DIME
Get the picksJoin

PROJECT · findings note

project_wnba_debate_chain_blind

WNBA debate/blog chain refused ALL picks because player agents got 0 odds_lines — extract_player_bundles read odds from the wrong shape AND joined on null player_id (+ a 5000-char prompt truncation cut prop_diagnostic). FIXED, MERGED TO MAIN + PUSHED (sport-agents 5029278) 2026-06-03; added B5 odds-aware selection. Data-layer verified. Remaining: deploy cron image + full LLM chain run.

← All findingssource · memory/project_wnba_debate_chain_blind.md

FIXED 2026-06-03 (code change in core/hydrator_client.py, NOT yet committed/ deployed). Root cause was ODDS, not prop_diagnostic. In extract_player_bundles:

  1. odds_data = _resolve_typed_or_raw(snapshot,"odds") returned {} for WNBA — typed odds is null; real data is raw_sections.odds, a LIST of per-game odds objects (whole slate: 4 WNBA + 8 MLB). The resolver only unwraps dicts → the list fell through to {} → zero props read.
  2. Props carry player_id: null (correlation_status:"pending") so the id-keyed lookup dropped every one → must join by player_name.
  3. Raw feed is the full alt-line firehose (~300 quotes/player) → added _condense_player_odds() (consensus main line per prop_type per book). Verified at data layer: NBA path untouched (pb.props short-circuits before condense). prop_diagnostic actually flowed fine for the players who had one; with 0 odds the agent SKIPs regardless, so that was secondary noise.

MERGED TO MAIN + PUSHED 2026-06-03 as 5029278 (fast-forward; rode along with 3985bae matchup-defender-context, which is safe — graceful .get() defaults + dead-code removal). Also added the audit's B5 follow-up — odds-aware player selection (extract_player_bundles now ranks the lineup by (has_odds, is_starter, projected_minutes, season_mpg)), because the WNBA lineup is a roster fallback and the old starters-first sort dropped priced stars (Ionescu, Jones) for alphabetical bench. Re-verified: NYL 6/8, TOR 4/8 players now carry odds_lines; NBA confirmed-starter selection unchanged; 16/16 tests pass. Remaining: push, deploy to sport-agents cron image, and a full LLM chain run (-m core.blog_generator --league wnba --model haiku) to confirm non-empty best_bets.

ORIGINAL DIAGNOSIS BELOW — Ran the real production pick chain python3.13 -m core.blog_generator --league wnba --date 2026-06-02 --model haiku (see reference_sport_agents_run_python). All 4 games completed every phase (team→coach→matchup→player→synthesizer + multi-persona debate) but returned SKIP / empty best_bets / empty parlay_of_day on every game.

Root cause (verified): player agents get no data to bet on. Uniform per-player skip reasons in output/blog/2026-06-02/wnba/haiku/*.json:

  • "No odds_lines provided — cannot price any prop"
  • "No prop_diagnostic field — cannot run lockdown gates"
  • "Coach projected minutes: 0 with 0% confidence"

The chain sources odds_lines + prop_diagnostic ONLY from the snapshot via hydrator_client.extract_player_bundles (:277 odds_lines = pb.props or odds_by_player[pid]; :286 prop_diagnostic = pb.prop_diagnostic). Two breaks:

  1. [internal detail removed]
  2. prop_diagnostic comes through empty for WNBA even though daily_props.py reads it from the SAME hydrator snapshots fine. Exact reason (snapshot shape vs fetch params vs player_id key mismatch) NOT yet confirmed — verify during the fix.

NBA/NCAAB/NFL/MLB adapters all explicitly attach odds_lines/prop_diagnostic in summarize_player_data; WNBA subclasses NBAAdapter so it inherits the SAME code — the gap is in the DATA reaching the snapshot/bundle for WNBA, not the adapter code.

Significance: answers "what's the point of all this if we're not using it" — the lockdown (project_methodology_lockdown, feedback_active_tonight_gate) and the personas (project_debate_persona_lockdown) both WORK, but for WNBA this chain runs blind to the odds + diagnostic that project_daily_props_tool joins correctly. Two consumers of the same data; only one is wired. Chain also emits no tagged picks for WNBA → never enters project_multi_formula_clv_harness measurement.

Fix (not yet done): thread odds-service props into per-player odds_lines for WNBA (build odds_by_player from OddsClient and pass into extract_player_bundles, or merge into snapshot.odds_data.player_props before extract), and confirm prop_diagnostic survives the WNBA snapshot→bundle path. Then rerun + verify non-empty best_bets.