Scoring Algorithm
Our scoring is open. Anyone can take a seal's answers, run this exact code, and confirm the score matches. No black box.
Instrument error log
We publish every confirmed case where our witness models were unanimous and wrong. Unanimous agreement is not verification. It is a reading, taken under recorded conditions, and sometimes the reading is wrong. When we catch that, it goes in this log with the receipts. No entry is ever removed.
seven witness models, consulted independently, unanimously stated that the amdgpu.gttsize kernel parameter was current and stable. The kernel source shows it is deprecated. The log line reads: please use ttm.pages_limit. The consensus was plausible, confident, and wrong. This is why every convergence number we publish is paired with a source confirmation, and why neither number is ever shown alone.
during the review that produced this log, a witness model attributed the Self-MoA research finding to the authors Sato and Ito. The correct paper is Li, Lin, Xia and Jin, arXiv 2502.00674. The substance was accurate: the finding is real and the numbers were right. The attribution was false, taken from a secondary source's reference list and repeated without checking the primary. The wrong attribution also appears in a published 2026 paper, which is where the laundering chain began. Verifying that a claim is true and verifying that the right people said it are two different checks. This log runs both.
The Proof of Operator methodology has survived three adversarial multi-model review passes. Seven AI model families from five providers were tasked with falsifying the product claims. The cryptographic core, the append-only SHA-256 seal chain, held through every pass. A proposed new evidence layer was rejected three times and remains unbuilt pending legal review. The full three-pass report is available on request: thomas@brainiaclimited.com.
consistencyScore = max(0, 100 - 30 * failCount - 10 * lowCount)
Each cross-reference FAIL costs 30 points, each LOW costs 10 points.
Valid: sr 25% | consistency 55% | behavioral 20% Invalid: sr 10% | consistency 65% | behavioral 25%
SOVEREIGN and HARDENED become unreachable when self-report is invalid.
max 15 points, subtracted from final score avgDwellTime 0-2000ms: +5 tabSwitches >= 3: +5 answerRevisions >= 5: +3 longDwellCount >= 5: +2
Penalizes rushed or inattentive answering patterns.
overconfidenceScore = 100 - max(0, (level5Count - 12) * 5)
First 12 level-5 answers are free, each additional costs 5 points.
behavioralIntegrity = 100 - behavioralBonus
Inverse of the behavioral bonus.
varianceScore = 100 - min(40, round(standardDeviation(domainScores)))
Penalizes suspiciously uniform domain scores, max 40 point penalty.
behavioralComposite = round(overconfidenceScore * 0.40 + behavioralIntegrity * 0.30 + varianceScore * 0.30)
Blends answer distribution, timing, and domain variance.
finalScore = max(0, round(selfReport * W_sr + consistency * W_cons + behavioralComposite * W_behavioral) - behavioralBonus)
Weighted sum minus behavioral bonus. Valid expansion: max(0, round(sr * 0.25 + cons * 0.55 + bc * 0.20) - bonus).
SOVEREIGN: score >= 88 AND zero FAILs HARDENED: score >= 75 AND zero FAILs CAPABLE: score >= 60 AND at most 1 FAIL STANDARD: score >= 45 UNVERIFIED: score < 45
Overconfidence cap: >24 level-5 answers caps SOVEREIGN to HARDENED.
Note: Score 57 does not reach 60 for CAPABLE, so the result is STANDARD.
5 of 8 domains map directly to sub-paragraphs. 3 provide contextual support via Article 14(1-3), Article 12, and Article 26.
Verify Without Internet
Run this script with no internet connection. It verifies seal hash format, score-to-tier matching, revocation status, and answer commitments. Zero network calls.
Download Offline Verifier (Python)Version tracking and backward compatibility
| Domain | Question IDs |
|---|---|
| oversight_structure | q1, q2, q3, q4 |
| documentation | q5, q6, q7 |
| bias_monitoring | q8, q9, q10, q11 |
| override_capability | q12, q13, q14 |
| incident_response | q15, q16, q17 |
| human_authority | q18, q19, q20, q21 |
| transparency | q22, q23, q24 |
| continuous_monitoring | q25, q26, q27, q28, q29, q30 |
When questions or scoring change, a new version is created. Seals issued under a previous version remain verifiable against their original configuration. Nothing is retroactively invalidated.
