bugfix — Per-Task Aggregation

Generated: 2026-05-12T10:00:20Z

Methodology

• Tools under test: 8
• Blind labels: 24
• Layout: single round per (artifact, judge); judge files read flat from each label dir. ^round[0-9]+$ subdir support is retained for multi-round stability studies (pilot/sample dirs excluded).
• Judges: opus, grok420, glm51, gpt54pro, mimo25pro (5-judge panel; each artifact scored once by every judge)
• Rubric: 20 items × 0–10 pts = 200 pt max
• Canonical score per judge file: sum(scores.values()) (not the stored total field)
• Reported tool mean: weighted mean of per-judge means (weights: opus×3, gpt54pro×2, grok420×1, glm51×1, mimo25pro×1)
• Total judgments aggregated: 120

Caveats / threats to validity

• Judge weights are pre-registered, not derived. The 3 / 2 / 1 / 1 / 1 weighting is stored as judges.*.weight in versions.lock.json (committed 2026-05-12) and reflects the operator’s prior trust in the Anthropic (opus) and OpenAI (gpt54pro) reviewers. An equal-weight aggregation is emitted alongside this report as final-report.equal-weight.md; the in-report Pooled Mean column is also the equal-weight comparator and lets readers verify rank-stability without leaving this file.
• Judge scorer asymmetries. gpt54pro is consistently the harshest scorer in the panel (lowest mean across labels). mimo25pro is the most lenient and occasionally emits 200/200 saturations; its weight of 1 dilutes the impact, but right-tail scores should be read in that context.
• σ decomposition. The per-tool standard deviation column is split into within_σ (trial-to-trial within-judge spread — mean of the per-judge stdev across the 3 trials) and between_σ (judge base-rate spread — stdev of per-judge means). Within > between would indicate the tool’s output is genuinely unstable trial-to-trial; the reverse means most variance is judge disagreement. The within_σ label is preserved from the multi-round harness era — in this single-round cohort it measures within-judge trial-to-trial noise.
• Judge sampling not pinned. Temperature is fixed to 0 where the provider exposes it (OpenRouter, OpenCode Go). Claude CLI and OpenAI /v1/responses do not expose temperature/seed, so residual sampler variance is absorbed in per-judge σ rather than eliminated.
• R1 mechanical-fact override. Rubric items with deterministic answers (e.g. tsc_errors == 0) are rewritten post-hoc from auto-metrics.json to remove LLM arithmetic / classification drift. Items locked per task: feature 12/13/16/20, bugfix 14/15, refactor 13/14. Pre-override scores are preserved under scores_pre_r1 on every judged file (scripts/aggregate-results.sh runs an idempotent R1 sweep before aggregating).
• Blind eval is structural, not semantic. Tool identity is hidden via NATO labels and a path-/content-level scrub of tool-specific directories (.omc/, _bmad/, _bmad-output/, _bmad-core/, docs/bmad/, docs/superpowers/, plans/, .claudekit/, .gstack/, .superpowers/, .compound-engineering/, .ecc/, CLAUDE.md.original). auto-metrics.json is anonymised by stripping plugin_versions and collected_at. A skilled judge could still infer identity from idiosyncratic code style; we don’t claim semantic anonymity.
• Cohort span: 2.7h (2026-05-10 → 2026-05-10). scripts/audit-cohort-symmetry.py flags spans >24h as a soft warning; this cohort completed within that window.

Aggregate Scores per Tool

Ranking (Weighted Mean)

1. claudekit — 184.33/200
2. ecc — 181.38/200
3. pure — 175.00/200
4. bmad — 173.83/200
5. superpower — 168.50/200
6. compound — 167.33/200
7. omc — 165.08/200
8. gstack — 159.71/200

Per-Judge Means