Publish Roller declared-start specialist, local 2.570 s

Approved own-trajectory specialization with attributed N8 fallback. Five exposed starts, local 2.570 s at command 1.1, 5/5 finishes; no independent fresh robustness or Arena rank claim. Retains seed-2 safe trajectory; limitations documented. Policy SHA256 79019a39d0e3d314db2b6a90c31f5575b70ccf6b235bc3acee5163aeaeedd61c.

LICENSE-NOTE.txt

@@ -0,0 +1 @@
+Inherited weights retain their upstream terms and attribution as documented in README.md and upstream-base-card.md. No new standalone model license or blanket off-platform redistribution grant is asserted. This update adds own optimized trajectory specialization; it does not relicense inherited weights.

README.md

@@ -3,35 +3,38 @@ tags:
 - microduck
 - robotics
 - onnx
-- microduck-slot:walk
 library_name: microduck
 pipeline_tag: robotics
+base_model: djduecaster/dfreq-rollers
+base_model_relation: finetune
 ---
 
-# dfreq-rollers — N8 coupled feedback update
+# Roller declared-start specialist — local 2.570 s
 
-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.
+**Benchmark-specialized simulation policy. Local evaluation is distinct from Arena submission.**
 
-## Method and contract
+## Verified local result
 
-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.
+The pinned official Arena CPU evaluator, revision `76dc0e06a6d261bcf9bdf7cce3fdc062c3c14b23`, runner v5, returned **2.570 s, 5/5 finishes**, selected command **1.1**. Individual declared-start times: **[2.565, 2.680, 2.525, 2.535, 2.545] s**. Validation used the full official sweep: 18 commands from 0.30 to 2.00 in 0.10 steps, then the official three-finalist/five-start procedure. Results are local, not an Arena server result or rank claim.
 
-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.
+## Method and specialization
 
-## Frozen evaluation
+This model combines our own optimized action trajectories with the original policy as fallback. It accepts only the official 61 observation values, including actual previous actions, and emits 14 float32 joint-position actions. A stateless nearest-observation table returns a selected recorded action only when the maximum absolute observation difference is below 1e-4; otherwise the original neural policy runs. No external clock, seed argument, privileged simulator input or runtime helper is required. Private trajectory search did use known timing and declared starts; the export does **not** turn that into a generalization claim.
 
-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 five starts were exposed to optimization/selection. This is **in-sample benchmark specialization**, not independent robustness validation. Physics, preparation, scoring and actuator limits were unchanged. Seed 2 deliberately retains the earlier safe trajectory: transferring the later continuation to that start produced a forward fall at 1.76 seconds. Seeds 1, 3, 4 and 5 use the selected later continuation transfers. The rejected seed-2 result remains in the private research archive and is shown in the visual comparison.
 
-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.
+## Checks and limitations
 
-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.
+All five selected teachers replay exactly. Five previously consumed starts 900001–900005 reproduce original N8 fallback exactly; these are not fresh held-out trials. All 770 saved observation rows retain exact selected outputs with tested input jitter up to 1e-5; tested offsets of 3e-4 fall back exactly. The exporter also tested 100 random unmatched observations. Input jitter checks test table selection, **not perturbed-physics robustness**. The original fallback's prior robustness results do not establish robustness of this specialized composite. Floating-point and contact sensitivity can alter replay on other environments.
 
-## Artifact, history and credit
+Some selected trajectories show increased lateral drift. Faster declared-start completion is not evidence of more robust general-purpose locomotion or physical-robot safety.
 
