Certificate of AI Acceptance SEAL-seal-llm-qa-20260729-2159-run-1

Certificate of AI Acceptance

SEAL-seal-llm-qa-20260729-2159-run-1 · session seal-llm-qa-20260729-2159 · SEAL Protocol 0.2

This certificate verifies itself. Every figure below can be recomputed from the evidence chain embedded in this file, and RFC 3161 timestamps establish that the test data was revealed only after the model was sealed. Ground-truth labels never left the examiner, and the metrics were computed by the examiner's server.

It holds without trusting the issuer. The approval is signed with the approver's own key, pinned to the ledger when the session opened, and the issuer does not hold that key — the issuer alone cannot produce this document. The timestamps were issued by a third-party TSA. Cross-checking the ledger is optional: the verification above still holds if the examiner's server disappears.

— runs from this file alone, no network

Decision

Verdict	PASS contracted bar: abstention_rate/v1 <=0.10 numeric_match/v1 >=0.90 (fixed at session open)
Assurance	Ledger-Born Model Verified (Level A) execution E1 The prediction file was written by the applicant's own script. Inference ran on applicant hardware (labels withheld; scoring on the examiner's server). A model detecting that it is under examination and behaving differently is not excluded at this level — excluding it requires examiner-side execution (E2), which is not offered today.
Measured	abstention_rate/v1 = 0.0000 numeric_match/v1 = 0.9846 (computed by the examiner's server)
Exam method	One sample — 문항 1개 (지문 없이 독립된 문항) Label — 수치 정답 하나. 유리수로 정확히 파싱해 비교 — 자릿점·후행 0은 표기 변형으로 같게 본다 Submitted value — 모델이 낸 수치 답. 답하지 않기로 했으면 __abstain__ 3회차. 지표와 합격선은 1·2회차와 동일하게 유지했다 — 바뀐 것은 응시 모델뿐이다(claude-haiku-4-5 → claude-sonnet-5). 결과를 짐작하고 합격선을 내리는 것은 이 프로토콜이 막으려는 것이므로 기준은 손대지 않았다. 문항 297건은 새로 생성했고 앞선 세 세트(1회차 300, 2회차 300, 리허설 40)와 지문·본문 모두 겹치지 않는다. 2회차에서 모호성 탐침을 본시험 문항으로 돌린 실수가 있어 그 문항 집합은 폐기했다. 문항 출처: authored. 응시 대상은 claude-sonnet-5 + 고정 프롬프트 + 학습분할 few-shot의 묶음. 디코딩 온도는 CLI 기본값이라 고정하지 못했다 — 이 숫자는 정확히 재현되지 않는다. fixed at session open · fingerprint sha256:b6a263689d4b37fb2a6fbd29843352a5e48fc78d5ce96d5558de4a9b86a9ee41
Validity	Not declared — this certificate does not expire, but if the deployment has changed since the examination the result no longer describes it. Elapsed time is shown on the project badge and console.
Issuance	Self-examination (applicant = issuer — self-attestation is capped at Level A)
Test set history	First reveal in this examination · 1 global reveal(s) · 1 sealed candidate(s)
Cumulative attempts	reveal number 3 against data pool seal-llm-qa (drawing a fresh paper from the same pool does not reset this · the pool was declared by the applicant, so this count is self-reported)
Test sample selection	Hash assignment (examiner-committed seed, split/v1) · the statistical validity of the split is not assured
Model fingerprint	sha256:97381a8036e2d6c436b99da7275f3eea6f920bf78016966cb443547468fc949d
Test set fingerprint	sha256:644699a0d423664d38944dfce008e2443d3596195c6212549693d2e2efc3ca40
Approval	signed by a human approver (key b361a822518f1eb4, pinned at session open) · issuer key b281d3767e970a6d

Verdict

PASS contracted bar: abstention_rate/v1 <=0.10 numeric_match/v1 >=0.90 (fixed at session open)

Assurance

Ledger-Born Model Verified (Level A) execution E1

The prediction file was written by the applicant's own script.

Inference ran on applicant hardware (labels withheld; scoring on the examiner's server). A model detecting that it is under examination and behaving differently is not excluded at this level — excluding it requires examiner-side execution (E2), which is not offered today.

Measured

abstention_rate/v1 = 0.0000 numeric_match/v1 = 0.9846 (computed by the examiner's server)

Exam method

One sample — 문항 1개 (지문 없이 독립된 문항)

Label — 수치 정답 하나. 유리수로 정확히 파싱해 비교 — 자릿점·후행 0은 표기 변형으로 같게 본다

Submitted value — 모델이 낸 수치 답. 답하지 않기로 했으면 __abstain__

3회차. 지표와 합격선은 1·2회차와 동일하게 유지했다 — 바뀐 것은 응시 모델뿐이다(claude-haiku-4-5 → claude-sonnet-5). 결과를 짐작하고 합격선을 내리는 것은 이 프로토콜이 막으려는 것이므로 기준은 손대지 않았다. 문항 297건은 새로 생성했고 앞선 세 세트(1회차 300, 2회차 300, 리허설 40)와 지문·본문 모두 겹치지 않는다. 2회차에서 모호성 탐침을 본시험 문항으로 돌린 실수가 있어 그 문항 집합은 폐기했다. 문항 출처: authored. 응시 대상은 claude-sonnet-5 + 고정 프롬프트 + 학습분할 few-shot의 묶음. 디코딩 온도는 CLI 기본값이라 고정하지 못했다 — 이 숫자는 정확히 재현되지 않는다.

fixed at session open · fingerprint sha256:b6a263689d4b37fb2a6fbd29843352a5e48fc78d5ce96d5558de4a9b86a9ee41

Validity

Not declared — this certificate does not expire, but if the deployment has changed since the examination the result no longer describes it. Elapsed time is shown on the project badge and console.

