is verified.
Next proof / success condition
Verify a harmless prompt is encrypted to the HPKE key bound to the attested workload and the pinned TLS channel; reject a substituted recipient key. AMD report authentication passes as of 2026-11-15; the platform TCB is below AMD's minimum (Hardware).

Tracked work and acceptance tests: #26#30

Evidence scope: dated repository source/review; no new live route proof is implied.

verifiers/tinfoil.pyresearch/review-2026-10-13/review.mddata/routes.jsonproviders/tinfoil.mdverifiers/certs/amd_genoa_tcb_minimum.jsonresearch/operator-review-2026-11-15/tinfoil.md

Our freshness proof to do Reply key + freshness

Evidence we still need

Evidence reviewed
Authenticated session freshness and the encrypted reply remain unverified.
Next proof / success condition
Verify authenticated freshness for the exchange and a harmless encrypted reply; reject substituted reply keys, stale evidence and replay. AMD report authentication passes as of 2026-11-15; the platform TCB is below AMD's minimum (Hardware).

Tracked work and acceptance tests: #26#30

Evidence scope: dated repository source/review; no new live route proof is implied.

verifiers/tinfoil.pyresearch/review-2026-10-13/review.mddata/routes.jsonproviders/tinfoil.mdverifiers/certs/amd_genoa_tcb_minimum.jsonresearch/operator-review-2026-11-15/tinfoil.md

Our rebuild proof to do Code

Evidence we still need

Evidence reviewed
Sigstore release provenance and measurement agreement are checked. The verifier explicitly does not independently reproduce the build.
Next proof / success condition
Independently reproduce the exact released artifact and bind it to the authenticated measured deployment.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

verifiers/tinfoil.pyresearch/review-2026-10-13/review.mddata/routes.jsonproviders/tinfoil.mdverifiers/certs/amd_genoa_tcb_minimum.jsonresearch/operator-review-2026-11-15/tinfoil.md

Control-plane prompt rewriting Operator access

Finding in the reviewed system

Evidence reviewed
Read at tinfoilsh/confidential-model-router ef64384: manager/model_metadata.go:147-183 (reasoning_params from the control plane), autoroute/autoroute.go:271-304,343-359 (merge without a key allow-list), manager/manager.go:743 (router-config fetched every minute, no hash). Live TCB 2026-11-15. research/operator-review-2026-11-15/tinfoil.md
Next proof / success condition
Production reasoning_params and router-config contents; the live minimum_tcb Tinfoil publishes; the admin-SSH implementation in tinfoil-config.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

verifiers/tinfoil.pyresearch/review-2026-10-13/review.mddata/routes.jsonproviders/tinfoil.mdverifiers/certs/amd_genoa_tcb_minimum.jsonresearch/operator-review-2026-11-15/tinfoil.md

Our worker test to do Downstream

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

verifiers/tinfoil.pyresearch/review-2026-10-13/review.mddata/routes.jsonproviders/tinfoil.mdverifiers/certs/amd_genoa_tcb_minimum.jsonresearch/operator-review-2026-11-15/tinfoil.md

Our refusal test to do Updates

Evidence we still need

Evidence reviewed
The measured router configuration can pin worker repositories. Automatic release authority, client approval and a live worker request still require a scoped assessment.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback. Verify measured worker-repository policy on a live route.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

verifiers/tinfoil.pyresearch/review-2026-10-13/review.mddata/routes.jsonproviders/tinfoil.mdverifiers/certs/amd_genoa_tcb_minimum.jsonresearch/operator-review-2026-11-15/tinfoil.md

Venice

2 paths · 4/5 returned · Reference history

Hardware · Routes differ

E2EE via NEAR: Third-party check only — Evidence we still need

Evidence we still need

Evidence reviewed
data/evidence/teep-2026-11-15/runs/venice_e2ee-glm-5-3-flash_20261115T165858Z/report.txt, data/evidence/teep-2026-11-15/runs/venice_e2ee-qwen3-8-27b_20261115T165916Z/report.txt. The REPORTDATA key 0x3b753b44... of e2ee-glm-5-3-flash equals the signing address of NEAR z-ai/glm-5.3-flash (data/evidence/near-2026-11-15/glm-5.3-flash-verify.json). teep tee_boot_config and tee_hardware_config fail as allowed (RTMR0-1 not in teep policy). The models captured differ from the model list in data/routes.json.
Next proof / success condition
Retain a run of our Venice adapter with boot and hardware register pins.

Tracked work and acceptance tests: #30

E2EE via Phala: Inference CPU unattested — Several open obligations

Several open obligations

Evidence reviewed
data/evidence/teep-2026-11-15/runs/venice_e2ee-glm-5-3-p_20261115T165902Z/report.txt: gateway quote, Intel chain, debug 0 and event-log replay pass; nvidia_signature, nvidia_nonce_client_bound, nvidia_nras_verified pass; tee_quote_present fails. The raw attestation body (zed:/data/amiller/teep-refresh-20261115/captures/venice/e2ee-glm-5-3-p/20261115T165902Z/capture/venice_e2ee-glm-5-3-p_20261115_165916/responses/0001_api.venice.ai_api_v1_tee_attestation.body, not retained here) has "gpu_evidence_omitted":false; teep@6548996 docs/attestation_gaps/aci_gateway.md residual 11 records GPU evidence omitted on 2026-11-12, so its presence varies by date.
Next proof / success condition
Inference-host CPU evidence bound to the GPU evidence and the request; a retained run of our adapter.

Tracked work and acceptance tests: #30

Recipient key · Key attested; no prompt sent

E2EE via NEAR: Key attested; no prompt sent — Evidence we still need

Evidence we still need

Evidence reviewed
data/evidence/teep-2026-11-15/runs/venice_e2ee-glm-5-3-flash_20261115T165858Z/report.txt, data/evidence/teep-2026-11-15/runs/venice_e2ee-qwen3-8-27b_20261115T165916Z/report.txt: tee_reportdata_binding and nonce_match pass; e2ee_usable fails with HTTP 401, reproduced with curl (data/evidence/teep-2026-11-15/SUMMARY.md).
Next proof / success condition
Verify that a harmless prompt is encrypted only to a key bound to the attested workload and the live channel; reject a substituted recipient key before the prompt is sent.

Tracked work and acceptance tests: #30

E2EE via Phala: Key attested; no prompt sent — Evidence we still need

Evidence we still need

Evidence reviewed
data/evidence/teep-2026-11-15/runs/venice_e2ee-glm-5-3-p_20261115T165902Z/report.txt: aci_key_custody and gateway_tee_reportdata_binding (enclave key and nonce) pass; e2ee_usable fails with HTTP 401, reproduced with curl (data/evidence/teep-2026-11-15/SUMMARY.md). teep@6548996 docs/attestation_gaps/aci_gateway.md residual 7.
Next proof / success condition
Verify that a harmless prompt is encrypted only to a key bound to the attested workload and the live channel; reject a substituted recipient key before the prompt is sent.

