vyndr/specs/vyndr-roadmap.md

VYNDR Master Roadmap — Generated 2026-07-10 (Session 56)

The foundation document. Derived from the Session-56 data audit (specs/propline-audit.md, verified against live MLB Stats / ESPN / Odds APIs). Every session after this references it. Update it as sessions ship.

---

Current coverage (verified 2026-07-10)

Sport	Odds props	Grading	Features (l5/l20)	Outcomes settled	Status
MLB	✅ PropLine + odds-api	✅	✅ real (statsapi)	✅ (S55, expanded S56)	LIVE
WNBA	✅ odds-api/PropLine	✅	⚠️ Python offline → thin	❌	PARTIAL
Soccer	✅ odds-api only (no PropLine)	⚠️ extractor only	⚠️	❌	PARTIAL
NBA	⚠️ summer league only	✅	⚠️ offline	❌	OFF-SEASON (Oct)
NFL	⚠️ preseason	⚠️ mapped, not e2e	❌	❌	PRE-SEASON (Sep)
NHL	❌ off-season	⚠️ mapped	❌	❌	DARK (Oct)

The core product is MLB. It is the only sport that is live end-to-end (odds → grade → real features → settled accuracy). Everything else is a build-out target.

---

Gap analysis (from the audit)

1. Outcome settlement is MLB-only — WNBA/NBA/soccer grades never settle, so the accuracy record (the S55 trust engine) only reflects MLB. Highest-value gap. Needs the ESPN box-score settle path (espnStatsAdapter already parses the shape for season avgs).
2. NBA/WNBA features depend on an offline Python nba_api service — l5/l20 are often empty; only season avgs survive (ESPN fallback). Needs an ESPN game-log feature path so intel populates without the Python service.
3. Soccer has no PropLine source — odds-api backup only; no SPORT_KEYS entry. Soccer grading isn't wired into the snapshot pipeline end-to-end.
4. NFL mapped but not graded end-to-end — markets normalize, but no feature/ outcome path. Wire before September.
5. RBI silent-failure (rbis/rbi) — FIXED S56.
6. Under-requesting MLB markets — FIXED S56 (6 → 12 markets).
7. No pipeline alerting / retry / missed-cron — FIXED S56.
8. No live calibration — the accuracy record exists but doesn't yet feed back into grade confidence (spec 2.3 from S55).
9. batter_strikeouts (batter Ks) mapped but not feature/settle-wired.

---

Session plan (priority-ordered)

Session	Focus	Ships	Scope
57	WNBA/NBA outcome settlement	ESPN box-score settle path in outcomeService → accuracy for basketball; WNBA goes fully live	M
58	ESPN game-log features	l5/l10/l20 for NBA/WNBA without the Python service → real intel on basketball grade cards	M
59	Soccer end-to-end	soccer into the snapshot pipeline (odds-api source), feature extractor wired, ESPN settle	L
60	Live calibration	accuracy record feeds grade-confidence adjustment (under/over-confident tiers nudge); "Model health: Calibrated/Learning" indicator	M
61	NFL pre-season prep	NFL feature + settle path (ESPN NFL boxscore) so Week 1 (Sep) is live	L
62	Prop breadth	batter_strikeouts, triples, pitcher walks; NBA pra/turnovers requested; WNBA extra markets	S
63	Accuracy depth	per-archetype hit rates, per-stat hit rates, player-level record ("VYNDR on Judge: 12-5"); ledger UI	M
64	NBA regular-season readiness	verify NBA pipeline for the Oct tip-off; depth-chart/cascade live	M

Scope key: S ≈ ½ session, M ≈ 1 session, L ≈ 1–2 sessions.

NOTE (Session 58): the numbered plan above predates the overhaul work order. The work order's phases (0 = kill the lies ✅ S57, 1 = truth infrastructure ✅ S58, 2 = slate UX, 3 = mobile, 2.5 below, 4 = scan/ parlay, 5 = records, 6 = landing/content) take sequencing priority; the table's items slot in where they don't conflict.

