vyndr/specs/resolution-ceiling-and-the-duel.md

The resolution ceiling — calibration is complete, and it was never the lever

PHASE 0 — two honest truths, on record

1. The swap is a BET, not an OOS win

On identical held-out rows the isotonic map scored BETTER: hits +0.0028 (CI [0.0013, 0.0045]), rbi +0.0042 (CI [0.0003, 0.0092]), total_bases tied.

We serve the low-parameter map anyway, on the untestable prior that isotonic's in-window edge is daily structure shared between the fit and evaluation windows and will not transmit forward. At 19 dates no instrument here can test that prior — LODO has 1.4–9.3% power against it.

Named as a bet, logged, not evidence. Phase 1 makes it falsifiable.

2. The MIN_SLOPE catch, as a standing guard rationale

runs fitted a = −0.032. A near-zero or negative slope collapses the curve toward base-rate-for-everything, which lowers Brier — shrinking a miscalibrated forecaster toward its base rate always does — while destroying all resolution.

A metric win that guts the product. Any calibration layer must refuse a non-positive slope on principle, not on inspection. That is now MIN_SLOPE.

---

PHASE 1 — the duel, instrumented forward

Both corrections are computed on every hits/TB prop: p_win_lowparam (served) and p_win_isotonic_shadow (logged, never read by serving or by chainAcross). The shadow runs in its own try — it can never break serving.

calibrationDuel.adjudicate encodes the rule in code, before any forward date exists, so the bar cannot drift toward whichever answer arrives:

condition	verdict	action
≥10 forward dates AND isotonic wins, date-block CI excluding zero	REFUTED	revert hits/TB to isotonic, log the reversal
≥10 forward dates, isotonic does not win	UPHELD	keep serving low-param
<10 forward dates	PENDING	keep serving, no verdict

A date counts as forward only if neither map was fitted on it — scoring inside a fit window would ask which map memorised better. Rows lacking that provenance are dropped, never assumed forward. Tests lock all of it, including that a decisive shadow win at 5 dates is still PENDING.

Nothing swaps now. The season decides.

---

PHASE 2 — the resolution ceiling, quantified

Murphy decomposition: Brier = reliability − resolution + uncertainty. Reliability is what calibration fixes. Resolution is discrimination, and a monotone map cannot create it — it relabels bins without re-sorting the rows inside them.

stat	n	base	reliability	resolution	uncertainty	share of variance explained
hits	1,140	0.5684	0.01353	0.00252	0.24532	1.03%
total_bases	1,050	0.5819	0.01419	0.00442	0.24329	1.82%
rbi	630	0.6571	0.00654	0.03268	0.22531	14.51%
runs	597	0.6348	0.00788	0.00130	0.23182	0.56%

What calibration did, exactly as theory predicts

stat	reliability	resolution
hits	0.01353 → 0.00233 (−0.0112)	0.00252 → 0.00231 (−0.0002)
total_bases	0.01419 → 0.00527 (−0.0089)	0.00442 → 0.00414 (−0.0003)

Calibration removed 83% of hits' reliability error and moved resolution by essentially nothing. It did the whole of its job, and its job was never the one the grade product needs.

Unexpected: rbi has 13× the resolution of hits and is the one stat we do not serve corrected — it needs calibration least (reliability 0.0065) and discriminates most. Worth carrying into the factor arc.

---

PHASE 2 — the HITS factor-transmission diagnosis

Verdict: NOT-TRANSMITTED. Not weak — absent. This is a plumbing defect, and it is the highest-value finding in the order.

Evidence, traced in code rather than recalled:

1. sprayDefense.js and platoonSeverity.js are required by NOTHING in src/. Only by analysis scripts and their own tests. The two causally-correct atoms that passed the two-part gate have never been on the serving path.

2. The served p_win reads exactly four inputs (intelligence/probabilityEstimator.js:54): game-log frequency over the line, opp_rank_stat (±0.03), home_away (±0.015), and a cv consistency pull. Zero occurrences of spray, platoon, hard-hit or contact-profile.

3. Ordering makes it structural. snapshotService grades at line 454 (gradeAndCacheSlate) and only computes challenger/context at line 640+. Everything proven is computed downstream of the grade it would inform.

So the three proven hits factors — defense_by_direction, pitcher_contact_profile, platoon_severity — were measured on ledger rows by analysis scripts and have never once moved a served number.

That reframes every null in this programme's recent history. "Calibrated p_win does not separate within archetype" was never a statement about factors. The factors were not in the forecast.

---

PHASE 3 — bands on served values

stat	basis	eval rows	slots	slots with lift
hits	p_win_lowparam (served)	765	7	0
total_bases	p_win_lowparam (served)	625	7	0
rbi	raw p_win	425	7	0
runs	raw p_win	424	7	0

28 archetype slots across four stats. Zero show lift. Every slot is one band indistinguishable from its own base rate.

This is no longer an open shrug. It is the arithmetic consequence of resolution of 0.0013–0.0327 against uncertainty of ~0.23: a forecast explaining 1% of the outcome's variance cannot produce bands that separate, and no correction to its numbers will change that.

---

THE HEADLINE

Calibration is complete. It delivered honest numbers on two stats and ZERO grade separation, because the counter has no resolution — 1.03% of variance on hits, 0.56% on runs.

And the three proven hits factors are NOT WIRED INTO THE FORECAST AT ALL.

Those two facts together are the programme's position. The second is the reason for the first, and it is a plumbing defect rather than a modelling wall — which makes it the cheapest high-value fix available.

Per-archetype grades require proven factors that actually reach p_win. That is the next and central arc. This is the last calibration order.

---

Invariants

Serving unchanged from 74cf1ce — this order logs and diagnoses. p_win never mutated; p_win_lowparam served, p_win_isotonic_shadow logged. No Bonferroni slot. Counter and frozen clusters verified file-by-file.