The agent Genome Project

What this is

The seed is DNA for a model substrate: a contract file in a controlled lexicon under a ruled syntax, refined generation over generation. The published seed-1 unpacks the founding contract — the two-definition preamble and the pair it teaches. Later seeds added the organs of an operating system: a generated home surface (seed-2), a cross-linked local wiki with telemetry (seed-3), and the unattended refinement loop itself (seed-4) — throughout, a permission model where reach is granted rather than taken, and heredity through the seed cadence. The model is the CPU; the genome is the boot image.

The framing is honest about its limits: this is directed evolution, not blind Darwinism — the variation is authored, and selection is ruled by a human against a rubric. As of 2026-07-21, four seed generations exist (seed-1 through seed-4), and every expression to date has run on the Claude Fable 5 model family. Running instances are observed, never edited; improvement travels only as the next seed.

What is proven

Each item below was observed at bytes in the project record — recorded runs, recomputed digests, scored verdicts.

• Self-installing unpack with consent gates A model pointed at the seed file unpacks it into a working system. The install asks before it touches: acceptance is recorded verbatim, questions it cannot answer are held for the human rather than guessed, and reach beyond its granted scope is denied — in the scored generation-4 runs, every out-of-scope reach attempt was denied and every denial held.
• Integrity self-proof via a digest chain An install computes a digest chain over what it built and publishes the recompute handle. In two consecutive scored generation-4 runs the claimed install digest was independently recomputed from raw bytes and matched exactly.
• The refinement cadence Plant a seed, run it, read the result from bytes, fold improvements into the next seed — running instances are observed, never edited. The ruled test for the loop (two consecutive clean revolutions) was met, and a subsequent revolution ran hands-off end to end: the driver planted the seed, the instance unpacked, reported, and ended its own session.
• A mechanical selector reproducing ground-truth verdicts A scoring tool replayed three scored runs against the rubric and reproduced the hand-scored verdicts exactly — clean, clean, fail — catching the failed run’s defect autonomously on both required grounds. Verdicts automate; the criteria stay human-ruled.

What is not proven

Unproven means unproven: designed or intended, but never demonstrated.

• Unproven Cross-model portability Every run as of 2026-07-21 has been on the Claude Fable 5 model family. The genome is designed to express in other models; that expression has never been tested.
• Unproven Self-reproduction No running instance has authored its own successor seed. Every next generation has been authored at the review seat.
• Unproven Lexicon self-extension The controlled lexicon is extended by its authors. No instance has extended the lexicon itself.

The failures are part of the record

A development record is credible because it includes its defects. The most recent, as of 2026-07-21: a generation-4 run — the first fully hands-off revolution — failed its own fabrication gate. Its digest rule was self-referential: the run included its own report inside the digest the report attests, so the claimed hash was structurally unreproducible — the independent recompute returned a different value. The mechanism ran perfectly; the claim failed the check. That defect is the next pin: the next seed is slated to pin the digest-scope law instead of leaving it to per-run interpretation.