Repository navigation
J4: adherence and task-impact evidence for instructions (AC-J04-005, AC-J04-006) #96
Description
Parent: #95
J4 keeps four claims about instructions separate: correctness, reach, adherence and effect. Correctness (AC-J04-001) and reach (AC-J04-003, AC-J04-004) are enforced. The other two are open:
-
AC-J04-005 (proposed): adherence. No command measures it yet. We first need a design for tasks that exercise specific instructions, and for how adherence is reported apart from reach.
-
AC-J04-006 (partial): task impact.
repo run --arm guidance:none|guidance:rawandrepo comparealready exist. Nothing asserts an effect claim yet: no test, and no recorded run.- Option 1: an integration test that uses the fake harness. Its fidelity is limited and must be stated.
- Option 2: recorded live-agent evidence stored outside git (
kind: live_agent).
Paid runs must not be used just to turn a status green.
Done when each criterion is either enforced, with its evidence kind and stubs stated, or still proposed with the reason recorded here.
Activity
feat(research): every open issue accounted for; untracked list for en…
Add a research register for ideas, experiments and decisions (#139)
napetrov commentedon Nov 1, 2026
Related, from #342: the receipt in #335 reports
followed=null. Load is not effect, so following evidence stays here. The #334 prefilter (did the situation come up, did the agent act) is an input to it, not a verdict.