Skip to content

Harden Fable Mode with evidence-bound System 3 control loops - #8

Closed
REX-codebase wants to merge 6 commits into
mainfrom
zapia/fable-mode-system3-verifier-final-20260902
Closed

Harden Fable Mode with evidence-bound System 3 control loops#8
REX-codebase wants to merge 6 commits into
mainfrom
zapia/fable-mode-system3-verifier-final-20260902

Conversation

@REX-codebase

Copy link
Copy Markdown
Owner

Summary

This PR consolidates the accumulated Fable Mode work into one reviewable change set. It adds a hardened, evidence-bound System-3-style control loop around compatible MCP hosts; it does not claim to alter model intelligence or provide a complete OS sandbox.

Control and verification

  • Adds the strict fable_control_plane lifecycle: observe -> record prediction -> propose action -> record outcome -> request verification -> finalize.
  • Integrates ClaimGraph, VerifierPlanner, and ThreeValuedAdjudicator into ordinary finalization. Uncovered claims, unresolved UNKNOWN results, invalid provenance, and failed/counterexample-backed claims block finalization.
  • Enforces receipt-bound evidence, provenance resolution, verifier policy classes, quorum, diversity, independence, and signed process-attestation scope.
  • Preserves legacy fable_session compatibility through an explicit immutable compatibility mode rather than silently weakening strict paths.
  • Hardens counterexample integrity, claim decomposition/dependency DAGs, intrinsic risk classification, planner metrics, and mutation/metamorphic verifier hooks.

Runtime, MCP, and broker hardening

  • Adds receipt-bound System 3 observe/predict/act/outcome/update/verify/finalize handling.
  • Adds authenticated checkpoint/event reconciliation, immutable terminal state, graph closure, sealed axioms, typed measurements, and replay validation.
  • Adds broker handshake/cancellation, bounded worker admission, executable/shebang pinning, quotas, workspace accounting, and fail-closed unsafe Windows capabilities.
  • Adds session/CAS identity binding, sharded quotas, metadata and cancellation bounds, strict MCP protocol checks, adapter fixtures, and packaging/API drift checks.
  • Restores authenticated strict-System-3 state across signed active and finalized checkpoint round-trips, with tampering and unsigned replay rejection.

Validation

  • Full unittest suite: 212 passed
  • Red-team test suite (test_*redteam*.py): 47 passed
  • Server protocol and focused verifier tests included in the full suite
  • Documentation checker: 19 Markdown files passed
  • compileall: passed for source, tests, and docs
  • pip check: passed
  • Targeted strict checkpoint probes: active/finalized restore passed; tampering and unsigned authorization replay rejected

Explicit limitations

  • The broker is a policy boundary, not an OS/network sandbox; hostile workloads still require container/VM/MAC/seccomp/Job Object isolation.
  • In-process verifier claims are not a security boundary; strong process attestation requires an external authenticated verifier/host boundary.
  • V1 synchronous interruptive cancellation/deadlines remain unsupported; broker cancellation is bounded and active.
  • Release artifacts are unsigned; checksums provide integrity checking, not publisher authentication.
  • Claude Code and Antigravity adapters are fixture-tested, not certified against live installations.
  • Cross-session adaptation is disabled; persistence is auditable state, not authenticated self-improvement.

Please review the strict finalization, provenance/attestation binding, checkpoint restore path, and broker boundary assumptions in particular.

@REX-codebase
REX-codebase force-pushed the zapia/fable-mode-system3-verifier-final-20260902 branch from 8252606 to 0659a3f Compare September 2, 2026 02:42
@REX-codebase

Copy link
Copy Markdown
Owner Author

Superseded by v1.3.0 frontier release on main

@REX-codebase
REX-codebase deleted the zapia/fable-mode-system3-verifier-final-20260902 branch September 2, 2026 14:18
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant