# Suite27 implementation plan

Approved design: `DESIGN.md`. Build a native hard tree ensemble whose ancestor predicates condition chronological memory before split selection. Verify the complete runnable package. Do not infer new academic performance from software fixtures.

1. Implement the chronological cache and origin protocol; test missingness, strict history, and failed-origin audits.
2. Implement the native histogram grower, course reader, node evidence and optional surrogate co-adaptation; test actual memory splits, ancestor feedback, loss reduction and stale-reader rejection.
3. Implement complete origin selection/refit, six variants, checkpoints, reload and CLI reporting; test every variant, one-class stock probabilities, multi-seed selection, resume and fingerprints.
4. Reconstruct production Suite26 evidence tables and finalized design documentation.
5. Review independently, fix important findings, run the full tests and packaged smoke workflow. Save a complete downloadable archive.

Global constraints: primary scores are intercept plus hard-tree leaf increments; no external risk head/blend. Outcomes enter only strictly earlier histories. Auxiliary heads never score primary predictions. Reader changes require complete forest regrowth. All model selection uses earlier, supported single-semester origins. Historical F6/F7 are descriptive reused benchmarks. Defaults require every faculty support; unsupported origins fail with full audit. Allow simpler comparisons to win.

Final status: all five tasks completed; final frozen-source suite passed 116 tests and 28 subtests. Extracted-source all-six smoke passed. Independent review has no unresolved critical or important findings. See VALIDATION.md for evidence and empirical limits.
