| name | bench-oi |
| description | Use when an Agent needs to run a structured Oi conformance suite against an explicitly identified harness and model. |
Bench Oi
Use the colocated oi.mod and main.oi through the local using-oi bootstrap. Load exact Oi 0.0.2 execution.md once, close the manifested graph, and never reread the base specification.
Bind inputs to func main(suite bench.Suite, identity bench.RunIdentity). Map CompileTarget(target,input,observe,forbid) to tool.compile: sealed compile, target code zero, evidence absent. Map unchanged InvokeTarget(target,input,observe,forbid) to agent.invoke: sealed fresh run, actual receipt/digest. Both set LanguageVersion from target's nearest exact oi.mod; CompileTarget cross-binds target/version/source only. Before InvokeTarget returns, reject the first mismatch in Target→LanguageVersion→Artifacts(manifest/source order)→Inputs→Mappings→Authorities→AuthoritiesIdentity→Policy→Entry→Terminal→ReplyOperationID→Reply.Digest as BENCH_EVIDENCE_MISMATCH|runtime|target:1:1|sealed invocation mismatch: <FieldPath>|[]. Map Reply to caller.reply.
Host-invoke main once and expose its one typed BenchmarkReport Reply unchanged. Do not Oi-call main, infer a result, encode mode in text or fixture names, reuse one host effect then discard evidence, accept expected prose as evidence, or copy benchmark/runtime algorithms. Supports 0.0.1/0.0.2.
