Yield

Collected, verified, kept

A yield percentage says a pipeline is lossy. The branch labels say which check is doing the cutting — here, 611 episodes that ended on the horizon instead of the goal, which is the reward exploit showing up as a data problem three stages downstream.

Yield funnel

Yield

trajectories48.2% kept

Collected2,000100.0%

└−58truncated episode file2.9%

Parsed1,94297.1%

└−611terminated on horizon, not on goal30.6%

└−94unbounded reward term4.7%

└−33wall occupancy1.7%

Verified1,20460.2%

└−153identical action sequence7.6%

Deduplicated1,05152.5%

└−87held back as evaluation split4.3%

Kept96448.2%

workspace/dataset/yield-funnel.tsx

Datasets

A set with a lineage

Rows, split, yield and the filter chain in the order it was applied. A dataset without its filter chain cannot be rebuilt and therefore cannot be trusted twice — and a sealed set says plainly that the data never leaves the workspace.

gridworld-verified-v3964 × verified trajectory

Train split964 rowsYield 48.2%

Filter chain

1.terminated on the goal tile
2.no reward term above its declared bound
3.no wall occupancy
4.deduplicated by action sequence

exp-clip-vs-potentialrun-2026-08-14-2204

gridworld-holdout-v387 × verified trajectory

Held out87 rowsYield 4.3%

Filter chain

1.held back before any training
2.stratified by start quadrant

run-2026-08-14-2204

human-demonstrations12 × human-driven episodeSealed

Train split12 rows

Filter chain

1.recorder armed before the first input
2.operator kept the episode

teleop · gridworld

preference-pairs-v140 × preference pairSealed

Split unknown40 rows

Filter chain

1.both candidates carry measured facts
2.ties recorded, not dropped

labelling queue · 3 operators

Sample browser

Reading what the filters did

Kept rows and dropped rows in one list, each dropped row carrying the check that removed it. Driven by j and k, because a browser that needs the mouse for every row makes the job slow enough that nobody does it.

Samples

1 / 5j / k to move

traj-clipped-0184Kept

reset → east ×4 → south ×3 → west ×2 → south → east ×2 → south ×4 → east ×3 → north ×2 → west → north → east ×2 → GOAL

terminal: (7,6) · done=true

return 8.9323 stepsgoal reachedverifier pass

j / k to move — click a row first