Effect-bound
Permit binds tenant, release, model, policy, tool schema, effect digest, destination, nonce and expiry.
The gap
Safety must reach the destination.
Prompt controls, output filters, policy decisions and observability are important. They do not automatically prove that the exact external effect was authorized, current, unique and impossible to reuse.
CRLP is designed to verify before commitment, issue a signed single-use permit bound to the exact effect, enforce it at the destination, and preserve evidence.
Commit flow
From intent to verifiable effect—or no effect.
INTENT→PENDING_COMMIT→PRECOMMIT VERIFY→PERMIT / DENY→DESTINATION ENFORCEMENT→EFFECT / NO EFFECT→EVIDENCE
Single-use
Replay protection consumes the permit atomically and rejects duplicates or substitutions.
Fail-closed commitment
Unsafe commitment is denied while valid service paths may continue through reroute and recovery.
Evidence-first
Terminals, effects, manifests, hashes, receipts and recovery evidence are preserved.
Demonstrated on AWS
One live run. Seventeen final gates. Internally sealed evidence.
17/17final composite checks
2audited live runs: LOAD / UNLOAD
9/9host-network behavioral cases
67/67manifest-bound evidence artifacts
- PostgreSQL 18.4 transaction path with TLS 1.3, mTLS, verify-full, SCRAM-SHA-256 and channel binding.
- A production inference server (version 2.70.0) executed explicit model LOAD and UNLOAD operations with postcondition verification.
- Host-network v4 confirmed no host bindings through Docker Engine inspection.
- Runtime destruction, secret cleanup and post-run zero state passed.
- SHA-256 manifest and internal Ed25519 2-of-3 threshold verified.
Evidence SHA-256
Read the corrected audit
710f94b194d5ad7e9f3c480906eb45b7043cbbdc9aec46f9167769a14ecdf4d5Technical dossier · English edition
CRLP Strategic Technical Dossier
Commit authorization, containment and recovery for autonomous AI systems. The dossier presents demonstrated results, claim boundaries, market differentiation, the compromised-agent threat model, a technical integration vision and a measurable 90-day path.
Pre-production technical demonstration. Not externally certified, approved, adopted or supported.
Accelerated-stack integration candidate
Complement the stack at commitment time.
Potential integration surfaces include agent-tool middleware, execution guardrails, inference-server model management, API gateways and Kubernetes admission controls.
CRLP is not presented as a replacement for existing guardrails, inference runtimes, identity, policy engines or observability. It is a candidate destination-enforced commit authority.
Request technical reviewClaim boundary
Ambitious, but not exaggerated.
Demonstrated: a single-host pre-production live pilot with PostgreSQL and production inference server effects, final gate, cleanup and internal evidence seal.
Not yet demonstrated: production readiness, enterprise-wide no-bypass coverage, independent key custody, runtime-native enforcement, distributed federation, or external certification, approval or adoption.
Naming clarity. In the R3.5 product line, CRLP means Commitment, Containment and Recovery Layer. The historical Cache Refresh & Lineage Protocol branch is preserved separately and unchanged at CRLP Cache Sentinel.
CRLP / ECP relationship. CRLP explores pre-production commit authorization and external-effect containment. VSL-ECP is the focused enterprise product direction for authorization-to-effect control and evidence.
Public evidence snapshot: CRLP R3.5 v0.2.5 RC01 · evidence SHA-256 710f94b194d5ad7e9f3c480906eb45b7043cbbdc9aec46f9167769a14ecdf4d5
Continue exploring
BACK / MADRE
