[eval] Bounded-investigation evaluation harness (E0-E3)

[Aparnap2]

Child of orchestration (#66). Evaluation milestone, NOT agent optimization.

Architecture

Investigation case (inputs + expected evidence + constraints + acceptable findings/recommendations + mandatory escalations) -> agent under test (FakeModelClient for control plane, real model later for quality) -> evaluator -> grounded/bounded/useful verdicts.

Layers

• E0 control gates (100% pass required): invalid/undeclared/cross-tenant/fabricated-evidence/hypothesis-without-falsifier/invalid-recommendation rejected; budget exhaustion escalates; loops terminate; no mutation path. Mostly exists in orchestrate tests — wire as gates.
• E1 evidence behavior: relevant selection, disconfirming-seeking, contradiction handling, not-found vs failure distinction.
• E2 reasoning (G9 dimensions): hypothesis/falsification quality, grounding, unresolved identification, finding/recommendation appropriateness, escalation.
• E3 operational: turns, calls, latency, budget use, repeats, recovery, completion/escalation rates.

Corpus

Synthetic controlled cases A-M (agreement, amount/policy conflict, missing evidence, contradictory externals, ambiguous, stale, no-evidence, timeout, malformed upstream, cross-tenant/fabricated-evidence/loop temptations). Each: inputs, expected evidence, permitted tools, contradictions, boundary, acceptable findings/recommendations, escalation conditions. Semantic equivalence, never exact-wording match.

Rules

• FakeModelClient for deterministic CI; real model as separate backend later (distinguish orchestrator-broken from model-reasoned-badly).
• No values/PII in eval artifacts. No production wiring. No [test-isolation] Outbox tests state-sensitive on shared DB #65/[dist] Pull loop always-ACKs transient failures #57-[dist] Claim lifecycle lacks durability #61 mixing.