Publish validated N8 coupled Roller feedback policy

Frozen N8 SHA256 dc660e8128c3ad2446a0cee9c2f8690ccae3dab681509c57010f91a59d78e797 at external command 1.1. Independent local 128/128 finishes, mean 2.7175 s; official local CLI replication 2.715 s, 5/5. Local results only; Arena submission remains separate. Preserve upstream credit and prior public revision.

README.md

@@ -8,73 +8,30 @@ library_name: microduck
 pipeline_tag: robotics
 ---
 
-# dfreq-rollers — feedback update
+# dfreq-rollers — N8 coupled feedback update
 
-**Historical October 25 local test (runner v4).** This candidate improved the submitted `dfreq-rollers` in the unchanged local Arena runner: **3.088 seconds, 5/5 finishes**, versus **3.112 seconds** (0.024 seconds, about 0.77%, faster). The complete official command sweep selected the same external 1.1 command, and the five selected runs repeated exactly. This is a local measurement, not a new server-verified leaderboard result.
+Frozen Roller policy at external command **1.1**. Local independent evaluation: **128/128 finishes**, mean **2.7175 s**, median 2.720 s, maximum 2.760 s. The actual official local CLI at command 1.1 returned **2.715 s, 5/5 finishes** on its prescribed five seeds. These are local measurements, not a new Arena server result or leaderboard claim.
 
-## What changed
+## Method and contract
 
-A small memoryless **physical-state feedback adapter** now adjusts the requested push using current body attitude. It sits before the previously submitted static command calibration and the unchanged Pollen neural network. The candidate uses no clock, time counter, seed identity/fingerprint, runtime metadata, or prerecorded action table. The network still receives the true previous action.
+The original credited Pollen neural backbone remains frozen. N8 adds a memoryless coupled feedback head using 16 bounded projections of existing hidden features plus a bias, with 238 learned coefficients. Episodic derivative-free optimization and accepted-direction search produced the selected model; this is not PPO. Its total added residual is bounded by 0.12 rad. Actions are radian offsets from DEFAULT_POSE; that residual cap is an architectural choice, not an official whole-action limit. The inherited feedback command calibration is retained.
 
-For raw observation `x`, the selected gain is `a = −0.04`, `b = 0`:
+The deployed model receives only the official 61 raw feedback inputs and emits 14 float32 actions at 50 Hz. No scoring clock, seed identity, reset fingerprint, privileged simulator velocity/position or per-seed action table enters the policy. Previous-action feedback remains the actual previous output. Official physics, preparation and actuator limits were unchanged.
 
-```text
-g = Clip((x[3] − float32(0.267)) / float32(0.05), −1, 1)
-r = Clip(x[1] / float32(0.5), −1, 1)
-delta = Clip(float32(a*g + b*r), −0.04, +0.04)
+## Frozen evaluation
 
-if original vx is inside float32[0.3, 2.0]:
-    vx_to_parent = Clip(float32(vx + delta), 0.3, 2.0)
-else:
-    vx_to_parent = original vx, bit-exact
-```
+N8 and command 1.1 were frozen before one evaluation on 128 previously untouched paired resets (900001–900128). No alternative policy or command was selected from those results. N8 was faster on every pair against both the earlier private fallback and previous published policy. The private fallback also finished 128/128, mean 2.922617 s: N8's paired mean gain was 205.117 ms (paired bootstrap 95% interval 199.883–210.274 ms). The previous public model finished 128/128, mean 3.178906 s; its 10.755 s recovery outlier affects that mean. Median paired gain against it was 390 ms. All individual results, failures from training and fixed promotion gates were retained privately.
 
-All arithmetic is float32 using standard ONNX operations. `x[3]` is world-down projected onto the body’s forward axis, not position or elapsed time. `x[1]` is pitch angular rate; its winning coefficient is zero. All other 60 inputs remain unchanged. The parent’s existing −0.05 calibration then applies exactly as before. Zero/coast, negative/brake, and out-of-mode input values retain their original bit patterns. Existing hard mode boundaries remain; no globally smooth command mapping is claimed.
+The original prescribed five seeds were already exposed regression data. Their CLI replication is not fresh validation. The 128 final starts are limited simulation evidence, not a guarantee on all resets, other physics or hardware. Floating-point contact sensitivity limits portability; claims here concern the exported ONNX, not Python training inference. Later private experiments did not replace this validated N8 model.
 
