4c64328884
Two-pass hardening of the #690 demo so no epistemic state is proposer-mintable: Pass 1 (sealed corpus, from the reconciliation pass): - evidence items are references only (evidence_id + content_sha256) resolved against a committed content-addressed corpus (evidence_corpus.json) - support derived from committed subject/predicate match; independence from distinct provenance_root values; proposer support/independence labels are rejected by the closed schema - evidence-strength taxonomy (substrate-capability vs interface-contract) in PUBLIC-DEMO-ROADMAP / DEMO-PACKAGING-CHECKLIST; position-paper refinements Pass 2 (entailment-decided inference, closes a live false-status hazard): - the merged inferred leg required only that premise IDs resolve — a claim with no support citing one unrelated record as a premise was assigned 'inferred' (empirically confirmed). inferred is now decided by generate.proof_chain.entail (sound+complete ROBDD) over resolved corpus premises, cross-checked against the independent truth-table oracle (evals.deductive_logic.oracle, no shared code); proof keys + oracle verdict in the trace; disagreement => defensive refusal; refutation => contradicted - committed adversarial scenario unrelated-premise-still-undetermined proves the attack now lands on undetermined - corpus seal (corpus_sha256) pinned into every evaluated trace; scope refusals no longer consult the corpus; corpus cache hands out copies; claim subject/predicate constrained to atom-compatible snake_case - .gitignore: private packet dir, agent worktrees, tooling artifacts Validation: 30 demo tests; 96 across all three demo suites; 7/7 scenarios double-run byte-identical; 61 architectural invariants green.
2.5 KiB
2.5 KiB
Public Demo Packaging Checklist
This checklist applies to public CORE demos.
Scope
- Demo purpose is stated narrowly.
- Demo status is clear: merged, draft PR, proposed, or not yet implemented.
- Demo evidence class is stated: substrate-capability, interface-contract, simulation-only, or proposed.
- README includes "what this proves."
- README includes "what this does not prove."
- README avoids named-company outreach strategy.
- README avoids funding, sponsorship, runway, or executive-packet language.
Safety boundary
- No real external side effects.
- No network dependency.
- No model API dependency.
- No shell execution unless the demo explicitly exists to test shell-denial behavior.
- No production MCP claim.
- MCP-shaped interfaces are described as MCP-shaped, not production MCP.
- Simulation-only demos are labeled simulation-only.
- Proposed demos are not described as implemented.
Determinism
- Runner output is deterministic.
- Expected artifacts are pinned.
- Double-run behavior is byte-identical where applicable.
- Trace hashes are deterministic.
- No timestamp, random, host-path, or environment-dependent output appears in expected artifacts.
Authority boundary
- Proposer input cannot set final authority status.
- Proposer input cannot smuggle final action artifacts.
- Proposer input cannot indirectly decide the result through trusted support, independence, verdict, or clearance labels.
- Invalid payloads fail closed.
- Refusal and ask outcomes are first-class successes.
- Authorized outputs, if present, are inert artifacts unless a later production system explicitly implements execution.
Evidence strength
- Capability claims route through a real CORE operator, runtime path, sealed eval lane, or independently checked proof surface.
- Interface-contract demos are not described as substrate-capability demos.
- Fixture evidence is labeled as fixture evidence.
- Sealed-corpus or independent-oracle claims name the corpus/oracle and the no-shared-code boundary.
- Boolean/string policy demos are framed as boundary checks unless another substrate decision is actually present.
Public hygiene
- No named-person outreach planning.
- No named-company approach strategy.
- No private red-team personas.
- No speculative current-facts dossiers.
- No claims of robotics, vehicle, aerospace, defense, or safety-certified deployment unless directly implemented and independently verified.