Submission schemas
Version 1.0 public submission schemas — illustrative and compatible with the v1.0 protocol; no participant data. The paper is authoritative if there is a conflict; examples are in examples/.
These six schemas describe the shape of a Version 1.0 submission, from a single target instance through to a leaderboard reporting row. They are provided so a team can prepare a well-formed run; they do not by themselves grant a score, and no participant data ships with them. Schemas are illustrative — where wording differs from the paper, the paper governs.
The six submission schemas
Each file is a standalone JSON Schema. Follow the Schema link to read the raw file served from this site, or the GitHub link to view it in the repository.
participant.schema.json
A target instance: the record that identifies one target and the instance being predicted — the who and the what, with no outcome attached.
evidence_manifest.schema.json
The evidence tiers, observation window, and hashes: which evidence streams were available up to the sealed time t, at what tier, with content hashes for auditability. Evidence, never a scored state.
forecast.schema.json
The sealed probability vector over the discrete answer space A, emitted and hashed at time t before the outcome exists.
outcome.schema.json
The resolved label: the observable answer recorded at resolution time r under the pre-registered, deterministic resolution rule.
submission.schema.json
The run manifest: ties a set of sealed predictions to the system, track, split, and version that produced them, with the attestations a run requires.
Platform-layer schemas (release V1.0)
Release V1.0 also ships six public platform-layer schemas — the objects the end-to-end prototype (registry, sealing, scoring, evidence efficiency) reads and writes. Each link serves the raw JSON Schema from this site.
- benchmark-run.schema.json — a registered benchmark run: target type, evidence streams, horizon, baselines, cost axes, and lifecycle status.
- prediction-instance.schema.json — one prediction instance: the query, evidence cutoff, seal deadline, resolution time, answer space, and pre-registered resolution rule.
- answer-space.schema.json — the discrete operational possibility space: the mutually exclusive future states a prediction is scored over.
- sealed-prediction.schema.json — a sealed probability distribution over the answer space, with payload and answer-space hashes and the sealing timestamp.
- resolved-outcome.schema.json — the resolved outcome under the pre-registered rule, with resolution time and supporting evidence references.
- scoring-report.schema.json — the per-run scoring report: skill vs R1 and R2 in bits, Brier, calibration summary, and gate results.
How the abstract prediction unit maps to fields
The paper defines a prediction unit as (i, E≤t, q, A, r) — an instance i, evidence up to sealed time t, a query q about a future state, a discrete answer space A, and a resolution time r > t. In the v1.0 personal-track schemas that abstract unit maps onto concrete fields as follows.
| Abstract unit | Schema field | Lives in |
|---|---|---|
| target | participant_id | participant.schema.json |
| instance i | task_id | participant.schema.json / forecast.schema.json |
| question q | task_type | forecast.schema.json |
| answer space A | answer_space | forecast.schema.json |
| evidence E≤t | evidence_manifest | evidence_manifest.schema.json |
| label | observed_answer | outcome.schema.json |
Inclusion rule. A target state is scored only if it has a pre-registered observable resolution rule; otherwise it is evidence, not a scored state. Attention, affect, and inferred goals are evidence and are never scored states. Self-report may be evidence but is never the outcome label when it is also a model input.
Next steps
Run the harness
See how the schemas feed the reference checks — sealing, deterministic resolution, R1/R2 baselines, calibration, and the permutation gate — on synthetic data only.
Product quickstart
For teams mapping a product's longitudinal traces onto a well-formed v1.0 submission, from target instance to sealed prediction.
A high score under these schemas certifies one thing: validated target-specific predictive lift — calibrated, prospective skill distinguished from generic prediction and routine replay. Scope and governance limits are defined in the paper.