ONE SENTENCE
A bare task instruction.
REAL SKILL / VISITOR-OWNED EXECUTION / CONTROLLED COMPARISON
Use PARS. Then test what it adds.
PARS runs inside your own ChatGPT/Codex task, not inside this webpage. Use Rooke Poole’s pinned Skill directly, or inspect the validated Arkhē plugin that packages it with a reproducible comparison.
EXACT UPSTREAM SKILLYOUR CHATGPT / CODEXNO OWNER API KEYMECHANICAL RECEIPT
01 / USE PARS NOW
The shortest honest route is the real Skill.
Start a new Codex task, give it the pinned installation instruction, then invoke $apply-pars-deep on actual work. The hash below identifies the exact upstream SKILL.md audited here.
PASTE INTO A CODEX TASK
Install the root Codex Skill from https://github.com/rookepoole/PARS/tree/57ac9970fa324c7c2a6f613fe32f5cc0577d3c3e as apply-pars-deep. Confirm the source commit and the SHA-256 of SKILL.md before using it: 424bf5359795b899126f5ce0c8bf6cc99fb2675b800f1ab915b410dce240b602.THEN, ON REAL WORK
Use $apply-pars-deep on this task. Apply the Skill to the work itself; do not merely explain PARS.AUTHORITATIVE SOURCEInspect Rooke’s pinned repositoryCOMMIT · 57ac9970fa32 · MITARKHĒ PLUGIN / 0.1.0Download the inspectable pluginSHA-256 · 7366b9b9…c044f03
PUBLIC DIRECTORY STATUS
The skills-only plugin is structurally validated, installed and enabled in a real local Codex marketplace. Universal-directory publication still requires OpenAI’s verified-publisher submission and review, so this page does not pretend the download is a one-click public install.
02 / TEST THE CONTRIBUTION
Do not compare PARS to nothing.
The packaged $run-pars-comparison Skill runs the same frozen packet across four fresh, isolated Codex contexts. Only arm D receives PARS. Order is randomized; scoring is deterministic and does not ask the producing model to grade itself.
CHECKLIST
A short ordinary verification checklist.
VERIFIER KERNEL
The strongest compact prompt found in the audit.
FULL PARS
The exact pinned upstream Skill.
- 01Fresh contexts
No arm sees another arm’s output.
- 02Same frozen input
Task packet and hashes remain fixed.
- 03Mechanical scoring
Schema, answers and false support are checked in code.
- 04Local receipt
Results stay on the visitor’s machine unless they share them.
LIVE ROUTE CANARY / 16 AUG 2026
The visitor-owned path ran for real.
With provider API-key variables removed, a fresh Codex arm ran through ChatGPT authentication and produced a scored receipt. The bare one-sentence arm scored 10/11, missing one claim-status item. This proves the route works—not that PARS wins.
INSPECT CANARY RECEIPT →03 / KEEP THE BOUNDARY CLEAR
Three different things, three different proofs.
A working install proves packaging. A successful run proves execution. Only matched, repeated comparisons can estimate whether PARS contributes reasoning value beyond simpler prompts.
No model inference and no result upload.
Uses your account and your local task context.
Fresh arms, frozen inputs and a non-model scorer.
FULL REPRODUCTION
Need packages, hashes and protocol detail?
The research page preserves the larger cross-platform alpha harness. It is optional infrastructure, not the main finding and not required merely to use PARS.
OPEN RESEARCH REPRODUCTION →