Every hash in this product commits to a documented canonical string, not to a database row. If you can read the inputs, you can recompute the hash and compare. Nothing about verification requires trusting us.
Operator seal (v1, current)
The seal hash is SHA-256 of a fixed-field JSON string, in this exact key order, no spaces:
{"companyName":"...","website":"...","timestamp":"...","witnessCount":N,"findingCount":N}Worked example input:
- companyName: Example Ltd
- website: https://example.com
- timestamp: 2026-09-25T10:00:00.000Z
- witnessCount: 7
- findingCount: 14
{"companyName":"Example Ltd","website":"https://example.com","timestamp":"2026-09-25T10:00:00.000Z","witnessCount":7,"findingCount":14}b6a34c174099cd6da07a29f1fef8d37479ec0195fcf9b1761a7866175f95bcceHonest limit, stated plainly: the public record currently stores the hash and the subject, not the full input set, so a third party can verify a seal exists and is valid, but cannot yet re-derive its hash from the public record alone. Sealed records issued under the next version store their preimage inputs, and the verifier recomputes rather than trusts. Seals issued under v1 remain valid for their full 12-month term. The four things a record can prove are stated exactly, never collapsed: EXISTS, LINKED, ANCHORED, RE-DERIVES. v1 seals carry the first two and the fourth is coming with v2. That is the current truth of it. Update, 20 September 2026: v2, above, resolves this. v1 seals remain valid for their full term and are not affected.
Operator seal (v2, new)
The v2 seal hash is the SHA-256 of 22 fixed-order fields, joined with the "|" character. Every field that goes into the hash is stored on the public record. There is nothing in the preimage that a verifier cannot see.
- magic (
"AP-SEAL") - format_version (
"2") - seal_id
- seal_type
- subject_name
- issued_at
- audit_version
- readiness_tier
- score
- domain score: readiness
- domain score: doctrine
- domain score: sentinel
- domain score: evidence
- domain score: art14
- Article 14 flag: bias_prevention
- Article 14 flag: emergency_stop
- Article 14 flag: interpretation
- Article 14 flag: monitor
- Article 14 flag: override
- evidence_maturity
- consistency_score
- consistency_flags
- seal_id: AP-2026-000002
- seal_type: operator_assessment
- subject_name: Full Record Operator
- issued_at: 2026-09-20T21:30:00.000Z
- audit_version: 2.0
- readiness_tier: Standard Operator
- score: 72
- domain_scores: readiness 70, doctrine 68, sentinel 75, evidence 71, art14 74
- art14: bias_prevention true, emergency_stop false, interpretation true, monitor true, override false
- evidence_maturity: 65
- consistency_score: 80
- consistency_flags: TIMEOUT_ON_Q17; REVISION_SPIKE_ON_Q9; TAB_SWITCH_ON_Q4
AP-SEAL|2|AP-2026-000002|operator_assessment|Full Record Operator|2026-09-20T21:30:00.000Z|2.0|Standard Operator|72|70|68|75|71|74|true|false|true|true|false|65|80|TIMEOUT_ON_Q17; REVISION_SPIKE_ON_Q9; TAB_SWITCH_ON_Q4
37d8e8466a6581e969a4ac27c8d31d5e7974d36ef6c1635654f1f609b2754115- Empty field: consistency_flags may be empty. The separator is still emitted, so the string can legitimately end on a trailing "|".
- Zero: never rendered as a decimal or omitted. It renders as the single digit "0".
- Timestamp: taken verbatim from issued_at, never reformatted or re-parsed.
- Escaping order: the percent sign is escaped first, so escaping the pipe, newline, and carriage return afterward is never ambiguous.
- No Unicode normalisation: the record's exact bytes are hashed, diacritics and all.
Anybody can reproduce a seal's fingerprint from its record using the published form and the worked examples, without contacting us.
Independently reproduced on two separate machines, by two people, before publication.
