Small, sharp engineering pieces — each built to make one failure mode measurable, then measure it honestly. Eval-driven, CPU-only where it can be, reproducible by anyone: where a repo makes a numeric claim, that number is produced by its eval/run_eval.py, not asserted. Where a piece is shipped craft rather than a measured claim, it's labelled as such — never dressed up in a metric no eval produces.
These are the public, reproducible demonstrations behind Vorion — a governed-AI execution platform shipped as the open-source @vorionsys/* packages, where the same discipline (governance loops, provenance, measured claims) becomes infrastructure for autonomous agents.
| Repo | The claim it turns into a number |
|---|---|
| governed-agents · live | A governed multi-agent loop — Researcher → Reasoner → policy gate → execute-or-halt. Watch it allow one action and block another (an unverified external send) with a logged rationale, streamed live to a trace UI. |
| agent-memory-service · live | "Append everything to a vector store" memory returns stale facts when facts change. A flat store answers current-fact queries right 0% of the time; the full memory service reaches 100% with 0% staleness. |
| art-director | Most of "good design" we call subjective is checkable. Single-shot generation passes WCAG AA contrast on 12% of text pairings; the same generator inside a critic loop reaches 92%. |
| genealogy-graphrag | Dense + lexical retrieval can't answer relational questions ("maternal grandfather of X?"). Adding kinship-graph resolution takes relational recall@5 from 0.000 → 1.000, no regression elsewhere. |
| cason-heritage · flcason.com | Where the pieces compose — a governed autonomous heritage system. Eleven era-bounded ancestral-persona agents read an eleven-generation graph, research the open lines, and gate every claim through a typed policy with an NDJSON audit trace — propose, never publish, the top autonomy tier deliberately empty. Reuses the governance loop, supersession memory, and kinship-graph retrieval from the repos above, on a real family. (Shipped system; the governance invariants are self-tested, not asserted.) |
| brainmix | Browser-native discipline, shipped: a full neuroacoustic DJ studio — dual-deck Web Audio engine, BPM detection, live rooms and WAV export — running entirely client-side in Next.js, no plugin and no DAW. (Shipped app, not a measured claim — no eval number is asserted.) |
| gigtrip | Planning a trip around live events is constrained optimisation, not a sort. An exact, brute-force-verified optimiser captures +30% more preference-weighted value than sort-by-date at the same travel budget (5-scenario eval; honest +0% cases kept). |
How the repos compose is mapped in
docs/GOVERNANCE-FEDERATION.md; how the work runs is inCONTRIBUTING.md. The governance plane across these repos is concrete and self-tested — a typed policy gate with attested model-consensus independence and an adversarial Red Cell, weight-space + behavioral model attestation, an append-only signed provenance graph (PAG), post-commit reconciliation drift, and a BASIS privilege ladder whose top tier is provably unoccupied — each pinned by a test and graded provable / qualified / aspirational, never asserted.
Each repo takes a claim that usually gets hand-waved — "our memory is smart," "the design looks good," "retrieval works" — turns it into a reproducible number, and reports the lift honestly (what's measured vs. aspirational, stated plainly). When a piece is shipped craft rather than a measured claim, it earns its place on the strength of what it actually does — not a borrowed metric. Generate-then-critique loops, provenance and audit trails, and disciplined scoping are the same principles the Vorion governance platform applies to autonomous agents. cason-heritage is where they compose — the governance gate, supersession memory, and kinship-graph retrieval running together as one governed autonomous system, on a real family's record.
Built by Ryan Cason — AI systems builder (@vorionsys).