task01_picking_up_trash/code/EXPERIMENTS.md

Task 01 development record

Objective: mean official final Q > 0.7 across public instances 301–320 (indices 0–19).

Runtime policy inputs: RGB, linear depth, and the official 61-dimensional R1Pro proprioception only. Static robot URDF and camera calibration are permitted. No runtime Codex, LLM, or external model calls. Privileged poses may be read in separate development diagnostics, never by the final policy.

Experimental calibration revision: the official evaluator also supplies robot_r1::cam_rel_poses, three camera-to-robot poses. Revisions11–13 tested this evaluation-observation calibration field for depth projection; object/world poses remained excluded. The default has since been restored to the validated pack09 version, which reconstructs camera poses from proprioception and the URDF.

Development approach: inspect observations and scene geometry first, iterate on short diagnostics and a small representative subset, then freeze and evaluate all 20. If several evidence-driven rounds stop yielding useful improvements, report the blocker and ask the user.

Reference: 5080-1:/shared_work/markhsp/gpt6_behavior_solution. Its autonomous code targets Halloween decorations, not Task 01. Kinematics, sensor projection, and motion/navigation helpers were adapted from the local preserved r21 copy of that reference. No reference scene coordinates or task2 action replay are used.

Local simulator: existing behavior-392dev environment; BEHAVIOR v3.9.2 commit b1979916ec1549b10a4e65e630bc6504a9af1b00, clean checkout at inspection. Existing assets under /home/eai/code/b1k/datasets. GPU: RTX 5080 Laptop, 16 GiB.

Inspection rounds

1. First environment probe loaded, then hit an object-scope compatibility error when the robot appeared directly in the scope. Fixed in the separate diagnostic tools. No policy result.

2. debug/lab301/scan01: 458 diagnostic steps. RGB-D projection estimated the visible can and bin within about 1 cm horizontally of debug ground truth. A chair crossbar caused a false bin proposal; requiring near-floor sides removed it. Velocity-only odometry accumulated roughly 1.17 rad yaw error over a commanded full turn, producing duplicate can tracks. Q=0 is expected for a scan.

3. debug/lab301/scan02_visual: 558 diagnostic steps. Added ORB/RGB-D relative-pose registration every ten observations. The scan finishes with exactly three can tracks and one bin; the velocity-only duplicate-track failure is resolved. Q=0 is expected for a scan. Policy localization uses only onboard observations, not the separate geometry dump.

4. grasp01: reached the blue can but rejected its mostly silver lid in the wrist image. An open-hand false positive then incorrectly selected placement recovery. 1600 diagnostic steps, Q=0. Fixed by geometric refinement around an already identified can, an open-gripper rejection, and robot-CAD exclusion in held-object evidence. Both fixes were verified against the saved failing observation.

5. grasp02: first successful autonomous can lift, verified in all camera views and by closed finger width plus 48,225 non-robot colored pixels near the raised hand. Stopped immediately after lift at 1090 steps as planned; Q=0 because no deposit was attempted. No simulator poses entered the policy.

6. deposit01: carried the blue can successfully to the kitchen bin, but close-range bin recognition rejected the brown interior. The policy reached an unresolved placement state at step1433. The developer simulator was interrupted to avoid consuming the remaining diagnostic budget on idle actions. This interrupted diagnostic is not an official evaluation result. A depth-only refinement around the already identified bin was verified on the failing recorded view (5,837 surface points). The lab now supports immediate stopping, stops automatically on unresolved states, and can load another public instance without relaunching the simulator.

7. debug/lab301b/deposit02: 1460 diagnostic steps, Q=0. Transport again succeeded. Local-map translation drift exceeded the bin refinement's crop radius, so it saw only part of the bin. The lab stopped automatically. Broadened close-range matching and removed the color condition from the bin shape detector; its height/near-floor-side criteria reject the recorded chair/foot false positives.

8. release_geometry_probe + release_open_probe: two short developer-directed physical checks from the already-loaded state (60 and 45 steps). The release target was computed from the current RGB-D bin surface, not ground-truth poses. The official task metric reached Q=1/3 and the wrist image shows the can inside the bin. This verifies the release geometry; it is not a complete autonomous episode or benchmark result.

9. debug/lab301b/deposit03: autonomous success, Q=1.0, 4471 steps, within the official 7901-step horizon. All three cans were deposited using the policy's own RGB-D/proprio decisions. Two navigation/reacquisition recoveries occurred while locating the third can. Wall time was approximately 4m25s. This remains a one-instance development result, not the final 20-instance benchmark. Source is preserved under debug/revision_deposit03/.

