offdiagonal

Zeno teams

Evidence for the next read

What Offdiagonal can supply now, and what still needs a clock, a right to use, or a matched target. Zeno owns its training and score claims.

The rule

Decision time is a wall

Every episode is split at its decision time. Left of the line is a training input; right of the line is outcome or provenance only. Nothing crosses it.

decision time

Known at the decision — inputs

  • Every forecast posted before the decision, with its real post date
  • Each forecaster's own ranges and track record up to that moment
  • Pair miss-together counts from outcomes already known then
  • Jev judgments, only when judged before the decision (Jev began 25 Sep 2026)

Known only later — outcomes and provenance

  • The outcome and its first dated release
  • Revisions posted after the decision
  • Replay judgments about past decisions — evaluation only, never inputs
  • ForecastBench resolution files (first public commit — excluded from strict no-lookahead claims)
Fields gated to each row's own availability clock; archival-only fields are never model inputs. Unknowns stay unknown — nothing is zero-filled or backdated.

Market diagnostic

Crypto questions are not crypto forecasts

A small, live sample of title-tagged market questions. The market’s yes/no probability stays separate from a price-range forecast.

Reading question sample…

Open the paged diagnostic

Existing evidence

Four different readiness states

Use the recorded row-level admission reasons; counts from one run’s split cannot be inferred from total collected rows.

Weather

Compare at fixed lead

Use the revision clock and observation together. Count independent model families on the same site, target and lead; do not equate all captured rows with Zeno’s held-out cases.

Revision cases

Health

Admit per revision

FluSight, COVID and RSV are separate sources. Keep submission deadlines, forecast revisions and first outcome vintages; a captured row alone is not a training case.

Health cases

Macro

Philadelphia: raw only

The Philadelphia survey is downloadable. Survey deadlines and first-release outcome vintages still need evidence before it becomes a training example.

Survey records

Crypto

Baseline, not new voices

Hourly price observations and retrospective forecasts exist. This is not yet a set of independently issued price-range forecasts.

Crypto integrity

System One

Judgments, not labels of truth

Jev can direct review of shared methods, ambiguous settlements and stale descriptions. It cannot supply missing release clocks or rewrite outcomes.

Open Jev workspace·Machine handoff