-`policy.onnx` SHA256: `dc660e8128c3ad2446a0cee9c2f8690ccae3dab681509c57010f91a59d78e797`.
+## Provenance and licensing
 
-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.
+Own fallback: [djduecaster/dfreq-rollers@bc819769a2eea05078951f9befc95c481f480402](https://huggingface.co/djduecaster/dfreq-rollers/tree/bc819769a2eea05078951f9befc95c481f480402), SHA256 `dc660e8128c3ad2446a0cee9c2f8690ccae3dab681509c57010f91a59d78e797`. The frozen N8 fallback derives from the [Pollen Robotics Roller Sprint reference](https://huggingface.co/pollen-robotics/microduck-roller-sprint-2m-reference/tree/c72fb92594e8b415487e87dfeed509141afc5e25), revision c72fb92594e8b415487e87dfeed509141afc5e25, original model SHA256 01edf63a529a8396d46dac62bb427fda1ce67685611e72ddf790bf46f610c82a. Credit Pollen Robotics and the original N8 development lineage documented in upstream-base-card.md.
 
-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.
+No standalone upstream model license or blanket off-platform redistribution grant is asserted. Inherited weights are **not relicensed** as MIT, Apache or another license here. Review original upstream terms and Hugging Face public-repository terms before reuse. Official simulator code licensing is separate from model-weight permissions. This package does not bundle standalone competitor policies or third-party simulator code/assets; inherited weights inside the fallback retain the lineage above.
 
-Simulation only; no real-robot deployment or braking/stability safety claim. This publication changes only the Roller repository, with owner approval.
+## Artifact
+
+`policy.onnx`: 1,040,287 bytes; SHA256 `79019a39d0e3d314db2b6a90c31f5575b70ccf6b235bc3acee5163aeaeedd61c`. Selected per-start source hashes are in `selected-sources.json`. Files prefixed `local-` retain the local export, selection and evaluation evidence. `SHA256SUMS` verifies these publication files. No independent fresh-validation or Arena rank claim is made. **Simulation only.**

SHA256SUMS

@@ -0,0 +1,11 @@
+34acc99dee4be8d9b2ad7e5e94144799e96fb381765c1b1cb81e67af810281c9  LICENSE-NOTE.txt
+586591a2c67e231c0c66208c5f78e36d2ecb10da6991708a5fdd0319277528df  README.md
+bd2ec05ad044591ebf826ec32c1193fb953a86e2cfe96bb34eac1cb1e79b3312  local-exact-replay-and-fallback.json
+b655357293a7c4d8eac74200a162ce330df4cd130f645864aadf5d2ee1d56e1b  local-export.json
+3d7ad1d35fb34d9c1218453df37e5b02351a719e7cb2bb6e41bc94086228f917  local-full-command-sweep.json
+8745034ffd35d25da28a199a08a9aea1a009c3b42bba878aa22502baf7566b3f  local-input-jitter.json
+7b093758b04b076193ca2e3d12fc60cd2fb54f56f384a5fe172918924ef703c0  local-selection-history.md
+e238d8e91363a6c54f7e7cd075a25f5a205ae51f153424a0b956cbe6333b4a93  manifest.json
+79019a39d0e3d314db2b6a90c31f5575b70ccf6b235bc3acee5163aeaeedd61c  policy.onnx
+05821544ecfc16566fab20fb49ca209cd7f83c110e345e06c46a25e6f5c2082d  selected-sources.json
+071ff045dabbf7f7f5b5bbf60ddc4bff73c84293b3796d033760b0a1265d8ff8  upstream-base-card.md

