fix(science): external-audit response wave 1 — honest power analysis, earned L3, nominal α - #67
Merged
Conversation
…earned L3, nominal alpha Response to an independent external audit (validity 3.1/10). Accepted and fixed in this wave: 1. POWER ANALYSIS (the audit's arithmetic verified and confirmed): the preregistered H1 rule 'mean of n samples > bound' is a DECISION RULE, not an alpha-controlled test — at the exact boundary it flags ~50% by construction. The A6 CLI and doc now publish the rule's full operating characteristics (detection AND false-alarm rows: at worst sigma, 27% of cells 2 points on the passing side get falsely flagged at n=5) plus a proper one-sided alpha=0.05 power table: sigma=7.60, delta=5, n=5 -> power 0.431, n~15 for 80%. The old 'n=5 gives >=0.93 power' headline is retracted as a test-level claim and re-scoped to screening; corpus README, run-002 README, and the Phase A closure report all carry the amendment. H1 flags are screening signals published with mean/SD/CI, never confirmed violations alone. 2. EVIDENCE LEVEL L3 MUST BE EARNED: verificationEffective(total=0) let a citation-free response reach L3. New rule (deterministicLiftEarned): the L2->L3 lift requires >=1 positively verified reference AND 0 non-existent ones. Citation-free texts, all-network-error runs, and texts carrying a fabricated reference all stay L2 with the reason in the caveats. Route summary now carries the verified count; tests updated to the new semantics (a fake reference can never wear the L3 badge). 3. NOMINAL KRIPPENDORFF ALPHA implemented (the PIGA annotation protocol required it but only the interval metric existed) — with hand-computed anchor 4/9 and null-honesty cases. 4. PIGA CLI: --judge-provider flag decouples judge family from the subject's; loud warning when subject and judge share a family or are the same model (self-judging bias). 5. Annotation sheet piga_dataset_protocol: stale 0.2.0-alpha -> 0.3.0-alpha. Validation: web 411/411, core 213/213, cold tsc clean, sync no drift. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Y7wiMwFa2D4P8zD9RhCgsv
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
2 participants
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.
Response to the independent external audit — wave 1 (science fixes)
An independent audit of
main@8d5bd80scored global validity ≈3.1/10 and flagged two substantive scientific errors plus several gaps. Each claim was verified in-repo before acting; this PR fixes everything accepted from the science tier.1. Power analysis made honest (audit arithmetic verified: correct)
The preregistered H1 rule ("mean of n samples > bound") is a decision rule, not an α-controlled test — at the exact boundary it flags ~50 % by construction, and the old doc's "n=5 gives ≥0.93 power" presented rule-flag rates as test power. Now published:
2. L3 must be earned
verificationEffective(total=0) === truelet a citation-free response reach L3. New preregistered rule (deterministicLiftEarned): the L2→L3 lift requires ≥1 positively verified reference AND 0 non-existent ones. Citation-free texts, all-network-error runs, and texts carrying a fabricated reference all stay L2, each with its reason in the caveats — a fake reference can never wear the L3 badge (test asserts exactly that).3. Nominal Krippendorff α implemented
The PIGA annotation protocol requires nominal-scale agreement; only the interval metric existed. Added with a hand-computed anchor (4/9) and null-honesty cases.
4. PIGA CLI cross-family judging
--judge-providerdecouples the judge's provider family from the subject's; loud warning on same-family or same-model (self-judging bias) runs.5. Stale version
Annotation sheet
piga_dataset_protocol0.2.0-alpha → 0.3.0-alpha.Accepted but deferred to wave 2 (next PR)
Instrument renaming to match measurement (PIGA → ambiguity-handling scope + goal-alignment roadmap), related-work section for the preprint, evidence-profile expansion. Production P0s from the audit's engineering tier (Prisma 7 driver adapter, prod Dockerfile, missing migrations) are verified real and tracked — sequenced after the science per the project's declared ordering.
Validation
web 411/411, core 213/213, cold tsc clean, sync-core no drift, corpus validator green.
🤖 Generated with Claude Code
https://claude.ai/code/session_01Y7wiMwFa2D4P8zD9RhCgsv
Generated by Claude Code