MemKraft 3.0.3

MemKraft 3.0.3

MemKraft 3.0.3 adds a bounded deterministic execution path to ReasoningBank while preserving existing natural-language recall/injection APIs and schema-v1 trajectories.

What’s new

• Added reasoning_procedure_ref(), reasoning_build_authorization(), and reasoning_execute().
• Added six versioned, allowlisted exact procedure grammars for deterministic execution.
• Added strict schema-v1 provenance validation and process-local HMAC-sealed authorization.
• Anchored trajectory reads to no-follow directory descriptors, keeping validation, hashing, and parsing on bytes read from the same file descriptor.
• Enforced fail-closed, exact-once fallback for malformed, ambiguous, unauthorized, unsupported, oversized, or unsafe inputs.
• Kept lesson prose, dynamic code, shell commands, eval, and exec outside the execution path.

Evaluation

On the frozen 28-case product matrix:

• accuracy: 28/28
• deterministic executor routes: 24
• fallback routes: 4
• observed model calls: 28 → 4 (-85.714%)

In one live sequential provider campaign:

• model-only baseline: 305.406 s, 28/28 correct, zero endpoint errors
• product router: 12.894 s, 28/28 correct, zero endpoint errors
• observed campaign wall reduction: 95.778%

The wall-time result is descriptive evidence from one campaign, not a universal latency claim. Exact grammar coverage does not establish paraphrase, noisy-store, or broad semantic generalization.

Compatibility and security scope

• No destructive migration is required.
• Existing ReasoningBank recall/injection APIs and schema-v1 records remain compatible.
• Authorization trust is process-local and intentionally does not persist across restarts.
• Arbitrary same-process code/key access, debugger or fork compromise remain outside the current threat model.

Validation

• 112 focused execution tests passed
• 2,240 passed / 3 skipped in the full repository suite
• exact-commit Memory Gym CI passed for the product implementation
• independent review: Critical 0 / Important 0

Detailed benchmark methodology and limitations are documented in docs/bench/reasoning-injection-3.0.2.md. Raw credential-free evidence is included under benchmarks/results/.