Tracked work and acceptance tests: #30

Reply key + freshness · Our reply test to do

E2EE via NEAR: Our reply test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that the reply key is bound to the same attested workload; reject a substituted reply key, stale evidence and replay before the reply is disclosed.

Tracked work and acceptance tests: #30

E2EE via Phala: Our reply test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that the reply key is bound to the same attested workload; reject a substituted reply key, stale evidence and replay before the reply is disclosed.

Tracked work and acceptance tests: #30

Code · Our code proof to do

E2EE via NEAR: Our code proof to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace the deployed measured artifact to reviewed source/build and model identities, including mutable configuration; test an unapproved artifact.

Tracked work and acceptance tests: #30

E2EE via Phala: Our code proof to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace the deployed measured artifact to reviewed source/build and model identities, including mutable configuration; test an unapproved artifact.

Tracked work and acceptance tests: #30

Operator access · Routes differ

E2EE via NEAR: Runtime-deployed serving code — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
NEAR app_compose e8416a02 and action log (teep capture 2026-11-13, 114 deploys 10-21..11-13); compose-manager c7fbaaf; cvm-compose-files c330ab4a. research/operator-review-2026-11-15/venice.md
Next proof / success condition
Measurement of each deploy or a client-checked tag allowlist; a pinned log policy.

Tracked work and acceptance tests: #30

E2EE via Phala: Non-TEE-only gateway host — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
teep captures 2026-11-13; gateway e9b6ae8a; live sessions 2026-11-15. research/operator-review-2026-11-15/venice.md
Next proof / success condition
Proof that Venice sets aci_verified or uses a TEE-only host; relayed receipts.

Tracked work and acceptance tests: #30

Downstream · Routes differ

E2EE via NEAR: Our worker test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

E2EE via Phala: Downstream host self-asserted — Evidence we still need

Evidence we still need

Evidence reviewed
data/evidence/teep-2026-11-15/runs/venice_e2ee-glm-5-3-p_20261115T165902Z/report.txt: "Gateway downstream TLS: api.redpill.ai"; aci_backend_named fails. teep@6548996 docs/attestation_gaps/aci_gateway.md residuals 1 and 5.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Updates · Our policy test to do

E2EE via NEAR: Our policy test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback.

Tracked work and acceptance tests: #30

E2EE via Phala: Our policy test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback.

Tracked work and acceptance tests: #30

E2EE via NEAR NEAR backend

Reviewed 2026-11-15 by claude-code dated re-review 2026-11-15 of the cells whose text cites 2026-11-15 evidence; other cells carry the 2026-10-14 codex-source-review (neither is independent audit approval) · source revision 1eca00d00f17 · Route inventory and raw checks

Third-party check only Hardware

Evidence we still need

Evidence reviewed
data/evidence/teep-2026-11-15/runs/venice_e2ee-glm-5-3-flash_20261115T165858Z/report.txt, data/evidence/teep-2026-11-15/runs/venice_e2ee-qwen3-8-27b_20261115T165916Z/report.txt. The REPORTDATA key 0x3b753b44... of e2ee-glm-5-3-flash equals the signing address of NEAR z-ai/glm-5.3-flash (data/evidence/near-2026-11-15/glm-5.3-flash-verify.json). teep tee_boot_config and tee_hardware_config fail as allowed (RTMR0-1 not in teep policy). The models captured differ from the model list in data/routes.json.
Next proof / success condition
Retain a run of our Venice adapter with boot and hardware register pins.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/venice.md

Key attested; no prompt sent Recipient key

Evidence we still need

Evidence reviewed
data/evidence/teep-2026-11-15/runs/venice_e2ee-glm-5-3-flash_20261115T165858Z/report.txt, data/evidence/teep-2026-11-15/runs/venice_e2ee-qwen3-8-27b_20261115T165916Z/report.txt: tee_reportdata_binding and nonce_match pass; e2ee_usable fails with HTTP 401, reproduced with curl (data/evidence/teep-2026-11-15/SUMMARY.md).
Next proof / success condition
Verify that a harmless prompt is encrypted only to a key bound to the attested workload and the live channel; reject a substituted recipient key before the prompt is sent.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/venice.md

Our reply test to do Reply key + freshness

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that the reply key is bound to the same attested workload; reject a substituted reply key, stale evidence and replay before the reply is disclosed.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/venice.md

Our code proof to do Code

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace the deployed measured artifact to reviewed source/build and model identities, including mutable configuration; test an unapproved artifact.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/venice.md

Runtime-deployed serving code Operator access

Finding in the reviewed system

Evidence reviewed
NEAR app_compose e8416a02 and action log (teep capture 2026-11-13, 114 deploys 10-21..11-13); compose-manager c7fbaaf; cvm-compose-files c330ab4a. research/operator-review-2026-11-15/venice.md
Next proof / success condition
Measurement of each deploy or a client-checked tag allowlist; a pinned log policy.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/venice.md

Our worker test to do Downstream

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/venice.md

Our policy test to do Updates

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/venice.md

E2EE via Phala Phala backend

Reviewed 2026-11-15 by claude-code dated re-review 2026-11-15 of the cells whose text cites 2026-11-15 evidence; other cells carry the 2026-10-14 codex-source-review (neither is independent audit approval) · source revision 1eca00d00f17 · Route inventory and raw checks

Inference CPU unattested Hardware

Several open obligations

Evidence reviewed
data/evidence/teep-2026-11-15/runs/venice_e2ee-glm-5-3-p_20261115T165902Z/report.txt: gateway quote, Intel chain, debug 0 and event-log replay pass; nvidia_signature, nvidia_nonce_client_bound, nvidia_nras_verified pass; tee_quote_present fails. The raw attestation body (zed:/data/amiller/teep-refresh-20261115/captures/venice/e2ee-glm-5-3-p/20261115T165902Z/capture/venice_e2ee-glm-5-3-p_20261115_165916/responses/0001_api.venice.ai_api_v1_tee_attestation.body, not retained here) has "gpu_evidence_omitted":false; teep@6548996 docs/attestation_gaps/aci_gateway.md residual 11 records GPU evidence omitted on 2026-11-12, so its presence varies by date.
Next proof / success condition
Inference-host CPU evidence bound to the GPU evidence and the request; a retained run of our adapter.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/venice.md

Key attested; no prompt sent Recipient key

Evidence we still need

Evidence reviewed
data/evidence/teep-2026-11-15/runs/venice_e2ee-glm-5-3-p_20261115T165902Z/report.txt: aci_key_custody and gateway_tee_reportdata_binding (enclave key and nonce) pass; e2ee_usable fails with HTTP 401, reproduced with curl (data/evidence/teep-2026-11-15/SUMMARY.md). teep@6548996 docs/attestation_gaps/aci_gateway.md residual 7.
Next proof / success condition
Verify that a harmless prompt is encrypted only to a key bound to the attested workload and the live channel; reject a substituted recipient key before the prompt is sent.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/venice.md

Our reply test to do Reply key + freshness

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that the reply key is bound to the same attested workload; reject a substituted reply key, stale evidence and replay before the reply is disclosed.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/venice.md

