Make a controlled adaptation change.

Run a reproducible local study

Download the dependency-free Python starter. It requires Python 3.10 or later and no account, network call or package installation. Original LearnAI teaching code and synthetic data; see the distribution notice below. No separate open-source license is asserted. The supplied reference mode and learner TODO mode are distinct. Download the executed reference manifest and all-run results, including source SHA-256, byte count and Python version. Output files must be new; earlier runs are never overwritten.

1. Run python adaptation-study.py --reference --output reference.json. Inspect every one of the nine runs, including gate failures.
2. Implement learner_gradient and learner_buffer. Run python adaptation-study.py --output learner.json. The default TODO mode raises an explicit error until implemented.
3. Compare all thirteen gradients against finite differences. Verify stream membership, reproducibility, disjoint monitoring IDs and simultaneous updates. Record failures before fixing them.
4. Reproduce the browser reservoir run: seed 17, buffer 16, changed task frequency, 100 old updates and 100 adaptation updates. Compare parameters and monitoring metrics within 10⁻⁸; binary64 library differences may require a documented tolerance.
5. Predeclare one extension: a stratified buffer, a freshness weighting, or a different fixed replay ratio. Keep the old checkpoint, total fitting-row budget and task partitions fixed. Explain any unavoidable mismatch.
6. Report all three seeds and curves every ten updates. Include no-adaptation and new-only controls, regression gates and rollback checkpoints. A new buffer seed is not a new task family.

Build evidence beyond a passing script

Submit selected code, test commands, parameter checkpoints, data/configuration versions, all-run outputs and a retention-versus-adaptation decision in Project 14. Keep complete source files locally. Ask another person to replay the study and inspect your reasoning when review is available.

The starter's finite-difference checks are public software diagnostics. The project has separate objective transfer cases; neither establishes independent implementation proficiency or human review. Broader task-family evidence comes from the adaptation suite, whose support/query boundaries must remain explicit.