Evidence-record contract¶
Status: proposed, authored 2026-09-08 under an operator ruling. Not yet ratified and not yet mechanically witnessed. It selects a repair direction for the evidence record; it does not attest OBPI-0.35.0-05 and does not authorize Stage 5.
Origin¶
OBPI-0.35.0-05 ran eleven Step-4b adversarial rounds. No round from 8 through 11 demonstrated a production defect. Every finding was against the evidence record, and rounds 10 and 11 found their defects inside the agent's own repairs to that record — the defect reproducing within its own fix.
Operator ruling 2026-09-08, verbatim, correcting the agent's proposed remedy:
tool-generated evidence does not make false claims structurally impossible. A command can faithfully reproduce a tautology or measure the wrong thing. Interpretation still needs independent review.
Do not prohibit numbers or classifications merely because they appear in analysis—the requirement is traceable support and honest qualification.
The agent had claimed that publishing raw artifacts would have made rounds 10 and 11's findings "structurally impossible." That claim is withdrawn as untested.
The five requirements (binding once ratified)¶
- Preserve raw artifacts and bind them to the revision, command, environment, and actual result. An artifact floating free of the tree that produced it is not evidence.
- Each acceptance claim must identify its supporting artifact and explain exactly what the observation establishes. Merely naming a command is insufficient.
- Counts, locations, and failure classifications must come from inspected results. Claims of isolation, independence, exclusivity, or completeness require specific demonstrations; they cannot be inferred from a green run.
- Keep agent interpretation explicitly distinguishable from observed evidence. Numbers and classifications are permitted in analysis; the requirement is traceable support and honest qualification, never silence.
- Step 4b must judge whether the evidence supports the claim, including whether the measuring method itself is sound.
Tested against the actual failures of rounds 10 and 11¶
Per the ruling, the contract is tested rather than asserted. Round 12
(arb-step-codexadversary-3e166363a03243adb5fc29f6d71bd7db) refuted the first version of
this section for conflating two different things under one "prevented" column, and for
asserting an unmeasured rate reduction. Its correction is adopted: three distinct columns.
- Auto-enforced — a mechanism fails closed without author cooperation.
- Excluded by compliant authoring — if the author actually performs what the requirement demands, this defect does not occur. Nothing forces compliance.
- Retrospectively detectable — a reviewer holding the artifacts can fault it. All five Yes cells here are historical detections that did happen; they are not a guarantee of future detection.
| # | Actual failure | Auto-enforced | Excluded by compliant authoring | Retrospectively detectable |
|---|---|---|---|---|
| 1 | Census said 144; actual 145 (whitelist dropped self.assertIsInstance) |
No | No — R1 requires disclosing the method; disclosure does not stop the author choosing an unsound one. The 144 was a faithful run of a wrong method | Yes — round 10 recounted |
| 2 | Literal comparison attributed to :940; it is :939 |
No | Yes — R3 requires locations come from inspected results; inspecting rather than recalling excludes it | Yes — round 10 checked the line |
| 3 | Wrote "All 17", enumerated 16 | No | Yes — R3 requires a completeness demonstration; counting the enumeration against the inventory excludes it | Yes — round 11 counted |
| 4 | test_ownership.py:241 misfiled under the shared-parser bound |
No | Partial — R4 excludes presenting a reading as observation; it does not make the reading correct. A wrong-but-labelled interpretation survives | Yes — round 11 traced operands |
| 5 | 103/17/25 published from an undisclosed heuristic |
No | Yes for the disclosure defect — R4 requires traceable support and qualification. It does not make the classifier correct | Yes — round 11 produced 121/20/4 |
Conclusion, stated at the strength the evidence carries. The contract auto-enforces nothing — no requirement here fails closed, and no validator scope is proposed. Three of five failures would be excluded by compliant authoring, one partially, one not at all. All five were detected retrospectively by review, which is an observation about rounds 10-12, not a property of this contract.
Effectiveness is an UNTESTED HYPOTHESIS, not an established consequence. An earlier revision of this section asserted that the contract "reduces the rate of unverified claims reaching attestation." That claim is WITHDRAWN: no rate was measured before, none has been measured since, and asserting it repeated the exact remedy-effectiveness overclaim this document exists to address — the second time in one session (the first was "structurally impossible"). Whether requiring published methods and artifacts changes the rate at which unsupported claims reach Gate 5 is testable, and untested.
Named residual. A claim whose method is disclosed, plausible, and wrong in a way review does not catch still passes. R5 is a review obligation, not a mechanism. Round 12's own Weakest point states the residual precisely: "Publishing a reproducible count establishes the count under that method; neither a caveat nor earlier successful reviews establishes that the method supports the intended acceptance claim."
Amendments required to existing governing surfaces¶
| Surface | Current state | Amendment needed |
|---|---|---|
gz-obpi-pipeline SKILL.md § Stage 4 template |
REQ table has Proof location and Proof columns |
Add a what this does NOT establish column. (An earlier revision called R2 "unsatisfiable" without one; round 12 correctly faulted that — the existing Proof fields CAN carry the explanation. The amendment is to PROMPT it structurally, not to make it possible.) |
gz-obpi-pipeline SKILL.md § Step 4a-v |
Replays $ transcripts; its own § Reach already disclaims completeness and interpretation |
Add: the packet must separate OBSERVED from INTERPRETATION. 4a-v cannot check this and must say so rather than imply coverage |
gz-obpi-pipeline SKILL.md § Step 4b dispatch contract |
Requires re-deriving the claim from REQs and repo | Add R5 explicitly: the adversary must judge whether the measuring method itself is sound, not only whether the claim reproduces. Rounds 10 and 11 did this without being asked |
.gzkit/rules/tests.md § Verification exit-code integrity |
Governs exit codes of verifiers | No amendment. Orthogonal — it governs whether a result is real, not whether a claim about it is supported |
AGENTS.md § Governance doctrine surfaces "A value written in a Markdown doc is ILLUSTRATIVE, never authoritative" |
Governs thresholds and state read by execution | Tension to surface, not resolve. An evidence packet is prose carrying authoritative measurements. This contract's R1 binding (revision + command + environment + result) is what distinguishes a bound measurement from an illustrative value; whether that satisfies the existing rule is an operator reading, not the agent's |
What this contract is not¶
- Not mechanically witnessed. Every requirement is a reading. No validator scope is
proposed here, and inventing one would repeat the doctrine-declared-without-mechanism
family named in
AGENTS.md. - Not a completion gate. It changes what a packet must contain, never whether Gate 5 fires.
- Not a substitute for Step 4b. Per the ruling: "Interpretation still needs independent review."