Know which version you are trying.

7 older releases still need correction.

The corrected source for these plugins is not deployed in this snapshot. Their live servers still use earlier versions. Start with synthetic data and inspect every output.

The catalog marks these pages Corrections pending. They remain available for inspection. The separate browser CSV comparer is an option for exact-key CSV changes and has its own input rules.

The nine recent plugins have real host observations.

The recorded test accounts called a tool and displayed its widget for all nine in ChatGPT and Claude on 11 October 2026. Later control and download checks are dated on each plugin page. They test specific tasks; they do not certify all functions or availability in your account.

A fix for widgets that restarted.

The shared widget code was corrected so a repeated host result does not reset work already entered. The fleet records that fix in the 48 reviewed releases. The seven older versions awaiting correction do not have it in this snapshot. Four follow-up control cases passed across the two hosts; this does not mean every control was retested.

Read the review scores with their scope.

45 releases have passing independent review results. Three further releases use documented administrator rulings about a difference between the old live server and corrected source, followed by recorded deployment checks. Their raw scores remain visible below.

A review score summarizes that review's cases. It is not a safety certificate, customer rating, or guarantee of correctness for every input.

Directory review is a separate route.

The dated portal records show 49 in review at OpenAI, with no recorded OpenAI publication. Claude records show 2 published, 52 in review, and 1 queued. Each plugin page gives the observation date. This website update did not sign into either directory.

All 55 public endpoints answered anonymous metadata discovery, exposing 124 functions. That check initializes the server and reads its function list; it does not invoke a task or test installation in a host.

What the measurements do not establish.

The fleet's native-host trials found that the hosts could already perform most of the nine tasks. Its speed comparisons used different prompts, so they cannot establish a controlled speed advantage. Two undeployed engine trials are excluded from the catalog. No revenue or customer adoption claim follows from these checks.

Where these observations came from.

The source records are the owner's fleet-7 and fleet-8 states, real-host result files, recorded release checks, and factory directory snapshots. Conflicting dashboard summaries were reconciled against those records. The public dataset includes source-file hashes for provenance, but omits account details, private conversation links, screenshots, and logs.

The HTML plugin pages are the main reference. Browse reviewed releases or inspect releases awaiting correction.