# CAIN42 — ATLAS Evidence Specification (Prompt 15, sections 7-8)

## EvidenceObject (`CAIN42/ATLAS-EVIDENCE/v1`)

evidence_id, subject_agent, organization, issuer, kind, issued_at, ttl, scope_action_classes,
consequence_ceiling, revocation_state, independence_root, depends_on (dependency dimensions), confidence,
attestation (`SIGNED_ASSERTION` or `CRYPTOGRAPHIC_PROOF`), payload_digest, signature.

Evidence is **typed**, not scored. `verify_evidence` checks digest, signature, trusted issuer, revocation and
freshness.

## Categories and capabilities

IDENTITY, AUTHORITY, EXECUTION_PATH, BEHAVIOR, CONTINUOUS, EFFECT, INDEPENDENT_ATTESTATION, FEDERATION, plus
RUNTIME, MODEL, MEMORY, SUPPLY_CHAIN, WORLD_STATE, SECURITY, OPERATOR, COMPLIANCE. Each maps to an assurance
capability token.

## Evidence independence

`EvidenceIndependenceGraph` groups evidence by **shared dependency** across the dimensions model, runtime, host,
database, operator, telemetry, signing_authority, upstream, world_state_source. `EvidenceCorrelationDetector`
reports correlated groups and flags high-risk correlation (same signing authority, operator, runtime, model, host,
upstream).

**Rule:** 2 correlated validators != 2 independent validators. An evidence-independence requirement is satisfied
only by distinct roots.

## Composition

`FederatedTrustCompositionEngine` accepts an evidence object only if it is signed, in scope, fresh, unrevoked and
from a trusted issuer; it contributes at most its capability token; the capability set maps to a level. It never
sums numeric scores (`sums_scores: false`).

## Status

`IMPLEMENTED`, `TESTED`. Selection-disclosure / zero-knowledge is a research direction only (no ZK implemented);
attestations here are **signed assertions**, not cryptographic proofs of hidden statements.
