rtrtwrw/granular-material-pour-rollouts

granular-material-pour -- policy rollouts

Every rollout recorded on this station, as the scorer wrote it. These are the evaluations audited in the accompanying paper; the demonstrations they were trained on are published as one repository per session alongside this one.

Camera-ready evaluation (camera_ready_2026-11/)

The camera-ready version of the paper (November 2026) re-scores every rollout under one written rule and adds the three models trained only on the recovery demonstrations. Every number in its results table and appendix comes from this folder.

file
logs/	the per-model logs as the station wrote them, one row per attempt
rollout_results_golden.csv	the first evaluation session (2 October), before the per-model logs
plans/	the target plan each model was run against
all_runs.csv	every attempt from both sources, with the status the rule gave it
results.json	the per-model results: counts, 95% intervals, slope, contact runs
analysis.py	the rule and the statistics; python3 analysis.py recomputes everything

The rule. A pour is scored when the station ran the policy and read a settled weight (the median of the last second). A pour is not scored when the pan was moved or emptied before the reading, when the policy never started pouring, or when the station ran with a wrong camera map. Pre-load and remove-20-g runs are closed-loop probes, reported apart, and targets outside a model's trained target range are reported apart. One camera-only ACT run (380 g, 2 October) is scored from its plateau, 437.3 g, as described below.

Take-back labels. takeback/ holds the detector that the paper uses to count recovery pours, and its output for all 1,839 pours (takeback_labels.csv).

The slope is the Theil-Sen estimate of delivered against requested weight: 1 means the policy follows the request, 0 means it pours the same amount whatever is asked.

Two directories, on purpose

audit_2026-10-03/ holds exactly the 197 rollouts the submitted version of the paper analyses, frozen at a cutoff of 14:04 on 3 October 2026. Every count in the paper's evaluation section can be reproduced from these files and nothing else.

full/ holds 264 rollouts: the same runs plus everything recorded after the cutoff, because the bench kept running while the paper was written. Use this one to see more of the models' behaviour; use the audit snapshot to check the submitted version; the camera-ready version uses camera_ready_2026-11/.

plans/ holds the plan each set was run against, so a row number resolves to the target, the bench setup and what was expected.

Columns

column
row	the plan row this run executed
model	the checkpoint
target_g	what was asked for
final_g	the delivered weight, the median of the settled plateau
err_g	final_g - target_g
peak_g	the highest instantaneous reading, which is not always material
start_g	the scale before the run
crossings	times the weight crossed the target
outcome	OK, LOW, HIGH, OSC, or SUSPECT where the trace could not be scored
notes	written by hand where a run was recovered or rejected

Read this before computing anything

peak_g does not measure material. When the arm touches the pan the load cell reads the arm. Peaks above 1000 g appear in seven of these runs and are contact, not grain.

Not every outcome is the policy's. 29 of the 197 audited runs were set by measurement or protocol: a pan moved before the reading, a run scored on paper with no trace kept, a control loop that never sent an observation, a board that reset mid-run. The paper counts and locates each one. A run's notes field is the first place to look.

One row is knowingly wrong and left in place. The camera-only ACT run at 380 g is logged as -0.6 g. Its weight trace holds flat at 437.3 g for thirty-five seconds and falls to zero only in the final second: the pan was cleared after the run and the scorer read the last sample instead of the plateau. The correct value is +57.3 g. It is left as recorded, and reported in the paper, rather than quietly fixed.