Skip to main content
Read-only snapshot: model data as of 2026-07-16; paper desk record through 2026-10-22. No live data; paper units only.
CourtVision
Ask Scout

Measurement / State-conditioned reliability

Where calibration changes during the game.

Calibration can differ between early and late game states. This view groups published forecasts by both game phase and probability band. Use the grids to compare the recorded model and reference rows without assuming their matching labels identify the same predictions.

For aggregate reliability bins and bootstrap intervals, see Calibration reliability. Those bins summarize a different published grouping.

Artifact date: 2026-09-16

Data integrity

These MLB/soccer numbers are revision 2, computed on the segment-clean corpus (2026-09-16). Revision 1 values are withdrawn and kept in the regeneration receipt. Segmentation excluded 51,635 of the 78,986 MLB ticks (65.4 percent); the specifically identified label mismatches were a smaller 27,076 (34.3 percent).

  • state_conditioned_calibration
Read the full finding

Reading trail

Read first: Read the calibration reliability bins before comparing state-conditioned cells.

Next question: Which published game states carry the largest recorded absolute residuals?

Read the analysis: How to read a CourtVision paper (sources regenerated)

PaperSupporting sourceDoes calibration hold up over time and across market types? (sources regenerated)Cross-sport | as_of 2026-09-16PaperSupporting sourceHow to read a CourtVision paper (sources regenerated)Cross-sport | as_of 2026-09-16PaperSupporting sourceWhen a lead lasts (sources regenerated)Cross-sport | as_of 2026-09-16PaperSupporting sourceWhere forecasts match outcomes (sources regenerated)Cross-sport | as_of 2026-09-16PaperSupporting sourceWhere the forecast errors live (sources regenerated)Cross-sport | as_of 2026-09-16ModuleSupporting sourceState Conditioned CalibrationCross-sport | as_of 2026-09-16FindingSame source fileMLB in-game join integrityMLB | as_of 2026-09-16InspectorNext questionCalibration reliabilityCross-sport | as_of unrecordedPrerequisite: Read the reliability finding before comparing published bins.

MLB: 27,351 forecast observations across 15 cells; 668 skipped without a state field.

Matching time and probability labels do not establish paired membership. Model and reference support are shown in aligned source grids; their populations can differ.

Showing Signed gap (pp) for MLB. Signed gap is mean_y minus mean_p; absolute gap is the published calibration_error.

Signed gap scale: observed minus forecastSupport scale: published observation countnot published

Model source

Model source by time and probability bucket for MLB
Time0-.2.2-.4.4-.6.6-.8.8-1
early(inn1-3)not publishedn not published-11.61 ppn 1,033-2.15 ppn 5,754-0.29 ppn 2,702not publishedn not published
mid(inn4-6)-13.79 ppn 1,088-11.78 ppn 2,2850.27 ppn 2,4687.33 ppn 2,44911.45 ppn 1,462
late(inn7+)-1.18 ppn 2,884-17.77 ppn 617-2.91 ppn 8204.33 ppn 7523.97 ppn 2,369

Reference source

Reference source by time and probability bucket for MLB
Time0-.2.2-.4.4-.6.6-.8.8-1
early(inn1-3)-3.00 ppn 531-7.08 ppn 1,210-6.43 ppn 4,207-0.97 ppn 2,8477.80 ppn 694
mid(inn4-6)-2.01 ppn 2,173-10.52 ppn 1,472-1.45 ppn 1,5401.78 ppn 1,9590.41 ppn 2,608
late(inn7+)-3.16 ppn 2,851-11.66 ppn 640-14.51 ppn 6385.22 ppn 6471.75 ppn 2,666

Inspect published rows

Filter or sort the detail table. The grids above retain all published cells. Support counts forecast observations, not independent games; matching labels do not pair model and reference populations.

Showing 28 of 28 published rows for MLB.

CSV keeps this row order, raw probabilities, percentage-point gaps, and source context. The artifact date is included; observation dates are not published. Missing cells are omitted, not exported as zero.

All published state-conditioned rows for MLB
SourceTime bucketProbability bucketnMean forecastObserved frequencySigned gap (pp)Absolute gap (pp)
Modelearly(inn1-3).4-.65,75450.97%48.82%-2.15 pp2.15 pp
Referenceearly(inn1-3).4-.64,20751.78%45.35%-6.43 pp6.43 pp
Modelearly(inn1-3).6-.82,70265.46%65.17%-0.29 pp0.29 pp
Referenceearly(inn1-3).6-.82,84768.27%67.30%-0.97 pp0.97 pp
Referenceearly(inn1-3).8-169487.44%95.24%7.80 pp7.81 pp
Modelmid(inn4-6).6-.82,44969.31%76.64%7.33 pp7.33 pp
Referencemid(inn4-6).6-.81,95970.55%72.33%1.78 pp1.79 pp
Modelmid(inn4-6).8-11,46285.06%96.51%11.45 pp11.45 pp
Referencemid(inn4-6).8-12,60890.89%91.30%0.41 pp0.41 pp
Modellate(inn7+).6-.875273.46%77.79%4.33 pp4.33 pp
Referencelate(inn7+).6-.864770.98%76.20%5.22 pp5.22 pp
Modellate(inn7+).8-12,36991.01%94.98%3.97 pp3.97 pp
Referencelate(inn7+).8-12,66692.06%93.81%1.75 pp1.75 pp
Modelmid(inn4-6).4-.62,46850.62%50.89%0.27 pp0.27 pp
Referencemid(inn4-6).4-.61,54049.57%48.12%-1.45 pp1.45 pp
Modelmid(inn4-6).2-.42,28529.72%17.94%-11.78 pp11.78 pp
Referencemid(inn4-6).2-.41,47229.27%18.75%-10.52 pp10.52 pp
Modellate(inn7+).2-.461726.20%8.43%-17.77 pp17.78 pp
Referencelate(inn7+).2-.464027.60%15.94%-11.66 pp11.66 pp
Modellate(inn7+)0-.22,8847.84%6.66%-1.18 pp1.18 pp
Referencelate(inn7+)0-.22,8517.93%4.77%-3.16 pp3.16 pp
Modelearly(inn1-3).2-.41,03335.62%24.01%-11.61 pp11.62 pp
Referenceearly(inn1-3).2-.41,21030.30%23.22%-7.08 pp7.08 pp
Modelmid(inn4-6)0-.21,08816.18%2.39%-13.79 pp13.79 pp
Referencemid(inn4-6)0-.22,1739.60%7.59%-2.01 pp2.01 pp
Modellate(inn7+).4-.682050.84%47.93%-2.91 pp2.91 pp
Referencelate(inn7+).4-.663852.13%37.62%-14.51 pp14.52 pp
Referenceearly(inn1-3)0-.253112.79%9.79%-3.00 pp3.00 pp

Source fields: state_conditioned_calibration.json -> sports[sport].n_records, n_skipped_no_state_field, and buckets[] -> time_bucket, prob_bucket, source, n, mean_p, mean_y, calibration_error.