Public Registries

Claims should point to evidence, limits, and attack routes.

These registries are intentionally small. They make the public layer machine-readable without exposing private product code, private finance execution, customer data, or unpublished commercial orchestration.

Claim Registry JSON Evidence JSON Claim-to-Evidence Table Schema Notes Repro Scope Read Boundaries

Evidence map connecting public claims, artifacts, and attack routes.

Claims Bounded statements

Evidence Artifacts and limits

Counterexamples Repair routes

Actions Warranted packets

Downloads

Four registries, one public contract.

claim_registry_v0.json Download claims

evidence_registry_v0.json Download evidence

counterexample_registry_v0.json Download counterexamples

action_registry_v0.json Download actions

claim_to_evidence_table_v0.html Read claim-to-evidence table

claim_to_evidence_table_v0.json Download claim-to-evidence table

schema_notes_v0.html Read schema notes

artifact_repro_scope_v0.html Read artifact and reproduction scope

P02-C1

Learning from failure

Language agents may succeed on first exposure while failing to improve under repeated failure and feedback.

Evidence: E-P02-3600. Attacks: leakage, stronger baseline, scoring bug, reproduction failure.

PCA-C1

Proof-carrying action

High-risk AI action should require a warrant and receipt closure before it earns action credit.

Evidence: E-PCA-NOGO-001, E-DEMO-RECEIPT-001. Attacks: false no-go, missing receipt, credit leak.

P24-C1

Relational observability

Adaptive systems need observable relations, constraints, control debt, and evidence half-life.

Evidence: E-RO-PUBLIC-001. Attacks: missing relation, unmeasured debt, transfer failure.

P20-C1

Evidence integrity

Safe physical AI should not convert detector confidence directly into action under degraded evidence.

Evidence: E-P20-SYNTH-001, E-P20-PUBLIC-LOG-001. Attacks: independent dataset failure, baseline underfit.

Reading Rule

A registry entry is not a slogan. It is a repairable contract.

Each entry is expected to survive three questions: what is the exact claim, what evidence supports it, and what would make it narrower, false, or incomplete? When an attack succeeds, the correct response is not defensiveness; it is a boundary update, an artifact fix, or a new negative result.

Submit Public Attack Run Public Demo Evidence Map