Status: draft for the final experimental 0.x public preview.
Not a stable release. No capability is promoted to "stable".
MathEvidence is published as an experimental open computational-evidence
platform for Lean. This preview packages protocol, checkers, adapters, Agent
API v0.1.0, Studio surfaces, registry, Foundry samples, benchmark/conformance
corpora, and replayable evidence under explicit limitation documentation
(KNOWN_TRUST_GAPS.md,
STATUS.md).
The theorem-promotion rule is candidate-bound and fail-closed: a theorem-level Certification Record requires the exact submitted candidate to pass the registered production verification path. Fixtures, nearby theorems, adapter booleans, and benchmark scores cannot grant theorem status.
Five owned capability fragments are theorem-CR eligible under exact candidate binding:
algebra.ideal_membership_witness— witness identity only;algebra.linear_algebra— exact rationalinverse_witness,system_solution,kernel_vector, anddet_identity;logic.finite_counterexample— exact witness establishesrefuted;algebra.formal_rational_calculus— registered formal/algebraic operations;analysis.analytic_calculus— strict theorem-form whitelist with explicit hypotheses.
algebra.rational_equality remains an experimental checker/soundness/bridge
capability, but theorem Certification Record promotion is disabled for the
pinned Lean 4.14 public-preview path. The candidate-specific checker proposition
does not currently elaborate through the production native-reduction path
without an unacceptable sorryAx dependency, so the release fails closed
instead of substituting fixture evidence.
Federated SAT/PB/SMT metadata remains non-CR-eligible in this repository.
- Candidate Bundle: v0.3.
- Certification Record for exact theorem promotion: v0.4.
- Legacy records retain their original semantics and must not be silently upgraded.
- Offline bundle replay and offline kernel theorem replay are separate maturity properties; the stronger release-wide offline-kernel property is not claimed.
- Trust posture explicit: untrusted adapters propose; checker/Lean authority is capability-specific and proposition-scoped.
- Production exact gate: CR-eligible paths are exercised through
scripts/ci/run_cr_exact_lean_e2e_production.pyand declaration identity is read fromLean.Environmentrather than inferred from source presence. - Agent API v0.1.0: public bundle open/inspect/replay accepts opaque
bundleIdonly; raw filesystem paths are rejected. - Capability separation: formal rational calculus is
algebra.formal_rational_calculus; analytic calculus is a separate strict whitelist capability. - Forensic suite:
tests/forensic/guards exact-binding, tamper, policy, adapter/checker, and assurance-boundary regressions. - Benchmark discipline: conformance/regression scores never grant theorem Certification Record eligibility.
- No stable capability promotion.
- No universal solver soundness or broad mathematical completeness claim.
- No claim that the frozen benchmark corpus estimates population false-accept probability or generalization.
- No live external federation agreements.
- No completed external user-confirmation / workflow-win / usability study counts invented for this draft.
- No attested immutable CI green on a release tag claimed in-tree before that tag is actually created and checked.
- Dev receipt HMAC/Ed25519 material is not production PKI; production release signing remains deferred unless separately established by release evidence.
- Repository branch/ruleset configuration is operational governance, not mathematical assurance evidence for this experimental preview.
- Prefer Agent
bundleIdflows; do not pass filesystem paths to public open / inspect / replay endpoints. - Treat Candidate Bundle v0.3 and Certification Record v0.4 as the current exact-promotion protocol surface.
- Use registry ID
algebra.formal_rational_calculus; treat legacysymbolic_calculuspath names underevidence/conformance/as fixture directory names only. - Do not treat rational-equality fixtures or bridge theorems as authority for an arbitrary submitted candidate; theorem CR is disabled for that capability in this pinned Lean 4.14 preview.
- Read
docs/security/KNOWN_TRUST_GAPS.mdbefore relying on any experimental capability.
External confirmations, independent domain/trust review, federation agreements,
usability evidence, multi-area review, and other checklist items remain future
requirements for a stable lifecycle promotion. They are not fabricated or
relabelled as completed by this experimental release.