# Jeu de données / Data pack: « Un modèle plus cher pour l'étape du plan : ce qu'il apporte, ce qu'il coûte »

**FR.** Ce jeu de données contient uniquement des chiffres. Il ne contient ni énoncé de cas, ni consigne, ni texte de notation, ni note rédigée, ni identifiant interne de cas : les six cas sont numérotés de 1 à 6 dans un ordre sans lien avec l'ordre interne. Pour recalculer les chiffres de l'article :

```
python recompute.py
```

Le script n'utilise que la bibliothèque standard de Python 3 et lit les fichiers `.csv` de son propre dossier.

**EN.** This data pack contains numbers only. It has no case statement, prompt, grading text, drafted note or internal case identifier: the six cases are numbered 1 to 6 in an order unrelated to the internal one. To recompute the article's figures:

```
python recompute.py
```

The script uses only the Python 3 standard library and reads the `.csv` files in its own folder.

## Fichiers / Files

- `arms.csv`: one row per run (`terra_1` 28 Oct 2026, `terra_2` 29 Oct 2026: the plan on GPT-6.1 Terra; `deepseek_1`, `deepseek_2` 27 Oct 2026: every step on DeepSeek v4.1-flash), the grader's summary of its six notes: mean score, share at light edits or better, verdicts, defect counts by class, and `grader_drift_mean_delta` (the grader's mean error on six reference notes in the same session).
- `terra_plan_calls.csv`: one row per call of the plan on Terra: run, anonymised case, call number inside the case, seconds, prompt, completion and reasoning tokens, tokens read from and written to the cache, cost in USD as billed. Run `terra_1` has 13 calls (one is a retry), run `terra_2` has 12.
- `deepseek_plan_calls.csv`: the 10 plan calls of the DeepSeek-only run of 27 Oct 2026 (seconds, prompt, completion and reasoning tokens; the log of that day has no cost field).
- `prices.csv`: the listed and billed prices per million tokens used in the article (GPT-6.1 Terra, DeepSeek v4.1-flash-1017).
- `cache_probes.txt`: the log of the cache tests on synthetic texts.
- `billed_costs.csv`: OpenRouter's own per-request figures summed over each run's time window (requests, tokens, cache, USD), by model family. `deepseek_1` is the DeepSeek-only run (the second DeepSeek-only run has no isolated billing window): the window ends at 08:56 (lower bound) or 08:57 (upper bound, which also contains the first calls of the next run).

## Limites / Limits

- **FR.** Six cas de travail, un seul correcteur automatique, un passage par cas et par configuration, deux passages par modèle. Le code de la chaîne a changé entre les passages sur DeepSeek (27 octobre) et ceux sur Terra (28 et 29 octobre). Les coûts du passage `deepseek_1` ne comptent que les appels du modèle (le correcteur n'est pas inclus).
- **EN.** Six working cases, one automated grader, one run per case and configuration, two runs per model. The chain's code changed between the DeepSeek runs (27 October) and the Terra runs (28 and 29 October). The cost of run `deepseek_1` counts the model calls only (the grader is not included).

Énoncés des cas et consignes sur demande / case statements and prompts on request: hello@lucapaper.com.