39  Recorded Evaluations

40 Recorded Evaluations

Proof by Difference becomes credible only when weak and repaired prompts are run against the same fixture and scored with the same rubric. This section preserves those runs, including non-wins when they occur.

Generated at: 2026-07-02T12:44:28.207381+00:00

40.1 Summary

Case Surfaces Weak Outcome Repaired Outcome Outcome Delta Reviewability Delta
Refactor Without Breaking Behavior claude-code-safe, codex-cli-default 2.0 2.6 repaired prompts satisfied 0.6 more outcome checks on average repaired prompts scored 3.4 reviewability points higher on average
Incident Diagnosis Without Fake Certainty claude-code-safe, codex-cli-default 2.6 3.0 repaired prompts satisfied 0.4 more outcome checks on average repaired prompts scored 3.0 reviewability points higher on average
API Design Without Hidden Compatibility Traps claude-code-safe, codex-cli-default 3.0 3.0 weak and repaired prompts tied on outcome checks repaired prompts scored 2.9 reviewability points higher on average
Online Migration Without Data Loss claude-code-safe, codex-cli-default 2.9 3.0 repaired prompts satisfied 0.1 more outcome checks on average repaired prompts scored 0.6 reviewability points higher on average
Test Generation Without Overfitting To Implementation claude-code-safe, codex-cli-default 2.6 3.0 repaired prompts satisfied 0.4 more outcome checks on average repaired prompts scored 2.4 reviewability points higher on average
Performance Tuning Without Micro-Optimization Drift claude-code-safe, codex-cli-default 2.5 3.0 repaired prompts satisfied 0.5 more outcome checks on average repaired prompts scored 3.8 reviewability points higher on average

40.2 Rubric

Outcome scores count case-specific obligations such as passing fixture tests, naming planted causes, preserving invariants, avoiding dirty-data traps, and including rollback boundaries.

Reviewability scores use 0-2 per criterion: artifact boundary, invariants, output contract, verification, failure behavior, and assumption control. The reviewability score is secondary and review-oriented. It partly rewards prompt echo because repaired spells contain words such as invariant, verify, rollback, and assumption. Read it as an inspectability signal, not as direct work quality. Transcripts and outcome checks remain the primary evidence.