compass-verify-assumptionslisted
Install: claude install-skill Anselmoo/werkstoff
# compass-verify-assumptions
Resolve **one** assumption in a bounded loop, then **validate with the guard**. The
step cap, the one-per-invocation rule, and the confidence gate are enforced in
code.
`GUARD="python3 ${CLAUDE_PLUGIN_ROOT}/scripts/compass.py"`
## Input
One uncertainty entry: `{element, default_interpretation, confidence,
other_readings, blocking}`.
## Rules
- **Resolve exactly one assumption per invocation.** For multiple uncertainties,
invoke this skill once each — **never share the 3-step budget across entries.**
- Run a **Reasoning / Action / Observation** loop, **at most 3 steps total** (hard
cap). Each action closes a named gap.
- **Stop as soon as any outcome is reached**, even if steps remain in the budget.
- Report **exactly one outcome**:
- `confidence_raised` — new confidence **must be >= 90**; **cite the source(s)**
(file:line or URL).
- `reading_changed` — the default interpretation was wrong; state the new one and
**cite the source(s)**.
- `still_unresolved` — use the refusal template from `compass-ground-evidence`
(`The available [sources] do not contain sufficient information to …`).
## Validate
```
echo '{
"assumption":"target env is prod",
"steps":[
{"reasoning":"need to know which env the config targets","action":"read config","observation":"env=production"}
],
"outcome":{"kind":"confidence_raised","confidence":95,"citations":["(config.yaml:3)"]}
}' | $GUARD verify -
```
The guard refuses: more than 3 st