Issuance

Self-examination (applicant = issuer — self-attestation is capped at Level A)

Test set history

First reveal in this examination · 1 global reveal(s) · 1 sealed candidate(s)

Cumulative attempts

reveal number 3 against data pool seal-llm-qa (drawing a fresh paper from the same pool does not reset this · the pool was declared by the applicant, so this count is self-reported)

Test sample selection

Hash assignment (examiner-committed seed, split/v1) · the statistical validity of the split is not assured

Model fingerprint

sha256:97381a8036e2d6c436b99da7275f3eea6f920bf78016966cb443547468fc949d

Test set fingerprint

sha256:644699a0d423664d38944dfce008e2443d3596195c6212549693d2e2efc3ca40

Approval

signed by a human approver (key b361a822518f1eb4, pinned at session open) · issuer key b281d3767e970a6d

Evidence chain

#	Event	entry_hash
75	SESSION_OPENED	sha256:21246b313180588a457c3016ad0c44878a9dd2ee7060b3cd502d4cde235c6448
76	SAMPLES_COMMITTED	sha256:f0c3afe6296dc1604bd79cda9eba80a70f419b50b6d277e61a91cea25bb89e0e
77	RELEASE_REGISTERED	sha256:185a7f20fed0735360e1dc9a6d75b77031fe9f17c35ffeb7d2985d7e2bf6e67f
78	ESCROW_LOCKED	sha256:adc705f9a9ae391a6951c9755009576e2662d1a5fbef02dcc15a916a2e46011b
79	RUN_STARTED	sha256:4c8ffa40e24ec4d8e6a6e16bc04cbb55c36261619c5ce4a65eaab243deff8b15
80	TRAIN_PULLED	sha256:ab77c8973940dbe741e7ea2090c309f0ecc09d09de637ea95871b9b7bdad7d26
81	MODEL_SEALED	sha256:9de4618e19524ce43bc6892123324a6f38b6922ce5f23518deaa2e2862a8c19b
82	TEST_REVEALED	sha256:b1da641cbb00434f1f0a5b9881e586db26972a33b9147a1c2e596b66e124d387
83	PREDICTIONS_SUBMITTED	sha256:133eee09ab86bc7b19996d3297777692c3cff56a180b1f04e5713fdec3f84b0d
84	EVAL_SCORED	sha256:6bd9ba1d4410e7e2ebcbcc96069221a6057fe228dc9c8298ae416d625db1bb0f
85	GRADE_DECIDED	sha256:5cd9eced3908a318cf4a00b45f78da6c1fcd59b68b58a28751e9b46eb8577f95
86	APPROVED	sha256:0831b767ba7557c0acadc523ddf9326065e0a6446bdcbbd4269607a5dca8476f

Event

entry_hash

SESSION_OPENED

sha256:21246b313180588a457c3016ad0c44878a9dd2ee7060b3cd502d4cde235c6448

SAMPLES_COMMITTED

sha256:f0c3afe6296dc1604bd79cda9eba80a70f419b50b6d277e61a91cea25bb89e0e

RELEASE_REGISTERED

sha256:185a7f20fed0735360e1dc9a6d75b77031fe9f17c35ffeb7d2985d7e2bf6e67f

ESCROW_LOCKED

sha256:adc705f9a9ae391a6951c9755009576e2662d1a5fbef02dcc15a916a2e46011b

RUN_STARTED

sha256:4c8ffa40e24ec4d8e6a6e16bc04cbb55c36261619c5ce4a65eaab243deff8b15

TRAIN_PULLED

sha256:ab77c8973940dbe741e7ea2090c309f0ecc09d09de637ea95871b9b7bdad7d26

MODEL_SEALED

sha256:9de4618e19524ce43bc6892123324a6f38b6922ce5f23518deaa2e2862a8c19b

TEST_REVEALED

sha256:b1da641cbb00434f1f0a5b9881e586db26972a33b9147a1c2e596b66e124d387

PREDICTIONS_SUBMITTED

sha256:133eee09ab86bc7b19996d3297777692c3cff56a180b1f04e5713fdec3f84b0d

EVAL_SCORED

sha256:6bd9ba1d4410e7e2ebcbcc96069221a6057fe228dc9c8298ae416d625db1bb0f

GRADE_DECIDED

sha256:5cd9eced3908a318cf4a00b45f78da6c1fcd59b68b58a28751e9b46eb8577f95

APPROVED

sha256:0831b767ba7557c0acadc523ddf9326065e0a6446bdcbbd4269607a5dca8476f

Timestamp anchors: seq 76 ← http://timestamp.digicert.com · seq 81 ← http://timestamp.digicert.com · seq 82 ← http://timestamp.digicert.com · the tokens themselves are in the embedded JSON.

What is not assured

How to re-verify independently

1. From this file alone (no server — this is the verification)
Extract the embedded JSON and run seal verify cert.json — all six steps (issuer signature · approver signature · hash chain · third-party timestamps · ordering invariant · grade re-decision). It passes with no network, and it passes if the examiner's server no longer exists. The button above performs the same checks in the browser, leaving only the TSA step (4) to the CLI.

2. Ledger cross-check (optional — whether the ledger changed after issuance)
seal verify --ledger https://seal-ledger.fly.dev --auth 706a7a4c999fe29e6e9f9364507ca5bcae710fd8c0cbf397126866a8428fb917 cert.json. If this step fails or the server cannot be reached, the conclusion of step 1 is unchanged.

Keep this file. Badges and web pages are conveniences; the evidence is this file.

QR = ledger address · cross-check token: 706a7a4c999fe29e6e9f9364507ca5bcae710fd8c0cbf397126866a8428fb917