Politique éditoriale
Les 20 premiers originaux comptent 40 cas réels conservés, avec évaluation croisée par IA. Le propriétaire a approuvé cette bêta limitée en attendant la relecture humaine. Descriptions et traductions assistées par IA n’étendent pas cette approbation. Instructions et sorties anglaises sont conservées ; aucune autre langue ou modèle n’est présenté comme testé. Tout modèle nouveau ou modifié exige ses propres preuves.
Qu’a-t-on réellement testé ?
Chaque original anglais a deux exécutions réelles dans Codex : cas normal et cas limite, évalués par une IA de la même famille. L’identifiant exact du modèle est inconnu. Relecture humaine et autres modèles restent non vérifiés.
Original anglais
Les instructions et exemples enregistrés restent en anglais. Les descriptions traduites facilitent la lecture ; les tests concernent uniquement l’original anglais.
Original anglais
The current edition
The first 20 templates are original and AI-assisted. They have structural checks and two actual synthetic model executions each: 40 separate Codex subagent turns. Another agent cross-evaluated each output. Human sampling is still pending. The site owner approved this limited public beta with the pending status visible.
The displayed excerpts come from retained actual executions, not imagined outcomes. Execution used the OpenAI Codex family; the exact runtime model ID was not exposed. This is same-family AI cross-evaluation, not an independent provider benchmark or human review. Other models remain untested.
What a release requires
Each template must pass schema, variable, source, duplication, and publication checks. The first release requires at least two actual model cases: a normal input and a missing-information or conflicting-constraint case. Each must meet the internal quality threshold and avoid hard failures.
The original launch plan requires human inspection of the first ten trial templates. The owner explicitly approved an exception for the current 20 tested template versions so the public beta can launch while human review remains pending. This exception does not approve new or materially changed templates. Later batches require the sampling described in the project plan, with broader review when a problem appears. An automated or AI-assisted check is never labelled human review.
What testing can tell you
A test record identifies the provider, requested model, date, cases, method, and evidence reference. A passing example does not guarantee future answers. The author and evaluator may be correlated; independent checks and human judgement remain useful.
We check task completion, useful structure, clear inputs, missing-information handling, natural English, and a distinct task purpose. Fabricated facts, unapproved source copying, exposed secrets, and unresolved variables block release.
Sources and corrections
The initial collection is original rather than copied from public prompt lists. Publicly readable material is not automatically reusable. Third-party additions require permission evidence, attribution, and license conditions.
For corrections or rights concerns, email zdage300@gmail.com with the affected URL. We remove affected page content, downloads, and index entries together when needed.
Mise à jour