PREREG.md — pre-registered selection and kill rules (template)

Proposed location once adopted: hive/PREREG.md. Commit by 2026-10-20 (before the first fingerprint-model LB results). Append-only after that date.

rule_version: 1
committed_utc: 2026-10-20T00:00Z
commit_hash: <git sha of this file at commit>
val_set_hash: <sha256 of sorted held-out inchikey14 list>
val_set_spec: "held-out inchikey14; NP-like; timsTOF/QTOF spectra; n=<N> molecules"

final_two_rule: |
  Eligibility: scored on public LB without error; stress-set rerun <= 6h; code at or before 2026-12-04 freeze.
  Pick A (val-pick): highest local structure-disjoint MRR@25 among eligible versions whose public LB is
    within 0.05 of the best eligible public LB.
  Pick B (LB-pick): highest public LB among eligible versions. If identical to Pick A, instead the eligible
    version from a different scorer family with local val within 0.02 of Pick A.
  Ties (< 0.005 on either axis) break toward shorter rerun time.
  Applied on 2026-12-11 by the PI. Changes require an entry under `amendments`.

kill_criteria_ref: research/q7_compute_execution/ninety_day_plan_milestones_risks_and_kill_criteria.json  # table == kill_criteria

predictions:            # the contemporaneous "confidence statement": written BEFORE the public score is seen
  - version_id: <kaggle notebook version>
    predicted_private_mrr25: 0.00
    interval_90: [0.00, 0.00]
    written_utc: <timestamp>
    public_lb_seen_after: null   # filled with the score once known; never edited before

amendments: []          # {date_utc, what_changed, triggering_information, author}

selection_record:       # filled 2026-12-11
  pick_a: {version_id: null, clause: null}
  pick_b: {version_id: null, clause: null}