STATE: Session 73 — park factors derived + composable; pipeline gap documented

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VCNgGSt5qvcLxaeQqa7Zpj
This commit is contained in:
Kev
2026-07-21 01:06:49 -04:00
parent 3ac91c3d96
commit 63e4bce858
+75 -1
View File
@@ -1,5 +1,5 @@
# VYNDR — STATE OF THE WORLD
### As of `927e867` (main, DEPLOYED), 2026-07-21. This file opens every future session. **Start with the CURRENT STATUS + OPEN ITEMS block below.**
### As of `3ac91c3` (main, DEPLOYED), 2026-07-21. This file opens every future session. **Start with the CURRENT STATUS + OPEN ITEMS block below.**
---
@@ -131,6 +131,80 @@ exist locally; harmless, the data restores completely.
`specs/snapshot-retention.md` (Phase 2 report + schema) ·
`specs/audit-data/grade-collapse.md` · `specs/audit-data/gate-simulation.md`.
---
# 🟡 SESSION 73 — PARK FACTORS (derived, composable) (2026-07-21)
**Derivation + composition BUILT and proven. PIPELINE WIRING INCOMPLETE — see
the gap.** Layer 3 Step 4.
## 🔴 PHASE 0 — A STATIC PARK TABLE ALREADY EXISTED (nearly built a duplicate)
`src/data/parkFactors.js`**static FanGraphs table, Session 15**, consumed by
`computeFeatures.js:374`. I almost created a second source of truth.
- **NEW derived service = `src/services/parkFactors.js`** (distinct path).
- **Checked:** the static table reaches `computeFeatures` but **NOT
`probabilityEstimator`** → it does **not** currently touch `p_win`.
- **This reframes the order:** if the champion ever does see park, adding it to
the challenger is DOUBLE-COUNTING — precisely what the S72 harness catches.
Park therefore ships as a **nominated challenger to be TESTED**, not an
assumed win.
## Derivation (ours, from real games)
statsapi `schedule?hydrate=linescore,venue,scoringplays` — **one call per date
range**. Every HR scores ≥ the batter, so **HR totals are fully recoverable from
scoringPlays**. Derived from **5,055 real games, 20222025**:
| Park | HR base | RUN base |
|---|---|---|
| Dodger Stadium | **1.106** | 1.023 |
| Coors Field | 1.029 | **1.099** |
| Oracle Park | **0.923** | 0.969 |
| PNC Park | **0.917** | 0.994 |
**18 parks present · 18 honestly absent** (below the 150-game regime floor).
*(Sample was a partial pull — month-chunked with `-31` end dates drops days — so
absences are inflated. Re-derive with proper date ranges before trusting counts.)*
## COMPOSABLE — the architectural point
Emits a **multiplier ~1.0**, never an additive nudge, because weather must
modulate it next: `effective = park_base × weather_mod × …`. Additive terms
don't compose (5% park + 8% wind = 1.05×1.08, **not** +13%). The challenger
converts via `Math.log(env)` into the same log-odds space the archetype nudges
use. A test multiplies a placeholder weather term on to prove the shape composes
**with no rearchitecting**.
## Honest states (three, distinct)
`absent` (thin → adjust nothing) · `present` (adjust) · **`weather_na`** (dome —
**the park factor STILL APPLIES**; the flag exists so next order's weather does
nothing there). **N/A ≠ absent** — conflating them either drops a valid park
factor or applies wind indoors.
## Directional by prop-owner
`home_runs` and `home_runs_allowed` both key off `hr_base` **in the same
direction** — the sign lives in the STAT, not the park.
## Structural breaks + shrinkage
A deviating season starts a new regime **only if the NEXT season confirms** (one
odd year = noise). Only post-break seasons used. Factors **regress toward
neutral by sample size**; fine conditioning (park×hand, park×batted-ball) stays
**unavailable** until its own larger floor (400 games).
## 🔴 THE GAP — park is NOT yet reaching live props
`adjust()` accepts `environment`, and `parkMultiplier` produces it — but
**nothing in `snapshotService` looks up a venue per prop and passes it**. Props
carry `game_time` (via gameBinder) but **no venue**. Remaining wire-up:
1. Attach venue to each prop (statsapi schedule already returns it).
2. Persist derived factors (a table or cached blob) + a refresh job.
3. Pass `environment` into `attachChallenger`.
**Until then park adjusts nothing live** — the derivation and composition are
proven, the plumbing is not.
## Next
1. Wire venue → prop → park lookup (the gap above).
2. **Weather modulation** — composes by multiplying `park_base`; domes already
flagged `weather_na`.
3. Still blocking all measurement: **no settled row carries `p_win` yet**.
---
# 🟢 SESSION 72 — TIER-1 MAPPINGS LIVE + TIER-2 HARNESS (2026-07-21)