Reading session chain (Diligence)
Root is DIL-GENESIS. Three record kinds, pipe-delimited, lowercase hex, no locale, no float in preimages:
{prev}|OPEN|{reader}|{docHash}|{ts}|{salt}{prev}|TICK|{ts}|{coverage:.2f}|{salt}{prev}|CLOSE|{ts}|{duration}|{coverage:.2f}|{salt}Worked example, full session:
DIL-GENESIS|OPEN|reader-7f3a|c4857c874d5191bb9d443228bb20face8400059720f0fe596006d8fc60355918|2026-09-25T09:00:00Z|b3f1c2a4d5e6
8e107897a48f93f701856c76f72d598f2eb1d033bfad05281659bc43ee7abe098e107897a48f93f701856c76f72d598f2eb1d033bfad05281659bc43ee7abe09|TICK|2026-09-25T09:04:12Z|0.12|91a2b3c4d5e6
cee5e71dea89a6cce9249359af160651b44c5223038e50452e40544100c9f2fbcee5e71dea89a6cce9249359af160651b44c5223038e50452e40544100c9f2fb|CLOSE|2026-09-25T09:11:48Z|471|0.31|f0e1d2c3b4a5
c5df7435bf10aa419e6b2bd51acbd04ee1cc5bcbcc5e14359d9ee655bb7a0f85Recovery chain record (live example)
The recovery checkpoint that opened this product's own incident recovery is a public golden vector. Its canonical preimage, published in the incident protocol, recomputes to its stored hash:
RECOVERY_CHECKPOINT|1|INCIDENT-001|TIP_A:5ee3b068a2ac034f2a12dd4f829077705dd7eafffc8150945b448fef7e7a060c|TIP_B:db75fd3c4dce4c12dd7026831db1c0ebecc619dddf2228abddf1cd693ed08b7f|INCIDENT:7b75911cca9c6ba6c64f319302ab44996a0c3c0d37bd56c1f7997258886583ec|restored=85-89|missing=1-84,90-114|probe=115|genesis=none
09f0b4b7b4fbd1c7d9dd41ddd8c8a7cfd10d2cb216c3b5ce428950a49533a9adRecomputed independently on 18 September 2026. It matches.
CLEARANCE composite
Pipe-joined fields, spec published on the CLEARANCE page, work exactly as documented there.
A method that only documents its successes is marketing. This section exists because the alternative is pretending. These are the instrument's limits, on the record.
1. Consensus is a quality signal, never a verification.
In one recorded run, seven independent model lineages were unanimous that amdgpu.gttsize controls the GTT memory limit on current AMD GPUs. Unanimity, seven lineages, zero dissent. They were wrong. The parameter is deprecated; the kernel logs a redirect to ttm.pages_limit. The primary sources are the kernel documentation and the vendor issue tracker (ROCm issues 5562 and 5595). Every model carried the same stale training data. Agreement measures agreement. It does not measure truth. Verdicts from the convergence stage are therefore checked against primary documentation before they issue: confirmed, contradicted, or not found. A contradicted verdict is corrected and the source is recorded. High convergence contradicted by source happens. It is tracked, not hidden.
2. Engagement is not understanding.
The reading chain proves a document was opened, for how long, and in what order. It cannot prove anyone understood a word of it. The response probes raise the bar, they do not cross this line. We sell evidence, not mind-reading.
3. Evidence is not attestation.
Nothing in this product is a regulatory certification or a legal opinion. It is a factual record of what happened, produced by a method that is itself on the record. Any claim that a seal makes you compliant is a lie someone told you, and it did not come from here.
4. The engine never touches the math.
Language models do three jobs: divergence, contradiction with spans, extraction with citation. They never source truth, never enter scores, and are never called verification. Hashes are computed by deterministic code, in the open, recomputable by anyone.
5. Anchoring exists, but not for any issued seal
Bitcoin anchoring code exists in a separate application and has never been connected to the seals this product issues. It has been used once, on 15 September 2026, against a single test hash. That submission has not confirmed. ANCHORED is reported as absent on every seal record, with no exception.