Our code proof to do Code

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace the deployed measured artifact to reviewed source/build and model identities, including mutable configuration; test an unapproved artifact.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/venice.md

Non-TEE-only gateway host Operator access

Finding in the reviewed system

Evidence reviewed
teep captures 2026-11-13; gateway e9b6ae8a; live sessions 2026-11-15. research/operator-review-2026-11-15/venice.md
Next proof / success condition
Proof that Venice sets aci_verified or uses a TEE-only host; relayed receipts.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/venice.md

Downstream host self-asserted Downstream

Evidence we still need

Evidence reviewed
data/evidence/teep-2026-11-15/runs/venice_e2ee-glm-5-3-p_20261115T165902Z/report.txt: "Gateway downstream TLS: api.redpill.ai"; aci_backend_named fails. teep@6548996 docs/attestation_gaps/aci_gateway.md residuals 1 and 5.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/venice.md

Our policy test to do Updates

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/venice.md

Chutes

1 paths · Stale / mixed · Reference history

1 provider reference surfaces differ from the review baseline. Determine route relevance before relying on the dated assessment.

Hardware · Our live run to do

TEE models: Our live run to do — Evidence we still need

Evidence we still need

Evidence reviewed
verifiers/chutes.py and verifiers/nvidia_nras.py (merge 83fe47a); tests/test_chutes_gpu_replay.py, tests/test_chutes_registers.py. teep 6548996 (data/evidence/teep-2026-11-15/runs/chutes_*/report.txt): Intel TCB UpToDate, debug 0, GPU SPDM signatures and client-derived nonce pass (1 and 8 Blackwell GPUs); cpu_gpu_chain fails (allowed); GLM-5.2-TEE tee_measurement fails on MRSEAM.
Next proof / success condition
Retain a live run of the merged adapter; set an MRSEAM (TDX module) policy and explain f40772d8.... This does not establish serving-code/model identity.

Tracked work and acceptance tests: #28

Recipient key · Our request test to do

TEE models: Our request test to do — Evidence we still need

Evidence we still need

Evidence reviewed
The daily adapter implements key checks. The retained route assessment does not contain a complete request test for this client path. teep 6548996 on 2026-11-15 completed ML-KEM-768 E2EE round trips for Qwen/Qwen3.8-27B-TEE (30 chunks) and zai-org/GLM-5.2-TEE (298 chunks) (data/evidence/teep-2026-11-15/runs/chutes_*/report.txt); no echoed nonce, so freshness rests on the REPORTDATA binding. Our adapter has not run a request.
Next proof / success condition
Verify a harmless prompt is encrypted to a key bound to the attested workload; reject a substituted recipient key before release. Reconcile the dated manual encrypted-completion evidence with the exact client/mode.

Tracked work and acceptance tests: #30#19

Reply key + freshness · Our reply test to do

TEE models: Our reply test to do — Evidence we still need

Evidence we still need

Evidence reviewed
The daily adapter implements nonce checks. The retained route assessment does not contain a complete request/reply test for this client path. teep 6548996 on 2026-11-15 completed ML-KEM-768 E2EE round trips for Qwen/Qwen3.8-27B-TEE (30 chunks) and zai-org/GLM-5.2-TEE (298 chunks) (data/evidence/teep-2026-11-15/runs/chutes_*/report.txt); no echoed nonce, so freshness rests on the REPORTDATA binding. Our adapter has not run a request. Decrypting a reply does not test reply-key substitution.
Next proof / success condition
Verify a harmless reply is bound to the same attested workload; reject substituted reply keys, stale evidence and replay. Reconcile the dated manual encrypted-completion evidence with the exact client/mode.

Tracked work and acceptance tests: #30#19

Code · Serving code unmeasured

TEE models: Serving code unmeasured — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
The reviewed Chutes base-TD measurements do not commit to serve.py or model weights. Golden-set membership is not approval of the serving workload.
Next proof / success condition
Establish authenticated commitments to the actual serving code/model and test that malicious code or weight substitution is rejected before payload release.

Tracked work and acceptance tests: #31

Operator access · Code can change unmeasured

TEE models: Code can change unmeasured — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
chutes cfsv_wrapper.py:54 at 08d79872; chutes-api e2e/router.py:140-152, invocation/router.py:505-937 at archived 3b5609f4. research/operator-review-2026-11-15/chutes.md
Next proof / success condition
Measured chute code and model revision; quote-bound key discovery; an append-only reference-value log; current control-plane source.

Tracked work and acceptance tests: #31

Downstream · Model not committed

TEE models: Model not committed — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
The requested model artifact and serving code are not established by the measured registers.
Next proof / success condition
Bind requested model/weights and serving code to the verified request; reject a model or artifact substitution.

Tracked work and acceptance tests: #31

Updates · Provider set only

TEE models: Provider set only — Several open obligations

Several open obligations

Evidence reviewed
The provider golden set is captured and compared; it is not an independently approved complete workload policy.
Next proof / success condition
Define independently approved complete workload expectations and test the actual client refuses unapproved code/model updates.

Tracked work and acceptance tests: #31

TEE models Intel TDX

Reviewed 2026-11-15 by claude-code dated re-review 2026-11-15 of the cells whose text cites 2026-11-15 evidence; other cells carry the 2026-10-14 codex-source-review (neither is independent audit approval) · source revision 1eca00d00f17 · Route inventory and raw checks

Our live run to do Hardware

Evidence we still need

Evidence reviewed
verifiers/chutes.py and verifiers/nvidia_nras.py (merge 83fe47a); tests/test_chutes_gpu_replay.py, tests/test_chutes_registers.py. teep 6548996 (data/evidence/teep-2026-11-15/runs/chutes_*/report.txt): Intel TCB UpToDate, debug 0, GPU SPDM signatures and client-derived nonce pass (1 and 8 Blackwell GPUs); cpu_gpu_chain fails (allowed); GLM-5.2-TEE tee_measurement fails on MRSEAM.
Next proof / success condition
Retain a live run of the merged adapter; set an MRSEAM (TDX module) policy and explain f40772d8.... This does not establish serving-code/model identity.

Tracked work and acceptance tests: #28

Evidence scope: dated repository source/review; no new live route proof is implied.

verifiers/chutes.pydata/audits/chutes_tee.jsondata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/chutes.md

Our request test to do Recipient key

Evidence we still need

Evidence reviewed
The daily adapter implements key checks. The retained route assessment does not contain a complete request test for this client path. teep 6548996 on 2026-11-15 completed ML-KEM-768 E2EE round trips for Qwen/Qwen3.8-27B-TEE (30 chunks) and zai-org/GLM-5.2-TEE (298 chunks) (data/evidence/teep-2026-11-15/runs/chutes_*/report.txt); no echoed nonce, so freshness rests on the REPORTDATA binding. Our adapter has not run a request.
Next proof / success condition
Verify a harmless prompt is encrypted to a key bound to the attested workload; reject a substituted recipient key before release. Reconcile the dated manual encrypted-completion evidence with the exact client/mode.

