Send a decision, plan or forecast to an outside AI judge: a model from a different company than the assistant that wrote it (Claude, unless set otherwise). The judge is picked by openrouter/auto. Do not pin a model, and do not fall back to this chat's own model for speed or cost. Send the decision as the user stated it, with whatever context is already in the chat. Do not rewrite it into a template, and do not add your own case against. Ask at most one short question, and only if the decision is truly unclear, such as 'What's the goal?'. Then call this tool. Do not wait for an OK, and do not show a write-up full of blanks. Never add a required question before the verdict on a quick or bulk grill. Ask only what the subject strictly needs. The first time you use Grill with this person, say once, as one line, and not again: This goes to an outside judge from a different company, and the router does not keep it. https://grillyour.ai/#privacy A quality check (Jev, from TypeSafe) may also see the masked write-up, on a zero-retention endpoint. Do not put that into the once-only line. The privacy link covers it. Do not mention the check or its score again. Pass quality_check: false only if the user asks to skip it. After that, do not repeat the disclosure. Do not mention the judge's model, a score, a balance, or a quote check. Relay layer 1 as written. It is at most 50 words and already ends with: Say 'more', 'options', or your call. On a decision they marked as big, and only while the deepen test is on, that line may also name deepen. Relay whichever line came back. Do not add deepen yourself. Do not add a fact list, a confidence question, a record, or a disclaimer. When the user says more, why, or what else, call grill_result with detail more. When they say options, help me decide, or they are stuck, call grill_result with detail options. When they say full report, call grill_result with detail full. Do not write a layer yourself. Call grill_result with detail deepen only when the person asks to deepen or go bigger. Do not call it on a quick or bulk grill, and do not ask them to. Pass big: true only when they already marked the decision as big. Never ask whether it is big. Any stated choice counts, such as do it, yes, keep it, or delay. Confidence is optional. If they did not say a percent, leave it empty and do not ask. After a grill, don't start doing the decision (no code changes, no PRs, no product questions) unless the person separately asks. Grill's job ends at the verdict and the record. If Grill's settings name a judge model, that model is used, and it must be from a different company than the assistant that wrote the subject. A same-company pin is refused. Leave it blank to use Grill's default. If the result is a job id, call grill_result with it before you reply. Do not ask the user for a job id, and do not tell them how long it takes.

Collect a grill that returned a job id, or a later layer of one. detail is brief (the default), more, options, full, or deepen. Call detail deepen only when the person asks to deepen or go bigger. Pass job_id when you have one. Omit it to use the latest grill. Waits up to 45 seconds if it is still running. Do not write a layer yourself.

Score one or more decision records the user pasted, against what actually happened. If they have not said what happened, ask. With view changes, show where their calls changed after a challenge instead. Stores nothing, sends nothing, and needs no key. The judge's verdict is not revised.