FedCourtsAI

Data

Alongside the predictions, the public repository commits a set of roll-ups under metrics/: counts and rates about the Court's dockets, and the scores of the models. A scheduled job regenerates and commits each one, so any figure can be traced to the commit that produced it. The links below point to the commit this site was built from. Before quoting a number, read the metrics guide: it explains what each figure measures and what can be claimed from it.

Updated Nov 9, 2026 · ledger commit fe890c9

The roll-ups

  • Docket pack

    Facts about the Supreme Court's dockets: what it is asked to take, from which court below, on which fee stream, after how many relists, and how it disposes of petitions, by October Term. It makes no claim about the predictions.

  • Stat pack

    The base rates the predictions are anchored on and scored against: cert petitions by Term and by salience band, the interim (applications) docket, and merits outcomes for granted cases.

  • Big-case board

    Each model's view of how consequential each pending case is: the data behind the front-page board.

  • Leaderboard

    The models ranked on their graded forecasts, in the pre-registered (frozen) scope only.

  • Claim scores

    Each model's mechanically scored claims, pooled per model and per stratum, published beside the leaderboard.

  • Backtest

    Simple baseline predictors replayed over cases decided long ago, each shown beside the always-deny floor. It measures recall of known history, not foresight.

  • Cert backtest

    The models replayed on decided cert petitions with the outcome hidden, scored against a base rate that could have been known at the time.

  • Weekly updates

    One short account per week of what the Court did and what the models predicted, generated from the ledger.

By October Term

A few counts for each Term, copied from the stat pack. Each column heading links to the stat pack section it comes from, which has the rest of that Term's figures and the caveats for reading them.

TermCert petitions filed (paid / IFP)Cert grantsApplicationsMerits cases (granted / decided)
OT2026600 / 1,0949649—
OT20251,432 / 2,7021061,46753 / 26
OT20241,329 / 2,5291201,29579 / 77
OT20231,375 / 2,8481391,17756 / 56
OT20221,252 / 2,9071091,13173 / 73
OT20211,611 / 2,66815187662 / 62
OT20201,829 / —16217880 / 73
OT20191,478 / —1581,07160 / 58
OT20181,593 / —1511,36981 / 78
OT20171,718 / 4,5951701,42558 / 57

Filings count docketed numbers, a slight upper bound on real petitions. Grants are those observed in the corpus. Applications include requests for more time. A merits case counts as decided once its judgment has been parsed. A Term still under way shows only what has happened so far.

By topic

The share of each stage's labeled cases under each topic, copied from the docket pack. Each column is a share of its own stage, never of the whole docket, and the columns are not steps of one funnel: they cover different cases, labeled to very different depths and from different inputs. Read each column's scope note below the table before quoting a share. Labeled rows on hand are in parentheses.

TopicCert petitions (est. n=9,635; 9,347 rows)Merits cases (n=341)Emergency applications (n=2,029)
criminal-law25.7% (2,292)12.9% (44)8.4% (171)
civil-procedure18.8% (1,787)13.5% (46)6.2% (125)
constitutional-rights12.6% (1,217)8.5% (29)0.6% (12)
habeas-and-postconviction7.5% (679)4.7% (16)23.8% (482)
employment-and-antidiscrimination5.7% (545)6.2% (21)1.5% (30)
business-and-financial-regulation5.6% (541)10.0% (34)1.0% (21)
first-amendment4.8% (452)7.3% (25)1.7% (34)
firearms4.3% (403)2.9% (10)0.3% (6)
administrative-law-and-benefit-programs3.8% (367)13.5% (46)2.0% (40)
unclassifiable3.4% (316)0.3% (1)44.1% (894)
sovereignty-and-foreign-relations2.4% (230)7.6% (26)0.5% (10)
intellectual-property2.2% (209)2.9% (10)0.1% (3)
tax1.0% (96)1.5% (5)0.1% (2)
election-law0.9% (86)1.8% (6)4.2% (85)
environment-energy-and-property0.9% (84)2.1% (7)1.4% (29)
immigration0.4% (43)4.4% (15)4.2% (85)