Tracked work and acceptance tests: #30#19

Evidence scope: dated repository source/review; no new live route proof is implied.

verifiers/chutes.pydata/audits/chutes_tee.jsondata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/chutes.md

Our reply test to do Reply key + freshness

Evidence we still need

Evidence reviewed
The daily adapter implements nonce checks. The retained route assessment does not contain a complete request/reply test for this client path. teep 6548996 on 2026-11-15 completed ML-KEM-768 E2EE round trips for Qwen/Qwen3.8-27B-TEE (30 chunks) and zai-org/GLM-5.2-TEE (298 chunks) (data/evidence/teep-2026-11-15/runs/chutes_*/report.txt); no echoed nonce, so freshness rests on the REPORTDATA binding. Our adapter has not run a request. Decrypting a reply does not test reply-key substitution.
Next proof / success condition
Verify a harmless reply is bound to the same attested workload; reject substituted reply keys, stale evidence and replay. Reconcile the dated manual encrypted-completion evidence with the exact client/mode.

Tracked work and acceptance tests: #30#19

Evidence scope: dated repository source/review; no new live route proof is implied.

verifiers/chutes.pydata/audits/chutes_tee.jsondata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/chutes.md

Serving code unmeasured Code

Finding in the reviewed system

Evidence reviewed
The reviewed Chutes base-TD measurements do not commit to serve.py or model weights. Golden-set membership is not approval of the serving workload.
Next proof / success condition
Establish authenticated commitments to the actual serving code/model and test that malicious code or weight substitution is rejected before payload release.

Tracked work and acceptance tests: #31

Evidence scope: dated repository source/review; no new live route proof is implied.

verifiers/chutes.pydata/audits/chutes_tee.jsondata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/chutes.md

Code can change unmeasured Operator access

Finding in the reviewed system

Evidence reviewed
chutes cfsv_wrapper.py:54 at 08d79872; chutes-api e2e/router.py:140-152, invocation/router.py:505-937 at archived 3b5609f4. research/operator-review-2026-11-15/chutes.md
Next proof / success condition
Measured chute code and model revision; quote-bound key discovery; an append-only reference-value log; current control-plane source.

Tracked work and acceptance tests: #31

Evidence scope: dated repository source/review; no new live route proof is implied.

verifiers/chutes.pydata/audits/chutes_tee.jsondata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/chutes.md

Model not committed Downstream

Finding in the reviewed system

Evidence reviewed
The requested model artifact and serving code are not established by the measured registers.
Next proof / success condition
Bind requested model/weights and serving code to the verified request; reject a model or artifact substitution.

Tracked work and acceptance tests: #31

Evidence scope: dated repository source/review; no new live route proof is implied.

verifiers/chutes.pydata/audits/chutes_tee.jsondata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/chutes.md

Provider set only Updates

Several open obligations

Evidence reviewed
The provider golden set is captured and compared; it is not an independently approved complete workload policy.
Next proof / success condition
Define independently approved complete workload expectations and test the actual client refuses unapproved code/model updates.

Tracked work and acceptance tests: #31

Evidence scope: dated repository source/review; no new live route proof is implied.

verifiers/chutes.pydata/audits/chutes_tee.jsondata/routes.jsondata/evidence/teep-2026-11-15/SUMMARY.mdresearch/operator-review-2026-11-15/chutes.md

Private Inference

3 paths · Not sampled · Reference history

3 provider reference surfaces differ from the review baseline. Determine route relevance before relying on the dated assessment.

Hardware · Routes differ

Standard: No attested-route claim — Scope of this route

Scope of this route

Evidence reviewed
The documented standard/ZDR path is not a client-verified TEE route.
Next proof / success condition
Keep this path outside an attested-confidentiality claim; evaluate any future private mode as a separate client/request path.

Zero data retention: No attested-route claim — Scope of this route

Scope of this route

Evidence reviewed
The documented standard/ZDR path is not a client-verified TEE route.
Next proof / success condition
Keep this path outside an attested-confidentiality claim; evaluate any future private mode as a separate client/request path.

Private beta via NEAR: Our hardware test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Retain a route-scoped replay authenticating hardware signatures, TCB/debug state and required CPU/GPU evidence under the actual platform policy.

Tracked work and acceptance tests: #30

Recipient key · Routes differ

Standard: Proxy sees prompts — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
Standard routing terminates prompt confidentiality at the provider proxy; it makes no client-to-attested-model encryption claim.
Next proof / success condition
Client-to-attested-recipient confidentiality requires a separately verified encrypted route; retention policy alone cannot establish it.

Tracked work and acceptance tests: #30

Zero data retention: Proxy sees prompts — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
Zero data retention is a retention policy, not client-to-attested-model encryption.
Next proof / success condition
Client-to-attested-recipient confidentiality requires a separately verified encrypted route; retention policy alone cannot establish it.

Tracked work and acceptance tests: #30

Private beta via NEAR: Our recipient test to do — Evidence we still need

Evidence we still need

Evidence reviewed
No fixed deployment or complete session was verified.
Next proof / success condition
Verify a harmless prompt is encrypted only to a key bound to the attested model workload; reject a substituted recipient key before the prompt is sent.

Tracked work and acceptance tests: #30

Reply key + freshness · Routes differ

Standard: No attested reply key — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
Standard routing makes no client-to-attested-model encryption claim; no reply key is bound to an attested workload.
Next proof / success condition
Reply confidentiality from the provider proxy requires a separately verified encrypted route; retention policy alone cannot establish it.

Tracked work and acceptance tests: #30

Zero data retention: No attested reply key — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
Zero data retention is a retention policy, not client-to-attested-model encryption; no reply key is bound to an attested workload.
Next proof / success condition
Reply confidentiality from the provider proxy requires a separately verified encrypted route; retention policy alone cannot establish it.

Tracked work and acceptance tests: #30

Private beta via NEAR: Reply key substitutable — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
The inspected client leaves the reply-key header substitutable by a relay. No fixed deployment or complete session was verified.
Next proof / success condition
Verify a fixed client/protocol/deployment rejects substituted reply keys before disclosure.

Tracked work and acceptance tests: #30

Code · Routes differ

Standard: No attested-route claim — Scope of this route

Scope of this route

Evidence reviewed
The documented standard/ZDR path is not a client-verified TEE route.
Next proof / success condition
Keep this path outside an attested-confidentiality claim; evaluate any future private mode as a separate client/request path.

Zero data retention: No attested-route claim — Scope of this route

Scope of this route

Evidence reviewed
The documented standard/ZDR path is not a client-verified TEE route.
Next proof / success condition
Keep this path outside an attested-confidentiality claim; evaluate any future private mode as a separate client/request path.

Private beta via NEAR: No approved baseline — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
The inspected browser checks quote consistency but does not compare compose/OS identity against an independently approved baseline.
Next proof / success condition
Authenticate and enforce an explicitly approved model-to-compose/OS/artifact baseline; reject a genuine quote for a different workload.

