tb-v1 report + plan: diagnosis on trial, branch pre-registered

Firing verified on a real prod snapshot: 10/10 total_bases props carry
proj_tb_p_over. The snapshot HTTP call returned 524 (Cloudflare's 100s
origin timeout vs a ~115s snapshot) but the work completed server-side --
confirmed from the ledger rather than assumed.

Face validity is good and diagnostic: means agree almost exactly with the
ladder (1.813 vs 1.833), so this is a SHAPE-ONLY intervention, which is
what was intended. Component rates are plausible, and Carroll's triples
rate (0.112, far above his peers) is a clean check -- he is a speed player
and the model sees it.

AN OBSERVATION I AM NOT RESOLVING BY EYE: tb-v1 reads systematically LOWER
than the ladder (0.424 vs 0.540 at the same mean). That is the expected
DIRECTION, since the NB overstates P(>=2) by treating a home run as four
accumulating events -- but whether 0.424 is right or an overcorrection is
not knowable from face validity. A ~1.8-TB hitter clearing 1.5 empirically
sits nearer 45-50%, between the two. I am not claiming tb-v1 is better; the
holdout decides.

BRANCH PRE-REGISTERED, before the result, so the verdict cannot be
reinterpreted afterward: improves -> family-mismatch HOLDS, similarity
stays off the critical path, hits is next; does not improve -> hypothesis
WRONG and the mean-weakness/similarity branch REOPENS.

Also recorded: I hit Number(null)===0 in my own new module -- a null
component rate treated as a measured zero, the difference between "never
triples" and "we don't know his triple rate". A test caught it. Sixth
appearance of this trap in this codebase, and it caught the person writing
the warnings about it.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QJs13VsyiSKYQP6rj3NNmc

specs/MASTER-PLAN.md

@@ -29,33 +29,27 @@ stay provisional until re-run** · documented ≠ verified.
 
 ## ▶ NEXT EXECUTABLE ORDER
 
-**TARGETED PER-STAT FIX for `hits` and `total_bases` in proj-v1.1.** Diagnosis:
+**AWAIT the tb-v1 verdict, then `hits`.** `specs/tb-v1-compound.md`.
-`specs/proj-v11-diagnosis.md`.
 
-The ladder's loss is **concentrated, not systemic**. On matched, direction-aligned
+**tb-v1 is LIVE as a challenger** — total_bases modelled as the compound outcome
-rows (n=437) the real gap is **0.252 vs 0.352** — not the 0.108 vs 0.331 previously
+it is (per-component Poisson rates convolved exactly) instead of one negative
-reported, which was **unaligned on direction across 31.4% under-graded rows**.
+binomial that treats a home run as four events. Firing on **10/10 TB props**;
+means agree with the ladder, **only the shape differs** — the intended
+intervention.
 
-**`hits` (n=245) and `total_bases` (n=49) are 67% of the sample and carry
+**The diagnosis is on trial, and the branch is pre-registered:**
-essentially no signal** (0.060 and 0.009). Everything else is fine or better:
+- **tb-v1 improves TB resolution** → family-mismatch **HOLDS**, similarity stays
-**`walks` 0.519 vs champion 0.544**, and on **`doubles` the ladder's MEAN beats the
+  **off** the critical path, **`hits` is next** (AB-bounded, mostly 0.5 lines, so
-champion's** (0.207 vs −0.062). Priors and plumbing are clean; compression was
+  it rides on P(0)).
-tested and **disproved** (spread ratio 0.80–0.94).
+- **tb-v1 does not improve** → hypothesis **WRONG**, the **mean-weakness /
+  MLB-similarity branch REOPENS**.
 
-**Hypothesis to test in that order:** the negative binomial mismatches those two
+**Verdict n-blocked:** 0 settled TB rows carry tb-v1 yet; first settle is
-stats' structure — `total_bases` is a **weighted sum**, not an event count, and
+tomorrow. `scripts/tb-compound-holdout.sql` is TB-rows-only and direction-aligned.
-`hits` is **bounded by at-bats**, so both violate the unbounded-count assumption
-that `walks`/`runs`/`doubles` satisfy.
 
