Story-2498 · functions lane

A model score cannot overrule document bytes

Mental model: the panel is an untrusted witness. The eval replays what it proposed, what Sol accepted, what the scheduler covered, and what the provider charged. Deterministic code decides whether any of that evidence is admissible.

The hard-zero wall

captured AgentCore result
      ↓ strict RecordedRun contract
deterministic envelope assembler
      ↓
schema · V-CITE · coverage · material-finding · normalization
      ↓
clean result == poisoned-twin result
      ↓
quality receipt

Gate	What fails
V-CITE 100%	One quote differs from canonical UTF-8 bytes.
Coverage	No real scheduler batch receipt or an incomplete page set.
Collapse, never drop	A known material finding is absent from the admitted envelope.
Assertions	A labeled date is not ISO or a labeled amount is not numeric.
Injection refusal	Embedded instructions change grades, coverage, or findings against the clean twin.

Value is exclusive and cost is actual

Candidate IDs join each admitted finding back to its proposer. The scorer groups semantically identical findings across the matter-16 runs. A proposer receives incremental credit only when no other evaluated proposer supplied that finding. Shared findings and Sol additions receive no proposer credit.

Costs come only from provider-actual cost_micros receipts with unique generation IDs. Missing receipts remain visibly NOT MEASURED; static token-price estimates cannot pass the measured gate.

Model placement stays intact

The panel still runs only in the agents repository's AgentCore runtime. This functions-side harness consumes captured requests, effects, scheduler truth, and usage events. Muse Spark and Kimi K3 are report rows for eval-only challenger runs; they never enter the production lineup.