10. generalize308: invalid diagnostic setup, excluded. The developer lab switched instance snapshots before resetting the prior robot joint state, retaining the previous episode's arm/torso posture. The official evaluator resets before switching, so this was a lab bug, not an official evaluation. Stopped this run and corrected the lab order. A fresh simulator will verify instance308 from the official initial posture. Also excluded the robot's own close floor footprint after a self-pixel false can detection; no object poses are used for this exclusion.

11. debug/lab308/clean308: fresh simulator and valid default initial posture, 1961 steps, Q=0. First can was lifted and transported, but placement stopped because the hand/can occluded the close bin. Developer-directed backup and raised-hand probes exposed the bin; RGB-D center differed from diagnostic ground truth by about 2 mm. These probes are not autonomous episode results.

12. debug/lab308/clearview308: stopped at 2667 steps, Q=0. A farther approach exposed a furniture collision and loss of localization, eventually tipping. The initial torso change lowered the parked right hand near the floor. Inspection also found the obstacle map excluded surfaces above 0.85 m; the exact first contact is not established by the saved snapshots. Expanded mapping to 1.5 m, increased furniture clearance, and restored the right-hand transport pose after lowering the torso. Source before this fix is preserved in debug/revision_generalize03/.

13. clearance308: scan hit its motion budget while correcting centimetres of translation drift; no complete episode result. Scanning now rotates by heading alone. clearance_scan308 then reached Q=2/3, 4445 steps, with two autonomous deposits. It exhausted its remaining attempts on narrow furniture parts classified as cans. Tightened observed can footprint dimensions; recorded-view checks retain all three actual cans and reject the problematic thin components. Also tightened the bin footprint to reject chair-leg proposals. Source for the Q=2/3 run is preserved under debug/revision_clearance_scan308/; the new shape checks are in debug/revision_shape05/.

14. The first shape320 command was canceled before any action by a pending lab stop flag; its empty metrics are excluded. shape320b is the actual next development trial, with the default initial arm and torso configuration verified after switching to instance320.

15. shape320b: Q=0, 491 steps, stopped after the first scan because the bin was hidden by dining furniture. Added bounded exploration from visible cans and observed floor viewpoints. explore320 then found the bin and reached Q=2/3, 4144 steps. It also detected the third can, but three navigation attempts stalled while moving sideways near kitchen furniture. The two completed deposits were autonomous. Source: debug/revision_explore06/.

16. Replayed the 308 perception observations against separate debug ground truth. The narrower can footprint retained 19 true detections and left one low furniture false positive. A minimum observed top height of 0.105 m removed that final false positive; the true detections ranged from 0.116 to 0.135 m. These are recorded-view regression checks, not extra rollout scores.

17. forward320: Q=0, 3081 steps. Facing every travel segment and continuously replanning rejected the first approach after a new view, then exploration took long detours without finding the bin. Reverted that navigation change. The next revision retains the route logic that completed two deposits, adding stop/backoff/camera inspection only when a navigation segment fails. Failed revision is preserved as debug/revision_forward07/; the targeted recovery is debug/revision_recovery08/.

18. recovery320: Q=2/3, 5263 steps. All three cans were autonomously carried and released without a navigation failure in this run. The final image and separate debug geometry showed a vertical stack: can centers at approximately 0.069, 0.204, and 0.331 m above the robot base. The third can was above the bin rim. Changed release targets to three positions on a 4.5 cm radius triangle relative to the observed bin, using only episode odometry to keep their directions consistent.

19. packing_hover, packing_regrasp, packing_close, packing_lift, packing_offset, packing_release: short developer-directed checks from the stacked state. The protruding can was localized from RGB-D, regrasped, and released 5.5 cm beside the center stack. The unchanged official task metric reached Q=1.0 / success=true after seven release steps. This is a physical placement diagnostic, not an autonomous episode result. Next validation uses the standalone runner on indices19 and7 (320 and308) before any full sweep.

20. runs/validation_pack09: standalone official validation, indices19 and7, unchanged default horizon and evaluator, frozen source under the run's source/. Instance320 succeeded autonomously, Q=1.0, 5196 steps. No manual actions or runtime model/backend calls. Its official video was opened through its last frame. Instance308 is the second planned validation; the full 20-instance mean remains unmeasured.

