switchboard-100.pages.dev
Human + agent control center for LLM provider budgets, priced model catalogs and routing policy. Agent-ready via WebMCP.
First seen 2026-11-05 · Last seen 2026-11-05 · Source: webmcp.com
Answer 11 · Act 3 · Transact 1 · Score: Not yet scored
Tool inspector
simulate_policy
Project what a candidate routing policy would cost and deliver, WITHOUT applying it. Classes you omit keep their current routing. Definitions: deliveredRatePct is the chain-level probability that some link succeeds, i.e. 1 minus the product of every link failing — this, not a single model's requestSuccessRate, is what a class's minSuccessRate is judged against. expectedLatencyMs is the probability-weighted median across the chain. worstCaseLatencyMs is the SUM of p95 across the WHOLE chain — what a request pays if every link but the last fails. It is reported unconditionally, so it never understates the tail; read tailRiskProbability alongside it, which is the chance of getting past the primary at all. HOW A LATENCY CEILING IS JUDGED, because the two rules interact: a fallback whose own p95 exceeds the ceiling is a WARNING, not a blocker, since a slow answer beats no answer — so the class stays permissible. But the chain's summed worst case is checked too and raises a CHAIN_LATENCY warning when it exceeds the ceiling. So a chain can be permissible and still carry a flagged tail. Say which you mean: "the primary meets your 4s ceiling; the chain can reach Xms in the Y% of cases where the primary fails" is the honest phrasing. Cost counts every attempt, since a failed call is still billed. The candidate's compliance verdict is returned here too and is the same computation check_compliance runs, so you do not need both — call check_compliance only when you want to audit the LIVE policy. One side effect: this repaints the projection on the operator's screen so they can follow your reasoning, so iterate deliberately rather than sweeping dozens of candidates.
Input schema
| Property | Type | Required | Description |
|---|---|---|---|
rules |
array | yes | Candidate routing rules. One entry per traffic class you want to change; classes you omit keep their current rule. |
label |
string | no | Short human-readable name for this candidate, shown on screen. |
Raw JSON schema
{
"type": "object",
"properties": {
"rules": {
"type": "array",
"description": "Candidate routing rules. One entry per traffic class you want to change; classes you omit keep their current rule.",
"items": {
"type": "object",
"properties": {
"classId": {
"type": "string",
"description": "Traffic class this rule routes."
},
"primaryModelId": {
"type": "string",
"description": "Model tried first, e.g. 'opencode-go/minimax-m3'."
},
"fallbackModelIds": {
"type": "array",
"items": {
"type": "string"
},
"description": "Models tried in order if the primary fails. Pass [] for no fallback."
}
},
"required": [
"classId",
"primaryModelId",
"fallbackModelIds"
]
}
},
"label": {
"type": "string",
"description": "Short human-readable name for this candidate, shown on screen."
}
},
"required": [
"rules"
]
}
Similar websites
reloadium.com
Paste a topic, URL, or PDF/TXT/MD — get a multi-episode AI podcast with audio, transcripts, timelines and quizzes. 60+ languages.
api.lmm.best
Connect AI models, tools and applications with LMM Forge. Explore model access, developer guides and open-source collaboration.
deskintab.dgkhan08.workers.dev
An operating system in your browser, where you and your AI agent share apps, files, and projects through WebMCP.
webmcp.agentsdesk.dev
Meridian Ops