reports: independent audit of the regressions

chess-sim/reports/round2/independent_audit.md

@@ -0,0 +1,189 @@
+# Independent audit of the baseline and recovery regressions
+
+This audit reads the local code and evaluation JSON, plus the actual baseline and
+both recovery datasets' action tables/metadata from `Machanize/playful`. It also
+reads the published baseline's config and saved normalization tensors. No model
+was trained, no pod started, and no remote artifact changed.
+
+## Conclusion
+
+The evidence supports a narrow, orientation-dependent grasp tolerance and an
+inaccurate learned approach. It does not establish that changing the recovery
+teacher alone caused both regressions. The fine-tuning recipe also changes the
+normalization and restarts optimization. Several diagnostic labels overstate
+what was measured. Control these confounds before collecting more recovery data.
+
+## 1. The paired regression is real; precision did not improve enough
+
+Recomputed from `baseline_dev/eval_results.json` and `v2_dev/eval_results.json`:
+
+| Metric, same 60 trials | Baseline | Recovery v2 |
+|---|---:|---:|
+| Full success | 31 | 18 |
+| Both models succeed | 12 | 12 |
+| Unique successes | 19 | 6 |
+| Closing error <= 2 mm | 9 | 9 |
+| Target ever lifted > 8 mm | 44 | 37 |
+| Another piece displaced > 2 mm at end | 11 | 11 |
+| Pawn lift / success, 42 trials | 26 / 20 | 20 / 12 |
+| Knight lift / success, 18 trials | 18 / 11 | 17 / 6 |
+
+The two-sided exact paired McNemar test on 19 losses versus 6 gains gives
+p = 0.01463. This is evidence of a regression on this test set, not an estimate
+of performance on arbitrary boards. Repeated checkpoint selection on this set
+would require a separate confirmation set.
+
+The drop is not solely failure to initially lift: 17/18 knights lift under v2,
+but only 6 finish. Transport, release and disturbance need their own phase labels.
+The 26/37 toppled-target counts must not be confused with the 11/11 displaced-
+neighbor counts, or interpreted as proof of when each toppling occurred.
+
+## 2. Fine-tuning changes the coordinate system as well as the examples
+
+`sim/train_policy.sh` loads weights through `--policy.path`; it does not resume
+optimizer state or specify a reduced fine-tuning learning rate. The published
+baseline config has optimizer peak LR 1e-4 and decay LR 2.5e-6. The installed
+LeRobot 0.4.4 scheduler scales its schedule to the requested run length. A new
+4,000-step run with these settings ramps back toward 1e-4, rather than continuing
+the baseline's final low learning rate.
+
+More importantly, installed `lerobot/scripts/lerobot_train.py:251-287` passes
+the new dataset's statistics as overrides to both normalizer and unnormalizer.
+`lerobot/processor/normalize_processor.py:171-200` explicitly gives supplied
+statistics precedence over the saved tensors. This occurs when loading a
+pretrained policy, including a fine-tuned baseline.
+
+The baseline's published action normalization agrees with its original dataset
+statistics to < 1e-5. Combining the recovery tables with the original changes
+those statistics. To isolate the output-scaling effect, for each original action
+`a` compute:
+
+```
+z = (a - baseline_mean) / baseline_std
+a_reinterpreted = z * merged_std + merged_mean
+```
+
+This holds the normalized output fixed. It is NOT a forward pass of the policy:
+the changed state normalization would also affect network predictions, and
+training can adapt to the new statistics.
+
+| Output-scaling diagnostic | Recovery v1 mixture | Recovery v2 mixture |
+|---|---:|---:|
+| Median abs shoulder-lift target shift, all baseline rows | 2.120 deg | 1.146 deg |
+| Median abs wrist-flex target shift, all baseline rows | 0.540 deg | 1.066 deg |
+| Median XY shift of gripper reference site at closing | 4.877 mm | 1.532 mm |
+| Median 3D shift of gripper reference site at closing | 5.556 mm | 2.773 mm |
+| 95th percentile 3D reference-site shift at closing | 6.748 mm | 3.393 mm |
+
+FK used the actual `gripperframe` site in `sim/chess_scene.xml`, evaluated at the
+first recorded command below 3 after a command above 4 in each of the 300 baseline
+episodes. It is a reference point near the fingertips, not a measurement of the
+actual contact surface or policy rollout. All FK results were finite. The merged
+statistics were calculated directly from 89,008 baseline rows plus 36,645 v1 or
+29,044 v2 rows.
+
+These are meaningful perturbations relative to the grasp tolerance. This is a
+confirmed property of the current recipe and saved datasets, not proof that it
+caused the historical policy regressions. The v2 checkpoint/config was not saved
+on the Hub, so its exact runtime processors cannot be independently inspected.
+
+## 3. The simple teacher is not a single canonical trajectory per move
+
+The 300 baseline episode records include these grasp orientations:
+
+| Move | Expert jaw yaw counts |
+|---|---|
+| e7-e5 | -45 deg: 11; +45 deg: 12 |
+| g8-f6 | -45 deg: 14; +45 deg: 17 |
+| g1-f3 | -45 deg: 16; +45 deg: 17 |
+| c2-c4 | -135 deg: 15; +135 deg: 14 |
+
+Eight of the ten moves have two recorded yaw choices. Home joint jitter remains
+0.06 radians and motion-speed scaling remains 0.8-1.2 in the inherited config.
