# Bench run — bench-kimi-k3 (moonshotai/kimi-k3, open-charts)

13 fixtures, frozen inputs, open chart vocabulary. Writer: `moonshotai/kimi-k3`, `max_tokens: 32000` (raised from the 16k spec: reasoning bills inside the completion budget and capped shorter runs), `response_format: json_object`, no temperature (endpoint-unsupported). Judges: qwen3.7-max, gemini-3.5-flash, grok-4.5 (blind, non-contestant). Two fixtures (hottest-june, trap-disagreeing-toll) come from a same-day retry after a socket drop on the first pass; all params identical.

| fixture | words | sect | quo | charts | ground | recall | facts | traps | formfit | banned | limit-viol | writer $ | src |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| trap-unit-magnitude | 472 | 4 | 0 | 2 | 4/4 | - | 100% | 1/1 | 4.5 | 0 | 1 | $0.1657 | primary |
| hottest-june | 569 | 5 | 2 | 3 | 9/9 | 100% | 100% | - | 3.0 | 0 | 0 | $0.2903 | retry |
| tour-stage-4 | 503 | 4 | 4 | 2 | 7/7 | 100% | 100% | - | 3.5 | 0 | 0 | $0.3929 | primary |
| greek-polls | 593 | 5 | 0 | 4 | 32/33* | 50% | 100% | - | 4.0 | 0 | 0 | $0.3025 | primary |
| trap-prior-override | 470 | 4 | 0 | 2 | 8/8 | - | 100% | 1/1 | 4.0 | 0 | 0 | $0.1405 | primary |
| trap-derived-total | 459 | 5 | 0 | 2 | 3/3 | - | 100% | 1/1 | 4.5 | 0 | 0 | $0.2707 | primary |
| italy-oecd-wages | 473 | 4 | 3 | 2 | 12/12 | 100% | 100% | - | 3.0 | 0 | 0 | $0.2978 | primary |
| iran-ceasefire-collapse | 601 | 4 | 3 | 3 | 7/7 | - | - | - | 3.7 | 0 | 0 | $0.4301 | primary |
| trap-disagreeing-toll | 480 | 4 | 0 | 2 | 5/5 | - | 100% | 2/2 | 4.0 | 0 | 0 | $0.3855 | retry |
| easyjet-apollo | 553 | 5 | 3 | 2 | 3/3 | 100% | 86% | - | 4.5 | 0 | 0 | $0.1511 | primary |
| france-morocco-worldcup | 488 | 5 | 3 | 2 | 4/5* | - | - | - | 4.0 | 0 | 0 | $0.3253 | primary |
| sk-hynix-listing | 536 | 4 | 3 | 4 | 8/8 | 50% | 100% | - | 4.0 | 0 | 0 | $0.2873 | primary |
| volkswagen-model-cuts | 602 | 5 | 3 | 4 | 6/6 | - | - | - | 3.8 | 0 | 0 | $0.3544 | primary |

## Aggregates

- **Grounding gate:** 108/110 numeric chart values matched directly; the two unmatched (greek `may=5.5`, france `minute=66`) are disclosed derivations from sourced operands, stated in the chart notes and adjudicated benign in the adjudication log. **Gate: pass.**
- **Extraction recall** (6 gold-set events): 83% mean (100%, 50%, 100%, 100%, 50%, 100%).
- **Key-fact coverage:** 98% mean (easyjet 86%, rest 100%). Metric still coarse; reported, not ranked.
- **Traps:** 5/5 checks across all four synthetic fixtures (disagreeing-toll scores both figures).
- **Form-fit (judged):** 3.88 mean of per-fixture chart medians, range 3.0–4.5.
- **Length:** 459–602 words (spec 450–600; iran 601 and volkswagen 602 marginally over).
- **Banned vocabulary:** 0 across all fixtures. **Limit violations:** 1 (trap-unit-magnitude).
- **Writer cost:** $3.7941 for 13 fixtures (~$0.29/fixture; reasoning-heavy). Judge panel logged separately in the cost ledger.

`*` grounding denominator includes a disclosed derivation; see adjudication log.