Every experiment, one trajectory.
Development math accuracy · fixed 512-example cohort
Follow the experiments, the decisions, and the work behind the result.
Development math accuracy · fixed 512-example cohort
Reconstructed phases · UTC
Awaiting selection
Train a smaller auxiliary model, then use its learned corrections while the 32B target stays frozen.
Development selection is provisional. Independent hidden evaluation and final compliance review are separate.
GPU utilization over the research window
Hypotheses, development outcomes, and the candidate Astra kept.
| Trial / hypothesis | Math | Retention | Eval time | Status | Completed · UTC |
|---|
Visible messages, sandbox commands, waits, and tool results.
This is the visible execution record. Provider reasoning and encrypted payloads are excluded. All times UTC.
Live documents written by the researcher. Statements here are the agent’s own reports.