fix(exp49): H3 credit attribution on timestamp ties — arm C corrects to 0.969, gates clean - #466
Merged
Merged
Conversation
… corrects 0.915 → 0.969 (PASSES) Event-level reconstruction REFUTED the published 'one-turn-stale capture frame' residual: the measured before/after pairs were frame-clean (each credit's before chains exactly from the previous after). The liar was the harness extractor — the credit event is emitted at END-of-execute, and under back-to-back substrate actions the NEXT turn's motion event lands within the 10 ms JSONL timestamp rounding; the <= matcher attributed those credits to the wrong turn, scoring honest negative credit as sign-flipped. Fix: strictly-before matching (a credit can never tie its OWN motion — the >=0.3 s measurement settle guarantees separation; rounding is monotonic so the next motion never rounds below its predecessor's credit). Re-scored: arm C H3 125/129 = 0.969 (gate >=0.95 PASSES; remaining 4 are noise floor — 0.3 deg true progress under sigma=0.03), arm B unchanged 1.00. Zero confirmed honesty-gate leaks; the planned backend 'stale-frame fix' is unnecessary and hardware H3 claims are no longer instrument-blocked. Experiment doc corrected with a dated CORRECTION entry. Guard: test_credit_tied_with_next_motion_attributes_to_own_turn. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Phase 0 outcome, opposite of the plan: the intended backend 'stale-frame fix' is CANCELLED because event-level reconstruction refuted the hypothesis — the measurement honesty gates were clean all along. The extractor's credit-to-motion matcher mis-attributed ~5% of arm C credits on 10 ms-rounded timestamp ties under back-to-back actions (the credit event is emitted at end-of-execute, tying with the NEXT turn's motion event). Strictly-before matching is provably exact (own-motion separation >=0.3 s; monotonic rounding). Arm C H3 re-scores 0.915 → 0.969 (pre-registered >=0.95 gate PASSES); arm B unchanged at 1.00; remaining 4/129 mismatches are noise-floor (0.3° true progress under σ=0.03). Scripts + tests + a dated CORRECTION in the experiment doc; no src/ changes.
🤖 Generated with Claude Code