Show the work.
Keep the caveats.

Editorial practice

DF9 Labs publishes independent reporting and original work from Datum Foundry Nine. Research articles should identify their question, method, source evidence and limitations. Commentary and explanation are labelled as such.

Articles pass through an authenticated draft and publisher review workflow. Corrections can be recorded on the article itself. Editing history is retained in the editorial system.

60% technical. 40% human.

Each core challenge (B01–B05) gets two independent assessments, both out of 100. An AI judge scores the technical requirements of the brief against the evidence. The reviewer then judges the result itself: functionality for the software challenges, visual quality for the Blender model, playability for the games, each from a short published rubric, and chooses a quality band.

Final = min(technical × 0.60 + human × 0.40, band cap)

Unusable / caps at 40

It does not do what the brief asked in any meaningful way.

Weak / caps at 60

It works in part, with serious problems.

Acceptable / caps at 80

It does the job, without excellence.

Good / no cap

A strong result.

The band is chosen by the reviewer, not computed, so a submission that looks complete on paper but does not hold up in use cannot score highly. The two Wildcards (B06, B07) are creative challenges with no AI judge: the reviewer scores them out of 100.

Functionality sets the technical ceiling.

20 / cannot launch

The submission cannot be launched.

40 / primary function unusable

The main functionality is unusable.

70 / critical failure

Core functionality works, but a critical requirement fails.

100 / critical requirements satisfied

No demonstrated critical failure applies.

The final technical score cannot exceed its ceiling. A possible score is an upper bound while evidence is unresolved; it is not an awarded score.

Approval and verification are different.

Judged means the AI judge and the reviewer have both assessed every challenge. A model’s results are published only then, all at once; nothing is shown for a model that is part-way through.

Provisional marks the earlier reviews published under methodology 1.0 (70% technical, 30% human out of ten). They are kept for the record and are not ranked.

Experimental research describes an ongoing investigation. It must not be read as a completed or verified finding.

Comparisons that mean something.

Models are compared within the same challenge, brief version, prompt hash and methodology. Ties share a rank. The overall standing is the B01–B05 total out of 500, and only models judged on every challenge appear in it; partial programmes are not averaged.

A controlled public record.

Results are sent by the Model Gauntlet itself once a model is fully judged, through a single authenticated channel that can do nothing else. Only public fields and each model’s main render are sent. Private databases, local addresses, judging credentials and internal evidence paths are not part of the publication.