Public Release - 2026-05-27
Evidence is not a paragraph that supports a claim. For high-risk AI, evidence should be an inspectable object: provenance, assumptions, artifact, replay path, boundary, failure case, timestamp, and hash.
Definition
Many AI systems leave explanations. Fewer leave receipts. The difference is not rhetorical. A receipt is machine-readable enough to be indexed, replayed, bounded, expired, challenged, and repaired.
This is the narrow point of today's release: evidence is not attitude, confidence, or a polished narrative. Evidence has a structure that can fail.
Receipt Schema
claim_idWhich claim is supported?Evidence floats without target.Maps artifacts to statements.
sourceWhere did the evidence originate?Post-hoc contamination hides.Preserves provenance.
assumptionWhat must be true?The claim silently overreaches.Exposes dependency.
artifactWhich file, log, trace, or table supports it?No inspectable object exists.Connects claim to material.
replay_pathHow can another reader inspect it?The evidence cannot be checked.Turns trust into procedure.
boundaryWhat does this not prove?A local result becomes a universal claim.Prevents overclaiming.
failure_caseWhat would break it?The claim becomes unfalsifiable.Invites useful attacks.
timestamp/hashWhen and which exact object?Versions blur together.Fixes identity.
Supports a bounded claim with provenance, artifact identity, replay path, and an explicit failure condition.
Records why the current evidence is not sufficient: missing nulls, stale provenance, incomplete execution certificate, or no clean denominator.
Shows where the evidence chain breaks and turns that break into a repair target rather than a hidden weakness.
Lets an external reader submit the claim, required receipt, missing field, and wrong-action risk in a compact format.
Bad Evidence Shapes
A screenshot without provenance is weak. A table without denominator is incomplete. A log without replay path is opaque. A claim without boundary is too large. A post-hoc explanation without timestamp can reward the system for inventing a story after the fact.
The purpose of a receipt schema is not to make the system sound more rigorous. It is to make weak evidence fail in public before it becomes action authority.
Minimal Interface
{
"claim_id": "agent_improves_after_failure_rounds",
"source": "public benchmark artifact",
"assumption": ["same task family", "same scoring rule", "no post-hoc relabeling"],
"artifact": ["task table", "raw score file", "scoring script"],
"replay_path": "run scorer on released artifact",
"boundary": ["does not prove human-like wisdom", "does not prove deployment safety"],
"failure_case": ["stronger baseline wins", "score leak found", "task labels inconsistent"],
"timestamp": "2026-05-27",
"hash": "artifact identity"
}
Counterexample Route
A useful counterexample is not a vague objection. It identifies the original claim, the receipt that should support it, the missing or broken field, and why that gap could lead to a wrong action.
That is the public standard we want: do not trust the project first; inspect the shape of its evidence.
Boundary