DIME
Get the picksJoin

PROJECT · findings note

project-board-warnings

/verdict/card + board warning surface — shipped 2026-08-25, closes the "engine knows but never says" gap

← All findingssource · memory/project_board_warnings.md

The gap: the verdict engine could say "she hasn't played in 16 days" and "Pinnacle has -106" only through POST /verdict/slip, which a reader reaches only by BUILDING a slip. The board — the site's main page — surfaced nothing. Live HTML carried no stale_window, inactive_player OR windows_disagree; the last two had shipped months earlier and were invisible the whole time.

Shipped 2026-08-25 (committed + pushed + deployed)

  • grading-service 8937615 + b2d4c05: GET /api/v1/verdict/card?league=&date=
  • nightly-picks a7a4cf3: board-verdict.ts / board-verdict-join.ts, BoardView wiring, Slip rendering, dime.css

Why a new GET, not the frontend POSTing /verdict/slip — all three MEASURED

  1. POST is uncacheable at every layer. Board views map 1:1 onto verdict builds. Measured 3.42s cold / 0.25s warm on 17 legs, cold cliff recurring at every 60s odds-TTL rollover (BuildBoardAsOf refetches; EnrichWindowProvenance loads a snapshot per game).
  2. game_started would paint every card red ~5h nightly. Tips 23:00Z, todayET() holds the date to 04:00Z. Verified on the completed 2026-08-24 slate: 6 of 6 legs returned game_started first. It's a SLATE fact → hoisted to tipped/tipped_text.
  3. slip_issues would fire on a board. 17 live rows produced same_player ×5 + same_game_same_side → envelope red for being a board. CheckSlip is never called by /verdict/card.

Facts worth keeping

  • Join key player_id|prop_type|line|side. Live card: 85 picks → 69 distinct; 16 duplicated across core/list tiers (0 conflicting verdicts, but separate formulas CAN carry different captured prices) → deduped server-side, core wins.
  • Leg.Book is left EMPTY on purpose. The card records the price, not the book. CheckPrice then skips its self-exclusion so every book votes on consensus, and BestOffer still reports a better number. Correct input, not a missing one.
  • no_market fires on ~60% of card rows and is GENUINE — the line moved (card wants Engstler assists 1.5; board offers 2.5). Most actionable thing on the page, not noise.
  • better_book is rare on the served board: 6 of 85 rows, and mostly kelly_stake_units: 0 rows the board doesn't serve. Keep it (+2.84pp is the only validated price edge) but don't build the value story on it.
  • Render every non-green reason, in engine order — a live card returned [amber stale_window, red bad_price], so highest-severity-only would delete the staleness flag. CheckLeg runs structural rungs before the price rung.
  • The site has ONE theme. color-scheme: light in globals.css:33 is the only match for prefers-color-scheme|data-theme|color-scheme in src/.
  • Scope: nightly CORE board only. /clv/picks carries no player_id, so the ?lens=/?formula= slates have no safe key — a name match would hand one player's warning to another.
  • Failure → null → board renders as today plus "Risk check unavailable". Silence must never read as absence of risk.

Related: project_window_provenance, project_conditional_minutes_refuted, project_roi_guard_nightly_tiers, feedback_pinnacle_default