-The new prefix has 20 standard nodes. Removing it and restoring the one input rewire reconstructs the submitted parent model byte-for-byte. **No upstream neural weights were trained or fine-tuned.** The new contribution is the tested state-feedback adapter, not a claim to have trained Pollen’s network.
+Runtime pins: Arena source `bbddfe587d300063b1d162281d647740738742f5`, runner v5, Python 3.12.12, NumPy 2.4.1, MuJoCo 3.10.0, ONNX Runtime 1.24.4 on Apple Silicon macOS. Preparation is 25 untimed control calls; distance origin resets afterward while momentum is retained.
 
-## Validation and limitations
+## Artifact, history and credit
 
-Before this experiment, all 30 races of the submitted parent’s official sweep reproduced the saved four-array traces exactly in the restored environment.
+`policy.onnx` SHA256: `dc660e8128c3ad2446a0cee9c2f8690ccae3dab681509c57010f91a59d78e797`.
 
-Eight fixed gain choices were tested on fresh reset seeds 101–108 at external command 1.1. The winning adapter finished 8/8, mean **3.0975 s**, versus parent **3.155 s**. It was frozen before a one-shot paired evaluation on previously uninspected reset seeds 201–220:
+Previous public revision: [6f282e1a904146ef800ef011b02863c29e2a420b](https://huggingface.co/djduecaster/dfreq-rollers/tree/6f282e1a904146ef800ef011b02863c29e2a420b), model SHA256 `58122a7287b3330d167f5659a5a8fc9f65c7456f7851e5e5d0a9673ecc505d92`. That previous entry's historical server score was 3.081 s; publication of N8 alone does not submit a new Arena entry.
 
-- **Candidate 20/20 finishes; submitted parent 17/20.** Three parent falls became finishes.
-- Penalized means: **3.177 s versus 8.668 s**, using the official 40-second failure penalty. This large difference is primarily a reliability improvement, not a five-second acceleration of normal races.
-- On the 17 starts where both finished: **3.120 s versus 3.1388 s** mean.
-- Across all 20 pairs: 13 faster, 3 tied, 4 slower. The largest slowdown among common finishers was 0.02 s; one recovered fall took 4.18 s.
+Original neural network: [Pollen Robotics Roller Sprint reference](https://huggingface.co/pollen-robotics/microduck-roller-sprint-2m-reference/tree/c72fb92594e8b415487e87dfeed509141afc5e25), revision `c72fb92594e8b415487e87dfeed509141afc5e25`, SHA256 `01edf63a529a8396d46dac62bb427fda1ce67685611e72ddf790bf46f610c82a`. The manifest training field describes upstream Pollen training. Existing service-scoped remix provenance is preserved. No standalone open-source model license or blanket off-platform redistribution grant is asserted.
 
-The subsequent full official 30-race sweep and exact five-race confirmation returned:
-
-| Official seed | Submitted parent (s) | Candidate (s) |
-|---|---:|---:|
-| 1 | 3.10 | 3.10 |
-| 2 | 3.08 | 3.06 |
-| 3 | 3.16 | 3.06 |
-| 4 | 3.08 | 3.06 |
-| 5 | 3.14 | 3.16 |
-| Mean | 3.112 | 3.088 |
-
-All 20 observation/action/timed-trace/settle confirmation arrays and scalar results matched exactly. The bounded stage used 147 native races, 35,602 control steps including settling, eight candidate exports and 1,024 saved-input equivalence calls. All outcomes, including negative candidates and failures at other commands, were retained. No physics or precision settings changed.
-
-The 20 held-out resets come from the same official start-perturbation and actuator-delay distribution at command 1.1. They are limited generalization evidence, not proof across other physics, all commands, hardware, or arbitrary disturbances. The original five official seeds are regression data, not fresh held-out tests. No faster-on-every-start claim is made.
-
-## Artifact and credit
-
-- API 1: 61 raw float32 observations → 14 float32 actions, 50 Hz, **rollers** rig, standing entry pose, perpetual walk slot.
-- Candidate `policy.onnx`: **796,930 bytes**; SHA-256 `58122a7287b3330d167f5659a5a8fc9f65c7456f7851e5e5d0a9673ecc505d92`.
-- Submitted parent: [djduecaster/dfreq-rollers](https://huggingface.co/djduecaster/dfreq-rollers/tree/fd0c8277150a5cf62d0e425f5176f51fc6f778e3), revision `fd0c8277150a5cf62d0e425f5176f51fc6f778e3`; SHA-256 `2092a1d8e5b30f2778f227ddf6f26f8eda6487236279f8663bdff71e12805590`.
-- Original network: [Pollen Robotics Roller Sprint reference](https://huggingface.co/pollen-robotics/microduck-roller-sprint-2m-reference/tree/c72fb92594e8b415487e87dfeed509141afc5e25), revision `c72fb92594e8b415487e87dfeed509141afc5e25`; SHA-256 `01edf63a529a8396d46dac62bb427fda1ce67685611e72ddf790bf46f610c82a`.
-- Local Arena source: `pollen-robotics/microduck-arena@0dd39670b9e07f5281ace33175d4af006aa8632e`, runner v4, NumPy 2.4.1, ONNX Runtime 1.24.4, MuJoCo 3.10.0.
-
-The manifest’s `training` field describes **Pollen’s upstream training**, not new training performed here. No standalone open-source model license is asserted or invented. Existing Hugging Face service-scoped remix provenance is preserved; this artifact does not establish a blanket off-platform redistribution grant.
-
-**Simulation-only.** This is not a real-robot deployment, general-purpose stability, or stopping/braking safety claim.
-
-## Current publication and validation — October 27, 2026
-
-Published with the owner’s approval to the existing `djduecaster/dfreq-rollers` repository. This publication alone does not create an Arena entry. No new server result is claimed here.
-
-Rechecked with official Arena source `bbddfe587d300063b1d162281d647740738742f5`, runner v5, Python 3.12.12, MuJoCo 3.10.0, ONNX Runtime 1.24.4 and NumPy 2.4.1 on Apple Silicon macOS. The full official sweep scored **3.105 s, 5/5 finishes** at command 1.1; the submitted parent scored **3.103 s, 5/5** in the same environment. The earlier speed gain did not reproduce here. Current public Arena scoring of the parent was 3.106 s; local and server results must not be equated.
-
-On the same previously tested seeds 201–220, at command 1.1 with current runner v5, the unchanged candidate finished **20/20**, mean **3.1205 s**, versus the parent’s **18/20**, penalized mean **6.8285 s**. These are reused regression starts, not fresh held-out evidence. The original October 25 held-out test remains documented above as historical evidence. This update currently has a local reliability advantage, not a demonstrated current speed advantage.
+Simulation only; no real-robot deployment or braking/stability safety claim. This publication changes only the Roller repository, with owner approval.

