nc://gonogo

Go/No-Go: turns a pilot eval score into a ship-or-not verdict. A Nodecode rewrite of keppy/hermes-plugin-gonogo by keppy.

Get it

Tell the agent "add the gonogo cell", or run nodecode add nc://gonogo.

Set it up

Nothing to answer. The cell reads one optional member of its own section in ~/.nodecode/config.jsonc:

"gonogo": { "default_target_percent": 95 }

That is the accuracy a call that names no target is measured against, in percent; a call may always pass its own. No key, no network, no dependency: the statistics run in Lisp, in the session.

Use it

A pass rate is not a decision. 47 of 50 reads as 94%, but the 95% interval on that is [83.8%, 97.9%] — so against a 90% bar you cannot yet say you cleared it, and against 95% the honest answer is that 50 cases cannot settle it. The cell answers with the interval, the verdict, and the sample the claim would need:

(gonogo:decide :passed 47 :total 50)
;; => (:VERDICT "ASSIST ONLY" :REASON "pass rate 94.0% [83.8%, 97.9%] is below the 95% target ..."
;;     :PASS-RATE (:POINT 0.94d0 :LOW 0.838d0 :HIGH 0.979d0 :TEXT "94.0% [83.8%, 97.9%]")
;;     :NEEDED-N NIL :OPERATING-POINT NIL :MARKDOWN "**ASSIST ONLY** ...")

(gonogo:decide :passed 47 :total 50 :target 0.9)
;; => :VERDICT "INSUFFICIENT EVIDENCE" :NEEDED-N 204

Three verbs, and (help :gonogo) has their shapes:

The five verdicts

INSUFFICIENT EVIDENCE is the reason the cell exists: it is the verdict a careful consultant gives and a dashboard never does.

The results file

One JSON object per line; only passed is required, so a domain expert can build one in a spreadsheet.

{"id": "inv-001", "passed": true,  "confidence": 0.94, "expected": "Acme", "output": "Acme"}
{"id": "inv-006", "passed": false, "confidence": 0.42, "expected": "Acme", "output": "Cinder Works"}
{"id": "inv-007", "passed": false, "error": "TimeoutError: 30s"}
{"id": "msg-014", "passed": true,  "expected_action": "no_reply", "action": "no_reply"}

• confidence (0–1) is optional for a whole file, but mixed missing and present values are refused; it is what unlocks the review threshold, since a threshold can only be chosen from a per-case score.
• error is a crashed case — a failure, not an aborted run. abstained is an explicit abstention, which lands at zero confidence so a threshold defers it first.
• expected_action (absent means answer) and action grade decisions not to act; a hold-back case passes exactly when they match.
• group on every case folds correlated checks into one trial per group, passing only if every case in it passed — several fields from one document are one draw of the system under test.
• (gonogo:report ... :save-report "report.json") writes the report JSON, and gonogo:compare pairs two of those by case id (or by group id with matching case membership), so a run measured today can be checked against one measured last week.

What it touches

The paths you name: a results file read, a report JSON or HTML page written when you ask for one. No key, no network, no environment variable, no other cell's state.

/gonogo prints the target in force and the three verbs; (help :gonogo) answers the manual.

Credits

The statistics and the verdict rules are gonogo's, by the same author; this folder carries them in Common Lisp. See NOTICE. MIT — see LICENSE.