Tracked work and acceptance tests: #30

Operator access · Routes differ

Standard: Proxy sees prompts — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
The documented standard route gives the proxy and selected supplier plaintext access outside a client-verified computation.
Next proof / success condition
Use a separately verified confidential route if the proxy/supplier must not read prompts; document any retained plaintext holders.

Tracked work and acceptance tests: #30

Zero data retention: Proxy sees prompts — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
The documented ZDR proxy still processes plaintext outside a client-verified computation; retention enforcement was not verified.
Next proof / success condition
Use a separately verified confidential route if the proxy/supplier must not read prompts; document any retained plaintext holders.

Tracked work and acceptance tests: #30

Private beta via NEAR: Our access review to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace operator credentials, debug access and unmeasured inputs to their effects on protected data in the exact deployed version; retain checks and counterexamples.

Tracked work and acceptance tests: #30

Downstream · Routes differ

Standard: Our worker test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Zero data retention: Our worker test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Private beta via NEAR: Our model test to do — Evidence we still need

Evidence we still need

Evidence reviewed
The browser requests a model, but rejection of a substituted attested model was not established.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Updates · Routes differ

Standard: No attested-route claim — Scope of this route

Scope of this route

Evidence reviewed
The documented standard/ZDR path is not a client-verified TEE route.
Next proof / success condition
Keep this path outside an attested-confidentiality claim; evaluate any future private mode as a separate client/request path.

Zero data retention: No attested-route claim — Scope of this route

Scope of this route

Evidence reviewed
The documented standard/ZDR path is not a client-verified TEE route.
Next proof / success condition
Keep this path outside an attested-confidentiality claim; evaluate any future private mode as a separate client/request path.

Private beta via NEAR: No approved baseline — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
The inspected client lacks an independent approved workload baseline; self-consistency does not require client approval of code changes.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback. Establish an authority independent of self-consistency.

Tracked work and acceptance tests: #30

Standard Provider routing

Reviewed 2026-10-14 by codex-source-review (not independent audit approval) · source revision ffb5b54e8cb7 · Route inventory and raw checks

Not assessed on the comparison page: no automated adapter in this repository. Cells below are from the dated source review.

No attested-route claim Hardware

Scope of this route

Evidence reviewed
The documented standard/ZDR path is not a client-verified TEE route.
Next proof / success condition
Keep this path outside an attested-confidentiality claim; evaluate any future private mode as a separate client/request path.

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/privateinference.mddata/routes.json

Proxy sees prompts Recipient key

Finding in the reviewed system

Evidence reviewed
Standard routing terminates prompt confidentiality at the provider proxy; it makes no client-to-attested-model encryption claim.
Next proof / success condition
Client-to-attested-recipient confidentiality requires a separately verified encrypted route; retention policy alone cannot establish it.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/privateinference.mddata/routes.json

No attested reply key Reply key + freshness

Finding in the reviewed system

Evidence reviewed
Standard routing makes no client-to-attested-model encryption claim; no reply key is bound to an attested workload.
Next proof / success condition
Reply confidentiality from the provider proxy requires a separately verified encrypted route; retention policy alone cannot establish it.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/privateinference.mddata/routes.json

No attested-route claim Code

Scope of this route

Evidence reviewed
The documented standard/ZDR path is not a client-verified TEE route.
Next proof / success condition
Keep this path outside an attested-confidentiality claim; evaluate any future private mode as a separate client/request path.

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/privateinference.mddata/routes.json

Proxy sees prompts Operator access

Finding in the reviewed system

Evidence reviewed
The documented standard route gives the proxy and selected supplier plaintext access outside a client-verified computation.
Next proof / success condition
Use a separately verified confidential route if the proxy/supplier must not read prompts; document any retained plaintext holders.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/privateinference.mddata/routes.json

Our worker test to do Downstream

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/privateinference.mddata/routes.json

No attested-route claim Updates

Scope of this route

Evidence reviewed
The documented standard/ZDR path is not a client-verified TEE route.
Next proof / success condition
Keep this path outside an attested-confidentiality claim; evaluate any future private mode as a separate client/request path.

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/privateinference.mddata/routes.json

Zero data retention Provider routing

Reviewed 2026-10-14 by codex-source-review (not independent audit approval) · source revision ffb5b54e8cb7 · Route inventory and raw checks

Not assessed on the comparison page: no automated adapter in this repository. Cells below are from the dated source review.

No attested-route claim Hardware

Scope of this route

Evidence reviewed
The documented standard/ZDR path is not a client-verified TEE route.
Next proof / success condition
Keep this path outside an attested-confidentiality claim; evaluate any future private mode as a separate client/request path.

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/privateinference.mddata/routes.json

Proxy sees prompts Recipient key

Finding in the reviewed system

Evidence reviewed
Zero data retention is a retention policy, not client-to-attested-model encryption.
Next proof / success condition
Client-to-attested-recipient confidentiality requires a separately verified encrypted route; retention policy alone cannot establish it.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/privateinference.mddata/routes.json

No attested reply key Reply key + freshness

Finding in the reviewed system

Evidence reviewed
Zero data retention is a retention policy, not client-to-attested-model encryption; no reply key is bound to an attested workload.
Next proof / success condition
Reply confidentiality from the provider proxy requires a separately verified encrypted route; retention policy alone cannot establish it.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/privateinference.mddata/routes.json

No attested-route claim Code

Scope of this route

Evidence reviewed
The documented standard/ZDR path is not a client-verified TEE route.
Next proof / success condition
Keep this path outside an attested-confidentiality claim; evaluate any future private mode as a separate client/request path.

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/privateinference.mddata/routes.json

Proxy sees prompts Operator access

Finding in the reviewed system

Evidence reviewed
The documented ZDR proxy still processes plaintext outside a client-verified computation; retention enforcement was not verified.
Next proof / success condition
Use a separately verified confidential route if the proxy/supplier must not read prompts; document any retained plaintext holders.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/privateinference.mddata/routes.json

Our worker test to do Downstream

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/privateinference.mddata/routes.json

No attested-route claim Updates

Scope of this route

Evidence reviewed
The documented standard/ZDR path is not a client-verified TEE route.
Next proof / success condition
Keep this path outside an attested-confidentiality claim; evaluate any future private mode as a separate client/request path.

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/privateinference.mddata/routes.json

Private beta via NEAR Intel TDX + GPU

Reviewed 2026-10-14 by codex-source-review (not independent audit approval) · source revision ffb5b54e8cb7 · Route inventory and raw checks

Not assessed on the comparison page: no automated adapter in this repository. Cells below are from the dated source review.

