Capture table for trained-19c7e19a

Capture trained-19c7e19a rollout, task 2, seed 3

Capture trained-19c7e19a rollout, task 2, seed 2

Capture trained-19c7e19a rollout, task 2, seed 1

Capture trained-19c7e19a rollout, task 2, seed 0

Capture one opencode rollout trace

Capture one opencode rollout trace

Capture table for base-19c7e19a

Capture base-19c7e19a rollout, task 2, seed 3

Capture one opencode rollout trace

Capture base-19c7e19a rollout, task 2, seed 2

Capture base-19c7e19a rollout, task 2, seed 1

Capture base-19c7e19a rollout, task 2, seed 0

Capture one opencode rollout trace

Capture one opencode rollout trace

Capture one opencode rollout trace

Capture one opencode rollout trace

Capture one opencode rollout trace

Capture table for trained

Capture trained rollout, task 1, seed 3

Capture trained rollout, task 1, seed 2

Capture trained rollout, task 1, seed 1

Capture trained rollout, task 1, seed 0

Capture one opencode rollout trace

Capture one opencode rollout trace

Capture one opencode rollout trace

Capture one opencode rollout trace

Capture table for base

Capture base rollout, task 1, seed 3

Capture base rollout, task 1, seed 2

Capture base rollout, task 1, seed 1

Capture base rollout, task 1, seed 0

Capture one opencode rollout trace

Capture one opencode rollout trace

Capture one opencode rollout trace

Capture one opencode rollout trace

Add the recording rig: serve one model behind a public URL and hold it

Forward ROLLOUT_SEED so the same task can be sampled more than once

Separate the dataset seed from the rollout seed: one seed meant changing the seed changed the task

Capture one opencode rollout trace

Screen base vs trained to pick the before/after task

Screen base vs trained to pick the before/after task

Screen base vs trained to pick the before/after task

Screen base vs trained to pick the before/after task

Screen base vs trained to pick the before/after task

Screen base vs trained to pick the before/after task

Screen base vs trained to pick the before/after task

Screen base vs trained to pick the before/after task

Screen base vs trained to pick the before/after task

Screen base vs trained to pick the before/after task