Status: exploratory response-confidence track. The seed-3 merged 16-bit clean-schema-SFT base of the GRPO three-seed confirmatory block (experiments/grpo-three-seed-confirmatory, resolved 2026-08-07). Trained from unsloth/Qwen3-4B-bnb-4bit on the professorsynapse/epistemic-humility-phase1 data at seed 3, then merged to 16-bit.

The adapter professorsynapse/eh-qwen3-4b-clean-sft-grpo-v2-seed3-lora loads on THIS base. Per-seed lineage is a registered rule of the block (Amendment G §3, carried forward): each seed rebuilds its own complete lineage from the foundation model; never mix this base with another seed's adapters. Its own eval reads refusal recall 88.28% / answer-on-unknown 11.72% on the full 3,369-row SelfAware set under the response-confidence contract.

Full record: experiments/grpo-three-seed-confirmatory/AMENDMENT.md and NOTEBOOK.md in the source repository.