Handbook

L1.2 — Function with agent-authored tests

Status: Demonstrated — retro-tagged L1.2b (Blueprints Example 2, L1-a-broken-link; closest evidence match, predates the scheme)

Updated

Sub-level at a glance
Parent level L1 — Function
Autonomous unit Function + agent-authored unit tests
Distinct assay evidence New test file in diff; tests pass
Grade today b (retro-tag)

Canonical policy, grades a-d, and promotion criteria: Autonomy levels (Blueprints). Per-project scoring: Autonomy maturity framework (Blueprints).

Gates and evidence

All gates and core evidence of the parent level (L1 — Function) apply unchanged. Declaring this sub-level adds the distinct check:

  • New test file present in the diff, authored by the agent, and all tests pass

A run that fails the sub-level check may still be reported at the plain parent-level grain if the parent's evidence holds — sub-level claims are strictly additive.

Prompt pack

Execution rubric (repo path, nav-hidden): prompts/autonomy/sublevel-l1-2.md — workcell-skeleton prompt declaring this sub-level, its evidence-to-return shape, and human gates.

Worked case

Field Value
Repo forge-dark-factory PoC — lenses-production campaign; granite-harness probes under runs/granite-harness/
Goal Fix broken relative markdown link ./no-such-page.md in README.md on a docs-health fixture
Evidence produced runs/campaigns/lenses-production/L1-a-broken-link/run-20260702T093821-599c26/{machine,human}/ (passing run); granite-harness iterations at runs/granite-harness/readme-link-green/ and run-20260702T122800Z-readme/
Assay pass rationale Core L1 passes (machine/assay.json ok: true). Retro-tagged L1.2b: run predates sub-level recording; _check_sublevel for L1.2 requires an agent-authored test file in changed_files — this run changed only README.md, so a strict L1.2 declaration would fail with L1.2 evidence: no agent-authored test file in changed files