What it will not do

The invariants are longer than the features. This is the deep-tech signal: each limit is a response to a documented failure, and each one is enforced structurally, not by a "please review carefully" label.

It will not edit its own referee.

An optimizer with write access to its own scorecard edits the scorecard. The Darwin Gödel Machine, scored by a function counting marker tokens, improved its score by deleting the markers. So the validator, the QA battery, the eval corpus, the tests, and every human-authored agent and skill are off-limits — enforced in-process by protected.js and again out-of-process by CI. ↗ doc ↗ commit

src/spec/**
src/qa/**
src/eval/**
src/evolve/patch.js
src/evolve/protected.js
test/**
.github/workflows/**
_fleet/local/scripts/validate-handoff.sh
docs/milestones/v0.5.0-self-evolution.md
docs/research/self-evolving-agents-2026-08.md
docs/architecture/multi-user-context.md

It will not change a handoff contract without a human.

Handoff contracts are rails. Instruction bodies evolve freely; declared inputs/outputs change only via an explicit, separately-reviewed mutation. The contract-change op is refused by default and never auto-applied. ↗ commit ↗ doc

It will not gate on a model.

A judge may advise; it may never gate. Deterministic checks carry every gate: the validator, the QA battery, the eval corpus. There is exactly one model call in the whole system — the proposer — and it is the one cell that cannot replay. ↗ doc

It will not merge unattended outside the whitelist.

--apply merges only proposals whose every op is on the auto-apply whitelist (update-bullet-counter, add-validator) — ops whose correctness a validator fully decides. Everything else queues for --review, and queuing is the point: a reviewer handed a steady stream of mostly-fine changes stops reading them. ↗ doc

What this page does not yet claim: the 19% result.

The judge-calibration result (the judge disagrees with itself on 19% of verdicts) is a planned negative-result section. It is cut from this release because docs/research/judge-calibration.md does not exist at the pinned v0.5.0 commit — and the site does not link to evidence that isn't there. It will appear when the pin moves past the commit that introduced it.

It is not for everyone.

The prompt-marketplace browser looking for "500 agent templates" will find this site mildly disappointing on purpose: no gallery, no templates promise, no breadth claims. Serving them would blur what the tool is.