docs/prequential-cutoffs.md

Prequential cutoffs for the live tier

Status

Implemented. TaskCard.cutoff_schedule accepts monotone ISO timestamps, the validator rejects non-monotone schedules, and the scorer evaluates scheduled tasks cumulatively at each cutoff before emitting one backward-compatible aggregate TaskScore.

Why

The v1 scoring collapses every prediction for a live task into one aggregate. That makes a model's position indistinguishable from an average-over-time; readers cannot tell whether the model was strong early and drifted, or strong late and is still climbing. Impermanent's prequential protocol makes rank drift visible by bucketing each evaluation at a rolling cutoff and reporting the rank per cutoff window.

FUTURE-TS already has richer as-of semantics than Impermanent (available_at, first_published_at, last_updated_at, source_release_id). Rolling cutoffs are the evaluation-time surface that operationalises those semantics.

Contract

A task card can optionally declare a cutoff schedule:

"cutoff_schedule": [
  "2026-05-01T00:00:00Z",
  "2026-05-15T00:00:00Z",
  "2026-06-01T00:00:00Z"
]

Each entry is a wall-clock moment at which the task is scored using only actuals whose available_at <= cutoff. Predictions are unchanged; only which actuals count toward scoring at each cutoff changes. The schedule MUST be monotone non-decreasing.

Internally the scorer computes one per-cutoff TaskScore with the same shape as today plus a cutoff field populated. The public score_submission API still returns a single BenchmarkReport: scheduled tasks are aggregated into one task-level score with secondary_metrics.prequential_cutoff_count set. This preserves existing report consumers while making the prequential window count explicit.

Scoring semantics

For a task with cutoffs c_0 < c_1 < ... < c_n:

• At cutoff c_i, a prediction p contributes iff the matching actual's available_at <= c_i. Predictions whose actuals land later are excluded from the c_i window but may enter c_{i+1}.
• Each cutoff produces an internal per-task TaskScore with the existing raw/skill/CI fields and an added cutoff field matching the schedule entry.
• The task's aggregate score across cutoffs is the arithmetic mean of the normalized per-cutoff skills. Cohort-level rank averaging remains a leaderboard concern because ranks require multiple submissions.
• Bootstrap CIs are computed per cutoff; aggregate CI is the envelope of the per-cutoff CIs, not a fresh bootstrap — so rank movement between cutoffs is visible rather than smoothed out.

Backward compatibility

Tasks without a cutoff_schedule (i.e. almost every task today) are scored exactly as before: one evaluation window ending at evaluation_as_of or the wall clock. Adding cutoff_schedule to an existing task is a breaking benchmark change and MUST bump the task's version hash in the benchmark manifest.

Reporting

The report format accepts an optional cutoff field on TaskScore for diagnostic per-slice reports, but the default score_submission output emits the aggregate task score only. Mean aggregates per tier and per track remain the headline numbers; downstream dashboards that need rank drift can compute cohort ranks over the same cutoff schedule from archived reports.

Open questions

• Should cutoff boundaries align with source update cadences (daily / weekly / monthly) automatically, or should task authors declare them explicitly? Current leaning: declared explicitly, so drift windows are stable even if a source's cadence changes.
• What happens when a submission is re-scored at a new cutoff that postdates its frozen_at? Answer: allowed — freezing is per-submission, not per-cutoff. The integrity guarantee is "no future-label contamination at prediction time," not "no re-scoring after submission."