Archive Kimi K3 and Gemini 3.1 Pro paired evaluation results
Add Kimi K3 and Gemini 3.1 Pro Preview to the existing results/<model>/<baseline|deepagents>/ archive, with 1,000 tasks per arm (4,000 new records).
Each arm includes JSONL/CSV task records, a summary, configuration and a README. Model directories include paired comparisons, provenance, failure-stage counts and export verification. The archive index and checksums are updated to include both models.
| Model | Setting | Compilation success | Partial replication | Coefficient direction | Significance level | Local perfect |
|---|---|---|---|---|---|---|
| Kimi K3 | Baseline | 52.6% | 36.6% | 49.4% | 43.1% | 23.0% |
| Kimi K3 | DeepAgents | 86.7% | 63.5% | 80.4% | 70.9% | 39.0% |
| Gemini 3.1 Pro Preview | Baseline | 54.3% | 36.6% | 50.6% | 43.7% | 24.6% |
| Gemini 3.1 Pro Preview | DeepAgents | 85.0% | 63.2% | 81.1% | 71.8% | 39.1% |
All denominators are fixed at 1,000. Kimi retains 5 baseline and 4 DeepAgents task-level unknowns; Gemini has none. Metric-level nulls are counted separately. Historical parent/child selection, source commits and unequal model/tool budgets remain explicit.
The four dimensions use the local hf-leaderboard-v1 implementation; local perfect is a separate local-paper-v1 metric. Official scorer parity is unverified. This PR archives results using the layout introduced in #1; the displayed leaderboard's root results.csv and page code are unchanged.
Raw prediction triplets are unavailable in the accepted score snapshot. New records therefore carry null prediction fields and an explicit prediction_export_status marker; this is an export limitation, not evidence that every model prediction was missing. Full prompts, API traces, reference answers and credentials are excluded.
Validation: all 4,000 task IDs, flags and result hashes match the accepted export; 36 metric aggregates were independently recomputed from the stored flags; JSONL/CSV and paired rows agree; archive hashes pass. Run python3 -B results/verify_kimi_gemini.py after checkout. This verification does not rerun models or independently rescore raw predictions.