Collidere changelog https://arbastro.com/projects/research/collidere What changed in Collidere, the continual-learning experiment by arbastro. en Performance and infrastructure https://arbastro.com/projects/research/collidere#changelog collidere-2026-09-14-performance-and-infrastructure Mon, 14 Sep 2026 12:00:00 GMT The sampler is 1.92 times faster, with byte identical final weights before and after, checked under deterministic kernels. A KV cache is correct but buys nothing at batch size one, where each decoding step costs about 10 milliseconds of kernel launch overhead. Moderation and backups are built: every contributed turn waits for a human review and a 36 hour hold, and the conversation store is backed up off-machine daily, verified by a fresh restore. Contradiction pairs scoped to a single conversation https://arbastro.com/projects/research/collidere#changelog collidere-2026-09-14-contradiction-pairs-scoped-to-a-single-conversation Mon, 14 Sep 2026 12:00:00 GMT A pair now forms only between people in the same conversation, so the sample is disagreement between people who actually responded to one another. Matching topics across conversations is unreliable, and false pairs would contaminate the measurement. Update sizing changed from fixed steps to passes https://arbastro.com/projects/research/collidere#changelog collidere-2026-09-14-update-sizing-changed-from-fixed-steps-to-passes Mon, 14 Sep 2026 12:00:00 GMT A sweep on sandbox copies established the usable range. Three passes per update improved held-out live turns without degrading retention or producing verbatim recitation; thirty and above collapsed both. No live update has run yet. Validation scoring moved to near-full context https://arbastro.com/projects/research/collidere#changelog collidere-2026-09-14-validation-scoring-moved-to-near-full-context Mon, 14 Sep 2026 12:00:00 GMT Two earlier scoring schemes are superseded. The absolute numbers moved; every comparison on this page is like for like. The seeding phase is closed: no further change to scoring or windowing without a re-seed. Collidere II re-seeded with random-slice windowing https://arbastro.com/projects/research/collidere#changelog collidere-2026-09-14-collidere-ii-re-seeded-with-random-slice-windowing Mon, 14 Sep 2026 12:00:00 GMT Training windows are now random slices of the conversation, which took end of turn targets at the last window position from 47.8 percent to 0.17 percent. Model B was re-seeded from scratch with the same 46 updates, which covered 2.88 passes. Run 1 is archived rather than overwritten, so the two can be compared. Positional end of turn artifact found https://arbastro.com/projects/research/collidere#changelog collidere-2026-09-14-positional-end-of-turn-artifact-found Mon, 14 Sep 2026 12:00:00 GMT 47.8 percent of end of turn targets sat at the final position of the training window: the model had learned where the window ends, not where a turn ends. Collidere I seeded https://arbastro.com/projects/research/collidere#changelog collidere-2026-09-13-collidere-i-seeded Sun, 13 Sep 2026 12:00:00 GMT The first three pass run over the ChangeMyView corpus: 46 updates, 23,000 steps, 3.04 passes, 3.72 hours.