SLM Mimir™
Seven AI specialist stages build, challenge and review work. Mathematics checks Engineering's logic; a blocking finding sends the work back for repair and a fresh review.
What Mimir is designed to do
SLM Mimir is a plugin workflow for a compatible AI coding host. It coordinates separate specialist agents to build, repair or review an authorised project using the host's selected model, tools and permissions. The workflow is intended to examine the same frozen candidate at each review stage and retain evidence for the decisions made.
The seven specialist stages
1. Engineering
Builds or repairs the implementation, runs relevant checks and freezes the candidate for review.
2. Mathematics
Checks the implementation's logic, assumptions, calculations and edge cases against the requirements.
3. Auditor
Reviews the candidate, earlier checks and supporting evidence for correctness and gaps.
4. Red Team
Challenges the candidate and its security and operational boundaries using authorised review and reproduction where appropriate.
5. Optimiser
Reviews potential improvements and records measured benefits and trade-offs where evidence is available.
6. Verification
Checks material claims against the candidate and records what was verified, partially verified, unverified or refuted.
7. Strategist
Assesses the result, remaining limitations and practical next steps.
Evidence: what can be checked
The beta includes local evidence tools designed to connect the review to defined project files and the task requirements. Depending on what actually runs and is recorded, a run can retain:
Candidate identity
A record of selected source scopes and requirements, with SHA-256 hashes used to identify the candidate and detect changes within those scopes.
Specialist reports
Stage verdicts, findings, host agent identifiers, times and references to the assessments and other evidence supplied for each stage.
Checks and outcomes
Recorded commands, results, failures, blocked checks and unresolved findings, where these were actually performed and documented.
Changed work and recovery
Saved run state can be checked on resume. A changed candidate needs a fresh review; failed or blocked reports can be retained for diagnosis.
What we have tested
These results apply to Build 5, 0.2.0-beta.5-obfuscated.1, on Windows x64 with Node.js 24.19.0. They distinguish package checks from testing a complete AI workflow.
All 14 packages · local checks passed
Package integrity and hashes were checked. Each package passed 22 local helper regression tests: 308 checks in total. All 14 preflight checks and 14 local free-beta activation checks passed. A corrupted protected payload was rejected.
Codex CLI · install and discovery
The OpenAI package was installed in an isolated profile; Codex discovered the coordinator and its installed helper passed preflight. This did not exercise seven agents on a real task.
Claude Code · plugin loading
The protected plugin loaded and the host inventoried one coordinator and seven agent profiles. Manifest validation passed with a missing-icon listing warning.
Gemini CLI · component loading
Extension validation and native component loaders found the coordinator and seven agent definitions. A complete interactive installation was not repeated.
Copilot CLI and OpenCode · partial
Copilot listed the plugin and found its coordinator; discovery of all seven agents was not verified. The OpenCode registration bridge produced seven definitions, but full host loading was not tested.
Nine other packages · not host-tested
Kimi, ZCode, Cursor, Antigravity, Grok, Factory, Qwen, Mistral Vibe and Perplexity passed the local package/helper checks above. Their native hosts were unavailable for this validation.
Earlier development tests
The following experiments tested earlier Mimir-related development workflows. They provide evidence of real specialist work; they are separate from the protected Build 5 plugin checks above.
Test 6 · FloodGate atomic authorisation engine
A specialist-led development run built and reviewed a local engine candidate. Its Auditor, Optimiser, Engineer and renewed reviewers documented and repaired an output-size limit defect. The frozen final candidate passed all 126 supported primary assertions; the 140-row record also contains 2 unsupported-profile failures, 9 blocked checks and 3 observations.
The full Red Team stage was blocked, so this experiment was a restricted workflow, not an unrestricted seven-stage pass. It did not qualify the engine for production deployment.
OpenCode · seven-stage development test
In a separate development test, seven actual specialist sessions reviewed a pinned candidate using a local model. Two small live QA cases passed: a fresh implementation and a seeded defect-and-repair case. The latter also recorded one stale false-positive revision.
This was a bounded backend test on one local configuration, not qualification of every host or the protected Build 5 release.
Test 2 · software engineering
Three selected software tasks were evaluated. Final native checks passed on two of three; one of three patches was accepted end-to-end within the tested scope. The other two remained unresolved or rejected.
Test 4 · completed
Specialist reviewers reconstructed and checked mathematical components on local CPU cases. Executed evidence covered nine of ten claim families; the remaining family was unverified. This did not reproduce full model training.
Test 5 · completed
Seven separate specialist contexts reviewed mathematical components and recorded their findings. After both candidates were frozen, the same independent common checks passed 15 of 15 for both the single-model and seven-stage workflow candidates. The results are bounded to those local checks and do not establish a general performance advantage.
Try the beta
Choose the package for your AI coding host and follow its installation instructions. A capable host must provide genuine separate-agent execution and the required file and command tools. The bundled helper requires Node.js 20 or later; account and host restrictions apply.
The beta has no SLM Mimir subscription fee. Your AI provider's own usage charges or account requirements may still apply. SLM Mimir is proprietary evaluation software.
Choose your package → Send feedback