Observed baseline
Mission state
Find repeatable model-quality alpha before GPU or mainnet spend.
Research active
Current gate
Model discriminator V1
Scorer fidelity is exact. Run identical model arms through the preregistered 12 / 60 / 180 progression on the frozen screen board.
- Official source
d4129ed- Screen board
- 180 bundles
- Replication
- 120 disjoint
- Fresh holdout
- Untouched
Evidence ledger
What the data supports now
Corpus composition
Primary vs supplemental
Primary
Supplemental
Bank recovery
Joinable supervision remains sparse
51
scored rows mapped exactly to preserved task content
Useful for audit and supplemental analysis. Too small to claim a primary fine-tuning dataset.
| Evidence item | Status | Result | Decision |
|---|
Execution path
Shortest credible route to deployment
Decision register