0.2.7 Test Policy¶
The test level follows the risk of the change. A green lower level does not substitute for a required higher level.
| Level | Use | Typical commands/evidence | When required |
|---|---|---|---|
| T0 | Smoke and import health | package import, CLI help/version, minimal maintained example | every work session or handoff |
| T1 | Targeted changed-code tests | focused unit/integration tests for modified modules and contracts | every code or harness patch |
| T2 | Focused checkpoint | cross-family, cross-route or gate-specific replay with manifests | coherent lot end, cross-cutting change, before promotion |
| T3 | Full regression and coverage | full pytest, coverage, package/docs/release checks | important WP close, major integration, WP14 final sweep |
Execution rules¶
- During ordinary WP work, run T0 and T1. Do not run global coverage or full pytest after every small patch.
- Run T2 when a change crosses registry, descriptors, V&V harness or several element-analysis routes.
- Run T3 only at a declared checkpoint. Record the exact command, environment, counts, skips and failures.
- A test that is
PLANNEDis not executable evidence. AREADYcase is not a qualified case until its result and acceptance policy are recorded. - External tools, slow campaigns and resource-scale probes may be skipped only with an explicit reason and a non-PASS classification.
- Any functional numerical source change requires targeted tests first and a risk-based decision on T2/T3. A documentation-only change does not trigger numerical regression by itself.
Level-Up setup policy¶
Installing the Level-Up portfolio is a governance-only operation. It runs
documentation, registry, anti-forgetting, Ruff and compile checks as
appropriate, but it does not execute a large-model benchmark, promote a
capability or run full regression. WP16 and WP18 retain their own execution
and evidence requirements; PLANNED, RESOURCE_LIMITED and BRONZE are not
successful solve verdicts.
Evidence minimum¶
Every T2/T3 result must record source SHA, dirty state, version, command, environment, configuration, policy identifiers, result counts and artifact digests. A failed or resource-limited run remains part of the evidence. No policy may be weakened because a higher level is expensive.
Initial foundation state¶
The foundation pack itself changes no numerical source and requires no full
regression. WP03 and WP04 establish their own T0/T1 records. The planned full
regression remains NOT_RUN until WP14 or an explicitly approved high-risk
checkpoint.