methodology/EVAL_PROTOCOL.md

# WildTrace — Evaluation Protocol (full reproduction spec)

Everything needed to reproduce the leaderboard end-to-end. The runnable harness is in
`eval/` (`run_eval.py` → model answers, `run_judge.py` → rubric scores); this document is
the exact specification those scripts implement, so a clean-room reimplementation matches.

## 1. Evaluation is evidence-withheld
The model receives ONLY the document and the question. The supporting clues, the gold
answer, and the rubric are **never** shown to the model under test. The exact prompt is:

```
Answer using ONLY the document below. Include every specific detail from the text.

Question: {question}

Document:
{document}
```

`{question}` = the task's `question_text`. `{document}` = the full corpus file
(`corpus/<corpus_file>`), truncated to the model's context cap (§3). No system prompt, no
few-shot examples, no chain-of-thought instruction.

## 2. Generation settings
- `temperature = 0.1` for models that accept it. **Omit temperature** for models that reject
  it (Anthropic Opus 4.6/4.8, GPT-5.4) — set `send_temperature: false` in config.
- `max_tokens` (completion budget): **32768 for reasoning models, 8192 otherwise.** Reasoning
  models truncated below this lose ~7pp — do not under-budget.
- One user turn, no retries on content (transient HTTP/rate-limit errors retry up to 5× with
  backoff; only a real refusal or empty output is a failure).

## 3. Context caps and out-of-context scoring
Each model is evaluated **at its native context window**. The document is measured in
**characters**; if `len(document) > cap` the task is **out_of_context_scope** and scored
**0** — the document is NOT sent (a system that cannot ingest the evidence fails the task).
This is the single most important convention: bottom-of-leaderboard models score low mainly
because 40–50% of documents exceed their window.

CJK detection: if the first 4000 chars contain >20 CJK ideographs (`一`–`鿿`), use the **cjk**
cap, else the **en** cap (CJK packs more tokens per character).

Per-model caps used in the paper (characters; en / cjk). ~3.3 chars/token EN, ~1.5 CJK:

| model | en cap | cjk cap |
|---|---|---|
| gpt-4.1, gpt-5.4, gpt-5.5, qwen3.5-plus, qwen3.6-plus, qwen3.7-max, qwen3.7-plus, gemini-2.5-pro, gemini-3.1, deepseek-v4, glm-5.2, claude-opus-4.6, claude-opus-4.8 | 2,850,000 | 850,000 |
| deepseek-v3.2 | 1,200,000 | 400,000 |
| gpt-5.1 | 1,050,000 | 320,000 |
| kimi-k2.6 | 1,000,000 | 320,000 |
| doubao-seed-2.1 | 1,000,000 | 320,000 |
| qwen3-max | 690,000 | 210,000 |
| minimax-m2.7 | 570,000 | 220,000 |
| **default (new/unlisted model)** | 2,850,000 | 850,000 |

For a NEW model, **probe its real window** and add an entry — do not inherit an older
version's cap (an earlier release of this benchmark mis-ranked Qwen3.7 by giving it
Qwen3-Max's smaller cap).

## 4. Judging — 3-judge non-contestant panel, averaged
Each answer is graded by **three judges that are NOT on the leaderboard**, and the three
0–1 scores are **averaged** (simple mean, no same-family exclusion — we verified family
judges show no measurable bias on this set). Panel used in the paper:

| role | family | route used in the paper |
|---|---|---|
| judge 1 | Claude-Sonnet-4.6 | `mr.claude-sonnet-4-6-20260217` |
| judge 2 | Qwen3.5 | `qwen3.5-plus` |
| judge 3 | Gemini-2.5-Flash | `gemini-2.5-flash` |

(Qwen3.5 is a *judge* here, hence it is not a graded contestant.) out_of_context_scope
answers stay 0 and are not sent to judges.

Exact judge prompt (per answer):

```
STRICT grader. Only award points for SPECIFIC details present.
QUESTION: {question[:600]}
RUBRIC:
P1 ({points}pts): {correct_criterion[:260]}
P2 (...): ...
RESPONSE:
{response[:5000]}
Reply JSON: {"points_awarded":[<pts>],"total":<sum>}
```

The rubric is the task's `ground_truth.scoring_rubric` (a list of `{points, correct_criterion}`).
Judge settings: `temperature = 0.1`, `max_tokens = 16384`. Parse the `total` field
(a 0–100 sum), score = `min(total / 100, 1.0)`. A task's final score = mean of the available
judges' scores.

## 5. Aggregation
- Per task, per model: mean of the 3 judges for valid scored responses; out-of-context rows
  are assigned 0 and are not sent to judges.
- The paper reports two denominator views. `Scored` is valid-response quality: the mean over
  valid scored responses for that route, with `n` reporting the number of valid scored rows.
  `All481` is coverage-sensitive quality: all 481 attempted tasks are the denominator, and
  missing, failed, and out-of-context rows are zero-filled.
- A matched-panel variant (tasks scored by all models) is reported only as a robustness check.
  It is not the headline denominator because it removes many long-context access failures.

## 6. Reproducing with the provided scripts
```bash
export API_KEY=...                      # bearer token for your endpoint(s)
cd eval
# 1) edit config.json: base_url + model (+ caps entry if your model is new)
python run_eval.py  --config config.json --data ../data/wildtrace_strict481.with_answers.json \
                    --corpus ../corpus --out ../results/mymodel.responses.json
# 2) edit config.json "judges" to your three judge endpoints
python run_judge.py --config config.json --data ../data/wildtrace_strict481.with_answers.json \
                    --responses ../results/mymodel.responses.json --out ../results/mymodel.scores.json
# overall % is printed and stored at results/mymodel.scores.json -> "overall"
```
LLM judges are non-deterministic, so a fresh run should be reported with the
model route, endpoint date, context policy, decoding settings, and judge panel
used for that run.