21. The same validation_pack09 policy failed on308: Q=0.0, 7902 steps, official timeout after navigation drift toward the window and an unresolved placement state. Complete two-instance validation mean was0.5. This result is retained, not replaced by earlier development scores.

22. Recorded-view odometry comparison on clearance_scan308, pairs801→901,901→1001,1001→1101,1101→1201: the old above-floor feature mask gave too few stable matches. Including textured floor points with the same ORB matcher gave72,71,57,67 geometric inliers; translation errors against separate debug poses were0.015,0.005,0.011,0.001 m, and yaw errors0.010,-0.002,-0.001,-0.001 rad. Changed only the feature mask, retaining robot exclusion and registration quality checks. Also generalized bounded target search to work when the initial scan sees the bin but no cans. validation_floor10 tests indices7 and0 with these changes.

23. validation_floor10 was interrupted during308 after repeated navigation/grasp failures; no completed official result was produced, and301 was not run. The recorded-image improvement alone did not resolve live navigation. Inspecting the unchanged official evaluator revealed its cam_rel_poses observation field, which the earlier sensor filter discarded. A separate short lab now compares that provided calibration with URDF reconstruction. The initial static poses agree to displayed precision; moving-frame comparison is still needed.

24. debug/labcal308/calibration_scan: static camera poses agree; moving-frame camera orientation differs by about0.02–0.04 rad from instantaneous URDF reconstruction. Using the provided calibration still left scan drift. compact308 then used folded default arms for transport but failed before collecting a can, Q=0,2295 steps. It stalled, tipped briefly, and then exhausted route attempts.

25. visual308 tested bounded virtual-joint velocity prediction and accepting strong visual corrections outside the odometry agreement gate. It was stopped after diagnostic poses showed severe tipping and falling during contact. This did not establish an improvement; it is not an official completed evaluation. All experimental source is preserved in debug/revision_final_experimental13/ and earlier revision directories.

26. Paused as allowed by the user after several navigation revisions failed to improve308. Restored the default policy/configuration from the completed standalone runs/validation_pack09/source/ snapshot. Later experiments remain separate. The two-instance official mean is0.5; the requested 20-instance mean >0.7 remains unachieved and unmeasured. No full20 sweep was launched. The persistent diagnostic simulator was shut down cleanly. The next development direction is a stronger visual-navigation module or a proven observation-only navigation baseline.

No full 20-instance evaluation has been run. The >0.7 mean-Q objective is not yet demonstrated. A standalone copy is maintained at 5080-1:/shared_work/markhsp/gpt6_b1k; measured simulator results here come from the laptop.

New visual navigation, requested after the plateau

The user selected building a new visual navigation module. Development resumed in visual_navigation.py, while policy.py retains the validated controller. The module uses official RGB-D, proprioception, and camera-relative calibration; dense obstacle points, observed-floor routing, compact transport posture, short forward segments, and a current-depth stopping corridor. It excludes robot surfaces using static URDF/mesh bounds. No object or world pose is read by the policy.

• debug/labnav14/nav308a: stopped before translating,650 steps. The initial scan accumulated33cm false visual translation, leaving no allowed route from its estimated position.
• nav308b:334 steps; retaining a single long-lived visual anchor lost feature correspondence. The policy stopped before translating because localization was unavailable.
• nav308c: navigation-only route completed in1710 steps,3.069m base travel. Final position error11.39cm, yaw error0.0642rad; robot upright (roll/pitch under0.001rad). Q=0 because this probe intentionally did no manipulation. Source saved in nav308c_source/. The tracker combines short-frame registration with older-view loop closure and constrains translation during zero-translation commands with low measured base velocity.
• task308d: full pickup/delivery development test started; report only its completed result when available.

Thirteen contract/navigation checks pass, including unknown-floor gaps, narrow clearance, thin-leg detection, robust RGB-D registration, and preserving trusted anchors when registration fails.

task308d was stopped at4905 steps for placement diagnosis, Q=1/3 (diagnostic, not a complete official episode). It navigated5.406m and made two releases. The first can missed because the bin moved about10cm during the arm motion; the second delivery counted. At the first release, the observed bin centre agreed with debug geometry to roughly2mm before the motion, so the failure was physical target motion, not initial recognition.

placement308e_from_d is a reset=false continuation from that partial state, solely to test higher horizontal transport, vertical Cartesian descent, rechecking the bin before release, and re-queuing visible misses. It must not be reported as a fresh Task308 evaluation. The already missed can was occluded/merged with the bin in the initial head views; the policy selected the remaining distant can.

