Files
Oleg Maslov 0c3e2ead3b
Some checks failed
Installer Smoke / installer-smoke (push) Has been cancelled
moving
2026-09-02 10:10:29 +02:00

3.8 KiB
Raw Permalink Blame History

§1.3g Judge Swap κ Analysis (4-Candidate Roster)

Date: 2026-04-24 · Sample: N=20 stratified 4-per-cell from Stage 2-Retry κ-calibration set.

Consensus reference: Opus 4.7 + GPT-5.4 agreement.

Opus≠GPT splits (excluded from κ): 0 / 20

Consensus subset size (denominator for κ): 20 / 20

Per-candidate results (ranked by κ descending)

Rank Candidate Model ID Routing κ Raw agreement n compared Parse OK / 20 Verdict
1 kimi kimi-k2.6 direct 1.0000 100.00% (20/20) 20 20/20 PASS (excellent)
2 minimax minimax/minimax-m2.7 openrouter 1.0000 100.00% (20/20) 20 20/20 PASS (excellent)
3 deepseek deepseek-v4-pro direct 1.0000 100.00% (18/18) 18 18/20 PASS (excellent)
4 zhipu glm-5.1 direct 1.0000 100.00% (19/19) 19 19/20 PASS (excellent)

Aggregate verdict: MULTI_PASS

  • PASS (κ ≥ 0.70): 4
  • BORDERLINE (0.60 ≤ κ < 0.70): 0
  • FAIL (κ < 0.60): 0
  • INCONCLUSIVE (parse < 18/20): 0

Sample-selection caveat (transparent)

Full 100-row κ-calibration set has 7/100 = 7% Opus-GPT split rate. Our 20-instance deterministic first-4-per-cell sample caught 0/20 splits — consistent with binomial variance at low end but signals selection bias toward "easier" instances where the two primary judges already agree. κ = 1.0 on all 4 candidates should be read as "all candidates match the consensus on a biased subset", not "all candidates match Opus+GPT on arbitrary LoCoMo instances". Recommended follow-up (Task 2.6): stratified-on-split re-probe at N=40-60 with split-inclusion oversampling to derive κ values that discriminate among the 4.

Secondary criteria (tie-break at κ=1.0)

Cand Parse OK/20 p50 latency p95 latency avg out tokens retries routing
zhipu 19/20 9 s 22 s 357 0 direct (api.z.ai)
minimax 20/20 12 s 28 s 483 0 openrouter
deepseek 18/20 12 s 29 s 465 0 direct (api.deepseek.com)
kimi 20/20 32 s 110 s 1002 (near cap) 4 direct (api.moonshot.ai)

Revised ranking by secondary criteria (κ tie-break via speed × parse-reliability × direct-routing):

  1. Zhipu GLM-5.1 — fastest p50 (9 s), direct, 19/20 parse. Best operational profile for N=400 (Stage 3 × 5 cells × ~1 judge triple ⇒ ~2000 Gemini-replacement calls at ~9 s each ≈ 5 hr judge-thread wall-clock).
  2. DeepSeek V4-Pro — direct, 18/20 parse, clean retries.
  3. MiniMax M2.7 (via OR) — 20/20 parse but OpenRouter fallback = extra dependency.
  4. Kimi k2.6 — 20/20 parse but 3.5× slower than Zhipu; N=400 judge thread would be >20 hr on Kimi's p50. Operational deal-breaker for Stage 3 timeline.

Revised recommended primary: ZHIPU Revised recommended backup: DEEPSEEK

Marko's original preference (Kimi/MiniMax) preserved as 3rd/4th picks — both PASS κ-wise but lose on speed (Kimi) or routing-stack (MiniMax). PM adjudicates final pick at manifest v6 brief time.

Per-cell breakdown (top-2 by κ)

kimi (kimi-k2.6)

Cell n Agree Raw %
agentic 4 4 100%
full-context 4 4 100%
no-context 4 4 100%
oracle-context 4 4 100%
retrieval 4 4 100%

minimax (minimax/minimax-m2.7)

Cell n Agree Raw %
agentic 4 4 100%
full-context 4 4 100%
no-context 4 4 100%
oracle-context 4 4 100%
retrieval 4 4 100%