0006: QA answers three questions: tiers, expect, oracles
Context: Agents report "done" on green runs that ran the wrong subset, collected nothing, or produced plausible garbage.
Decision: QA.md answers:
| question | mechanism |
|---|---|
| did I run the right thing | tiers + selection table |
| did it actually run | expect per tier: floors, skip list, warnings max, duration range |
| does it make sense | oracles: reference, invariant, bounds with units, diff vs main, negative control |
Oracles carry status: agents add proposed, only humans set confirmed.
Explicit "not evidence" list. Every task ends with a QA report.
Consequences: "Done" becomes checkable; expect must move in the same commit as an intended
baseline change. Humans own domain truth.
→ reflected in architecture: QA.
Alternatives:
| option | why not |
|---|---|
| tests green = done | misses 0 collected, skips, wrong-but-passing output |
| exact counts in expect | breaks on every added test; floors catch the real failure |
| agents confirm oracles | agent grades its own homework |