Technical buyer guide · Two independent controls
One asks whether information was available when a decision was made. The other asks whether outcome knowledge influenced which candidate features were allowed to enter. Treating them as one vague “leakage check” makes both harder to inspect.
01 / Availability-time boundary
A record can be dated before a decision and still have become available afterward. Event time alone does not establish usable knowledge.
02 / Pre-target selection boundary
A feature set can absorb label knowledge through manual curation, permissive naming, or an undocumented schema pass before model fitting begins.
Control comparison
| Control dimension | Temporal Boundary Auditor | Target-Blind Feature Freezer |
|---|---|---|
| Decision | Was context available, complete, ordered, and fresh enough at each decision time? | Which declared fields satisfied one policy before target values were read? |
| Reads | Local CSV or JSON Lines decision and context records under a strict temporal contract. | A strict values-free JSON column schema and freeze contract; no dataset rows or target values. |
| Retains | Aggregate JSON, streaming findings JSON Lines, Markdown, and exact input byte hashes. | Ordered manifest, selected-feature CSV, exclusion ledger, Markdown, and checksums. |
| Primary failure | Unavailable, incomplete, off-grid, ambiguous, duplicate, or stale temporal context. | Protected roles, forbidden names, disallowed types, pattern exclusions, or count-policy failure. |
| Does not prove | Source truth, causal validity, complete lineage, absence of every leakage path, or model fitness. | Correct roles, target-blind upstream computation, harmless proxies, feature usefulness, or model fitness. |
| Price and delivery | $499 USD once · one legal entity · private local delivery. | $249 USD once · one legal entity · private local delivery. |
Synthetic public evidence
The published adversarial report contains one EVENT_ONLY_JOIN_WOULD_LEAK error, one PERIOD_NOT_COMPLETE error, and one SAFE_CONTEXT_TOO_OLD warning.
The published exclusion-focused manifest records a six-column declaration with one selected feature and five exclusions. It explicitly records dataset_values_read: false and target_values_read: false.
These are wholly synthetic evaluation records. They demonstrate public contracts and expected outputs; neither archive contains executable audit or freeze logic, a commercial license, customer material, private schemas, or private implementation.
Optional sequence
A team may freeze a candidate-feature declaration before target values are read, then separately audit the availability boundary of temporal context used to compute or join those features.
The records stay distinct. A frozen manifest does not establish that upstream values were computed without future knowledge. A temporal PASS does not establish that the candidate list was selected without label knowledge. Each contract must be reviewed on its own terms.
Combined boundary
Together or separately, these tools do not prove correct source data, honest role declarations, target-blind upstream computation, complete lineage, absence of hidden proxies, absence of every future-information path, causal validity, statistical validity, compliance, safety, or fitness. They make two narrower decisions explicit and leave the remaining assurance work visible.
Choose the route
If timing is the risk
Use the guide, synthetic reports, offline pack, and closed-choice fit check before sending a purchase request.
$499 USD · one timeIf candidate selection is the risk
Use the guide, synthetic manifests, offline pack, and closed-choice fit check before sending a purchase request.
$249 USD · one timeBuyer questions
No. Temporal leakage asks whether context was actually available at a decision time. Target leakage asks whether outcome knowledge influenced which candidate features were allowed to enter. Each requires a separate declared boundary.
No. Each tool evaluates a narrower declared contract. Neither proves source truth, correct operator declarations, complete lineage, absence of every leakage channel, model validity, causality, safety, or fitness.
No. They are separately licensed local tools with independent contracts and outputs. A team may sequence them deliberately, but neither automatically consumes nor validates the other product's records.
No. The public evaluation materials are synthetic and downloadable. The licensed tools run locally without runtime network requests; the Feature Freezer accepts values-free schema metadata, while the Temporal Auditor evaluates local CSV or JSON Lines records.
Still deciding?