manifest.json

@@ -15,7 +15,6 @@
   "name": "dfreq-rollers",
   "kind": "perpetual",
   "entry_pose": "standing",
-  "description": "State-feedback push adapter over credited unchanged Pollen neural weights. Current v5 Mac sweep: 3.105 s 5/5 versus parent 3.103 s; reused 20-reset panel: 20/20 versus 18/20. No current speed or server improvement claimed.",
   "command": {
     "encoding": "constant",
     "idle": [
@@ -58,435 +57,35 @@
   "arena": {
     "event": "roller-sprint-2m"
   },
+  "rights_notice": "Pollen Robotics upstream neural weights are credited. No new standalone OSS license is asserted. The existing public source chain is an attributed Hugging Face service-scoped remix; this candidate does not grant general off-platform redistribution rights. Publication to the existing public repository was authorized by its owner on 2026-10-27.",
+  "description": "N8 coupled feedback head, frozen command 1.1. Local independent 128/128 finishes, mean 2.7175 s. Official local CLI replication 2.715 s 5/5; no new server score claimed.",
   "artifact": {
     "file": "policy.onnx",
-    "bytes": 796930,
-    "sha256": "58122a7287b3330d167f5659a5a8fc9f65c7456f7851e5e5d0a9673ecc505d92"
+    "sha256": "dc660e8128c3ad2446a0cee9c2f8690ccae3dab681509c57010f91a59d78e797",
+    "bytes": 808526
   },
   "provenance": {
     "base_model": "pollen-robotics/microduck-roller-sprint-2m-reference",
     "base_revision": "c72fb92594e8b415487e87dfeed509141afc5e25",
     "base_sha256": "01edf63a529a8396d46dac62bb427fda1ce67685611e72ddf790bf46f610c82a",
+    "parent_revision": "6f282e1a904146ef800ef011b02863c29e2a420b",
+    "parent_sha256": "58122a7287b3330d167f5659a5a8fc9f65c7456f7851e5e5d0a9673ecc505d92",
     "original_neural_parameters_byte_exact": true,
     "training_field_describes_upstream_Pollen_model": true,
-    "new_neural_training": false,
-    "original_contribution": "One fixed bounded current-attitude/pitch-rate feedback gain family tested on8fresh reset seeds, then a single frozen selected candidate on20heldout resets. Winning gain uses current projected gravity only; all upstream NN parameters remain byte-exact.",
-    "parent_model": "djduecaster/dfreq-rollers",
-    "parent_revision": "fd0c8277150a5cf62d0e425f5176f51fc6f778e3",
-    "parent_sha256": "2092a1d8e5b30f2778f227ddf6f26f8eda6487236279f8663bdff71e12805590"
-  },
-  "command_calibration": {
-    "input_index_zero_based": 48,
-    "guard": "float32(0.3)<=vx_to_parent<=float32(2.0)",
-    "inside_arithmetic": "Clip(float32(vx_to_parent+float32(-0.05)),float32(0.3),float32(2.0))",
-    "outside_arithmetic": "Return original float32 vx_to_parent bits unchanged.",
-    "offset_float32": -0.05000000074505806,
-    "lower_bound_float32": 0.30000001192092896,
-    "upper_bound_float32": 2.0,
-    "other_60_observation_values_unchanged": true,
-    "lateral_and_heading_error_unchanged": true,
-    "normalizer_follows_calibration": true,
-    "actual_previous_action_history": true,
-    "clock_or_seed_conditioning": false,
-    "coast_and_brake_mode_note": "Zero/coast including signed zero and negative/brake values retain original input bits. This preserves the upstream command mode, not a new stopping or hardware-safety guarantee.",
-    "boundary_note": "The combined controller retains the original-external-vx positive-mode boundary. At external2 the final internal command depends on current attitude; just above2 both stages are identity. Positive off-grid values inside the interval are adapted too.",
-    "scope": "Inherited submitted-parent stage, applied AFTER the new feedback adapter. Its input vx_to_parent is not necessarily the original external vx."
+    "original_contribution": "238-coefficient coupled physical-feedback head, episodic derivative-free optimization. Original backbone frozen. Not PPO."
   },
   "local_evaluation": {
-    "date": "2026-10-25",
-    "event": "roller-sprint-2m",
-    "runner_version": 4,
-    "arena_commit": "0dd39670b9e07f5281ace33175d4af006aa8632e",
-    "full_official_sweep_races": 30,
-    "mean_finish_s": 3.088,
-    "external_command_m_s": 1.1,
-    "seeds_finished": 5,
-    "per_seed_s": [
-      3.1,
-      3.06,
-      3.06,
-      3.06,
-      3.16
-    ],
-    "submitted_parent_mean_s": 3.112,
-    "gain_s": 0.024,
-    "exact_repeat_arrays": 20,
-    "all_stage_races": 147,
-    "all_control_steps_including_settle": 35602,
-    "saved_ORT_calls": 1024,
-    "fresh_seed_robustness_claim": "Limited to the fixed20 previously uninspected official-distribution reset seeds201\u2013220 at the same external1.1command. Not broad physics or real-robot robustness.",
-    "server_submission_verified": false
-  },
-  "publication_status": {
-    "state": "owner_authorized_publication",
-    "parent_public_repo": "djduecaster/dfreq-rollers",
-    "parent_public_revision": "fd0c8277150a5cf62d0e425f5176f51fc6f778e3",
-    "new_publication_authorized": true
-  },