Our hardware test to do Hardware

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Retain a route-scoped replay authenticating hardware signatures, TCB/debug state and required CPU/GPU evidence under the actual platform policy.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/privateinference.mddata/routes.json

Our recipient test to do Recipient key

Evidence we still need

Evidence reviewed
No fixed deployment or complete session was verified.
Next proof / success condition
Verify a harmless prompt is encrypted only to a key bound to the attested model workload; reject a substituted recipient key before the prompt is sent.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/privateinference.mddata/routes.json

Reply key substitutable Reply key + freshness

Finding in the reviewed system

Evidence reviewed
The inspected client leaves the reply-key header substitutable by a relay. No fixed deployment or complete session was verified.
Next proof / success condition
Verify a fixed client/protocol/deployment rejects substituted reply keys before disclosure.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/privateinference.mddata/routes.json

No approved baseline Code

Finding in the reviewed system

Evidence reviewed
The inspected browser checks quote consistency but does not compare compose/OS identity against an independently approved baseline.
Next proof / success condition
Authenticate and enforce an explicitly approved model-to-compose/OS/artifact baseline; reject a genuine quote for a different workload.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/privateinference.mddata/routes.json

Our access review to do Operator access

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace operator credentials, debug access and unmeasured inputs to their effects on protected data in the exact deployed version; retain checks and counterexamples.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/privateinference.mddata/routes.json

Our model test to do Downstream

Evidence we still need

Evidence reviewed
The browser requests a model, but rejection of a substituted attested model was not established.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/privateinference.mddata/routes.json

No approved baseline Updates

Finding in the reviewed system

Evidence reviewed
The inspected client lacks an independent approved workload baseline; self-consistency does not require client approval of code changes.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback. Establish an authority independent of self-consistency.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/privateinference.mddata/routes.json

Maple

2 paths · Not sampled · Reference history

17 provider reference surfaces differ from the review baseline. Determine route relevance before relying on the dated assessment.

Hardware · Our hardware test to do

Via Tinfoil: Our hardware test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Retain a route-scoped replay authenticating hardware signatures, TCB/debug state and required CPU/GPU evidence under the actual platform policy.

Tracked work and acceptance tests: #30

Via Privatemode: Our hardware test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Retain a route-scoped replay authenticating hardware signatures, TCB/debug state and required CPU/GPU evidence under the actual platform policy.

Tracked work and acceptance tests: #30

Recipient key · Our recipient test to do

Via Tinfoil: Our recipient test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that a harmless prompt is encrypted only to a key bound to the attested workload and the live channel; reject a substituted recipient key before the prompt is sent.

Tracked work and acceptance tests: #30

Via Privatemode: Our recipient test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that a harmless prompt is encrypted only to a key bound to the attested workload and the live channel; reject a substituted recipient key before the prompt is sent.

Tracked work and acceptance tests: #30

Reply key + freshness · Our reply test to do

Via Tinfoil: Our reply test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that the reply key is bound to the same attested workload; reject a substituted reply key, stale evidence and replay before the reply is disclosed.

Tracked work and acceptance tests: #30

Via Privatemode: Our reply test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that the reply key is bound to the same attested workload; reject a substituted reply key, stale evidence and replay before the reply is disclosed.

Tracked work and acceptance tests: #30

Code · Our EIF proof to do

Via Tinfoil: Our EIF proof to do — Evidence we still need

Evidence we still need

Evidence reviewed
PCR0 publication and SDK policy were source-reviewed; no independent EIF rebuild or live source-to-deployment mapping was established.
Next proof / success condition
Reproduce the exact Nitro EIF on a remote builder, map all included binaries to source, and bind a live authenticated measurement to that artifact.

Tracked work and acceptance tests: #30

Via Privatemode: Our EIF proof to do — Evidence we still need

Evidence we still need

Evidence reviewed
The Nitro source includes a Continuum sidecar; no independent EIF rebuild or deployed source-to-binary mapping was established.
Next proof / success condition
Reproduce the exact Nitro EIF on a remote builder, map all included binaries to source, and bind a live authenticated measurement to that artifact.

Tracked work and acceptance tests: #30

Operator access · Our access review to do

Via Tinfoil: Our access review to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace operator credentials, debug access and unmeasured inputs to their effects on protected data in the exact deployed version; retain checks and counterexamples.

Tracked work and acceptance tests: #30

Via Privatemode: Our access review to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace operator credentials, debug access and unmeasured inputs to their effects on protected data in the exact deployed version; retain checks and counterexamples.

Tracked work and acceptance tests: #30

Downstream · Our worker test to do

Via Tinfoil: Our worker test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Via Privatemode: Our worker test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Updates · Our policy test to do

Via Tinfoil: Our policy test to do — Evidence we still need

Evidence we still need

Evidence reviewed
Embedded roots and signed PCR0 histories can admit old values. This is source-reviewed policy, not verified live update enforcement.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback. Include embedded-root and signed-history admission plus onward backend policy.

Tracked work and acceptance tests: #30

Via Privatemode: Our policy test to do — Evidence we still need

Evidence we still need

Evidence reviewed
PCR0 history and embedded policy were inspected; sidecar admission and onward update enforcement remain unverified.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback. Include embedded-root and signed-history admission plus onward backend policy.

Tracked work and acceptance tests: #30

Via Tinfoil Nitro → Tinfoil

Reviewed 2026-10-14 by codex-source-review (not independent audit approval) · source revision ffb5b54e8cb7 · Route inventory and raw checks

Not assessed on the comparison page: no automated adapter in this repository. Cells below are from the dated source review.

Our hardware test to do Hardware

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Retain a route-scoped replay authenticating hardware signatures, TCB/debug state and required CPU/GPU evidence under the actual platform policy.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/maple.mddata/routes.json

Our recipient test to do Recipient key

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that a harmless prompt is encrypted only to a key bound to the attested workload and the live channel; reject a substituted recipient key before the prompt is sent.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/maple.mddata/routes.json

Our reply test to do Reply key + freshness

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that the reply key is bound to the same attested workload; reject a substituted reply key, stale evidence and replay before the reply is disclosed.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/maple.mddata/routes.json

Our EIF proof to do Code

Evidence we still need

Evidence reviewed
PCR0 publication and SDK policy were source-reviewed; no independent EIF rebuild or live source-to-deployment mapping was established.
Next proof / success condition
Reproduce the exact Nitro EIF on a remote builder, map all included binaries to source, and bind a live authenticated measurement to that artifact.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/maple.mddata/routes.json

Our access review to do Operator access

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace operator credentials, debug access and unmeasured inputs to their effects on protected data in the exact deployed version; retain checks and counterexamples.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/maple.mddata/routes.json

Our worker test to do Downstream

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/maple.mddata/routes.json

Our policy test to do Updates

Evidence we still need

Evidence reviewed
Embedded roots and signed PCR0 histories can admit old values. This is source-reviewed policy, not verified live update enforcement.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback. Include embedded-root and signed-history admission plus onward backend policy.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/maple.mddata/routes.json

Via Privatemode Nitro → Privatemode

Reviewed 2026-10-14 by codex-source-review (not independent audit approval) · source revision ffb5b54e8cb7 · Route inventory and raw checks

Not assessed on the comparison page: no automated adapter in this repository. Cells below are from the dated source review.