The continuation stopped at its inherited official timeout (4905+2997=7902 total environment steps), retaining Q=1/3. Its first revised hover was IK-infeasible; no release occurred in this continuation. Offline IK checks found a feasible .76m-forward/.60m-high hover, with orientation changed only after extension. A static interpolation check of the original first-drop pose found a gripper-finger bounding box reaching z=.244m over the bin, below its roughly .28m rim; the revised sequence kept the corresponding bound above .416m. These are conservative static geometric checks, not a replacement for simulation. task308f starts a fresh reset with that corrected sequence. Fourteen tests pass locally and on5080-1.

task308f completed its7901-step diagnostic horizon at Q=2/3, carrying the third can at the bin approach. Both releases were successful; pre-release bin shifts were0.64mm and4.66mm. No navigation or grasp recovery was logged. Base travel8.418m. A fresh task308g tests skipping already settled posture commands, two-frame joint settling, and a modestly higher forward cap (.36 action units), retaining dense depth stopping and adding a measured folded-arm check.

task308g was stopped at2942 steps (Q=0, holding the first can) after diagnosing a grid-goal convergence edge case: a rounded waypoint lay within.075m but the continuous goal remained just outside.10m, causing repeated zero-progress segments. The new endpoint uses the exact goal; an end-to-end kinematic regression reproduces this case and now reaches it. Fifteen tests pass. task308h is a fresh reset with this correction and the timing improvements.

task308h stopped at5028 steps, Q=1/3, retaining its second can after an unreachable high-hover target (.974m forward, .60m high). Offline checks from the recorded proprioception confirmed that .974m forward at .55m high was reachable. The subsequent revision adds short precise translations (with observed-map and directional depth checks) for final alignment and selects the highest feasible hover above the rim. Sixteen tests pass, including the sideways stopping corridor. task308i is a fresh reset with these changes.

task308i: fresh-reset autonomous development run, success=true, Q=1.0,6758 steps,9.145m base travel. All three cans were delivered under the official horizon. Pre-release bin shifts were1.01mm,0.73mm,and5.76mm. No task recovery was logged. Source frozen in debug/labnav14/task308i_source/. This uses the official evaluator/observations/metrics inside the development harness; it is not yet a standalone20-instance score. The identical policy is now being checked on320 (task320j) before the broader official evaluation.

task320j: Q=0 at2042 steps; untargeted exploration went away from the cans, then a newly observed obstacle invalidated its route. task320k executed zero steps due to a stale developer stop flag and is excluded. task320k2 used can-directed exploration: it found the bin, made two deliveries, and ended at7901 steps with Q=2/3. Offline map/pose analysis showed that the collected cans remained as ghost obstacles across the corridor to the last can. Final pose error12.2cm, max16.5cm; no collision/tipping issue. The next revision separates low/tall occupancy and clears low cells when floor is observed, while preserving taller furniture. It also handles invalidated exploration routes by choosing another viewpoint. Seventeen tests pass; task320l is the fresh simulation check.

task320l: fresh diagnostic reset, Q=2/3 at 7901 steps. Clearing observed low obstacles opened the corridor, but the last-can approach followed a false furniture proposal. Offline replay classified 121 true proposals and three false proposals using separate debug geometry. A stricter discovery-only body-color filter retained all 121 true proposals and rejected the three false proposals. Known-target reacquisition still accepts mostly silver lids. This replay result does not establish performance on unseen instances.

continuous_nav308m: stopped at 610 steps before translating (base travel 0.010 m), Q=0. No visual registration was accepted, including before the new follower started. The initial reset image differed from subsequent frames and left a stale initial anchor. Added ten stationary startup frames before any visual anchor is initialized. The continuous follower passes grid tests for unknown-space and corner cutting, but has not yet completed a simulator route. The upgraded developer harness now archives each command/source and observed map, and reloads the path follower between trials. Current local CPU checks: 21 passed. debug/labnav15/nav308a tests startup and continuous navigation on a fresh simulator.

