Demonstration data

Every figure on this site is deterministic demonstration data. It is not a live source, not a forecast, and not a recommendation. No wagering market is connected, and nothing here should be acted on.

Methodology

How outputs are produced, the point-in-time rules that constrain them, and the states every displayed value can be in.

The point-in-time rule

The rule everything else rests on. Break it and every measurement downstream becomes meaningless in a way no test result will reveal.

An output may only use information that existed before the race it is about. This sounds obvious and is violated constantly, because the violation is invisible: a model trained on data that includes the result of the race it is predicting looks spectacular and is worthless.

A forecast’s timestamp must be strictly before the scheduled start. Equality is a leak, not a boundary case — an output produced at the instant the race starts had access to everything up to the start.
A runner’s record is reconstructed as atthe date it is being read for, not as it stands today. A horse’s record in a race last March did not include what it did in April.
Post-race information may never appear inside a pre-race display. The contract rejects an envelope that violates this, and it is a schema-level refusal rather than a convention.
Settlement is kept structurally separate from feature construction, so a result cannot reach a feature by accident.

How an output would be produced

The intended pipeline. None of it is built — the stages are listed so the sequence is reviewable before it exists.

Retrieval. Raw payloads captured immutably with a retrieval receipt and a content hash. Nothing derived is stored at this stage.
Validation and quarantine. Records failing a validation rule are held back from the canonical archive entirely. Nothing derived from a quarantined record reaches an output.
Canonicalisation. Source-shaped records become canonical global records, retaining every original value alongside its converted form.
Snapshot. An immutable, content-addressed archive with training, validation, evaluation and untouched-holdout partitions fixed at capture time.
Formula measurement. Each formula measured for conditional contribution after controlling for the market, for collinearity, and for stability across segments. A formula is a hypothesis until it survives all three.
Council. Independent models produce independent opinions. Their disagreement is reported rather than averaged away.
Publication gate. An output is published, withheld or quarantined. Withheld outputs are recorded with the reason.
Settlement and scoring. Results collected separately, scored with proper scoring rules, and appended to the accuracy ledger — hits and misses alike.

None of this pipeline exists

Task 000 builds the interface and the contracts. Stages one to four arrive with Task 001, formula measurement with Task 002, the council with Task 005, the publication gate with Task 006, and scoring with Task 009.

Data states

Every state a displayed record can be in, generated from the contract that enforces them.

Every data state with what it means for the value shown beside it
State	What it means
Complete. Every expected evidence family was available for this record.	Every expected evidence family was available for this record.
Partial data. Some evidence families were unavailable. The missing families are listed by name; none has been imputed.	Some evidence families were unavailable. The missing families are listed by name; none has been imputed.
Reduced confidence. Published with a wider uncertainty interval because the supporting evidence was thin.	Published with a wider uncertainty interval because the supporting evidence was thin.
Insufficient evidence. Below the evidence threshold required to publish a research output. No estimate is produced.	Below the evidence threshold required to publish a research output. No estimate is produced.
Unrateable. No output can be produced for this record. It is shown so that its absence is visible.	No output can be produced for this record. It is shown so that its absence is visible.
Stale. The underlying data is older than the freshness threshold and may no longer reflect the source.	The underlying data is older than the freshness threshold and may no longer reflect the source.
Source delayed. The source acknowledged the request but has not yet returned current data.	The source acknowledged the request but has not yet returned current data.
Source unavailable. The source could not be reached. The failure is recorded with its reason.	The source could not be reached. The failure is recorded with its reason.
Region unavailable. No entitled source family could serve this region on this date.	No entitled source family could serve this region on this date.
Quarantined. Held back from the canonical archive by a validation rule. Excluded from every output.	Held back from the canonical archive by a validation rule. Excluded from every output.
Corrected. The source revised this record after first publication. The original is retained alongside the correction.	The source revised this record after first publication. The original is retained alongside the correction.
Abandoned. The race was abandoned. Any pre-race output is retained but no result exists.	The race was abandoned. Any pre-race output is retained but no result exists.
Postponed. The race was postponed. The scheduled time shown may not be final.	The race was postponed. The scheduled time shown may not be final.
Cancelled. The race was cancelled and will not be run.	The race was cancelled and will not be run.

