ground-truth-gates: rerun is new evidence — audit the cited evidence identity, never substitute a fresh run - #226
Merged
Conversation
…itute a fresh run New ground-truth-gates rule, placed immediately after the producer-side preservation sentence as its consumer-side dual: preserve so it can be re-read; when auditing historical evidence, re-read that evidence identity — a rerun is a new evidence row, not a substitute for the old one. Locked semantics: evidence identity is content-verified, never pathname-bound (a verified content-identical archival copy IS the cited record); an unrecoverable record leaves the historical claim UNVERIFIED; fresh runs stay fully legitimate for current-state health, reproducibility, corroboration of an inspected record, regenerate-and-diff, and verify-by-reconstruction — the sole prohibition is silently substituting the rerun for the evidence identity under audit. Adjudicated PARTIAL-GAP from a nine-passage falsification (36/36-ledger counterexample passes every existing rule); design review r1 dual-blind luna-max + sol-max = 2/2 PROCEED, zero findings, incl. the dedicated archival-misread check; landed byte-verbatim from the reviewed blocks (adaptations NONE); one new unprobed marker, behavioral probe joins the #115 queue. Evidence: reviews/2026-08-29-rerun-is-new-evidence/. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132RthrKSsMkywcEwtkXhkx
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
One new
ground-truth-gatesrule — rerun is new evidence — placed immediately after the producer-side preservation sentence as its consumer-side dual, plus its evidence package. One canonical file (+51, pure additions); 6 evidence files.Spine: preserve so it can be re-read; when auditing historical evidence, re-read that evidence identity. A rerun is a new evidence row, not a substitute for the old one.
Why (adjudicated PARTIAL-GAP)
A falsification-first orientation showed nine adjacent passages each pass this counterexample: an auditor never opens a cited 36/36 ledger, re-runs the recipe, gets a fresh 36/36, and declares the historical report verified — while a 35-row ledger, a crashed original run, a different config, a mistranscribed result, or a vanished artifact all survive that fresh PASS. Existing rules are producer-side (preserve; per-run persisted rows) or state-side (verify-by-reconstruction proves STATE, not history); the consumer-side evidence-identity duty was missing.
The rule (locked semantics)
Review
Design gate, dual-blind, mutually blind, isolated dirs, identity from tool banners: r1 = gpt-5.6-luna (max) PROCEED + gpt-5.6-sol (max) PROCEED — 2/2, zero blocking findings, all ten mandatory axes line-anchored by both — including the dedicated check that "cannot recreate the missing history" cannot be misread as forbidding durable-copy verification (both confirmed the archival carve-out is visible in the same rule). Both confirmed: no runtime tooling, no delegation-and-review mirror, no collateral rewrite.
Landing fidelity
Reviewed rule + Provenance blocks landed byte-verbatim (machine-proven; DECLARED-LANDING-ADAPTATIONS: NONE). Neighbors byte-unchanged by machine check: the preservation sentence, regenerate-and-diff, verify-by-reconstruction, persisted-run and replication rules; delegation-and-review and operational-rigor zero-byte. Static controls 8/8 + 2 archival variants (HISTORICAL-MISSING/-MISMATCH · ARCHIVAL-RECOVERY · FRESH-HEALTH · REPRODUCIBILITY · CORROBORATION ×2 · GENERATED-ARTIFACT · STOCHASTIC · EXACT-RECONSTRUCTION).
checks.pygreen (invisible-Unicode sweep included).Probe debt
Behavioral transmission/effectiveness remains unprobed. Exactly one new
unprobedmarker; provenance class = counterexample-derived / design-reviewed; the future behavioral probe routes to the standing #115 queue.🤖 Generated with Claude Code
https://claude.ai/code/session_0132RthrKSsMkywcEwtkXhkx