Publishing a checkpoint without losing its evidence

From a closed run to a public checkpoint

A completed laboratory run is a research record. A public milestone also needs a clear experiment, measured results, original evidence and an explanation of what the result does and does not establish.

The publication workflow now accepts editorial notes and explicitly selected evidence together. Incoming submissions remain private until review. Acceptance of an upload is not publication.

Evidence stays attached

For a verified Search milestone, the review checks the original run manifest, recorded actions, measured score and survival, replay comparison and captured screenshot hashes. Search remains distinct from Perception/Qwen: a Search run has zero model calls. A provisional accelerated candidate is not presented as a verified real-time record.

Repeated submissions do not create duplicate articles. Updates preserve a stable article identity and its revision history; an outdated checkpoint cannot silently replace a newer one. Published milestones appear in Lab Notes, while the project dashboard and timeline remain reviewed checkpoints rather than live counters.

Media needs its own decision

A video or image can explain a milestone, but its technical integrity does not settle redistribution rights. Gameplay candidates receive a separate editorial and rights review before they appear publicly. Captures must identify their run, branch and limitations. YouTube publication remains a separate operator decision.

We distribute our experiments, not the game. ROMs, BIOS, game dumps, extracted assets and laboratory secrets are excluded from publication packages.

What this change establishes

This is a publication infrastructure update. It reports no new game score, no new benchmark measurement and no new world-record claim. The current scientific checkpoint and its supporting artifacts remain linked below.