Category guide

Evidence-backed AI.

A useful AI answer should not become trustworthy merely because it sounds coherent. Evidence-backed AI keeps the path from observation to conclusion visible enough to inspect, challenge, and revise.

Plausibility is not evidence.

Given a metric change, a general-purpose model can generate several reasonable explanations. That ability is valuable for hypothesis generation, but it does not establish which hypothesis is supported. WITNESS treats explanation as one stage inside an investigation rather than as the finish line.

Minimum standard

Source visibility

The evidence used by the reasoning process remains identifiable.

Observation before inference

What was measured or observed is kept distinct from why the system thinks it happened.

Competing explanations

Alternative causes are tested rather than silently discarded.

Uncertainty

Unknown, missing, weak, contradictory, or disputed evidence is allowed to constrain the answer.

Change conditions

A strong finding states what new evidence could materially change the conclusion.

Traceability

The reasoning path remains attached so the conclusion can be audited or revisited.

Signal quality

Better reasoning cannot repair bad evidence.

WITNESS Signal Integrity evaluates whether an input deserves trust before downstream reasoning relies on it. A more capable model is not a substitute for provenance, structural consistency, contradiction detection, and evidence-quality controls.

The WITNESS rule

MODEL_OUTPUT_CANNOT_DIRECTLY_ESTABLISH_TRUTH.
A model can propose, compare, summarize, or infer. Its output does not become established truth simply because it was generated.