Evidence model

aefi never returns a bare boolean as the product. The unit of intelligence is an envelope.

Envelope

Field	Meaning
`summary`	One-line narrative a human or agent can read first
`result`	Tool-specific payload (search page, payments, job, …)
`confidence`	`high` · `medium` · `low` · `unverified`
`confidence_reasons`	Controlled fact codes, not free-form marketing copy
`confidence_model_version`	Which composer produced the label
`evidence`	Typed refs (source, reference, claims)
`coverage`	`complete` / `substantial` / `partial` / `minimal` / `unknown` plus `known_gaps`
`graph_refs`	Node ids when the API read Neo4j

Graph objects (Wave A)

Projected into Neo4j when they are agent-relevant:

Agent — ERC-8004 identity and/or provisional wallet agent
Wallet — chain + address
Job — ERC-8183 commerce job
Payment — settled USDC (and 8183 PaymentReleased)
TransferEvent — system USDC transfer
MemoEvent — same-tx memo annotation
Outcome — completed / rejected / expired
Evidence — supporting node with typed claims

Edges include CONTROLS, PROVIDER / REQUESTER / EVALUATOR, SETTLED_BY, FROM_WALLET / TO_WALLET, ANNOTATED_BY, FOR_JOB, HAS_OUTCOME, SUPPORTS.

Example fact codes

payment_settled / payment_only_observed
exact_job_id_memo
erc_8183_job_lifecycle
escrow_release_after_acceptance
erc_8004_identity_match
erc_8004_reputation_event

Provider scoring (studio / search)

Search ranks with completion (scaled by sample size so one perfect job does not dominate), volume, optional recency, then a stable provider_id tie-break. Semantic recall (when query is set) is fused with that graph score.