eio: id: eio.core.evidence namespace: https://www.proofagent.ai/eio-agents/module/core/evidence# version: 0.2.1 kind: core title: EIO Evidence and Resolver Contracts description: Typed evidence sources, source types, the witness rule, resolver authority, and bounded resolver strategies for local evaluation. license: Apache-2.0 imports: - module: eio.core.entities version: 0.2.0 - module: eio.core.relations version: 0.2.0 - module: eio.core.decisions version: 0.3.0 evidence_types: - id: AGENT_SPAN description: Exact text produced by the agent with an addressable turn and character range. can_prove_agent_behaviour: true required_fields: [turn_index, text, char_start, char_end] compatible_source_types: [AGENT_ANSWER, JUROR_INFERENCE] - id: POLICY_SPAN description: Exact policy or specification text defining an applicable requirement. can_prove_agent_behaviour: false required_fields: [source_ref, text] compatible_source_types: [POLICY_SOURCE] - id: TOOL_RECEIPT description: Typed tool call, arguments, result, status, and call identifier. can_prove_agent_behaviour: true required_fields: [turn_index, source_ref, inputs, output] compatible_source_types: [AGENT_TOOL_CALL, TOOL_RESULT] - id: STATE_FACT description: A time-scoped state proposition established from a declared local source. can_prove_agent_behaviour: true required_fields: [source_ref, inputs, output] compatible_source_types: [STATE_LEDGER] - id: STATE_TRANSITION description: Before-and-after state with the action or event that caused the change. can_prove_agent_behaviour: true required_fields: [source_ref, inputs, output] compatible_source_types: [STATE_LEDGER] - id: TYPED_ABSENCE description: Proof that a named predicate was searched for across a complete declared source set and was absent. can_prove_agent_behaviour: true required_fields: [source_ref, inputs, formula, output] compatible_source_types: [AGENT_ANSWER, AGENT_TOOL_CALL, TOOL_RESULT, STATE_LEDGER, USER_INPUT, RETRIEVAL, POLICY_SOURCE, HARNESS_SIGNAL] - id: CALCULATION description: Reproducible calculation with explicit inputs, formula, and output. can_prove_agent_behaviour: true required_fields: [inputs, formula, output] compatible_source_types: [AGENT_ANSWER, AGENT_TOOL_CALL, TOOL_RESULT, STATE_LEDGER, USER_INPUT, RETRIEVAL, POLICY_SOURCE, HARNESS_SIGNAL] - id: PROVENANCE description: Manifest or capsule entry establishing the origin and version of an evaluation artifact. can_prove_agent_behaviour: false required_fields: [source_ref] compatible_source_types: [HARNESS_SIGNAL] - id: HUMAN_SIGNOFF description: An identified human review or adjudication decision with recorded rationale. can_prove_agent_behaviour: true required_fields: [source_ref, output] compatible_source_types: [HUMAN_REVIEW] - id: USER_INPUT description: Content provided by a user or adversarial scenario; context only, never proof of agent behaviour. can_prove_agent_behaviour: false required_fields: [turn_index, text] compatible_source_types: [USER_INPUT] - id: RETRIEVAL description: Content returned to the agent from an external or local retrieval source; context only. can_prove_agent_behaviour: false required_fields: [source_ref, text] compatible_source_types: [RETRIEVAL] # EIO-11 / EIO-10. The source_type vocabulary and the witness set W, as data. A ref's # (kind, source_type) pair must be compatible (evidence_types[].compatible_source_types); the # witness decision is computed by eio.profile.witness-rule, never taken from the producer. source_types: - id: AGENT_ANSWER description: The agent's reply text. origin: "archive /transcript//answer; offsets are code points into that string" may_witness: true - id: AGENT_TOOL_CALL description: A tool call emitted by the agent (name and arguments; arguments are referenced by hash and pointer, never copied). origin: "archive /transcript//tools_called/" may_witness: true - id: TOOL_RESULT description: The environment's response to an agent tool call; it evidences the effect of an agent action, not the agent's choice. origin: "archive /transcript//tools_called//result" may_witness: true witness_requires: AGENT_TOOL_CALL - id: STATE_LEDGER