Demo T-001 seed is a recurring F1 instance (description authorises a command whose collateral the AC doesn't ratify)

[samkeen]

What this is

A tracker for the F1 friction ("the seed is the contract but nothing validates the seed", proposals/frictions-2026-05-26.md) as it shows up concretely in the demo's T-001. Not a proposal to patch the seeder for uv init specifically — see Explicitly out of scope below.

The shape

T-001 ("Bootstrap project skeleton"):

• Description authorises uv init to create pyproject.toml.
• uv init also creates README.md, .python-version, and a root main.py stub — none acknowledged by the description.
• Acceptance criteria require todo_cli/__main__.py as the entry point (from todo_cli.__main__ import main succeeds). They say nothing about the collateral files, and nothing about cleaning up the now-orphaned root main.py.

So the seed relies on the worker inferring that the auto-generated root main.py is dead weight to remove, and that README.md / .python-version are fine to keep. The contract is underspecified, not internally validated.

Where seen

• 20260526-173822-f411ea — original F1 example.
• 20260527-174931-25fa4c — Phase 1 demo run. Evaluator rejected; worker thrashed iters 24-35 (incl. an attempted [tool.ruff] exclude at iter 33 — F12 gaming) before user interrupt.

Why we're tracking, not fixing in v1

The run-time symptom (evaluator over-rejecting hygiene/scaffolding files) is being addressed separately by softening tilth/prompts/judge.md — redrawing the scope-creep bright line at cross-task interference and letting the evaluator apply judgement to hygiene/tooling/scaffolding collateral. That should let this seed succeed at run time with one corrective iteration (worker removes the dead main.py, keeps README.md/.python-version).

But the seed itself is still unvalidated — the judge softening compensates at run time; it doesn't make the contract self-consistent. The structural fix is contract negotiation at seed time (Anthropic harness pattern: evaluator reviews proposed AC before any code is written), which is v2 territory per proposals/v1-worker-evaluator-dialogue.md ("What v1 is deliberately not").

Value as a fixture

This seed is a good reproducible anchor for:

• Phase 3 (submit_case.work_arounds): does the worker now name the scaffolding collateral as a work-around, and does the evaluator accept/reject that specific claim instead of inferring it from the diff?
• v2 contract negotiation: does the evaluator catch the description↔AC underspecification before the worker runs?

Explicitly out of scope

• Hardcoding uv init awareness into tilth/seed/prompts.md (overfits — next time it's git init / npm init / cookiecutter).
• A narrow "scaffolding-command discipline rule" in the seeder prompt — considered and dropped; the judge softening is the better intervention layer for v1, and seed validation is the real fix for v2.

Related

• F1 / F2 / F4 / F12 in proposals/frictions-2026-05-26.md
• proposals/v1-implementation-plan.md (Phase 1 shipped; Phase 3 = submit_case)
• Judge can't review seed acceptance test files; rejects on file-existence criteria #16 (judge can't review seed test files), prep-feature: ask_user bundles multiple questions but options only cover one #18 (seeder compound questions)