# CAIN-42 — Master Formal Assurance (E20)

Formal-language is not decoration. Every invariant below is a **runtime check**, executed and
recorded, and classified honestly. None is a machine proof.

| Invariant | Meaning | Classification |
|---|---|---|
| I13 | temporal authorization cannot outlive its validity interval | TESTED (runtime check) |
| I14 | revoked capability cannot execute | TESTED |
| I15 | delegation cannot exceed delegated scope | TESTED |
| I16 | child authority cannot exceed parent authority | TESTED |
| I17 | simulation cannot directly create authority | TESTED |
| I18 | prediction cannot become observed fact | TESTED |
| I19 | uncertain evidence cannot silently become certain | TESTED |
| I20 | federation cannot silently transfer authority | TESTED |
| I21 | rollback cannot bypass authorization | TESTED |
| I22 | policy change cannot silently bypass release governance | TESTED |
| I23 | cross-tenant authority cannot cross tenant boundaries | TESTED |
| I24 | historical replay cannot alter authoritative history | TESTED |
| I25 | reputation cannot create authority | TESTED |
| I26 | adaptive learning cannot directly modify protected authority | TESTED |

Source: `cain/e20/formal/invariants.py` · Evidence: `artifacts/maximum-evolution/FORMAL_RESULTS.json`
(TESTED 14/14).

**Honest classification scale:** PROVEN > MODEL-CHECKED > TESTED > ANALYTICALLY-SUPPORTED >
UNVERIFIED. These invariants are **TESTED**. They must not be promoted to PROVEN without a
corresponding model/theorem artifact. No z3/Coq/TLA+ model of the E20 layer exists yet.