Publication states

What a publication decision does and does not assert.

Every publication state with what it asserts
State	What it asserts
Not published. No research output has been published for this race.	No research output has been published for this race.
Monitoring. Under observation. Evidence is being gathered and is not yet sufficient to publish research.	Under observation. Evidence is being gathered and is not yet sufficient to publish research.
Standard research. Published as standard research. A single evidence family supports the output.	Published as standard research. A single evidence family supports the output.
High evidence. Multiple independent evidence families agree. This indicates agreement between methods — it does not indicate certainty, it is not a guaranteed outcome, and it is not a prediction of the result.	Multiple independent evidence families agree. This indicates agreement between methods — it does not indicate certainty, it is not a guaranteed outcome, and it is not a prediction of the result.
Withheld. An output exists but was deliberately not published, because it did not meet the publication threshold. The reason is recorded.	An output exists but was deliberately not published, because it did not meet the publication threshold. The reason is recorded.
Quarantined. The underlying records failed validation. Nothing derived from them is published.	The underlying records failed validation. Nothing derived from them is published.

Rules about displaying a value

Enforced by the display contract and its test suite, not by convention.

A missing value is never converted to zero.An absent measurement renders its reason — “not collected”, “below sample threshold” — so a reader can always tell “we did not measure this” from “we measured this as zero”.
Absent evidence is never shown as a low score.A Career DNA dimension with too few qualifying runs is hatched across its full width, not drawn short — a short bar reads as “bad at this”.
A rate below its sample threshold is withheld.The counts are still shown. Two wins from three runs is a fact; “67%” beside a figure drawn from four hundred runs is an invitation to a comparison that misleads.
An estimate never appears without its interval. A point estimate alone implies a precision no model here has.
A rollup never reports better than its worst constituent. A meeting containing one quarantined race is not a complete meeting.
Stale data is shown and marked, not hidden. Data older than 15 minutes carries its age wherever it appears.
The original source value always remains available. A converted distance, weight or currency is shown beside what the source actually published, and a currency conversion always carries its rate source and timestamp.

Where the limits and the record are stated.

Model limitations — what has not been demonstrated.
Accuracy ledger — every published forecast against its outcome.
Data sources — provenance, entitlement and refresh behaviour.

State

What it means

Complete

Every expected evidence family was available for this record.

Partial data

Some evidence families were unavailable. The missing families are listed by name; none has been imputed.

Reduced confidence

Published with a wider uncertainty interval because the supporting evidence was thin.

Insufficient evidence

Below the evidence threshold required to publish a research output. No estimate is produced.

Unrateable

No output can be produced for this record. It is shown so that its absence is visible.

Stale

The underlying data is older than the freshness threshold and may no longer reflect the source.

Source delayed

The source acknowledged the request but has not yet returned current data.

Source unavailable

The source could not be reached. The failure is recorded with its reason.

Region unavailable

No entitled source family could serve this region on this date.

Quarantined

Held back from the canonical archive by a validation rule. Excluded from every output.

Corrected

The source revised this record after first publication. The original is retained alongside the correction.

Abandoned

The race was abandoned. Any pre-race output is retained but no result exists.

Postponed

The race was postponed. The scheduled time shown may not be final.

Cancelled

The race was cancelled and will not be run.

State

What it asserts

Not published

No research output has been published for this race.

Monitoring

Under observation. Evidence is being gathered and is not yet sufficient to publish research.

Standard research

Published as standard research. A single evidence family supports the output.

High evidence

Multiple independent evidence families agree. This indicates agreement between methods — it does not indicate certainty, it is not a guaranteed outcome, and it is not a prediction of the result.

Withheld

An output exists but was deliberately not published, because it did not meet the publication threshold. The reason is recorded.

Quarantined

The underlying records failed validation. Nothing derived from them is published.

Racing

Evidence

About

Methodology

The point-in-time rule

How an output would be produced

None of this pipeline exists

Data states

Publication states

Rules about displaying a value

Related

Racing

Evidence

About

Methodology

The point-in-time rule

How an output would be produced

None of this pipeline exists

Data states

Publication states

Rules about displaying a value

Related