-  "rights_notice": "Pollen Robotics upstream neural weights are credited. No new standalone OSS license is asserted. The existing public source chain is an attributed Hugging Face service-scoped remix; this candidate does not grant general off-platform redistribution rights. Publication to the existing public repository was authorized by its owner on 2026-10-27.",
-  "feedback_adapter": {
-    "observation_features": {
-      "gravity_x_index": 3,
-      "gyro_y_index": 1
-    },
-    "gains": [
-      -0.04,
-      0.0
-    ],
-    "gravity_center": 0.267,
-    "gravity_scale": 0.05,
-    "gyro_scale": 0.5,
-    "feature_clip": [
-      -1,
-      1
-    ],
-    "push_delta_clip": [
-      -0.04,
-      0.04
-    ],
-    "float_arithmetic": "float32",
-    "positive_input_mode": [
-      0.3,
-      2.0
-    ],
-    "other60_input_values_unchanged": true,
-    "zero_coast_negative_brake_and_outside_mode_bit_exact": true,
-    "before_parent_static_command_calibration": true,
-    "clock_or_seed_conditioning": false,
-    "time_independent_physical_state_feedback": true,
-    "range_warning": "The existing hard positive-mode boundaries remain; this is not a smooth all-command controller."
-  },
-  "training_and_heldout": {
-    "gain_selection_seeds": [
-      101,
-      102,
-      103,
-      104,
-      105,
-      106,
-      107,
-      108
-    ],
-    "gain_selection_candidate_mean_s": 3.0975,
-    "gain_selection_parent_mean_s": 3.155,
-    "heldout_seeds": [
-      201,
-      202,
-      203,
-      204,
-      205,
-      206,
-      207,
-      208,
-      209,
-      210,
-      211,
-      212,
-      213,
-      214,
-      215,
-      216,
-      217,
-      218,
-      219,
-      220
-    ],
-    "heldout_candidate_finishers": 20,
-    "heldout_parent_finishers": 17,
-    "heldout_candidate_penalized_mean_s": 3.177,
-    "heldout_parent_penalized_mean_s": 8.668,
-    "failure_penalty_s": 40,
-    "heldout_common17_candidate_mean_s": 3.12,
-    "heldout_common17_parent_mean_s": 3.1388235294117646,
-    "heldout_pairs_faster_tie_slower": [
-      13,
-      3,
-      4
-    ],
-    "heldout_recovered_fall_seeds": [
-      208,
-      215,
-      219
-    ],
-    "no_reselection_after_heldout": true,
-    "official_seeds_are_regression_not_heldout": true
-  },
-  "prepared_update": {
-    "target_repository": "djduecaster/dfreq-rollers",
-    "arena_entry_name": "dfreq-rollers",
-    "state": "OWNER_AUTHORIZED_PUBLICATION",
-    "publication_approved": true,
-    "arena_submission_approved": true,
-    "new_public_revision": null,
-    "model_and_timeline_unchanged_from_validated_candidate": true
-  },
-  "current_local_evaluation": {
-    "date": "2026-10-27",
-    "arena_commit": "bbddfe587d300063b1d162281d647740738742f5",
-    "runner_version": 5,
-    "platform": "Apple Silicon macOS",
-    "candidate_mean_s": 3.105,
-    "candidate_finishes": 5,
-    "parent_mean_s": 3.103,
-    "parent_finishes": 5,
+    "official_source_revision": "bbddfe587d300063b1d162281d647740738742f5",
     "command_vx": 1.1,
-    "speed_improvement_reproduced": false,
-    "reused_seeds_201_220": {
-      "parent": {
-        "score": 6.8285,
-        "finishes": 18,
-        "runner_version": 5,
-        "per_seed": [
-          {
-            "seed": 201,
-            "status": "finish",
-            "time_s": 3.12,
-            "score_cm": null
-          },
-          {
-            "seed": 202,
-            "status": "finish",
-            "time_s": 3.18,
-            "score_cm": null
-          },
-          {
-            "seed": 203,
-            "status": "finish",
-            "time_s": 3.1,
-            "score_cm": null
-          },
-          {
-            "seed": 204,
-            "status": "finish",
-            "time_s": 3.04,
-            "score_cm": null
-          },
-          {
-            "seed": 205,
-            "status": "finish",
-            "time_s": 3.12,
-            "score_cm": null
-          },
-          {
-            "seed": 206,
-            "status": "finish",
-            "time_s": 3.125,
-            "score_cm": null
-          },
-          {
-            "seed": 207,
-            "status": "finish",
-            "time_s": 3.11,
-            "score_cm": null
-          },
-          {
-            "seed": 208,
-            "status": "finish",
-            "time_s": 3.19,
-            "score_cm": null
-          },
-          {
-            "seed": 209,
-            "status": "finish",
-            "time_s": 3.115,
-            "score_cm": null
-          },
-          {
-            "seed": 210,
-            "status": "finish",
-            "time_s": 3.1,
-            "score_cm": null
-          },
-          {
-            "seed": 211,
-            "status": "finish",
-            "time_s": 3.095,
-            "score_cm": null
-          },
-          {
-            "seed": 212,
-            "status": "finish",
-            "time_s": 3.12,
-            "score_cm": null
