Yesterday
-
Apply all-source values admission: verified bounded queries, contextual reviews and whole-query quarantine
-
Add source-grounded Taiwan civic curriculum and disclosed sovereignty/self-determination policy; hold unreviewed topic queries
-
Quarantine readiness query defects; preserve retained answers and Barbet supervision
-
Use Barbet tokenized training by default; remove MiniCPM tokenized exports
2 days ago
-
Append independently verified finite-record memory and tool queries; preserve frozen v26 alternative-answer rule
-
Append independently verified finite-record memory and tool queries; preserve frozen v26 alternative-answer rule
-
Append IFM coding winners selected by minimum frozen native-v26 answer PPL; retain full source and first-hold evidence
-
Append full-context IFM math winners selected by frozen v26 minimum answer PPL; retain all prior query holds
-
Exclude 33 episodes in 21 queries with unresolved source tool contracts; preserve retained answers and v26 selection
-
Append source-qualified OpenCode with complete frozen-v26 answer selection and full format proofs
-
Append contextual Natural Instructions with frozen-v26 minimum-PPL references and full format proofs
-
Publish cumulative new-only SFT with minimum-v26-PPL answer selection and exact budget proofs
-
Publish new-only memory, skills and tool SFT v2 wave 1
Replace active legacy configs with verified new v2 data; retain history. 10M target remains in progress. No model training or efficacy claim.
