Methodology notes, integrity incidents, and review-log entries. Falsified or corrected claims stay public; the change is recorded here.
**Date:** 2026-07-18 **Type:** presentation integrity **Status:** accepted Hypothesis pages must show the **first-spec git commit** (true pre-registration timestamp), not repository HEAD, with a clickable GitHub link and the run generation timestamp. Inverted ordering is rejected by automated audit. This closes a class of issues where the site could appear to invert its own Invariant 1 to a casual reader. ## Historical inversions A small minority of result cards have run timestamps earlier than the current path's first git-add commit (rebases, renames, or pre-git local runs). The site still displays the true first-spec commit for the path and the run time; it does not stamp repository HEAD. `scripts/audit_prereg_display.py` inventories these rows in `data/prereg_audit.json`.
**Date:** 2026-07-18 (retrospective entry) **Type:** quality audit **Status:** accepted internal audit A substantial share of result cards record primary estimator failure (e.g. rank deficiency) with fallback OLS, or report "effect magnitude effectively zero." These must not be treated as clean directional confirmations. Follow-up: CI estimator floor rejects directional SUPPORTED/PARTIAL when the primary estimator fails to converge or the reported effect is machine-epsilon.
**Date:** 2026-07-18 (retrospective entry) **Type:** methodology / integrity **Status:** accepted internal audit Social-outcome hypotheses that selected favourable indicator subsets without enumerating the canonical literature basket are capped at `supported_subset` rather than full `SUPPORTED`. Costa Rica wellbeing is the flagship example: adding the safety leg (homicide) reverses a prior positive reading. This entry seeds the public review log with an integrity incident already documented in methodology and how-it-works materials.
See also contribute and methodology.