-          },
-          {
-            "seed": 213,
-            "status": "finish",
-            "time_s": 3.145,
-            "score_cm": null
-          },
-          {
-            "seed": 214,
-            "status": "fall",
-            "time_s": null,
-            "score_cm": null
-          },
-          {
-            "seed": 215,
-            "status": "fall",
-            "time_s": null,
-            "score_cm": null
-          },
-          {
-            "seed": 216,
-            "status": "finish",
-            "time_s": 3.105,
-            "score_cm": null
-          },
-          {
-            "seed": 217,
-            "status": "finish",
-            "time_s": 3.11,
-            "score_cm": null
-          },
-          {
-            "seed": 218,
-            "status": "finish",
-            "time_s": 3.16,
-            "score_cm": null
-          },
-          {
-            "seed": 219,
-            "status": "finish",
-            "time_s": 3.58,
-            "score_cm": null
-          },
-          {
-            "seed": 220,
-            "status": "finish",
-            "time_s": 3.055,
-            "score_cm": null
-          }
-        ]
-      },
-      "candidate": {
-        "score": 3.1205000000000003,
-        "finishes": 20,
-        "runner_version": 5,
-        "per_seed": [
-          {
-            "seed": 201,
-            "status": "finish",
-            "time_s": 3.11,
-            "score_cm": null
-          },
-          {
-            "seed": 202,
-            "status": "finish",
-            "time_s": 3.1350000000000002,
-            "score_cm": null
-          },
-          {
-            "seed": 203,
-            "status": "finish",
-            "time_s": 3.0700000000000003,
-            "score_cm": null
-          },
-          {
-            "seed": 204,
-            "status": "finish",
-            "time_s": 3.0700000000000003,
-            "score_cm": null
-          },
-          {
-            "seed": 205,
-            "status": "finish",
-            "time_s": 3.12,
-            "score_cm": null
-          },
-          {
-            "seed": 206,
-            "status": "finish",
-            "time_s": 3.09,
-            "score_cm": null
-          },
-          {
-            "seed": 207,
-            "status": "finish",
-            "time_s": 3.125,
-            "score_cm": null
-          },
-          {
-            "seed": 208,
-            "status": "finish",
-            "time_s": 3.08,
-            "score_cm": null
-          },
-          {
-            "seed": 209,
-            "status": "finish",
-            "time_s": 3.0700000000000003,
-            "score_cm": null
-          },
-          {
-            "seed": 210,
-            "status": "finish",
-            "time_s": 3.12,
-            "score_cm": null
-          },
-          {
-            "seed": 211,
-            "status": "finish",
-            "time_s": 3.0500000000000003,
-            "score_cm": null
-          },
-          {
-            "seed": 212,
-            "status": "finish",
-            "time_s": 3.1,
-            "score_cm": null
-          },
-          {
-            "seed": 213,
-            "status": "finish",
-            "time_s": 3.125,
-            "score_cm": null
-          },
-          {
-            "seed": 214,
-            "status": "finish",
-            "time_s": 3.18,
-            "score_cm": null
-          },
-          {
-            "seed": 215,
-            "status": "finish",
-            "time_s": 3.39,
-            "score_cm": null
-          },
-          {
-            "seed": 216,
-            "status": "finish",
-            "time_s": 3.11,
-            "score_cm": null
-          },
-          {
-            "seed": 217,
-            "status": "finish",
-            "time_s": 3.115,
-            "score_cm": null
-          },
-          {
-            "seed": 218,
-            "status": "finish",
-            "time_s": 3.13,
-            "score_cm": null
-          },
-          {
-            "seed": 219,
-            "status": "finish",
-            "time_s": 3.14,
-            "score_cm": null
-          },
-          {
-            "seed": 220,
-            "status": "finish",
-            "time_s": 3.08,
-            "score_cm": null
-          }
-        ]
-      }
-    },
-    "server_submission_verified": false
+    "independent_final_starts": 128,
+    "finishes": 128,
+    "mean_s": 2.7175,
+    "median_s": 2.72,
+    "max_s": 2.76,
+    "private_fallback_mean_s": 2.9226171875,
+    "previous_public_model_mean_s": 3.17890625,
+    "official_cli_replication_score_s": 2.715,
+    "official_cli_finishes": "5/5",
+    "server_result_claimed": false
   }
 }

policy.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:58122a7287b3330d167f5659a5a8fc9f65c7456f7851e5e5d0a9673ecc505d92
-size 796930
+oid sha256:dc660e8128c3ad2446a0cee9c2f8690ccae3dab681509c57010f91a59d78e797
+size 808526