# Five-model practical test evidence

Published article: https://kingy.ai/blog/qwen4-astra-sol-opus-fable-real-world-tests/

## Grading method and reviewer identity

The initial and completion grading passes used GPT-6.1 Sol (model identifier `gpt-6.1-sol`) with high reasoning effort, as recorded in the grading-session run records. The exact immutable served model build was not recorded. Separate reviewer sessions checked outputs against the supplied source packets, answer keys and rubric. Reviewers could see candidate identities: this was unblinded AI review, not a blinded human evaluation. The grading model is from the same provider as two of the tested models; its scores should be read with that potential bias in mind. Deterministic checks and website integration checks supplement the AI review.

## November 16, 2026 corrections

The article and downloadable guides now identify the grader and repair missing spaces in display text. FINAL-STATUS.json now identifies the published article. These are publication and method-disclosure corrections, not new model testing. The original 200 scored outcomes, model outputs, saved review records, accounting and testing dates are unchanged.

Use index.html to browse the 200 stages and observed-results.json for the saved results. Original review notes retain their original wording, including typographical errors; the article and guide display corrected spacing.
