# Artifact Contract

This file defines the stable artifact outputs that later phases must produce.

## Goal

Judges should be able to find:

- raw training outputs
- readable plots
- summary metrics
- baseline comparisons

without guessing file names.

## Reserved Output Files

Core environment artifacts:

- `artifacts/training_traces.json`
- `artifacts/train_report.json`
- `artifacts/reward_curves.json`
- `artifacts/baseline_report.json`

Submission-grade training evidence:

- `artifacts/loss_curve.json`
- `artifacts/loss_curve.png`
- `artifacts/reward_curve.png`
- `artifacts/training_summary.json`

Optional comparison outputs:

- `artifacts/baseline_vs_trained.json`
- `artifacts/baseline_vs_trained.png`
- `artifacts/belief_accuracy_curve.png`

## Required Semantics

`loss_curve.json`

- x-axis: training step or epoch
- y-axis: training loss
- include labels or schema keys that make plotting unambiguous

`reward_curve.png`

- readable axis labels
- clear title or caption
- if possible, compare baseline and trained runs on the same axes

`training_summary.json`

- model name
- training method
- scenario
- number of steps / epochs
- key metrics
- output model path or checkpoint path

## Rules

- Do not rename these outputs casually once later phases start depending on them.
- Save plots as committed image files, not only notebook cell outputs.
- Keep the artifact paths relative to repo root so README links remain stable.