0040 — Evidence-bounded status for runs, features and research claims

Status: proposed · Date: 2026-09-21 · Cites: sub-doctrine 10.f · Related: ADR-0039, ADR-0020 · Evidence: the cutoff-bounded Qwen storm extraction and its event logs

Owes: the conduct rule is sub-doctrine 10.f. This ADR records why status claims must be bounded by evidence; it does not replace the canon's rule.

Context

Autonomous software systems emit several signals that look like completion but are not equivalent: a model verdict, a runner marker, a terminal event, a clean test, a local commit, a pushed branch and an accepted pull request. They can arrive in different orders, disagree, or be absent because the process is still alive. A generated file can also look like an implementation while not being tracked by the repository or connected to its tests.

The Qwen storm record demonstrated each distinction. Some runs emitted a provisional verdict without the completion marker. One exhausted its turn limit and failed without a verdict. A fresh run directory existed at the cutoff with no events, while a storm process was still alive. The extraction therefore kept those states separate instead of converting partial activity into success.

Decision

All operational and research status is evidence-bounded. A status claim names:

  1. the object being claimed (item, run, repository, branch or pull request);
  2. the evidence source (tracked code, test result, remote response or event stream); and
  3. the scope and observation cutoff.

The status vocabulary keeps at least these distinctions:

unknown       evidence is absent, stale or contradictory
active        work has not reached a terminal event
blocked       a concrete condition prevents the next safe step
failed        a terminal event records unsuccessful work
verified      the stated checks passed for the stated revision
published     the stated remote head and pull request are evidenced

No weaker signal promotes a stronger one: a verdict does not promote a run to completed, a marker does not promote a feature to delivered, a local commit does not promote a branch to published, and a generated artifact does not promote a feature to implemented. Missing evidence is reported as missing, never inferred from silence or activity. When sources disagree, the narrowest supported status wins until a fresh check resolves the disagreement.

Consequences

Evidence extraction must be cutoff-bounded and reproducible. Live monitors must not call a still-running process terminal, and papers must distinguish tracked compact evidence from ignored raw logs. Completion gates become stricter, but the resulting failure is actionable: the next CDD iteration knows whether it needs a test, a source change, a clean-up, a fresh remote query or human input.