Skip to content

Latest commit

 

History

History
169 lines (133 loc) · 10.7 KB

File metadata and controls

169 lines (133 loc) · 10.7 KB

Known limitations and trust gaps

This document lists current, honest limitations of the MathEvidence public preview. It is part of the trust surface. Experimental capabilities must not be presented as stable, and human/external gates must not be invented.

All registry capabilities remain "status": "experimental" until the stable promotion checklist and GOVERNANCE.md requirements are met with real artifacts.

For machine-readable CR maturity, use registry/maturity-inventory.json. For the short public status, use docs/STATUS.md. Historical dated audits are evidence of their date, not current promotion authority.


Trust invariants

These do not change with backend, benchmark, or release status.

  • External backends, models, search procedures, and adapters are untrusted.
  • Lean/checker authority is capability-specific and must match the advertised proposition.
  • A backend Boolean answer is never sufficient theorem evidence.
  • A fixture or nearby theorem cannot certify a different submitted candidate.
  • Theorem-level Certification Records require exact candidate binding and live registry CR eligibility.
  • Assurance may not be escalated by an adapter, serializer, receipt field, user flag, benchmark result, or fallback path.
  • Unsupported exact modes fail closed.
  • Counterexample certification has polarity refuted, not proved.
  • Numerical agreement is not exact proof by relabeling.
  • Failure to find a counterexample is not a proof of universality.
  • Historical records retain the semantics under which they were created; they are not silently upgraded when a later version gains stronger assurance.

Forensic regressions under tests/forensic/ guard these properties.


Current engineering posture

Exact candidate binding is the theorem-CR rule (ADR 0005). CR eligibility is registry-backed, and the required lean workflow executes production-generated candidate modules for every CR-eligible capability. Structural source generation alone is not sufficient release evidence.

Area Honest status
Exact binding / CR Five owned capabilities are registry-eligible for exact CR (proved, except finite CEX refuted). Rational equality is deliberately non-eligible for theorem CR in the pinned Lean 4.14 public preview. Federated SAT/PB/SMT remain non-eligible.
Ideal membership Witness identity only (algebra.ideal_membership_witness); no Gröbner-basis, non-membership, radical, minimality, or completeness claim.
Rational equality Checker, soundness theorem, bridge, and exact-source generation remain. Candidate-specific theorem CR is disabled fail-closed because the pinned Lean 4.14 production native-reduction path does not admit the generated checker proposition without an unacceptable sorryAx dependency. Binary floating point is not silently promoted to exact arithmetic.
Linear algebra Exact rational inverse_witness, system_solution, kernel_vector, and det_identity; no broad linear-algebra completeness/rank/basis claim.
Finite counterexample Exact finite witness can establish refuted. No-witness or sampled search cannot establish the universal claim.
Formal calculus algebra.formal_rational_calculus is a formal/algebraic grammar, not general analytic calculus.
Analytic calculus analysis.analytic_calculus is a strict theorem-form whitelist, not arbitrary analysis. Exact ODE support retains its documented obligation/initial-condition restrictions.
Evidence bundles Candidate Bundle v0.3; Certification Record v0.4 for exact promotion. Legacy records must not be silently upgraded.
Offline bundle replay Available where declared: sealed candidate artifacts can be regenerated/validated without consulting the solver after materialization. This may end at theorem_pending.
Offline kernel replay Tracked separately as offline_kernel_replay_exists. It is currently false as a release maturity property; optional Lean execution succeeding on a machine is not the same as a required, network-isolated release gate.
Bundle verifier mathevidence-verify-bundle emits operational checker status only. It is not theorem Certification authority.
CI / local checks Local just check is useful feedback, not release attestation. Exact release claims require green remote gates on the exact release SHA.
Repository rules Branch protection/rulesets are operational governance choices for this experimental preview. They are not mathematical assurance evidence and are not a public-preview release prerequisite.
Stable promotion Blocked until the repository-defined human/domain/trust/external gates close. Experimental CR eligibility and stable lifecycle promotion are separate.
CODEOWNERS Single-owner incubation stub (@fraware). Multi-area dual review is not enforceable yet.
Signing / PKI Production receipt PKI and production release signing remain deferred. Dev keys are not production authority. The experimental release workflow records unsigned status explicitly rather than claiming a signature.