Phase 2.5 — Intraday line refresh + directional movement (LOGGED Session 58)

Authoritative scope from the Phase 1 GO prompt. Build after Phase 3 (mobile), before Phase 4. During slate hours (~noon–midnight ET):

• Lightweight ODDS-ONLY refresh every 15–30 min (no full re-grade run). Recompute the signed delta per graded prop RELATIVE TO THE GRADED SIDE — direction is the signal, not magnitude alone.
• Moved WITH the grade (market chasing our number): PropRow badge STEAM ▲ +N. Good for the record; entry edge compressed. No re-grade.
• Moved AGAINST the grade ≥ 1.0 (or odds-equivalent): auto re-grade THAT PROP ONLY.
  • Grade holds → badge VALUE ▲ better entry (better number, same read).
  • Grade drops → update the grade with revised_from_grade set (column already exists in ledger_entries) + a visible revision marker (original grade struck through — the ledger UI already renders it). Revisions are ALWAYS public — never silently regrade, per Ledger ethos.
• All displayed lines remain real book values from the refresh — the refresh CAPTURES market numbers, never computes them.
• After ship: drop SNAPSHOT_EXPECTED_INTERVAL to the refresh cadence — the HeartbeatBar SYNC badge (normal <1.5x · amber ≥1.5x · STALE ≥3x) goes genuinely live with zero further UI changes.

Phase 4.5 — WNBA settlement via ESPN box scores (DUE DATE, not "roadmap")

Logged Session 59. Pending-forever WNBA ledger rows are honest but become a credibility hole past ~2 weeks of accumulation. Due: within 2 weeks of day one of the public record (by ~Jul 24 2026). Scope: ESPN box-score settle path in outcomeService + ledgerService defaultGetPlayerStats (WNBA branch) so WNBA rows settle nightly like MLB. Build right after Phase 4 (scan/parlay polish).

Phase 5 addition — calibration by grade tier (logged Session 59)

Once settles mature: the MODEL tab shows A-tier hit% vs B-tier hit% SEPARATELY (same n≥20 rule PER TIER — a tier below threshold shows "building", never a %). The separation between tiers is the proof the grades mean something. ledger_entries.grade already carries the tier; this is an aggregate + UI change only.

---

Stat-type coverage target (fully built)

• MLB (batters): hits, total_bases, home_runs, rbi, runs, doubles, triples, walks, stolen_bases, batter_strikeouts. (pitchers): strikeouts, earned_runs, hits_allowed, innings_pitched, outs, walks. (Bold-new this session: doubles, triples, outs, runs, walks requested + settleable.)
• NBA/WNBA: points, rebounds, assists, threes, steals, blocks, turnovers, pra.
• Soccer: goals, shots, shots_on_target, tackles, cards, corners, saves, passes, clean_sheet.
• NFL: passing/rushing/receiving yards, receptions, pass/rush/rec TDs, anytime_td, interceptions, kicking_points.

Sport coverage target

• MLB ✅ live now (settled accuracy).
• WNBA → live after S57 (settlement) + S58 (features).
• Soccer → live after S59.
• NFL → ready for September (S61).
• NBA → ready for October (S64).
• NHL → ready for October (fast-follow once NBA path exists; shares ESPN pattern).

---

Slip Reader — vision-model upgrade (POST-REVENUE)

Session 9 (A1 board) shipped the zero-API Slip Reader: tesseract.js OCR (self-hosted WASM, free) + rigid per-book layout parsers (DK/FD/MGM/Caesars) with per-field confidence and needs_review honesty. That architecture is deliberately conservative — it reads clean screenshots of the four big books and refuses everything else.

The upgrade, when revenue funds it (zero-out-of-pocket rule): a vision model (Claude-class multimodal) replaces the OCR+layout-parser pair — one call reads ANY book, any theme, any crop, and returns structured legs with real confidence. Costs per-call money, so it is gated on paid tiers paying for themselves. The route contract (POST /api/slips/parse → { legs, needs_review, source: 'user_slip' }) is the stable interface; only the extraction engine behind it swaps. Keep the never-guess rule: model output below confidence threshold still nulls the field.

Operating invariants (do not regress)

• Three stat_type whitelists stay in sync: routes/analyze.js, routes/scan.js, python/utils/validation.py.
• A requested market MUST have a MARKET_MAP entry (else silent zero) AND, if MLB, a MLB_LOG_FIELD entry in BOTH featureCache and outcomeService (else no features / no settlement).
• The accuracy pill stays HONEST — "LEARNING" below MIN_SAMPLE, never a faked %.
• Pipeline alerts (ntfy) fire on success/failure/stale/overdue; the status probe (GET /api/internal/snapshot/status) exposes overdue.