-> **This diagnosis REMOVES the MLB similarity build from the critical path.** That
+**Then:** `archetype_x_archetype` → retire/rewrite `bayesianEngine`.
-> branch assumed a *global* mean weakness; the mean is fine or better on three of
-> six stat families. Similarity may be worth building later — **on evidence, not
-> on this.**
 
-**Then:** `archetype_x_archetype` (matchup's upper rung) → retire/rewrite
+### FIVE challengers accruing in parallel — do NOT re-run early
-`bayesianEngine` (keys on 7 non-live stat names; defaults count stats to Gaussian).
-
-### FOUR challengers accruing in parallel — do NOT re-run early
 Verified firing on a real prod snapshot (293 grades), not inferred:
 
 | axis | coverage | mean \|nudge\| | holdout query |
@@ -63,6 +57,7 @@ Verified firing on a real prod snapshot (293 grades), not inferred:
 | **environment** | **84.6%** | 0.057 | *(shares the arch-v1 pattern)* |
 | **matchup** (`batter_own_split`) | **82.9%** | 0.015 | `scripts/matchup-axis-holdout.sql` |
 | **opportunity** | 30.0% | 0.142 | `scripts/opportunity-axis-holdout.sql` |
+| **tb-v1** (compound total_bases) | **10/10 TB props** | *(shape, not a nudge)* | pre-registered: improves → `hits` next; doesn't → similarity reopens |
 | **proj-v1.1** (distribution ladder) | **94.2%** | *(forms the projection, not a nudge)* | 437 settled — aligned gap **0.252 vs 0.352**, concentrated in `hits`+`total_bases` |
 
