LIMEN13

Validation record

LIMEN13 0.10.1: technical validation

Software version 0.10.1. Verified on 14 November 2026 (America/Toronto). Run by the project itself: this is not an independent audit.

In plain words. Before each release we run more than a thousand automatic tests and a library of simulated attacks. This page gives the exact results for release 0.10.1, including the two attacks it does not stop.

Results

Verification results

Verification gates and results for release 0.10.1
GateResult
Full verification script1,086 checks passed, 0 failed
Python test suites29 suites
Console JavaScript checks7 passed (included in the total)
Reference acceptance test42 of 42 passed
Attack library73 of 73 expected blocks; 2 documented limits
Security scanClean: Bandit, dependency advisories, supply-chain check, Gitleaks, Semgrep
SBOMRegenerated from the locked dependencies (CycloneDX)
Container build and smoke testBuilt; the health endpoint, the console and the simple view answered; the image health check passed
Runtime hardeningNon-root, read-only root filesystem, all capabilities dropped, no new privileges

Categories

What the checks cover

  • Engine and gateway decisions
  • Console and access control, route by route
  • Threat cases and hardening
  • Model gateway and budgets
  • Alert delivery and recovery after restart
  • Discovery, collectors and bypass probes
  • MCP client and server sides
  • Cryptography and data at rest
  • Customer isolation with two engines
  • Supply-chain checks
  • Kill switch and containment
  • Evidence coverage and reporting

Definitions

How results are defined

Check passed
An automated assertion produced the expected result.
Expected block
An attack case that LIMEN13 refused, held or contained, as the case defines.
Documented limit
An attack case that LIMEN13 does not stop. It is kept in the library on purpose and described below.
Failed
An assertion that did not produce the expected result. Count in this release: 0.

Limits

The two documented limits, and the others

  • Rewriting an unsigned lab log. In lab mode without a signing key, a log can be rewritten. This is not a bypass of the signed, anchored production design.
  • An agent making a false promise. An agent can tell a customer something untrue inside an allowed message. LIMEN13 checks actions, not truthfulness.
  • Calls already admitted to an external service may finish after containment.
  • An alert accepted just before a crash can be delivered again.
  • Budget accounting is estimated; it is not a provider billing cap.
  • These tests establish correctness for the cases tested, not a sustained high-volume capacity.

Corrections

Issues found and corrected in this review

  • Invalid budget or delivery state no longer resets silently: startup refuses it with a recovery error.
  • A containment that happens during a slow tool check or an approval re-check now refuses the action before it executes.
  • Interrupted model streams keep conservative usage accounting instead of counting missing output as zero.
  • Alert delivery progress no longer skips an earlier event whose record arrives late.

Measured

Real-engine demonstration and local latency

Run on 15 November 2026 with the real engine (release 0.10.1) and the fictional reference client, on one development laptop. This is not the browser simulation, and it is not a production measurement.

A real agent run on the real engine
StepWhat the engine answered
A test agent asks for a tool in its role (list_new_orders)HTTP 200, decision: allow
The same run asks for a tool outside its role (delete_customer)HTTP 403, decision: deny, rule: role, reason: tool 'delete_customer' is not in this agent's role
Evidence writtenTwo signed records for the run: one allow, one deny, each with its rule and reason
Chain verificationok: true, signed: true
Latency of tool calls, local development measurement
WorkloadRequestsp50p95p99Throughput
Allowed calls, 1 client1,0004.95 ms8.49 ms9.93 ms177 per second
Refused calls, 1 client3004.71 ms5.62 ms6.44 msnot measured
Allowed calls, 4 clients at once40036.18 ms43.07 ms63.57 ms109 per second
Allowed calls, 16 clients at once1,600177.46 ms234.57 ms265.72 ms86 per second
  • Local development figures. Client and engine on one laptop (AMD Ryzen 5 4600H, 12 threads, 18 GB, Linux, Python 3.14), over the loopback interface, with a test tool that answers from local files. A real network and a real business system add their own time.
  • One engine handles calls one at a time. Throughput did not rise with more clients, and waiting time grew. Plan one engine per workspace and measure with your own load.
  • Not measured: memory and processor use, behaviour during a dependency failure, long runs, and model-gateway calls.
Reproduce
.venv/bin/python deploy/bench_gateway.py
Evidence artefacts
The raw results are given to pilot participants on request.

Reproduce

How to reproduce

The software is delivered as source, so pilot participants can run the same scripts on their own machine.

All checks
./deploy/verify.sh
Acceptance test and attack library
./deploy/verify_attacks.sh
Security scans
./deploy/security_scan.sh
Evidence artefacts
The verification report and the SHA-256 list of the reviewed files are given to pilot participants on request. The source repository is private.

Pending

Not yet done

  • An independent penetration test.
  • Validation with a customer's own single sign-on, SIEM, vault and MCP servers.
  • Load tests on a deployment.
  • Backup and restore objectives.
  • Published support commitments.