layout.md

Repository layout

Location	Responsibility
reasoning_core/tasks/, core modules	Procedural generators, discovery, scoring contract
reasoning_core/evaluation/	Zero-shot evaluation, metrics, batteries, intrinsic rewards, influence, composition
reasoning_core/evaluation/training/	Data mixing, arm execution, optimization, checkpoints
reasoning_core/generation/	Dataset builds: python -m reasoning_core.generation (see docs/release.md)
datasets/	Hub dataset cards (source of truth) and their generated stats
reasoning_core/integrations/	Optional Python adapters
reasoning_core/resources/	Shipped data: battery manifests and frozen eval legs
root integrations/	Independently packaged OpenEnv and Prime Intellect applications
reasoning_core/task_search/	Repository-only developer orchestration, excluded from the wheel
scripts/	Commands, smoke runs, and builders for new evaluation datasets
docs/, docs/results/	Protocols and maintained published results

There is one import path per module. Reach for a module where the table says it lives; nothing aliases anything else.

Import cost

Importing evaluation or its group/composition helpers does not load training. Metrics depend on Torch, while arm execution additionally needs the training extra (TRL/Transformers). Runtime adapters import the external framework only at its boundary, so core stays importable without reasoning_gym installed.

reasoning_core/__init__.py binds only the intended API. Registry discovery internals live in reasoning_core.registry and are imported from there.

Deliberate exceptions

• Battery manifest filenames keep their version numbers. A manifest pins its own name, and that name is hashed into the battery identifier recorded by every measured arm, so copyfree_battery_v8_tiny is an identity, not a label. Renaming those files to something more descriptive would either orphan the results that carry the old identifier or leave a filename that disagrees with the identity inside it. Three files also share one declared name: copyfree_battery_v8.json, v9, and v10 all announce copyfree_battery_v8, distinguished only by the digest. That is a wart, and it is frozen for the same reason.
• Engine and provenance identifiers keep their existing spelling. They are hashed into arm identities, so renaming them would silently split old and new results.

Manifests are addressed by path, though, including by consumers outside this repo, so code reaches them through battery._MANIFESTS rather than a path literal, and moving them is a coordinated change rather than a forbidden one.

Group fields default to values that spell themselves absent, so adding them left every pre-existing arm identity byte-identical. Any change here must preserve that: an ArmSpec must keep hashing to the same spec_id across a refactor, or previously measured arms stop pooling with new ones.

task_search is available from a checkout, editable install, or source distribution, not from the ordinary wheel; its dependencies are documented in its README. The independent applications under root integrations/ retain their own package structure and pyprojects.

Operational reports, results and G5K orchestration are private and live outside this repository.