CertifiedData.io

Learn · Prediction Integrity

What is prediction integrity?

Prediction integrity is the ability to prove when a prediction was created, what it contained, whether it changed, how it was generated, and how its outcome was resolved.

It is a property of the record, not of the prediction itself. A prediction can be wrong and still have integrity — meaning the record of the wrong prediction is preserved tamper-evidently. Likewise, a prediction can turn out to be correct yet lack integrity — meaning no one can prove it was issued in advance rather than reconstructed afterward.

The distinction matters because most platforms publishing AI predictions today have neither: predictions are stored in mutable systems, timestamps are display-only, and the record can be edited at any time. Prediction integrity makes the record itself the artifact — cryptographically bound, machine-verifiable, and independent of the platform that issued it.

The five components

Each component is independently necessary. Together they form a complete prediction integrity record.

01
Creation time

When the prediction was generated, captured with an ISO-8601 timestamp and bound into the signed payload. A prediction without a verifiable creation time cannot be distinguished from a post-hoc reconstruction.

02
Contents

What the prediction claimed — the input context, the predicted outcome, the confidence or probability, and the metadata identifying the platform and model. The contents are captured in a canonicalized payload that produces a deterministic hash.

03
Change history

Whether the prediction has been modified since creation. The Ed25519 signature over the canonical payload makes any change detectable — modifying a single byte invalidates the signature. The chain of modifications, if any are permitted, must be itself tamper-evident.

04
Generation lineage

How the prediction was generated — which model version, which input data, which approval events, which policy was applied. Lineage connects the prediction artifact to the certified training data and certified model artifact upstream.

05
Resolution evidence

How the outcome was determined when the prediction resolved. The resolution payload — outcome, resolver identity, evidence reference, timestamp — is itself certified, creating a verifiable record of the settlement decision.

Prediction integrity vs prediction accuracy

Integrity and accuracy are independent properties. A high-integrity record can be inaccurate; a high-accuracy prediction can lack integrity. They are measured in different ways, on different timescales, by different parties.

AspectPrediction integrity
What it measuresWhether the record can be trusted to reflect what was claimed in advance
When it can be measuredBefore the outcome is observed
How it is establishedCryptographic signing of the prediction at issuance time
Who can verify itAny party with the public signing key
What it depends onThe certificate and the underlying payload

Examples

AI forecaster publishes a daily pick

Integrity: Each pick receives a signed certificate at issuance with a pre-outcome timestamp. A daily manifest is published at midnight. Any reader can verify the picks match the manifest.

Accuracy: Tracked separately: hit rate, calibration, expected vs actual, by category. Accuracy can be measured only after enough outcomes have resolved.

Event market shows live odds

Integrity: Market state snapshots are certified at configurable intervals. The resolution payload at market close is certified with its evidence reference. Snapshots cannot be revised silently.

Accuracy: A separate question of market calibration — whether the displayed odds matched empirical outcome frequencies — measured over many resolved markets.

AI agent makes an autonomous classification

Integrity: The classification record (input context, output label, model version) is certified at decision time. The certificate is referenced in the downstream system that acts on the classification.

Accuracy: Reviewed in periodic audits comparing classifications to known-correct labels. Accuracy informs model retraining; integrity protects the historical record.

Machine-readable summary

{
  "concept": "Prediction integrity",
  "concept_type": "category-definition",
  "schema_type": "DefinedTerm",
  "canonical_url": "https://certifieddata.io/learn/what-is-prediction-integrity",
  "parent_concept": "Prediction Integrity",
  "definition": "The ability to prove when a prediction was created, what it contained, whether it changed, how it was generated, and how its outcome was resolved.",
  "components": [
    "creation_time",
    "contents",
    "change_history",
    "generation_lineage",
    "resolution_evidence"
  ],
  "distinct_from": "prediction accuracy (a separate statistical property)",
  "related_concepts": [
    "Pre-outcome proof",
    "Tamper-evident records",
    "Forecast lineage",
    "Resolution audit trail"
  ],
  "positioning": "CertifiedData proves the prediction record. DecisionLedger proves the prediction process."
}
What Is Prediction Integrity? Definition and Examples | CertifiedData | CertifiedData