WorkLedger
Interface study replay
Replay-only viewer for sanitized release-candidate samples. No provider transport, credentials, secrets, or model calls.
Release-candidate summary
- Qwen3-14B: exact replacement advantage replicated (U1 1/19, R 13/19).
- Qwen3-235B-A22B: exact replacement advantage replicated (U1 3/19, R 16/19).
- Qwen3-32B: external anchor, same direction (U1 0/19, R 9/19), not jointly randomized and not in the paid denominator.
- Qwen3-8B: directionally favorable (U1 2/19, R 5/19), paired coverage 18/19, no registered interface-advantage replication.
- DeepSeek V4 Pro: infrastructure-unassessable because 38/38 cells were privacy/data-policy route failures with empty final answers; this is not a coding-performance score.
- Strict native and one-outer-fence-normalized policies were identical in the new multimodel run.
- One attempt per cell; provider routes and inference settings are confounded; no production-readiness claim.
Public author: Confetti3. License: MIT for this Space.