You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Stage A of the go-live playbook: prove the whole product works, on a laptop, against a private test server. Nothing here needs the cluster — the bot dials out to Discord, so a laptop runs the real thing.
Roughly an afternoon. The playbook has the commands and the "how you know it worked" for every step.
Blocked on #10 — the keys and the Discord application are org-owned and created there.
This is the first gate: nothing goes to the real server until this is clean.
A5 The acceptance checklist: mention, ask channel, /ask, forum auto-reply, feedback buttons, rate limit, long answers, /issue registering, the apology path
A6 Calibrate Answering:MinScore (P-07) and commit the new default
A7 Generate the real Eval/baseline.json (optional, cheap)
A6 is the one that matters most. The refusal threshold ships as a guess: too low and Prompter answers questions the docs do not cover, too high and it refuses ones they do. It cannot be calibrated without keys, and it is the difference between a bot people trust and one they stop asking. It is also the single most likely thing to bite.
Note on A5: none of these surfaces has ever touched a live Discord gateway — they are spec-verified, not observed. Expect to find something here; that is what this stage is for. If a slash command does not appear at all, give it time before debugging — the first global registration can take up to an hour, which is why the test server (instant, guild-scoped) is the right place to iterate.
Stage A of the go-live playbook: prove the whole product works, on a laptop, against a private test server. Nothing here needs the cluster — the bot dials out to Discord, so a laptop runs the real thing.
Roughly an afternoon. The playbook has the commands and the "how you know it worked" for every step.
Blocked on #10 — the keys and the Discord application are org-owned and created there.
This is the first gate: nothing goes to the real server until this is clean.
dotnet run -- index) — and a second run reporting 0 embeddedask --verbose), plus an out-of-scope question that refuses/ask, forum auto-reply, feedback buttons, rate limit, long answers,/issueregistering, the apology pathAnswering:MinScore(P-07) and commit the new defaultEval/baseline.json(optional, cheap)A6 is the one that matters most. The refusal threshold ships as a guess: too low and Prompter answers questions the docs do not cover, too high and it refuses ones they do. It cannot be calibrated without keys, and it is the difference between a bot people trust and one they stop asking. It is also the single most likely thing to bite.
Note on A5: none of these surfaces has ever touched a live Discord gateway — they are spec-verified, not observed. Expect to find something here; that is what this stage is for. If a slash command does not appear at all, give it time before debugging — the first global registration can take up to an hour, which is why the test server (instant, guild-scoped) is the right place to iterate.
Next: #7.