ARKHĒPUBLIC RELEASE / 16 AUG 2026

OPTIONAL RESEARCH REPRODUCTION / VISITOR-OWNED INFERENCE

Reproduce the comparison.

This secondary harness compares the one-sentence baseline, checklist, short verifier and frozen PARS Skill using your own ChatGPT-authenticated Codex account. Start with the lightweight Try PARS page unless you specifically need hashes, scoring and receipts.

NO API KEYNO UPLOADLOCAL RECEIPTRESEARCH ALPHA
01 / WHAT HAPPENS

Optional infrastructure, not the main experience.

The runner refuses non-ChatGPT Codex authentication, removes common provider API-key variables, binds to a random loopback port and executes every selected condition once in a fresh directory and ephemeral thread. Arm order is randomized; scoring is deterministic.

  1. 01
    Verify locally

    The runner checks Codex version and ChatGPT login status without reading tokens.

  2. 02
    Select arms

    Run one condition as a smoke test or all four for the frozen comparison.

  3. 03
    Execute cleanly

    Fresh directories, ephemeral threads, identical flags and no automatic retry.

  4. 04
    Inspect the receipt

    Scores, tokens, latency, hashes, logs and failures remain on your computer.

02 / DOWNLOAD ALPHA

Choose your local runner.

This first build is unsigned and not notarized. Do not bypass an operating-system warning merely to test it; inspect the source or wait for the signed release. Every archive includes the protocol, frozen task packet, answer key, PARS Skill and local interface.

Verify the complete archive list before execution. A matching hash proves byte identity with this release—not safety or scientific validity.

SHA256SUMS.TXT ↗
03 / START

One local command.

Unzip the package, open Terminal or PowerShell in that folder and launch the included executable. The browser page is served only from your machine and closes when you stop the terminal process.

macOS / Linux./pars-lab
Windows PowerShell.\pars-lab.exe
USAGE BOUNDARY

The trivial non-API canary reported 23,617 aggregate Codex tokens. That confirms the route and exposes agent-runtime overhead; it is not a per-arm forecast or PARS cost measurement. The lab requires explicit acknowledgement before it starts any experiment arm.

04 / SCIENTIFIC STATUS

Useful replication input. Not crowd-voted truth.

Self-run receipts help people inspect PARS and find task-level failures. They do not enter the canonical audit automatically. Promotion requires locked versions, matched models and budgets, hidden tasks, many randomized replicates, deterministic validators and blind external semantic judging.

ESTABLISHES

Frozen input identities, local materialized outputs, deterministic scoring and the child process’s reported aggregate usage.

DOES NOT ESTABLISH

Global superiority, sampling equivalence, creation-path provenance, or an independent same-model verdict.

READ THE PROMOTION STANDARD →