Signals and Provenance
Signals describe the evidence structure produced by a completed validation. Provenance shows how the public result is connected. Neither layer changes the canonical determination.
Signals
Every completed validation response contains exactly four categorical signals:
| Signal | Values | What it describes |
|---|---|---|
grounding | strong, partial, absent, mixed | The support structure reflected by unit determinations and relationships. |
relevance | high, partial, low, mixed | Whether the evaluated relationships address the information. |
conflict | none, present, material, mixed | The conflict pattern in relationships and the determination. |
coverage | complete, partial, insufficient, mixed | How completely the evidence covers the information units. |
Signals are deterministic categorical projections of completed canonical state. They are not probabilities, confidence scores, source-quality ratings, or alternate verdicts. When interpreting the response, the top-level determination remains authoritative.
Source prestige does not substitute for support. Supplied text remains the evidence evaluated; its citation metadata is not independently checked. Explicit acquire yields inspected material with acquired provenance. Historical stored origin labels remain unchanged. Signals summarize the resulting canonical graph; see Source Inspection.
Provenance
The provenance object contains seven connected sections:
| Section | Lineage exposed |
|---|---|
validation | The canonical validation_id. |
information_units | Each unit ID and its supplied_information origin. |
evidence | Each supplied evidence ID, optional source metadata, and optional duplicate reference. |
relationships | Information-unit/evidence references and their canonical relationship. |
unit_determinations | Each unit determination and the relationship references behind it. |
determination | The aggregate value, reason code, and participating information-unit IDs. |
signals | Each signal value and the canonical structural categories from which it was derived. |
All references resolve to IDs already present in the public response. Duplicate evidence remains traceable, but duplicates are absent from independent relationship evaluations.
What provenance does not expose
Public provenance does not expose internal evidence ledgers, evaluator diagnostics, prompts, chain-of-thought, model or provider identity, retrieval state, confidence, weights, scores, sessions, or revisions.
The endpoint validates structural lineage before returning it. Broken or incomplete internal state produces an execution error rather than a repaired response or a fabricated uncertain determination.