Our hardware test to do Hardware

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Retain a route-scoped replay authenticating hardware signatures, TCB/debug state and required CPU/GPU evidence under the actual platform policy.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/maple.mddata/routes.json

Our recipient test to do Recipient key

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that a harmless prompt is encrypted only to a key bound to the attested workload and the live channel; reject a substituted recipient key before the prompt is sent.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/maple.mddata/routes.json

Our reply test to do Reply key + freshness

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that the reply key is bound to the same attested workload; reject a substituted reply key, stale evidence and replay before the reply is disclosed.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/maple.mddata/routes.json

Our EIF proof to do Code

Evidence we still need

Evidence reviewed
The Nitro source includes a Continuum sidecar; no independent EIF rebuild or deployed source-to-binary mapping was established.
Next proof / success condition
Reproduce the exact Nitro EIF on a remote builder, map all included binaries to source, and bind a live authenticated measurement to that artifact.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/maple.mddata/routes.json

Our access review to do Operator access

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace operator credentials, debug access and unmeasured inputs to their effects on protected data in the exact deployed version; retain checks and counterexamples.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/maple.mddata/routes.json

Our worker test to do Downstream

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/maple.mddata/routes.json

Our policy test to do Updates

Evidence we still need

Evidence reviewed
PCR0 history and embedded policy were inspected; sidecar admission and onward update enforcement remain unverified.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback. Include embedded-root and signed-history admission plus onward backend policy.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/maple.mddata/routes.json

Privatemode

3 paths · Not sampled · Reference history

2 provider reference surfaces differ from the review baseline. Determine route relevance before relying on the dated assessment.

Hardware · Our hardware test to do

Automatic manifest: Our hardware test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Retain a route-scoped replay authenticating hardware signatures, TCB/debug state and required CPU/GPU evidence under the actual platform policy.

Tracked work and acceptance tests: #30

Pinned manifest: Our hardware test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Retain a route-scoped replay authenticating hardware signatures, TCB/debug state and required CPU/GPU evidence under the actual platform policy.

Tracked work and acceptance tests: #30

Proxyless API: Our hardware test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Retain a route-scoped replay authenticating hardware signatures, TCB/debug state and required CPU/GPU evidence under the actual platform policy.

Tracked work and acceptance tests: #30

Recipient key · Routes differ

Automatic manifest: Our recipient test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that a harmless prompt is encrypted only to a key bound to the attested workload and the live channel; reject a substituted recipient key before the prompt is sent.

Tracked work and acceptance tests: #30

Pinned manifest: Our recipient test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that a harmless prompt is encrypted only to a key bound to the attested workload and the live channel; reject a substituted recipient key before the prompt is sent.

Tracked work and acceptance tests: #30

Proxyless API: No client attestation — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
The reviewed documentation describes no client-side attestation for this path; the client does not verify the recipient through the proxy protocol.
Next proof / success condition
Establish and test client-side recipient verification for any confidential claim; provider-side checks alone are a different trust model.

Tracked work and acceptance tests: #30

Reply key + freshness · Routes differ

Automatic manifest: Our reply test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that the reply key is bound to the same attested workload; reject a substituted reply key, stale evidence and replay before the reply is disclosed.

Tracked work and acceptance tests: #30

Pinned manifest: Our reply test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that the reply key is bound to the same attested workload; reject a substituted reply key, stale evidence and replay before the reply is disclosed.

Tracked work and acceptance tests: #30

Proxyless API: No client attestation — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
The reviewed documentation describes no client-side attestation for this path; the client binds no reply key to an attested workload.
Next proof / success condition
Establish and test client-side reply-key binding and freshness for any confidential claim; provider-side checks alone are a different trust model.

Tracked work and acceptance tests: #30

Code · Our code proof to do

Automatic manifest: Our code proof to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace the deployed measured artifact to reviewed source/build and model identities, including mutable configuration; test an unapproved artifact.

Tracked work and acceptance tests: #30

Pinned manifest: Our code proof to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace the deployed measured artifact to reviewed source/build and model identities, including mutable configuration; test an unapproved artifact.

Tracked work and acceptance tests: #30

Proxyless API: Our code proof to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace the deployed measured artifact to reviewed source/build and model identities, including mutable configuration; test an unapproved artifact.

Tracked work and acceptance tests: #30

Operator access · Our access review to do

Automatic manifest: Our access review to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace operator credentials, debug access and unmeasured inputs to their effects on protected data in the exact deployed version; retain checks and counterexamples.

Tracked work and acceptance tests: #30

Pinned manifest: Our access review to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace operator credentials, debug access and unmeasured inputs to their effects on protected data in the exact deployed version; retain checks and counterexamples.

Tracked work and acceptance tests: #30

Proxyless API: Our access review to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace operator credentials, debug access and unmeasured inputs to their effects on protected data in the exact deployed version; retain checks and counterexamples.

Tracked work and acceptance tests: #30

Downstream · Our worker test to do

Automatic manifest: Our worker test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Pinned manifest: Our worker test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Proxyless API: Our worker test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Updates · Routes differ

Automatic manifest: Provider updates auto-adopted — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
Documented default behavior may adopt the latest provider manifest after a validation failure; this does not require independent approval of the changed reference.
Next proof / success condition
Demonstrate independent approval before changed manifest admission if that guarantee is claimed; preserve the documented automatic mode as a distinct policy.

Tracked work and acceptance tests: #30

Pinned manifest: Our pin-refusal test — Evidence we still need

Evidence we still need

Evidence reviewed
A local manifest pin provides an approval boundary in the documented configuration. A mutated-reference fail-closed request test remains outstanding.
Next proof / success condition
Pin the exact client and manifest, then retain a harmless accepted request plus refusal of changed/unapproved manifest values and rollback.

Tracked work and acceptance tests: #30

Proxyless API: Our policy test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback.

Tracked work and acceptance tests: #30

Automatic manifest Attesting proxy

Reviewed 2026-10-14 by codex-source-review (not independent audit approval) · source revision ffb5b54e8cb7 · Route inventory and raw checks

Not assessed on the comparison page: no automated adapter in this repository. Cells below are from the dated source review.

Our hardware test to do Hardware

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Retain a route-scoped replay authenticating hardware signatures, TCB/debug state and required CPU/GPU evidence under the actual platform policy.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Our recipient test to do Recipient key

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that a harmless prompt is encrypted only to a key bound to the attested workload and the live channel; reject a substituted recipient key before the prompt is sent.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Our reply test to do Reply key + freshness

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that the reply key is bound to the same attested workload; reject a substituted reply key, stale evidence and replay before the reply is disclosed.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Our code proof to do Code

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace the deployed measured artifact to reviewed source/build and model identities, including mutable configuration; test an unapproved artifact.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Our access review to do Operator access

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace operator credentials, debug access and unmeasured inputs to their effects on protected data in the exact deployed version; retain checks and counterexamples.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Our worker test to do Downstream

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Provider updates auto-adopted Updates

Finding in the reviewed system

Evidence reviewed
Documented default behavior may adopt the latest provider manifest after a validation failure; this does not require independent approval of the changed reference.
Next proof / success condition
Demonstrate independent approval before changed manifest admission if that guarantee is claimed; preserve the documented automatic mode as a distinct policy.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Pinned manifest Attesting proxy

Reviewed 2026-10-14 by codex-source-review (not independent audit approval) · source revision ffb5b54e8cb7 · Route inventory and raw checks

Not assessed on the comparison page: no automated adapter in this repository. Cells below are from the dated source review.

