Basecode Sweep
Purpose: run the shared basecode/blockdiff toy metric contract across the
implemented sampler hooks before attaching real-model adapters.
Remote result:
| Job ID | Partition | Status | Output |
|---|---|---|---|
678190 |
normal |
completed | runs/basecode/latest_sweep.jsonl |
The sweep contains 12 rows:
- confidence remasking at block sizes 8, 16, and 32
- DMax-style reduced-remasking at block sizes 8, 16, and 32
- speculative hook at block sizes 8, 16, and 32
- multiblock decoding at 2, 4, and 8 blocks
This is still a toy contract check, not real-model parity. The next parity step
is to add a DFlash adapter that consumes the same metric fields as
runs/source_reproductions/dflash_qwen3_4b/*metrics*.json.