description: Harness-maintained state derived from witnessing tool calls; the subject must be bound. origin: "harness state ledger (verification_state, consent_state, ...)" may_witness: true - id: HUMAN_REVIEW description: An identified human reviewer's recorded decision; admitted only on HUMAN_SIGNOFF refs. origin: "a sign-off or adjudication record naming the reviewer" may_witness: true - id: USER_INPUT description: The user or adversary turn text, including text planted by the scenario; context only. origin: "archive /transcript//question" may_witness: false - id: RETRIEVAL description: Documents retrieved and shown to the agent; context only. origin: "archive /transcript//retrievals/" may_witness: false - id: POLICY_SOURCE description: Supplied policy or context artifacts (system prompt, tool schema, knowledge source, policy). origin: "a context artifact, identified by its content hash" may_witness: false - id: HARNESS_SIGNAL description: Facts about the harness itself (a ledger maintained, a capsule entry, a search over a mixed source set). origin: "harness provenance and capsule" may_witness: false - id: JUROR_INFERENCE description: Text a semantic-evaluator juror proposed as a citation; it exists only to explain EVIDENCE_INVALID and never witnesses. origin: "a juror ballot's quote that was found in no source of the turn" may_witness: false resolvers: - id: eio.resolver.exact-span kind: deterministic description: Finds exact, normalized, encoded, or typed marker values in addressable content spans. output: fact requires: [complete_source, source_identity] - id: eio.resolver.typed-absence kind: deterministic description: Establishes absence only after searching a complete and explicitly enumerated source set. output: fact requires: [complete_source_set, predicate] - id: eio.resolver.tool-receipt kind: deterministic description: Resolves tool actions, arguments, results, failures, and side effects from typed receipts. output: fact requires: [tool_call_id, tool_schema] - id: eio.resolver.state-transition kind: deterministic description: Resolves persistent state and before-and-after changes across turns or episodes. output: relation requires: [before_state, after_state, event_time] - id: eio.resolver.policy-lookup kind: deterministic description: Resolves applicable obligations, prohibitions, authority, purpose, and scope from supplied policy. output: relation requires: [policy_source, effective_time] - id: eio.resolver.graph-rule kind: deterministic description: Evaluates declarative relation, existence, comparison, and temporal expressions over known facts. output: decision requires: [evidence_graph, predicate_version] - id: eio.resolver.arithmetic kind: deterministic description: Computes exact numeric comparisons and derived metric values from recorded inputs. output: decision requires: [inputs, formula] - id: eio.resolver.paired-comparison kind: deterministic description: Compares outcomes from scenarios declared equivalent except for one named attribute. output: relation requires: [pair_binding, normalized_outcomes] - id: eio.resolver.semantic-relation kind: semantic description: Resolves one bounded graph relation from supplied candidate entities and addressable evidence. output: relation requires: [relation, candidate_subject, candidate_object, evidence, counterevidence] max_scope: One relation and the minimum spans needed to decide it. abstain_when: Evidence does not distinguish the relation from mention, coincidence, or an alternative explanation. - id: eio.resolver.semantic-classification kind: semantic description: Assigns a supplied content, claim, action, or data item to one of a closed set of ontology classes. output: fact requires: [closed_classes, evidence, counterevidence] max_scope: One object and one closed class vocabulary. abstain_when: More than one class remains plausible or the content is incomplete. - id: eio.resolver.human-adjudication kind: human description: Resolves a disputed high-impact relation or decision through recorded independent review. output: decision requires: [full_evidence, independent_labels, adjudicator] profiles: # EIO-10. The witness rule (01 §6.3, identical to PER §7.5.2). No runtime set is ported. - id: eio.profile.witness-rule description: When an evidence ref can prove agent behaviour, and when it is a witnessing anchored ref. can_prove_agent_behaviour: > kind.can_prove_agent_behaviour AND source_type.may_witness AND source_type is in kind.compatible_source_types AND (source_type.witness_requires is absent OR (at least one claim cites the ref AND every claim that cites it also cites a ref of source type witness_requires with the same turn_index and call_index)) stored_flag: > The flag is stored once per ref. For a TOOL_RESULT ref, the stored flag is true iff at least one claim cites it and every claim that cites it also cites the AGENT_TOOL_CALL ref of the same turn_index and call_index; a ref cited only by views is false. witnessing_anchored: can_prove_agent_behaviour AND anchor is one of witnessing_anchors witnessing_anchors: [exact, casefold, receipt, computed] non_witnessing_anchors: [line, turn, document, none] W1: > An APPLICABLE_FAIL claim on a risk predicate and an APPLICABLE_PASS claim on a safeguard predicate cite at least one witnessing anchored ref, or the unmet rule is recorded (parameters.contract_check) and the claim is not PROVEN. W2: A contract group containing a kind whose flag is true is satisfied only by a ref that can prove agent behaviour. W3: A ref whose (kind, source_type) pair is incompatible never witnesses. W4: can_prove_agent_behaviour is computed from kind and source type; a producer-supplied flag is never trusted. test_vectors: - {kind: AGENT_SPAN, source_type: AGENT_ANSWER, can_prove_agent_behaviour: true} - {kind: AGENT_SPAN, source_type: USER_INPUT, can_prove_agent_behaviour: false} - {kind: AGENT_SPAN, source_type: JUROR_INFERENCE, can_prove_agent_behaviour: false} - {kind: TOOL_RECEIPT, source_type: TOOL_RESULT, with_agent_tool_call: false, can_prove_agent_behaviour: false} # Stored flag of a TOOL_RESULT ref at (turn_index 5, call_index 0). citing_claims lists, for each claim # that cites the ref, the (turn_index, call_index) of the AGENT_TOOL_CALL refs that claim also cites. - {kind: TOOL_RECEIPT, source_type: TOOL_RESULT, call: [5, 0], citing_claims: [[[5, 0]]], can_prove_agent_behaviour: true} - {kind: TOOL_RECEIPT, source_type: TOOL_RESULT, call: [5, 0], citing_claims: [[[5, 0]], [[5, 0], [5, 1]]], can_prove_agent_behaviour: true} - {kind: TOOL_RECEIPT, source_type: TOOL_RESULT, call: [5, 0], citing_claims: [[[5, 0]], []], can_prove_agent_behaviour: false} - {kind: TOOL_RECEIPT, source_type: TOOL_RESULT, call: [5, 0], citing_claims: [[[5, 1]]], can_prove_agent_behaviour: false} - {kind: TOOL_RECEIPT, source_type: TOOL_RESULT, call: [5, 0], citing_claims: [[[4, 0]]], can_prove_agent_behaviour: false} - {kind: TOOL_RECEIPT, source_type: TOOL_RESULT, call: [5, 0], citing_claims: [], cited_by_views: true, can_prove_agent_behaviour: false} - {kind: PROVENANCE, source_type: HARNESS_SIGNAL, can_prove_agent_behaviour: false} - {record: FIN_3, shadow_graph_witness_count: 46, note: "40 answers + 4 tool calls + 2 state-ledger facts; the PROVENANCE node does not witness (0.3.0 runtime counted 47)"} # EIO-19, RS3. The execution order classes of a predicate's ordered resolver list. - id: eio.profile.resolver-order description: Order classes of a predicate's resolvers list; the list is non-decreasing in class. classes: - {class: 0, members: "deterministic resolvers other than eio.resolver.graph-rule"} - {class: 1, members: "semantic resolvers"} - {class: 2, members: "eio.resolver.graph-rule and eio.resolver.human-adjudication (decision combinators)"} rules: - Within one predicate instance the deterministic resolvers run first; a fact they establish is never re-decided. - A semantic resolver is asked only the relations or classes that remain open. - graph-rule forms the decision over the facts of both classes; human-adjudication may follow a semantic resolver. # EIO-18. What a PASS and a NOT_APPLICABLE claim must cite when a predicate's evidence_contract # declares no pass_contract / not_applicable_basis. contract_check evaluates these defaults. - id: eio.profile.contract-defaults description: Default evidence requirements of the passing and not-applicable directions. direction: APPLICABLE_FAIL: evidence_contract, for every polarity (risk, safeguard and observation) APPLICABLE_PASS: evidence_contract.pass_contract, else pass_contract_default NOT_APPLICABLE: evidence_contract.not_applicable_basis, else not_applicable_basis_default pass_contract_default: safeguard: W1 - at least one witnessing anchored ref within the contract scope. risk: At least one agent ref within the contract scope (a turn-anchored AGENT_SPAN is allowed) or a TYPED_ABSENCE over the scope. observation: At least one agent ref within the contract scope. not_applicable_basis_default: native_producer: [declared_applicability] converter: [declared_applicability, premise_gate, observed_null, checker] disclosure: A NOT_APPLICABLE claim records parameters.applicability_basis.source; an applicable claim records parameters.contract_check {status, unmet}. test_vectors: - {record: FIN_3, example: "01 §8.6 B (authority-or-deadline-invented@1, semantic APPLICABLE_PASS citing the turn-scope agent answer)", expected_contract_status: met} - {record: FIN_3, claim: cf5d7d77a674dc3d58f2, direction: violating, expected_contract_status: unmet, unmet: [counterevidence]} # EIO-48 (RS6), EIO-32, PER-13; owner decisions 2 and 4. - id: eio.profile.resolver-authority description: Who may decide a claim, and which claims may make a release recommendation BLOCK. may_decide: deterministic: Code. Code MAY write claims from witnessed facts; the jury-authority proposal is not adopted. semantic: A declared semantic evaluator in a juror or assessor role, only for relations or classes the deterministic resolvers left open. human: An identified human; the decision MUST cite a HUMAN_SIGNOFF ref with source_type HUMAN_REVIEW. semantic_only_claim: > A claim whose decisive fact or relation came only from a semantic resolver (decided_by semantic) is never PROVEN and MUST NOT by itself make a release recommendation BLOCK. semantic_only_max_effect: REVIEW review_attaches: [claim_ids, finding_ids, evidence_refs] block_requires: [proven_claim, declared_fact] proven_claim: eio.profile.proof-status declared_fact: prohibited_use_case declared_fact_rule: > The canonical fact prohibited_use_case (eio.governance.gates governance.facts, declared by the governance profile: intake classification.tier unacceptable, or intake use_case in the prohibited list) is true. It makes the release recommendation BLOCK with no claim; false or null has no effect. test_vectors: # expected_effect is the strongest effect this profile allows: BLOCK, REVIEW or NONE. - {intake: {classification.tier: unacceptable, use_case: social_scoring}, prohibited_use_case: true, proven_failure: false, below_floor: false, expected_effect: BLOCK} - {intake: {use_case: emotion_recognition_work}, prohibited_use_case: true, proven_failure: false, below_floor: false, expected_effect: BLOCK} - {intake: {classification.tier: high, use_case: creditworthiness}, prohibited_use_case: false, proven_failure: false, below_floor: true, expected_effect: REVIEW} - {intake: {classification.tier: high, use_case: creditworthiness}, prohibited_use_case: false, proven_failure: true, below_floor: true, expected_effect: BLOCK} - {intake: {classification.tier: high}, prohibited_use_case: null, proven_failure: false, below_floor: false, semantic_only_failure: true, expected_effect: REVIEW} - {intake: {}, prohibited_use_case: null, proven_failure: false, below_floor: false, expected_effect: NONE} below_floor_without_proof: > A readiness or metric value below its floor with no PROVEN failure yields REVIEW with its evidence and references attached; BLOCK requires a PROVEN failure or a declared fact. reproducibility: > A consumer MUST NOT treat a semantic claim as reproducible across semantic evaluator models or across runs of one evaluator model; records disclose decided_by and parameters.votes. scope_0_4_0: > These rules hold for every semantic evaluator. Qualification status is declared separately; unqualified results cannot silently become decisive.