Cert petitions: QP-bearing rows only — 9347 of 35546 in-scope ingested rows labeled; grant-enriched; primaries only; not docket-representative. Those two counts are raw rows; the bucket counts are denial-reweighted, and no reweighting recovers the docket — QP presence is itself outcome- and stream-correlated, so this stays a share of QP-bearing rows. Coverage is uneven across Terms and zero on the earliest of them, so the mix is not the whole slice's; the base-rate column is over a grant-enriched population and is not comparable to the sections above. The 26199 unlabeled row(s) span two gaps at once — rows carrying no stored questions-presented text, which no labeling batch can reach, and QP-bearing rows whose batch has not come up — and batch 11's extract sized the second: it counted 11871 QP-bearing row(s) in the frame it cut from, of which the artifact held 9351 labeled (78.8%). Both counts are batch 11's, taken at its own corpus vintage rather than this pack's: the frame grows with every pull while the labeled count moves only when a batch lands, so read that share as a ceiling on the share of the frame as it now stands, not as a reading of this blob. Labeling accrues in batches, so the labeled rows are two populations on different terms: 353 hand reference-set members, carried in every batch and so included with certainty, and 8994 drawn over this blob by a Term x fee-class-stratified, seeded-hash order from the QP-bearing frame batch 11's extract counted. Both count once here, so the reference block — grant-enriched by design and carrying no sampling weights — is over-represented by at least about 1.3x — row for row, how much likelier a reference row was to be included than a drawn one when batch 11 landed: 353 carried in with certainty against 8998 drawn from the 11518 QP-bearing row(s) outside the block that batch's extract counted. Measured against that frame rather than bounded over rows no batch could reach — and a floor rather than a ceiling at this pack's vintage, since the frame grows with every pull while the drawn count waits for the next batch. It closes to 1.0x as the frame is labeled, which a bound over unreachable rows never does, and this mix is not the frame's until every QP-bearing row is labeled. A naive share partly counts coordinated filing campaigns rather than subjects; no de-duplicated companion is published.

Merits cases: Granted QP-bearing petitions only — 341 labeled grant(s) of the 638 plenary grant(s) in the cert scope; primaries only; not docket-representative. Plenary review here is the `granted` disposition, which takes in probable jurisdiction noted on a direct appeal and leaves out GVRs recorded as such (a few older merits GVRs still carry `granted`). The petition frame is still being labeled, so the missing grants are both those with no stored questions-presented text and those whose batch has not come up; reference-set grants are in every batch, so they are over-represented until it clears. Questions-presented coverage is zero before Term 2023 and uneven after, so this mix is weighted toward the Terms with stored texts and is not the merits docket's.

Emergency applications: Labeled applications only — 2029 of 2234 in-frame live-slice applications labeled, 1993 of the labeled read from the docket alone; primaries only; not docket-representative. Those counts are raw rows (the second includes the few duplicate rows the labeling frame folds into one per docket number); the bucket counts are reweighted like every section here. The frame is the live slice's non-extension application dockets, substantive asks and asks whose kind was never read alike, read from different inputs than the petition cut, so the two are never pooled. A docket-only row whose caption and entries name no subject is `unclassifiable` by rule, so that bucket tracks how thin the dockets are as much as what the applications are about. Batch 4's extract counted 2029 application(s) in the frame it cut from, of which the artifact held 2029 labeled (100.0%); read that share as a ceiling on the share of the frame as it now stands, since the frame grows with every pull. The reference block rides in every batch, so its 241 row(s) are over-represented against the 1788 drawn from the 1788 outside it by no factor — the frame was fully labeled when that batch landed, and this mix is not the frame's until every application is labeled. Of the labeled rows here, 241 are reference-block members and 1788 were drawn.