Our hardware test to do Hardware

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Retain a route-scoped replay authenticating hardware signatures, TCB/debug state and required CPU/GPU evidence under the actual platform policy.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Our recipient test to do Recipient key

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that a harmless prompt is encrypted only to a key bound to the attested workload and the live channel; reject a substituted recipient key before the prompt is sent.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Our reply test to do Reply key + freshness

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that the reply key is bound to the same attested workload; reject a substituted reply key, stale evidence and replay before the reply is disclosed.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Our code proof to do Code

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace the deployed measured artifact to reviewed source/build and model identities, including mutable configuration; test an unapproved artifact.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Our access review to do Operator access

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace operator credentials, debug access and unmeasured inputs to their effects on protected data in the exact deployed version; retain checks and counterexamples.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Our worker test to do Downstream

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Our pin-refusal test Updates

Evidence we still need

Evidence reviewed
A local manifest pin provides an approval boundary in the documented configuration. A mutated-reference fail-closed request test remains outstanding.
Next proof / success condition
Pin the exact client and manifest, then retain a harmless accepted request plus refusal of changed/unapproved manifest values and rollback.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Proxyless API Provider-mediated

Reviewed 2026-10-14 by codex-source-review (not independent audit approval) · source revision ffb5b54e8cb7 · Route inventory and raw checks

Not assessed on the comparison page: no automated adapter in this repository. Cells below are from the dated source review.

Our hardware test to do Hardware

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Retain a route-scoped replay authenticating hardware signatures, TCB/debug state and required CPU/GPU evidence under the actual platform policy.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

No client attestation Recipient key

Finding in the reviewed system

Evidence reviewed
The reviewed documentation describes no client-side attestation for this path; the client does not verify the recipient through the proxy protocol.
Next proof / success condition
Establish and test client-side recipient verification for any confidential claim; provider-side checks alone are a different trust model.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

No client attestation Reply key + freshness

Finding in the reviewed system

Evidence reviewed
The reviewed documentation describes no client-side attestation for this path; the client binds no reply key to an attested workload.
Next proof / success condition
Establish and test client-side reply-key binding and freshness for any confidential claim; provider-side checks alone are a different trust model.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Our code proof to do Code

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace the deployed measured artifact to reviewed source/build and model identities, including mutable configuration; test an unapproved artifact.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Our access review to do Operator access

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Trace operator credentials, debug access and unmeasured inputs to their effects on protected data in the exact deployed version; retain checks and counterexamples.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Our worker test to do Downstream

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Our policy test to do Updates

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Confidential AI

1 paths · Not sampled · Reference history

5 provider reference surfaces differ from the review baseline. Determine route relevance before relying on the dated assessment.

Hardware · Our hardware test to do

Inference API: Our hardware test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Retain a route-scoped replay authenticating hardware signatures, TCB/debug state and required CPU/GPU evidence under the actual platform policy.

Tracked work and acceptance tests: #30

Recipient key · Our recipient test to do

Inference API: Our recipient test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that a harmless prompt is encrypted only to a key bound to the attested workload and the live channel; reject a substituted recipient key before the prompt is sent.

Tracked work and acceptance tests: #30

Reply key + freshness · Our reply test to do

Inference API: Our reply test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that the reply key is bound to the same attested workload; reject a substituted reply key, stale evidence and replay before the reply is disclosed.

Tracked work and acceptance tests: #30

Code · Our release check to do

Inference API: Our release check to do — Work in our verifier

Work in our verifier

Evidence reviewed
Release and allowlist bytes were captured, but signatures, live measurements and source-to-artifact mapping were not verified.
Next proof / success condition
Authenticate the exact release and allowlist, link them to measured live artifacts, and preserve source/build mapping with rejection controls.

Tracked work and acceptance tests: #30

Operator access · Operator access paths

Inference API: Operator access paths — Finding in the reviewed system

Finding in the reviewed system

Evidence reviewed
The dated threat-model review identifies operator key and unencrypted control-plane access paths; public attestation alone does not close them.
Next proof / success condition
For the reviewed deployment, establish that operator-key/control-plane access cannot alter or read protected computation; retain evidence of enforcement.

Tracked work and acceptance tests: #30

Downstream · Our worker test to do

Inference API: Our worker test to do — Evidence we still need

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Updates · Our refusal test to do

Inference API: Our refusal test to do — Evidence we still need

Evidence we still need

Evidence reviewed
Published allowlists and client pins exist in the design; actual update refusal and the same-session proof were not tested.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback. Bind the actual same-session client pins and release policy.

Tracked work and acceptance tests: #30

Inference API TDX mesh

Reviewed 2026-10-14 by codex-source-review (not independent audit approval) · source revision ffb5b54e8cb7 · Route inventory and raw checks

Not assessed on the comparison page: no automated adapter in this repository. Cells below are from the dated source review.

Our hardware test to do Hardware

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Retain a route-scoped replay authenticating hardware signatures, TCB/debug state and required CPU/GPU evidence under the actual platform policy.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Our recipient test to do Recipient key

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that a harmless prompt is encrypted only to a key bound to the attested workload and the live channel; reject a substituted recipient key before the prompt is sent.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Our reply test to do Reply key + freshness

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Verify that the reply key is bound to the same attested workload; reject a substituted reply key, stale evidence and replay before the reply is disclosed.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Our release check to do Code

Work in our verifier

Evidence reviewed
Release and allowlist bytes were captured, but signatures, live measurements and source-to-artifact mapping were not verified.
Next proof / success condition
Authenticate the exact release and allowlist, link them to measured live artifacts, and preserve source/build mapping with rejection controls.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Operator access paths Operator access

Finding in the reviewed system

Evidence reviewed
The dated threat-model review identifies operator key and unencrypted control-plane access paths; public attestation alone does not close them.
Next proof / success condition
For the reviewed deployment, establish that operator-key/control-plane access cannot alter or read protected computation; retain evidence of enforcement.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Our worker test to do Downstream

Evidence we still need

Evidence reviewed
This route has no retained proof establishing this assessment axis. A source or publication capture is insufficient.
Next proof / success condition
Bind the selected worker and requested model to the actual request; reject worker/model substitution and insecure fallback.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json

Our refusal test to do Updates

Evidence we still need

Evidence reviewed
Published allowlists and client pins exist in the design; actual update refusal and the same-session proof were not tested.
Next proof / success condition
Identify the actual client approval authority and policy revision, and test refusal of unapproved updates and rollback. Bind the actual same-session client pins and release policy.

Tracked work and acceptance tests: #30

Evidence scope: dated repository source/review; no new live route proof is implied.

research/review-2026-10-13/landscape.mddata/routes.json