Evidence

PING prefers a correct, well-supported answer over an impressive one. When the system makes an important claim, it shows where the claim came from. When it modifies architecture, it shows what changed. When it declares success, it shows what was verified. The more consequential the claim, the stronger the evidence required.

This is a product decision, not just an engineering virtue. Trust in an autonomous system is not a personality trait; it is a property of behavior. Systems that show their work get trusted with more. Systems that assert get audited.

Every observation PING records carries:

• The statement itself, in plain language.
• A source span, pointing at the exact material the statement was drawn from.
• A confidence grade, from confirmed to unknown. Never assumed, never hallucinated.
• Verification status, recording whether the claim was independently checked and what the check found.
• Contradictions preserved, because two sources disagreeing is information, not an error to be averaged away.

When PING researches something (a business's technology stack, a market signal, a technical claim), every finding is graded: confirmed, strong indication, weak indication, or unknown. The grade is part of the record. Downgrading a weak signal to unknown is a success of the system, not a failure: it means the verification gate did its job. Fail-closed beats confident-and-wrong every time.

Data crosses boundaries constantly: raw artifact to observation, observation to claim, claim to projection. At each crossing, PING asks what identity, ordering, provenance, and semantics must survive. Silent normalization that destroys meaningful distinctions is treated as a bug, not an optimization. The lineage from presented answer back to source event must remain walkable.