A verification-gated AI studio. Applied ML and agent tools that carry their own evidence — and the gates that make them checkable.

Run by Kiliaan Vanvoorden from a plant-filled loft in Riemst, Belgium. Ideas, code, culture, freedom.

We publish failure states unprompted. A guard that has never been watched to fail has not been shown to work — elohim ships tampered copies of its own instruments, and every project under this org publishes its own broken-on-purpose check that runs on every invocation.

Open to AI engineering roles in evaluation and agent infrastructure — remote from Belgium, employment or contract. 📧 [email protected] · one-page CV · Site

, Ideas · , Code · , Culture · , Freedom

• Agent systems — multi-agent tooling, MCP servers, CLI/TUI engineering agents
• Evaluation gates — deterministic checksum-gated measurement; tamper tests that must fail
• Local inference — local-first LLM tooling, Ollama pipelines, sidecar architectures
• Applied ML — industrial computer vision and yield-prediction experiments for food processing

• Every number in a README is re-derivable from the repo or from a recorded measurement.
• A claim we could not read is UNKNOWN. It is never FALSE, and never an all-clear.
• A confirmed defect is pinned as a failing test, not quietly fixed.
• A repository says what it cannot do.

• No customer count, revenue, or funding. None.
• No production deployments — there is nothing running for anyone but us. This is a measurement, not an ambition.
• No claim that our gates cover the applied-ML repositories, which are private and unreviewed.
• No guarantee against the owner rewriting history — audits come from re-deriving each claimed number against a fresh clone of the relevant repo, not from any promise that this README will always reflect the live project state.

This org is the studio's index — identity, thesis and distribution. The implementation lives in the public repositories below, maintained and built by the same one person. Built by one person means you know who answers for a broken gate — not that the work is small. Nothing here is a mirror or a fork.

The portfolio is five public repositories plus this index: elohim, harness, terminal221b, mcp-regression-lab and repotruth. Re-derive the count with gh repo list Quattro-Commas --limit 100 --json name --jq '.[].name'.

Five of five public repositories are accounted for above. Re-verified 2026-10-11: the four showcase repos (elohim, harness, terminal221b, mcp-regression-lab) all have green CI on their latest runs. repotruth is maintained in this org at Quattro-Commas/repotruth (transferred 2026-10-10 from Bakery-street-project/galacticfederation; the old URL returns a 301).

Every project sits at 0 stars and 0 forks. There are no customers and no deployments.

The applied-ML and local-inference work runs in private repositories and will not be published, so no client, dataset or repository name appears here — and no metrics are quoted, because none of them can be re-derived from a public repo. What can honestly be described is the shape of it, which is the only evidence behind the applied-ML and local-inference pillars above:

• a multi-task image classifier that takes one photo and returns both a class and a storage-age estimate, served from ONNX behind FastAPI with a browser demo in front of it
• a lot-level yield predictor that routes output to specification tiers
• a shelf-life-aware allocation advisor (FIFO/FEFO) paired with a local-first store assistant
• a Next.js chat copilot wired to internal tooling over MCP
• a MAP-Elites quality-diversity search over idea spaces

All of it is screen-shareable on request.

A guard that does not hold must fail the run.

Collected and explained in eval-gates.