 All three orthogonal (r ≈ 0 vs projection, `p_win`, line and each other). Each

specs/tb-v1-compound.md

@@ -0,0 +1,139 @@
+# tb-v1 — TOTAL BASES AS A COMPOUND OUTCOME (challenger)
+
+**Date:** 2026-08-02 · current ladder + champion `p_win` **byte-identical** ·
+ledger-only · **firing verified on a real prod snapshot**.
+
+**Gates:** 4,104 tests / 329 suites green · `next build` exit 0 · tb-v1 on
+**10/10 total_bases props** · holdout committed, verdict n-blocked.
+
+---
+
+## STEP 0 — COMPONENTS CONFIRMED ON REAL DATA
+
+statsapi has **no `singles` field**. But on a real 10-game log:
+
+```
+singles = hits − doubles − triples − homeRuns
+singles + 2·doubles + 3·triples + 4·homeRuns  ==  stored totalBases   ✓ exact
+```
+
+**The decomposition is exact, not an approximation.** Verified before any code
+was written — this fix depended on it.
+
+---
+
+## THE MODEL
+
+Each component gets its own per-game Poisson rate; TB is their weighted sum, and
+the PMF is built by **exact convolution** rather than simulated (TB support is
+small). It inherits the **same combined multiplier** proj-v1.1 computes, so the
+two models differ **only in structure**.
+
+### Why this is the fix, in one number
+
+With **identical mean TB of 1.0**:
+
+| hitter | P(TB≥2) | **P(TB≥4)** |
+|---|---:|---:|
+| pure home-run hitter | 0.221 | **0.221** |
+| pure singles hitter | 0.264 | **0.019** |
+
+**A 12× difference at the 4-base rung that an NB on TB alone cannot express**,
+because it treats one home run as four independent events. A test asserts this
+separation, and asserts that P(TB≥4) for a pure-HR hitter equals P(at least one
+HR) **exactly**.
+
+### The approximation, stated
+
+Components are modelled as **independent** Poissons. They are not: a plate
+appearance that becomes a double cannot also become a single, so they are weakly
+negatively correlated and **independence slightly overstates the tail**. Closer
+to the truth than what it replaces; **not a solved problem**, and labelled
+`independence_caveat: true` on every record.
+
+---
+
+## FIRING ON PROD — 10/10 TB PROPS
+
+| | |
+|---|---:|
+| rows written | 241 |
+| total_bases props | 10 |
+| **with `proj_tb_p_over`** | **10 (100%)** |
+| range / sd | 0.295–0.424 / 0.045 |
+
+Real records:
+
+| player | line | ladder mean | **ladder P** | **tb-v1 P** | tb-v1 mean | rates (1B/2B/3B/HR) | games |
+|---|---:|---:|---:|---:|---:|---|---:|
+| Ketel Marte | 1.5 | 1.813 | **0.540** | **0.424** | 1.833 | .614/.208/.028/.180 | 106 |
+| Corbin Carroll | 1.5 | 1.814 | 0.540 | 0.424 | 1.815 | .468/.206/**.112**/.150 | 107 |
+| Gabriel Moreno | 1.5 | 1.590 | 0.471 | 0.405 | 1.608 | .773/.221/.000/.098 | 84 |
+
+**The means agree almost exactly** (1.813 vs 1.833) — both use the same rate
+machinery. **Only the shape differs**, which is precisely the intended
+intervention. Carroll's triples rate (0.112, far above the others) is a good
+face-validity check: he is a speed player, and the model sees it.
+
+### An observation the holdout must settle
+
+**tb-v1 is systematically LOWER than the ladder** (0.424 vs 0.540 at the same
+mean). That is the expected *direction* — the NB overstates P(≥2) by treating a
+home run as four accumulating events. But **whether 0.424 is right, or an
+overcorrection, is not knowable from face validity.** A ~1.8-TB hitter clearing
+1.5 empirically sits nearer 45–50%, which is between the two.
+
+**I am not claiming tb-v1 is better. The holdout decides.**
+
+---
+
+## HOLDOUT — COMMITTED, VERDICT n-BLOCKED
+
+`scripts/tb-compound-holdout.sql`, with two guards baked in:
+
+1. **TB rows ONLY.** total_bases is 49 of 437 settled rows; averaging into other
+   stats would hide the effect entirely.
+2. **DIRECTION-ALIGNED.** `p_win` is P(graded side); both projection values are
+   P(over). 31.4% of rows are under-graded, and the unaligned comparison is the
+   artifact that accounted for **41%** of the ladder's apparent loss.
+
+**Settled TB rows carrying tb-v1: 0** — it went live today, on games not yet
+played. First settle pass is tomorrow.
+
+---
+
+## THE DIAGNOSIS IS ON TRIAL — recorded in advance
+
+> **If tb-v1 materially improves TB resolution** → the family-mismatch hypothesis
+> **HOLDS**, similarity stays **off** the critical path, and **`hits` is next**
+> (its own order: AB-bounded, mostly 0.5 lines, so it rides on P(0)).
+>
+> **If it does NOT improve** → the hypothesis is **WRONG**, and the
+> **mean-weakness / MLB-similarity branch REOPENS**.
+
+Written down now, before the result, so the verdict cannot be reinterpreted after
+the fact.
+
+---
+
+## A TRAP I HIT IN MY OWN CODE
+
+`Number(null) === 0` — again, and this time in the new module. A null component
+rate passed a naive finite check and was treated as a **measured zero**. That is
+the difference between *"this player never triples"* and *"we do not know his
+triple rate"*, and it would have silently narrowed the distribution.
+
+**A test caught it.** Both `tbPmf` and `tbMean` now reject `null`/`''`/boolean
+strictly. Recording it because this is the sixth time this specific trap has
+appeared in this codebase, and it caught the person who has been writing the
+warnings about it.
+
+## TAGS
+
+**VERIFIED:** the TB decomposition is exact on real logs · tb-v1 fires on 10/10 TB
+props · means agree with the ladder (shape-only intervention) · component rates
+face-valid (Carroll's triples) · 12× separation at the 4-base rung.
+
+**NOT CLAIMED:** that tb-v1 is better. n-blocked until TB rows settle.
+
+**STATED LIMITATION:** component independence overstates the tail.