Skip to content

Multi-judge panel — is the harness lift robust to the choice of judge?

The LLM-judged frontier reports use one judge at temperature 0 — quasi-deterministic, not exact. This panel re-scores the SAME stored responses with several independent judges and asks: do they agree on the lift (harnessed − baseline)? If they do, the relative comparison does not depend on any one judge — the real answer to the non-determinism concern, stronger than picking a single 'best' judge.

The judges may differ on absolute scores, but they agree on the LIFT. Krippendorff's α = 0.605 — only weak absolute agreement (inter-rater reliability of the absolute 0–10 scores); meanwhile the per-model lift is consistent across judges (mean spread ±1.12/10). We only ever claim the relative lift (the paired delta), and that is what is — and must be — judge-robust: absolute-score disagreement cancels in the pairing. This is the empirical version of read the delta, not the absolute score.

Per-model lift, by judge

Model n gpt-oss:120b glm-5.2 qwen3.5:397b kimi-k2.7-code deepseek-v3.2 Panel mean Judge spread
deepseek-v3.2 9 +5.11 +4.44 +8.89 +4.89 +5.83 ±1.78
gemma4:31b 9 +5.67 +7.11 +5.56 +3.72 +6.67 +5.75 ±1.17
glm-5.2 9 +4.56 +7.78 +3.83 +4.78 +5.24 ±1.51
qwen3-coder:480b 9 +5.67 +6.33 +4.17 +4.75 +5.23 ±0.83
qwen3.5:397b 9 +1.56 +2.0 +2.33 +2.25 +2.04 ±0.3

A — is a self-family exclusion: a judge never scores a response from its own model family (so GLM doesn't judge glm-5.2, etc.). n is the prompts per model with both arms scored — modest here (this is a balanced sample on the harder perdim subset, where baselines are weak so the absolute lift runs large). This panel's job is to show the lift is judge-robust, not to pin its magnitude; the larger-N magnitude estimates are the single-judge reports (harness_lift_report.md, comparative_results_llm_judge.md).

Reading this

  • Krippendorff's α (above) is the inter-rater reliability of the absolute 0–10 scores (1 = perfect, ~0 = chance, < 0 = systematic disagreement; ≥0.80 strong, 0.67–0.80 acceptable). A weak α together with a small lift-spread is the expected, acceptable pattern: judges can anchor their absolute scale differently yet still agree on how much the harness improved a reply — and the paired design uses only the latter.
  • Judge spread (last column) is the standard deviation of the per-model lift across judges. Small spread = the judges award the same relative improvement, so the headline lift is not an artifact of one judge.
  • We report the lift, not absolute scores, precisely because the lift is what survives judge disagreement (the paired design cancels each judge's scale).
  • Claude Opus can be added as a premium absolute-calibration judge via subagents (isolated context), but for this relative comparison the independent Ollama panel is sufficient and zero main-context. The deterministic per-dimension report is the judge-free, fully reproducible floor; the LLM judge is the primary holistic view.
  • Judges: gpt-oss:120b, glm-5.2, qwen3.5:397b, kimi-k2.7-code, deepseek-v3.2 — a diverse panel of large frontier models (gpt-oss, GLM, Qwen, Kimi, DeepSeek). Independence is preserved by self-family exclusion: a judge never scores its own family (e.g. GLM never judges a GLM candidate), so GLM / Qwen / DeepSeek can serve as judges for the other candidates while no model grades itself. Panel over 90 stored responses.