Recovered architecture-audit finding 11.2 (docs/plans/2026-08-07-audit-findings-recovery-and-open-work-execution-plan.md §7).
docs/guarantee-levels.toml is the source of truth for CALM's behavioral guarantees, and scripts/gen-status.sh --check already fails CI on format/presence drift (a stale entry vs the rendered docs/status.generated.md). But there is no contract test per guarantee ID — nothing asserts that the behavior a given ID describes still holds in code.
Example: guarantee::txn.begin_before_write (level enforced) claims a write is refused if the durable transaction journal can't start. Today only a human reading the TOML links that sentence to txn.rs; a refactor that silently weakened it would pass CI.
Scope
A small framework mapping each enforced guarantee ID to a test that exercises the specific behavior it promises, so a semantic regression fails the build the same way a format drift does. Start with the enforced-level entries (the ones users are told to rely on).
Meta-tooling / new test framework — related to the machine-readable-limitations idea (11.1b) and the durable-record theme in this plan's §3.
Recovered architecture-audit finding 11.2 (
docs/plans/2026-08-07-audit-findings-recovery-and-open-work-execution-plan.md§7).docs/guarantee-levels.tomlis the source of truth for CALM's behavioral guarantees, andscripts/gen-status.sh --checkalready fails CI on format/presence drift (a stale entry vs the rendereddocs/status.generated.md). But there is no contract test per guarantee ID — nothing asserts that the behavior a given ID describes still holds in code.Example:
guarantee::txn.begin_before_write(levelenforced) claims a write is refused if the durable transaction journal can't start. Today only a human reading the TOML links that sentence totxn.rs; a refactor that silently weakened it would pass CI.Scope
A small framework mapping each
enforcedguarantee ID to a test that exercises the specific behavior it promises, so a semantic regression fails the build the same way a format drift does. Start with theenforced-level entries (the ones users are told to rely on).Meta-tooling / new test framework — related to the machine-readable-limitations idea (11.1b) and the durable-record theme in this plan's §3.