rl/README.md

# Public AIM-AG RL companion

This directory is the policy-visible portion of a larger internal RL production package. It demonstrates how the ten AIM-AG research tasks are decomposed into exact tool exercises, targeted milestones, and expert-gated frontier episodes.

Included:

- 10 public task definitions;
- 53 public curriculum episodes;
- 33 public exact benchmark prompts;
- 10 public full-frontier evaluation prompts;
- public benchmark fixtures;
- submission, episode, task, artifact, and trajectory schemas; and
- reward, evaluation, curriculum, and trajectory policies.

`data/curriculum_episodes.jsonl` preserves the production-format public record shape. `data/curriculum_episodes_hf.jsonl` is a lossless type-stable view: `source_fields` preserves optional-field presence, `prompt_json` preserves the exact string-or-object prompt, and structured inputs are encoded in `input_json`. Convenience columns retain prompt text and episode-specific milestone/artifact fields without mixed Hugging Face column types. The `curriculum_{train,validation,test}_hf.jsonl` files expose its 18/14/21 internal train/dev/eval rows as native Hugging Face splits. Exact benchmarks are likewise exposed as 11/11/11 split files.

The client research records and these RL prompts intentionally use different instruction, deliverable, and allowed-tool wording. Their shared conjecture and definitions are synchronized in this release. All included prompts and fixtures are public; the split labels organize the demonstrator and do not create a held-out benchmark.

Excluded:

- hidden grader records and canonical candidates;
- hidden expected outputs and fixtures;
- calibration submissions and reviewer answers;
- deterministic grader implementation and internal readiness reports; and
- any claim that a public episode remains held out.

The exact episodes teach reproducible computation and evidence formatting. They do not certify the unresolved headline conjectures. Full proof or disproof claims require current literature review, at least two independent specialists, and adjudication.