# Exact-tree read-model counterexample probes

Use during final independent re-review of data-heavy dashboard/read-model trees.

## Fast sequence

1. Include untracked files; a base diff alone omits them.
2. Read production-path changes before accepting test claims.
3. Compare every new raw-SQL column with the migrated runtime schema (`PRAGMA table_info` for SQLite).
4. Exercise the minimum realistic domain row that enters each new SQL branch. Empty-database endpoint tests often never execute classified-account or linked-record queries.
5. Trace both selected period bounds end-to-end. Historical presets must use their requested period end, not the current request reference/as-of.
6. Probe absent-versus-zero semantics at provider, API, and UI boundaries. Empty containers and missing valuations are unavailable unless a zero is explicitly evidenced.
7. Test freshness thresholds exactly at their boundary using production's date-only/day-end convention; midnight-only tests can hide an off-by-one day.

## Deterministic probes

- Insert one classified performance account and one valuation, invoke the public read model, and catch schema/query drift such as selecting a derived property that is not a persisted column.
- Provide complete prior-year opening/closing valuations and cashflow coverage, set `as_of` in the following year, and require historical coverage to remain complete.
- Use a noncash container with no holdings or confirmed total and require omission/unavailable rather than `0.00`.

## Verdict shape

For verdict-only requests, output severity counts and actionable findings with file/line references. Avoid planning prose and non-actionable observations, but ground each finding in source tracing or an executed counterexample.