ToolBench toolbench v0.1.0
Run synthetic door cases to see how FragGate classifies them.
#toolbench #toolbench-plain #toolbench-suite #toolbench-run-case #toolbench-honesty
In-runtime engine (no separate product Worker). No download-tracker URL was invented. · runtime OpenAPI
Liveness. Does not increment download KV. The machine name for that step is health.
Richer liveness: synthetic suite, refuse list. Does not increment download KV. The machine name for that step is doctor.
List synthetic FragGate refuse / happy cases. Self-test ≠ third-party lab. The machine name for that step is suite.
Classify one named synthetic case against the live door table. Does not invent a pass. The machine name for that step is run_case.
Honesty labels: REAL door classification; SLOT for invent_completeness. The machine name for that step is limitation.
Return ToolBench skill markdown. Does not increment download KV. The machine name for that step is skill.
Machine call
POST /v1/fraggate/call { slug: "toolbench", op: "health" } — Liveness. Does not increment download KV.POST /v1/fraggate/call { slug: "toolbench", op: "doctor" } — Richer liveness: synthetic suite, refuse list. Does not increment download KV.POST /v1/fraggate/call { slug: "toolbench", op: "suite" } — List synthetic FragGate refuse / happy cases. Self-test ≠ third-party lab.POST /v1/fraggate/call { slug: "toolbench", op: "run_case" } — Classify one named synthetic case against the live door table. Does not invent a pass.POST /v1/fraggate/call { slug: "toolbench", op: "limitation" } — Honesty labels: REAL door classification; SLOT for invent_completeness.POST /v1/fraggate/call { slug: "toolbench", op: "skill" } — Return ToolBench skill markdown. Does not increment download KV.
curl -s -A 'Mozilla/5.0' -X POST https://godlock.uk/runtime/v1/fraggate/call \
-H 'content-type: application/json' \
-d '{"slug":"toolbench","op":"run_case","payload":{"id":"fg-halluc"}}'
FragGate only — same LIVE_OPS as MCP fraggate_call / POST /v1/fraggate/call. One backend, two surfaces.
Disabled verbs refuse FG-STUB. They stay named and are not a live public door.
Ready. ToolBench. The reply stays in plain language.