DIME
Get the picksJoin

PROJECT · findings note

project_methodology_lockdown

4 methodology discipline checks hard-coded into data-hydrator PropDiagnostic (commit 8354c43, live in prod 2026-05-20)

← All findingssource · memory/project_methodology_lockdown.md

4 fixes, surfaced per-stat on every PlayerBundle.PropDiagnostic.Stats[stat]:

  1. FIX 1 — Role anchor (flags.role_anchor_adjusted, current_season_min_avg, historical_min_avg). If recent minutes diverge from L20 minutes by ≥5, filter games to minute-matched window before computing recency/hit rates. Catches the Boston rebounds case (L20 had 30+min games, current role 23).

  2. FIX 2 — Line-vs-mean trap (line_edges[line].quality, line_edges[line].line_z_score). For every half-integer line in the projection band, classify edge: mispriced_over / stretched_over / no_edge / stretched_under / mispriced_under based on z-score distance from posterior mean. Lines within ±0.3σ flagged no_edge (50/50 traps). Catches the Hiedeman 2.5 reb case.

  3. FIX 3 — Minutes-source cross-check (flags.minutes_source_confidence). Compares lineup-service blended projected_minutes vs RotoWire raw vs bundle L5 actuals. Returns stable / injury_haircut / diverge_low_confidence / insufficient_data. Catches the J Williams case (blended 21.1, RotoWire 30, status "No" → can't trust blended).

  4. FIX 4 — Recent-vs-baseline (flags.recent_vs_baseline_flag, flags.season_avg). Flags slump / hot_streak / stable based on L10 vs season-avg deviation ≥15%. Catches the Vassell case (L10=11.9 ppg slump, season=13.6 → don't extrapolate L10 onto the line).

Read the flags BEFORE bet recommendations. If recent_vs_baseline_flag = slump AND line is set below season_avg, the line is mean-reversion bait — pass or take the OVER. If minutes_source_confidence = diverge_low_confidence, don't trust the projection's minutes — manually verify RotoWire/practice report. If line_edges[line].quality = no_edge, it's a 50/50 — pass.

Verified live 2026-05-20 on Hartenstein rebounds:

  • season_avg: 9.2, current_min: 24.3, historical_min: 23.6 (stable role)
  • minutes_source_confidence: "diverge_low_confidence" (caught at runtime)
  • line_edges 7.5 and 8.5 both = "no_edge" (z within ±0.3)

Tests in internal/snapshot/diagnostic_test.go (7 cases lock the rules).

See also: feedback_betting_discipline, project_scout_service_topology, project_wnba_launch_state.