# CAIN42 — ATLAS Assurance Specification (Prompt 15, section 5)

## Assurance ladder (CAIN-42 proposed; NOT an industry standard)

Each level requires the cumulative machine-verifiable capabilities in parentheses. A level is reached only when
every requirement is satisfied by **verified** evidence.

| Level | Meaning | Requires |
|---|---|---|
| A0_UNVERIFIED | nothing is verified | — |
| A1_IDENTITY_VERIFIED | identity established | IDENTITY |
| A2_AUTHORITY_VERIFIED | authority established | + AUTHORITY |
| A3_EXECUTION_PATH_VERIFIED | the execution path was verified | + EXECUTION_PATH |
| A4_BEHAVIORALLY_ASSURED | behavior within envelope | + BEHAVIOR |
| A5_CONTINUOUSLY_ASSURED | continuously reauthorized | + CONTINUOUS |
| A6_EFFECT_VERIFIED | durable effect verified | + EFFECT |
| A7_INDEPENDENTLY_ATTESTED | independent attestation | + INDEPENDENT_ATTESTATION |
| A8_FEDERATED_TRUST_VERIFIED | cross-organization federation proven | + FEDERATION |

Derivation is deterministic (`derive_level`). A revoked subject is A0. An independence requirement is satisfied
only by **distinct evidence roots**, never by evidence count.

## Transaction-scoped assurance

Trust is evaluated per transaction, not as a global score:

```
AGENT + AUTHORITY + ACTION + TARGET + CONTEXT + CONSEQUENCE + EVIDENCE + TIME -> TransactionAssuranceObject
```

The same agent may be HIGH for "read public documentation", LOW for "transfer $100,000", and UNAUTHORIZED for
"deploy production infrastructure". `TransactionScopedAssurance.required_level` maps (action class, consequence)
to a minimum level; `evaluate` returns AUTHORIZE / RESTRICT / DENY with machine-readable problems.

## Authority ceiling cap

An authority ceiling caps the assurance level an agent may claim (P107): LOW→A2, MEDIUM→A4, HIGH→A6, CRITICAL→A8.
Assurance can never exceed the authority that underlies it.

## Status

`IMPLEMENTED`, `TESTED` for the ladder and transaction engine; the levels are a **proposal**.
