🔒
// rmpl_core.py
Runtime Memory Persistence Ledger
The spine. Every memory write passes through a deterministic governance layer before it touches storage. Provenance tracked. Conflicts caught. Nothing slips through without a record.
100% context portability — 10/10 critical facts preserved across sessions
🌊
// z1_dam.py
Dam Layer
First gate. Stale context, conflicting facts, and bad writes hit the dam before they poison the ledger. Binary decision. No ambiguity. No negotiation with the model.
100% enforcement — 28/28 cases
⚔️
// z1_action_guard.py
Action Guard
Deterministic rules R0–R4 screen every proposed action first. The 3B auditor only sees boundary cases. This two-stage design is what hit 100% — model alone didn't get there.
100% enforcement — 31/31 cases
🗂️
// rmpl_silo_router.py
Silo Router
Deterministic keyword matching routes context to the correct memory silo. No model inference for routing — fast, exact, auditable. Tarpit detection lives here too, in pure Python.
100% routing accuracy — 45/45 cases
🔬
// 3B auditor model
Probabilistic Classifier
llama3.2:3b handles the fuzzy middle — conflict detection, stale context classification, boundary-case action gating. Fast. Local. 90.5% zero-shot. Fine-tuning is the R&D unlock funding solves.
90.5% zero-shot baseline — 200 examples
📋
// rmpl_audit_coordinator.py
Audit Coordinator
Phase 2 bridge. Silo-level auditing after every write. Tarpit flags written to per-silo audit_flags.jsonl. Human-only tarpit release. Observes and flags — never acts autonomously.
Phase 2 — loop closed, wiring in progress