+These alternatives can be valid demonstrations; they are not inherently wrong.
+However, changing between incompatible orientation choices across predicted
+chunks is a plausible failure mode that should be measured.
+
+The existing 20-episode grasp report shows median absolute yaw errors of 30.6 deg
+for toppled cases, 41.6 deg for closed/not-lifted cases, and 81 deg for the two
+lifted/dropped cases. It is not justified to reduce the diagnosis to XY error
+alone. These errors are relative to one expert grasp; a different orientation
+can sometimes be valid, so the measurements do not prove mode switching.
+
+The saved model has one observation step and executes 50 actions per prediction:
+1.67 seconds at 30 Hz. In the expert, closing plus holding is about 0.7 seconds
+before speed scaling. The prior 10-step replan result was only 4/15, on a different
+test protocol, so it does not establish that more frequent feedback is useless.
+
+## 4. Both recovery paths require more precision than the policy demonstrates
+
+`collect_recovery.py:73-106` chooses the lowest collision-free correction height
+for the exact simulated state. Of the 120 v2 records, 74 request zero additional
+rise. Nominal collision-free planning is not robustness to a 4-5 mm position
+error or a large wrist-angle error. Increasing jaw opening moves one jaw; it
+does not remove the fixed jaw's very small clearance (`fixed_gap_mm = 0.8`).
+
+The jaw test isolates fixed pose offsets with an expert orientation, descent,
+closure and lift. It does not reproduce the learned policy's simultaneous yaw,
+velocity, height and timing errors. Its tolerance percentages cannot be applied
+directly to a scalar median policy closing error.
+
+Also, 23,000 of v2's 29,044 rows occur after the first below-3 close threshold.
+Only 3,598 rows lie in the one second before it. With ordinary uniform frame
+sampling, adding 120 correction episodes is not equivalent to making precise
+pre-contact alignment the main training objective. This does not mean those
+later frames are useless; grasp, lift and placement must all be preserved.
+
+## 5. Measurement limitations and a unit error
+
+- `eval_policy.py:135-147` invokes the measurement callback AFTER applying and
+  stepping the action. The close event at lines 227-235 is a command crossing
+  below 3, not the physical onset of closing or first contact.
+- Five of the seven toppled cases in the older 20-episode inspection were already
+  tilted >45 deg at that recorded close event. Thus these records do not establish
+  whether approach, rotation, or early closing first caused the collision.
+- Closing position is compared to the initial expert grasp point, not a newly
+  measured grasp on a potentially displaced piece. Along/across axes use expert
+  yaw rather than the actual jaw direction. The yaw diagnostic also wraps modulo
+  180 deg, which loses fixed-versus-moving-jaw orientation information.
+- `jaw_tolerance.py` labels `1000 * (grasp.open_q - lo)` as `open_width_mm`.
+  This is an angle multiplied by 1000, not a physical width. It does not invalidate
+  the success outcomes or the input opening-margin settings, but those reported
+  width values must not be used for hardware calibration.
+- The exported feature shapes mention 256x256, but the policy's actual image
+  preparation resizes incoming images to 512x512 with padding. The absent third
+  camera is skipped with `empty_cameras=0`; neither fact alone is evidence of a
+  missing-camera or 256-pixel inference bug.
+- The baseline has both `freeze_vision_encoder=true` and `train_expert_only=true`.
+  Disabling only the first flag would still leave the VLM frozen in this version.
+
+## Recommended next experiment
+
+1. Before training, verify that the candidate fine-tuning pipeline reproduces
+   baseline outputs with zero updates: identical weights, observations, RNG,
+   action queue, saved normalization and processor settings. Separately evaluate
+   the effect of substituting merged normalization with unchanged weights.
+2. Hold the original state/action normalization fixed for a controlled continuation.
+   Make optimizer/schedule restart and LR explicit. Compare original-data-only
+   versus original-plus-corrections with the SAME initial checkpoint, LR schedule,
+   update budget and evaluation seeds. A baseline-only control under the current
+   recipe tests optimization but does not isolate the merged-normalization effect.
+3. Instrument the rollout before any new recovery-data design: actual jaw pose,
+   orientation, physical opening, target-relative error, first unintended contact,
+   first tilt, and action-chunk boundaries. Separate successful lift, retention,
+   transport and release.
+4. If the preserved-normalization correction run still fails, test a canonical
+   grasp orientation per move and an approach clearance tolerant of measured pose
+   errors. Change these independently and validate the expert before regeneration.
+5. Keep the baseline. Do not start a 1,000-episode regeneration or vision-unfreezing
+   run merely because the previous corrections failed. Choose the next change
+   from these diagnostic results.
+
+## Limits
+
+No GPU inference or new policy rollout was run in this audit. The actual v2 model
+and its optimizer/processor state were unavailable after pod deletion. Current
+local LeRobot source and saved artifacts support the findings, but reconstructing
+the precise historical training environment would require its code revision and
+full configuration. Private source tables and small model metadata were cached
+under `/private/tmp/chess-regression-audit`; no weight file was downloaded.