debug/labnav15/nav308a: visual initialization fixed; short probe stopped at 611 steps when a new view invalidated the route. nav308b adds approach/frontier recovery and route-coverage scans: navigation complete in 959 steps, 3.089 m base travel, 10.66 cm final position error (max 10.66 cm, median 2.70 cm). This is navigation-only, Q=0 by design. The earlier short-segment probe took 1710 steps for 3.069 m, with 11.39 cm final error. The two probes can choose different observed approaches and are not identical-path timing benchmarks. task320c tests the continuous follower and stricter discovery in a complete diagnostic episode.

task320c: Q=2/3 at 7902 steps, 14.831 m travel. Two releases by step 4550, but pose drift reached 54 cm near the third can and route-coverage scans consumed the remaining horizon. Replay of 25 older-view matches found maximum relative yaw error 0.00662 rad and translation error 1.03 cm; some useful reference views were older than the live 30-keyframe history. The next revision keeps up to 200 keyframes (approximately one horizon), preserves live object-position references during re-observation, and excludes in-place turning from the translation-stall window. task320d is the fresh test. No 20-instance sweep has been launched.

task320d: stopped at 5776 steps, Q=2/3, 10.922 m travel. Both releases counted. At step 4331 the next selection preferred a one-view furniture proposal at episode [-0.361, 3.840] over the true blue can observed twice at approximately [-0.639, -3.927]. The final candidate prioritizes multi-view-confirmed tracks before one-off proposals. It retains single-view candidates as a fallback rather than rejecting small / distant true cans. Current local and remote tests: 22 passed before this selection change; 22 local checks still pass afterward. A single frozen standalone all-20 evaluation is next; it will measure this candidate without further per-instance edits.

runs/visual_frozen20_16: launched the frozen standalone visual candidate for all 20 official indices, then stopped during the second instance after repeated failures. Instance 301 completed at Q=1/3, 7902 steps; its carry route became unreachable and placement recovery stopped at step 3199. Instance 302 was interrupted at policy step 4265 after target reacquisition failures, route invalidations, and localization loss. It has no completed official Q. Instances 303–320 were not run in this sweep. The completed result is retained; no all-20 mean is claimed. The simulator/evaluator and policy server were shut down. Paused under the user's allowance to stop and ask after several rounds without improvement. Latest source, manifest, summary, action/event logs and sensor recordings are preserved. Final local and remote CPU checks: 22 passed.

Fixed-scene prior redesign, user-authorized September 12

The user clarified that all Task 01 instances share the scene and explicitly authorized offline mapping and storing the map in the code. Implemented learned_localization.py and known_scene.py: local ALIKED/LightGlue reference matching, robust metric SE(2), frozen static occupancy with a rebuilt live depth layer, and map-aware target prioritization. Offline debug robot poses align RGB-D survey frames; debug task-object poses only exclude carried cans / bin from exported static data. Runtime reads official sensors plus explicit prior assets; no live truth or backend calls.

Survey: 279 stable frames, 57 reference views. Held-out replay after masking: 42/45 localized, 37 with at least 40 geometric inliers; median position error 3.23 mm, max 28.50 mm, median 34.4 ms per query. Torch URL weight loading was disabled during replay. debug/labprior17/nav301a: navigation-only 905 steps; after localization, max position error 26.63 mm, final 3.35 mm. Full task301b is running. CPU regression checks: 26 passed.

task301b ended at 5274 steps, Q=2/3, with exhausted target attempts. A repeated-floor alias caused exact 1.35–1.36 m global position jumps during close approaches; this explained the apparent reacquisition failures. The next revision requires at least eight non-floor geometric inliers for an absolute fix and rejects large discrepancies from a trusted tracking prediction. Replay on 35 sampled views from this failure run: 29 localized, max position error 28.52 mm, no accepted metre-scale aliases. A synthetic 100-match repeated-floor regression is rejected despite perfect planar geometric fit. 27 CPU tests pass. task301c is the fresh integrated test.

debug/labprior17/task301c: fresh-reset development success, Q=1.0, 5745 steps, 15.494 m base travel. All three cans delivered, no task recovery events. The structural-support and tracking-consistency gates removed the metre-scale floor aliases seen in task301b. This result uses the official evaluator metrics in the debug harness; standalone all-20 validation is still pending. The unchanged policy is now testing instance 320 as task320d.

debug/labprior17/task320d: Q=2/3,6137 steps,12.461m travel. Both early releases counted. The true third can was selected, but its final base alignment toward the fixed .72m reach was blocked by furniture clearance. Localization error was4.56mm median,46.35mm maximum,9.57mm final; this failure was not accumulated metre-scale map drift. A separate lost-localization interval consumed five rapid retries without changing the view. Offline IK/route replay showed a clear .92m-reach stance at step5399 and a reachable stationary .948m grasp at step6070. The next revision adds a prior-mode alignment fallback that checks hover/descent/lift IK without relaxing map clearance, and a zero-translation view search before retrying a lost-localization route. 29 CPU checks pass. Fresh-reset task320e is testing these changes.