Open human and governance gates — blocking stable

ID Limitation Where to record progress
H-1 ≥3 external Milestone 0 user confirmations docs/validation/user-confirmation.md (0 completed)
H-2 ≥1 external workflow-win confirmation (§21.10) docs/validation/workflow-win-log.md
H-3 Independent domain + trust-model reviews for stable promotion docs/validation/review-packets/, stable-capability-checklist.md
H-4 Live federation agreements with ≥2 external peers docs/validation/federation-live-checklist.md, docs/architecture/federation-agreements.md
H-5 Studio usability results with ≥3 completed sessions docs/validation/studio/usability/
H-6 Expert judgments for hypothesis interfaces / conjecture precision / TTP graph docs/validation/review-packets/
H-7 Real multi-area CODEOWNERS / dual approval .github/CODEOWNERS, GOVERNANCE.md

The semanticReview and trustReview registry fields refer to this stable-promotion human review layer. Their absent state is intentional and must not be presented as completed review. They are distinct from the mechanical exact-candidate CR gate used by this experimental preview.

Wave-8 human scaffolding remains blocked until real external artifacts exist; templates are not confirmations.


Open engineering and product gaps

ID Limitation Notes
E-1 Immutable all-green release commit The final tagged SHA must have the required assurance/security/replay/conformance gates green.
E-2 Repository governance hardening Optional operational hardening for this experimental preview; not a mathematical-assurance or release prerequisite.
E-3 Lean toolchain changes lean-toolchain is pinned; a bump requires a separately validated change.
E-4 LeanLink native Mathematica bridge Deferred; live Mathematica transport is wolframscript when configured.
E-5 Sage rational equality Declared/placeholder; not advertised as live Agent routing.
E-6 Analytic-calculus completeness Out of scope. Only the registered whitelist and explicit hypotheses are supported.
E-7 Production receipt PKI / release signing identity Deferred; no dev key or soft signing attempt may be marketed as production signing.
E-8 Foundry frontier / funding exits Tiny-suite tool-selection results do not establish frontier acceleration or maintenance funding.
E-9 Independent external reproduction Release artifacts are designed for it; third-party reproduction remains external work and must not be fabricated.
E-10 Ideal flagship adoption Exact candidate path exists, but live external adoption/held-out validation remains open.
E-11 Windows native Lake link Required workaround remains scripts/link_exe_via_rsp.py; degrade with dependency/setup status, never fake Certified.
E-12 Practical LA scale Exact determinant/checker cost and the IR size policy intentionally bound practical dimensions; this is not a completeness claim.
E-13 Lean internal expression identity Compiler-internal Expr.hash stability across revisions is not claimed as a protocol guarantee.
E-14 Rational theorem CR on pinned Lean 4.14 Disabled fail-closed for this public preview. Re-enabling requires a candidate-specific production theorem path that passes without sorryAx and is then requalified on an exact release SHA.

Environment-level Lean import/axiom audits are implemented through the mathevidence-import-graph / mathevidence-axiom-report drivers and CI; source scans remain defense in depth.


Capability naming and claim-scope notes

  • Public formal-calculus ID: algebra.formal_rational_calculus.
  • Public analytic-calculus ID: analysis.analytic_calculus; strict whitelist only.
  • Ideal-membership ID: algebra.ideal_membership_witness; witness identity only.
  • Rational equality must not be described as theorem-CR eligible in this pinned Lean 4.14 release, even though its checker/soundness/bridge code exists.
  • Linear algebra must be described operation-by-operation, not as generic verified linear algebra.
  • Legacy fixture/conformance directories may use historical names such as calculus or symbolic_calculus; directory names do not broaden the public mathematical claim.
  • Do not advertise a live registry capability that does not exist.

Benchmark interpretation

The frozen ideal-membership release corpus is a release conformance and assurance-regression corpus. It is useful for deterministic implementation checks, mutation testing, answer/evidence separation, and observed false-accept behavior on that corpus.

It does not by itself establish a population false-accept probability, universal solver soundness, broad mathematical generalization, or formal checker soundness. Formal assurance comes from the declared checker/soundness argument within its exact scope; empirical suites test the implementation and integration of that argument.

The critical failure cell remains:

answer incorrect + evidence verified

Any such deterministic release-corpus event is a release blocker.


Release truth rule

For a release claim, prefer evidence in this order:

  1. the mathematical proposition actually established;
  2. executable checker/verifier implementation;
  3. adversarial and contract tests;
  4. exact-SHA CI / replay evidence;
  5. machine-readable capability and maturity registry;
  6. current status documentation;
  7. historical audits and roadmap labels.

Documentation cannot strengthen a weaker checker.