Run timeline
What overlapped with what
A roster answers who is running. Only a timeline answers what overlapped — which is the question you actually have when two experiments disagree and you are looking for the shared cause. Specialists are indented under the replica they serve; the cyan line is now.
Timeline
7 lanesWindow 01:00workspace/lifecycle/run-gantt.tsxBudget
The line it is on
Spend-to-date tells you what happened; the projection tells you whether to intervene, which is the only reason to look at a budget while a run is live. The dashed line extrapolates the recent rate to the ceiling and names the time you have left.
Burn-down
108,600 / 250,000tokensCeiling in 1h 05m at this rateCeiling comfortably beyond the projectionBurn-down
108,600 / 140,000tokensCeiling in 14m at this rateUnder 30 minutes of headroom — the readout turns amberCheckpoints
Walking the workspace backwards
Chat has rewind; the workspace did not. Restoring is armed before it fires and names what would be discarded — moving a whole workspace backwards is not a click anyone should make by accident.
Checkpoints
5- workspace openedAuto22:04
clean tree at 7f2c9a4
- reward clipped, step cost addedManual22:19
envs/gridworld/reward.py, configs/ppo.yaml
- human took the controlsHuman22:41
124 human-driven steps recorded as demonstration
- exp-curriculum-v2 promotedPromotion23:02
4 commits merged into main
- sweep runningAutoCurrent23:14
runs/2026-08-15/*
Click Restore once to arm, again to confirm