Offline initial-state coverage check (debug/prior_instance_coverage.json): all20 robot starts lie in allowed map cells. All task-object initial points are within.20m of allowed cells except one can on317 at1.0m; this is diagnostic geometry, never read by the policy. These distances do not establish visual coverage, arm clearance, or episode success.

task320e: Q=2/3 at7902 steps,16.891m travel. The third can was successfully grasped at0.915m reach and lifted at7173, but the episode timed out while carrying it to the bin. Two earlier attempts were blocked by current depth despite map-clear stance proposals. Stationary view-search restored localization. The frozen candidate now tests hover/descent/lift reachability before any nominal base correction, avoiding unnecessary forward motion when the current stance is already reachable. 29 CPU tests pass; this final ordering change has offline evidence but is first undergoing integrated validation in the standalone sweep. The CLI default is now prior; older modes require explicit selection.

runs/prior_frozen20_18: standalone301 Q=1.0/5571 steps and302 Q=1.0/6283 steps, both with no task recovery. Stopped during303: after one release and the second successful lift, final bin alignment hit a depth obstacle at2815; the loaded-arm compact-posture retry failed at2867. No completed official Q for303;304–320 not run. Frozen source and interruption.json retained. Sensor-only IK replay at2770 found the existing stance could place with a.50m hover; at2815 a.55m hover was reachable. The new prior-mode bin alignment checks the entire vertical descent before base correction, and reassesses current reachability after an interrupted correction. 31 CPU tests pass. debug/labprior19/task303a is the fresh integrated test.

debug/labprior19/task303a: development Q=1.0, success,5533 steps,12.115m base travel. One second-delivery retry occurred because the preparatory wrist rotation changed the torso posture, making a single large extension fail its per-step joint-change bound. Sequential offline IK from the recorded2770 observation reached the1.022m-forward hover through four segments without changing absolute joint limits. Added a copied-state sequential extension/descent/retreat planner and short executed Cartesian segments in prior mode. 31 CPU tests pass. task303b validates the complete sequence on a fresh reset.

task303b: regression, Q=1/3 at2909 steps. Segmented high-hover extension made an endpoint reachable in kinematic planning but actual torso tracking failed (8.3cm EEF error) on the second delivery; compact-arm recovery also failed. The segmented extension/retreat overrides were removed. The next candidate retains the successful task303a stance policy and validates the actual preparation + single bounded extension + descent + retreat sequence on copied proprioceptive joint state. Offline replay rejects the unsafe far stance at2770 and accepts the actual depth-stopped stance at2815 with a.55m hover. No joint bounds or collision clearances were relaxed.31 tests pass. Fresh task303c is running.

debug/labprior19/task303c: Q=1.0, success,5026 steps,11.657m base travel,zero task recovery events. The second bin approach hit the depth stop, re-observed the bin at.965m, and continued with a reachable.55m hover from the actual stopped stance. Three releases completed. This is507 steps fewer than task303a; source and observations are preserved. Freeze this candidate for the next complete standalone20-instance evaluation. Full video encoding is now opt-in, matching the official default; official RGB-D observations and scoring are unchanged, and sampled observations/action logs are always retained.

runs/prior_frozen20_20/batch_000: completed301 Q=1.0/5596,302 Q=1.0/6123,303 Q=1.0/5136. During304, unexpected additional reset messages cleared shared policy state while predict was running, causing a websocket error;304 has no completed official Q. No policy source changes: the runner now selects a free loopback port, and the same frozen run resumes missing indices only. Transport failure details are archived in batch_000/transport_failure.json.

runs/prior_frozen20_20: complete frozen standalone evaluation, all20 official public instances301–320, mean final Q=0.766666667, 13/20 official successes; target >0.7 achieved. No horizon override, seed0, fullRGB-D wrapper. Batch000 completed301–303; identical policy resumed missing304–320 in batch001 after the documented transport interruption. All completed failures are retained. All673 frozen policy/asset/dependency hashes and the clean official evaluator were verified. Full table and remaining limitations: EVALUATION.md.