Grug 67B A2B — Campaign 4 adaptive-delta-08

Research artifact. This model has not been properly tested or evaluated and is not necessarily secure. Use at your own risk in production settings.

This is adaptive-delta-08, rank 1 by the final four-target sum across all 71 sealed Campaign 4 candidates. It transfers selected routed-expert rows from Reference-KL Step20 into Antidoom Step12 while preserving 90.53% of Base scalar weights bit-for-bit. The measured changed fraction is 9.4694%, below the campaign's 30.00% limit.

All scores are single-run development results from the pinned October 2026 Marin evaluation policy. The four-target sum uses unrounded component scores. The campaign used these target evaluations to select checkpoints, so the scores are in-distribution development results rather than an estimate of broad generalization.

The merge changes only selected rows of routed-expert gate, up, and down projection tensors. Its support covers 27.11% of Base scalar weights; unchanged values within that support account for the difference between support and the measured changed fraction. All weights outside the support, every non-weight file, the tokenizer, the chat template, model configuration, generation settings, and the weight index are copied byte-for-byte from Base.

Parent checkpoints are pinned as follows:

• Base: Antidoom RLVR1 Step12 at 50fe2f488aee7730bbccc721024b35229075602e.
• Extraction target: Reference-KL Step20 at fd07ef0f2b78a83099a0042e59c00de986e7b647.

The provenance/campaign4 directory contains the exact recipe and support, final readback audit, public copies of the five sealed evaluation records, final tracker snapshot, and experiment protocol. The public evaluation copies remove operator identity, internal job descriptions, and launch-host fields while recording each original file's SHA-256. All checkpoint source files are published byte-for-byte from the evaluated S3 checkpoint.

Licensed under OpenMDW 1.1.

We release our tools to the public with the intent that they are used safely, ethically, and responsibly. They are meant to contribute to the public's understanding of science, and aid in its advancement. By using this Marin model, you agree to comply with all applicable law.