vyndr/specs/rbi-resolution-anomaly.md

The rbi anomaly, decomposed — the model is out-resolved by a batting-order integer

PHASE 0 — the 14.51% is REAL

The figure came from the harness that produced a false null three times tonight, so it was re-derived with a paged pull asserted against an exact count:

stat	exact rows	paged rows	scorable	resolution	share
rbi	7,930	7,930 ✓	630	0.03268	14.51%
hits	11,690	11,690 ✓	1,140	0.00252	1.03%
total_bases	12,086	12,086 ✓	1,050	0.00442	1.82%
runs	6,440	6,440 ✓	597	0.00130	0.56%

Reproduces exactly. rbi deciles are monotone through the middle (0.552→0.577, 0.646→0.691, 0.746→0.784) with a genuinely low bin at 0.13→0.122, and 20 raw rows are printed in the artifact for hand audit.

---

The caveat that governs every number below

The naive forecasts are leave-one-out on the evaluation window itself — they see that player's performance in the very rows being scored, while the model is strictly point-in-time. They are therefore upper bounds on available resolution, not fair competitors. Every comparison is read that way.

---

PHASE 1 — the three-way split

stat	MODEL	(a) player base rate	(b) lineup slot	(c) within-stratum
rbi	0.03268	0.01167	0.03608	0.01908
hits	0.00252	0.00446	0.00100	0.00473
total_bases	0.00442	0.01331	0.03448	0.00607
runs	0.00130	0.00262	0.01156	0.01170

Finding 1 — rbi's resolution is LINEUP ROLE, almost exactly

Batting-order slot alone resolves 0.03608 against the model's 0.03268. A single integer — where he hits in the order — accounts for the entire anomaly and slightly more. That is real predictive signal and it is opportunity, not skill: the cleanup hitter bats with runners on, the 8-hole hitter does not.

Player base rate alone gives 0.01167, so ~36% of the model's rbi resolution is matched by knowing only who is batting.

Within strata of similar-base-rate players the model still resolves 0.01908 — 58% of its total, and higher than any other stat's entire model resolution. So rbi does carry genuine within-role discrimination on top of the role effect.

Finding 2 — on three of four stats the model is beaten by "he's a .270 hitter"

stat	model	player base rate alone
hits	0.00252	0.00446	base-rate-only resolves 1.8× the model
total_bases	0.00442	0.01331	3.0×
runs	0.00130	0.00262	2.0×
rbi	0.03268	0.01167	model wins, 2.8×

Even allowing that the naive forecast peeks at the window, a 1.8–3.0× gap is not explained by that advantage alone. The served counter — a frequency over the line, blended with the last five games, nudged by opponent rank and home/away — appears to destroy discrimination relative to the player's own rate. The recency blend and the ±0.03/±0.015 nudges move predictions in ways that do not track outcomes.

rbi is the one stat where the model beats the naive baseline.

Finding 3 — lineup slot out-resolves the model on THREE stats

stat	model	slot alone
total_bases	0.00442	0.03448 (7.8×)
runs	0.00130	0.01156 (8.9×)
rbi	0.03268	0.03608 (1.1×)
hits	0.00252	0.00100 (model wins)

Hits is the only stat where batting order carries less than the model — which makes sense: a hit is a hit whether you bat first or ninth, but runs, RBI and total bases all scale with opportunity.

---

PHASE 2 — which world

All three are partly true, in measured proportions:

• WORLD A — rbi resolution is real and role-driven: ~90% TRUE. Slot alone (0.03608) covers the model's entire rbi resolution. It is real, it is contextual rather than skill-based, and 58% survives within similar-player strata as genuine discrimination.
• WORLD B — base-rate-spread artefact: ~36% TRUE for rbi. Player identity alone accounts for about a third. Not the main story for rbi — but for hits, TB and runs it is the whole story and then some, since base rate alone out-resolves the model on all three.
• WORLD C — hits is intrinsically compressed: TRUE, and the ceiling is low. Total available spread resolution for hits is 0.00446 — 1.8% of variance even from a forecast that has seen the answers. Perfect factors cannot make hits a high-resolution grade.

---

PHASE 3 — the roadmap-deciding log

The next arc is not "strengthen hits factors." Hits has the lowest available resolution on the board (1.8% ceiling from an oracle-ish baseline) and last order's wiring already lifted it to 1.39% of a 1.8% ceiling. There is very little left there.

Named first factor order for next session: LINEUP SLOT / RISP OPPORTUNITY on rbi, through the two-part gate.

• The input is already ingested and prod-verified (lineup_context batting order, hitter_opportunity RISP share, S89).
• Its resolution is measured, not hypothesised: 0.03608 slot-only on rbi, 0.03448 on TB, 0.01156 on runs.
• The causally-correct unit is plate appearances with runners on, which is what RISP share measures directly — the crude version is the slot integer.
• It must clear the two-part gate like anything else. Measured availability is not a pass.

Second, and higher-value than either: the counter is out-resolved by a player frequency table on three of four stats. That is not a factor problem — it is a defect in the champion. Diagnosing whether the recency blend and the ±0.03/±0.015 nudges are destroying discrimination is the biggest single lever this decomposition found, and it costs nothing to test: they are three lines in probabilityEstimator.

The hits transmission win stands. The conduit is real and permanent — sign-verified, 75% coverage. This order changes only which stat has the most worth flowing through it.

---

Invariants

Diagnostic only. No factor wired, no serving path changed, no calibration refit, p_win untouched, all frozen modules byte-identical including the hits path from 43f65d3. No Bonferroni slot — this is resolution accounting, not a causal claim.