ワンクリックでManusで任意のスキルを実行

verify-estimator

スター0

フォーク0

更新日2026年7月12日 22:08

Verify a statistical estimator against numerical oracles (textbook/analytic, established package, seeded simulation). Use before shipping any ICC estimator or when the maintainer asks to "verify" or "check correctness".

インストール

Codex または Claude でインストールこの Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。

Manusで実行

ソース

jmgirard

jmgirard/intraclass

GitHub リポジトリを開く Creator のリポジトリを見る

ダウンロード

Manusで実行

Instructions

Implement PRINCIPLES.md #1 (oracle-first). Correctness is established by numerical agreement, never by assertion. Run everything on Opus — never route this to Fable (PRINCIPLES.md #19).

For the named estimator:

Assemble the oracle set (need ≥2 independent types):
- (a) closed-form / textbook value (e.g. Shrout & Fleiss 1979, Brennan 2001),
- (b) an established package on a case it supports (psych::ICC, gtheory),
- (c) a seeded simulation with known population variance components. Pull registered values from project/REFERENCES.md; if a needed value is not registered with provenance, stop — do not invent one (PRINCIPLES.md #4).
Run the comparison and report agreement to an explicit tolerance, per oracle.
Record results in project/REFERENCES.md (which oracle, tolerance, pass/fail). If the estimator's oracle block was scaffolded as planned / not yet asserted (by new-estimator), and it now passes, flip its status to asserted (M<n>) naming the test file and mark its provenance script committed — do not leave a green oracle registered as "planned". A planned status that outlives the work it describes is a silent tracking lapse.
If no oracle can pin the result: do not escalate on your own. Surface the gap plainly, name the specific oracle that would settle it, and recommend the maintainer approve a Fable review (§6 checklist), then stop and wait (PRINCIPLES.md #18, #19). A more capable model can reason about correctness but cannot establish it — oracles remain the source of truth.

このリポジトリの他の Skills

同じリポジトリ

add-decision

jmgirard/intraclass

Record an architectural or statistical decision as an ADR. Use when a non-trivial modeling or API choice is made, or the maintainer says "record this decision".

2026-07-120

finish-task

jmgirard/intraclass

Close out the current task. Use when the maintainer says "finish the task", "wrap up", or "close this out". Runs checks and only marks done on green.

2026-07-120

new-estimator

jmgirard/intraclass

Scaffold a new ICC estimator with its estimand block, test skeleton, and oracle checklist. Use when adding a new ICC variant (e.g. consistency, one-way, multilevel).

2026-07-120

start-task

jmgirard/intraclass

Begin the next task on the board. Use when the maintainer says "start the next task", "begin work", or "let's do the next thing".

2026-07-120

status

jmgirard/intraclass

Report current project state and the next action. Use when the maintainer asks "where are we", "what's next", or "status".

2026-07-120

name	verify-estimator
description	Verify a statistical estimator against numerical oracles (textbook/analytic, established package, seeded simulation). Use before shipping any ICC estimator or when the maintainer asks to "verify" or "check correctness".
allowed-tools	Read, Grep, Glob, Bash

Instructions

Implement PRINCIPLES.md #1 (oracle-first). Correctness is established by numerical agreement, never by assertion. Run everything on Opus — never route this to Fable (PRINCIPLES.md #19).

For the named estimator:

Assemble the oracle set (need ≥2 independent types):
- (a) closed-form / textbook value (e.g. Shrout & Fleiss 1979, Brennan 2001),
- (b) an established package on a case it supports (psych::ICC, gtheory),
- (c) a seeded simulation with known population variance components. Pull registered values from project/REFERENCES.md; if a needed value is not registered with provenance, stop — do not invent one (PRINCIPLES.md #4).
Run the comparison and report agreement to an explicit tolerance, per oracle.
Record results in project/REFERENCES.md (which oracle, tolerance, pass/fail). If the estimator's oracle block was scaffolded as planned / not yet asserted (by new-estimator), and it now passes, flip its status to asserted (M<n>) naming the test file and mark its provenance script committed — do not leave a green oracle registered as "planned". A planned status that outlives the work it describes is a silent tracking lapse.
If no oracle can pin the result: do not escalate on your own. Surface the gap plainly, name the specific oracle that would settle it, and recommend the maintainer approve a Fable review (§6 checklist), then stop and wait (PRINCIPLES.md #18, #19). A more capable model can reason about correctness but cannot establish it — oracles remain the source of truth.