What is drawn and what is built

Daedalus describes itself as a scaffold. README.md:42 is headed “v0.1 scope” and lists eleven components; RESEARCH.md:265 is headed “What We Built” and ends with nine numbered “Next Real Milestones”. Reading the source against those lists is the fairest way to assess it, and the result is consistent: the compile-and-record half of the system is complete and works, and the measure-and-act half does not exist.

Of the six clauses in that thesis, two are built. The strict local control plane exists and is genuinely strict — closed unions, path validation, a persisted ledger. Frozen evaluator snapshots exist, at manifest-identifier granularity. Artifact-level evals, replaceable coding agents and git worktree rollback are declared in types and documented in prose, with no executing code behind any of them.

Field survey — declared, never written

The project’s own checklist, checked

All ten v0.1 scope claims hold to the letter, though two hold narrowly: the evaluator pack manifests exist as files that nothing loads, and the adapter registry exists as data that nothing invokes. The claim that fails is the eleventh row, which is not from that list — “artifact archives”, listed under what Daedalus owns at README.md:30. The directory is created by init and no code path ever puts a file in it. Given that docs/ARCHITECTURE.md:114–119 makes artifact evidence the basis of acceptance, that empty directory is the single largest gap in the system.

What runs today

$ daedalus doctor
Daedalus root: /home/me/daedalus
DAEDALUS_HOME: /home/me/daedalus
# ensureDependencies(): looks for node_modules/.bin/tsx
Daedalus first run: installing dependencies...
# npm install  (only on first run)
# npm run typecheck  ->  tsc -p tsconfig.json --noEmit
Daedalus doctor passed.

$ daedalus build
# npm run build  ->  tsc -p tsconfig.json   ->  dist/

$ daedalus research:clone
# powershell -ExecutionPolicy Bypass -File scripts/clone-research-repos.ps1
# shallow-clones the 12 entries in research/repos.json into research/repos/
#   PowerShell-only; the runner forwards it on every platform

Repository inventory

Line counts for every tracked file, so the scale is not in doubt.

A large share of the TypeScript is declaration and data rather than behaviour: types.ts emits nothing at runtime, and adapters.ts plus the first 110 lines of evaluatorRegistry.ts are object literals. That is not padding — the registries are the product in a control plane whose job is to hold judgements — but it does mean the executing logic is closer to 800 lines than 1,400.

The verification budget is one file. scripts/smoke.ts is 52 lines of node:assert/strict covering a single happy path through the browser_game domain, plus one negative check on run-id validation. There is no test for classifyDomain on any other domain, none for scoreAttempt’s decision branches, none for selectEvaluators’ uncovered-proof path, and no CI workflow to run even this one.

Assessment

Daedalus is a well-argued scaffold. Its strongest asset is not code but the position it takes and records: that the durable part of an autonomous build system is the goal, proof and scoring vocabulary, and that everything which edits files should be replaceable. The proof union, the evaluator manifests with their mutation tests and anti-cheat notes, the deliberately split Autoloop entries, and the adapter failureModes lists all read like the output of someone who has watched agents fake progress and wants a schema that makes faking visible.

What exists is a compiler and a ledger. Given a sentence it produces a contract, a milestone list, an evaluator selection and a frozen hash, and it will faithfully record scores an agent hands it. What does not exist is anything that measures. The scoring input is a caller-supplied record of numbers; no evaluator is executed, no adapter is spawned, no artifact is captured, and no budget is enforced. The repository says as much — item 2 of its own outstanding list is precisely this — and the type system already has the shape waiting for it.

Judged as a v0.1 scaffold, which is how it labels itself, that is a coherent place to be. The danger it names in RESEARCH.md:63–66 — a loop that optimises a weak evaluator — is the danger it is currently exposed to, and it has written the schema for the fix before writing the fix.