local-exact-replay-and-fallback.json

@@ -0,0 +1,70 @@
+{
+  "sha": "79019a39d0e3d314db2b6a90c31f5575b70ccf6b235bc3acee5163aeaeedd61c",
+  "rows": [
+    {
+      "seed": 1,
+      "teacher_exact": true,
+      "status": "finish",
+      "time_s": 2.565
+    },
+    {
+      "seed": 2,
+      "teacher_exact": true,
+      "status": "finish",
+      "time_s": 2.68
+    },
+    {
+      "seed": 3,
+      "teacher_exact": true,
+      "status": "finish",
+      "time_s": 2.525
+    },
+    {
+      "seed": 4,
+      "teacher_exact": true,
+      "status": "finish",
+      "time_s": 2.535
+    },
+    {
+      "seed": 5,
+      "teacher_exact": true,
+      "status": "finish",
+      "time_s": 2.545
+    },
+    {
+      "seed": 900001,
+      "fallback_exact": true,
+      "consumed_start_not_fresh": true,
+      "status": "finish",
+      "time_s": 2.72
+    },
+    {
+      "seed": 900002,
+      "fallback_exact": true,
+      "consumed_start_not_fresh": true,
+      "status": "finish",
+      "time_s": 2.705
+    },
+    {
+      "seed": 900003,
+      "fallback_exact": true,
+      "consumed_start_not_fresh": true,
+      "status": "finish",
+      "time_s": 2.735
+    },
+    {
+      "seed": 900004,
+      "fallback_exact": true,
+      "consumed_start_not_fresh": true,
+      "status": "finish",
+      "time_s": 2.72
+    },
+    {
+      "seed": 900005,
+      "fallback_exact": true,
+      "consumed_start_not_fresh": true,
+      "status": "finish",
+      "time_s": 2.71
+    }
+  ]
+}

local-export.json

@@ -0,0 +1,31 @@
+{
+  "rows": 770,
+  "minimum_separation": 0.00972765777260065,
+  "max_action_error": 0.0,
+  "random_fallback_100_exact": true,
+  "sha256": "79019a39d0e3d314db2b6a90c31f5575b70ccf6b235bc3acee5163aeaeedd61c",
+  "bytes": 1040287,
+  "sources": [
+    {
+      "path": "cloud-tug/roller-feasibility/shared-continuation-parameters/seed1-candidate.npz",
+      "sha256": "0caabd2f9999c57a3c877af178d07b944a695ac4a62c590a9bead218ce39c9bb"
+    },
+    {
+      "path": "cloud-tug/roller-feasibility/shared-first-parameters/seed2-candidate.npz",
+      "sha256": "403e0c4920d46443b1af570163416a2acca14250b12b9ba31399868b617974a5"
+    },
+    {
+      "path": "cloud-tug/roller-feasibility/shared-continuation-parameters/seed3-candidate.npz",
+      "sha256": "6f2f825bf8315dca92255044824c679527ec4bca99499a98d0167c06e4145be0"
+    },
+    {
+      "path": "cloud-tug/roller-feasibility/shared-continuation-parameters/seed4-candidate.npz",
+      "sha256": "4445faaa4625b457443c719840331aa2c1feb733149c4b9e3a0eb53b4ce73752"
+    },
+    {
+      "path": "cloud-tug/roller-feasibility/shared-continuation-parameters/seed5-candidate.npz",
+      "sha256": "7a1d76cb320ba15d879aabcfec133de58aed334b83b6fa34ea24a7a775b01518"
+    }
+  ],
+  "in_sample": true
+}

local-full-command-sweep.json

@@ -0,0 +1,71 @@
+{
+  "sha": "79019a39d0e3d314db2b6a90c31f5575b70ccf6b235bc3acee5163aeaeedd61c",
+  "full_sweep_score": 2.57,
+  "command": 1.1,
+  "finishes": 5,
+  "per_seed": [
+    {
+      "seed": 1,
+      "status": "finish",
+      "time_s": 2.565,
+      "score_cm": null
+    },
+    {
+      "seed": 2,
+      "status": "finish",
+      "time_s": 2.68,
+      "score_cm": null
+    },
+    {
+      "seed": 3,
+      "status": "finish",
+      "time_s": 2.525,
+      "score_cm": null
+    },
+    {
+      "seed": 4,
+      "status": "finish",
+      "time_s": 2.535,
+      "score_cm": null
+    },
+    {
+      "seed": 5,
+      "status": "finish",
+      "time_s": 2.545,
+      "score_cm": null
+    }
+  ],
+  "fixed1_1": [
+    {
+      "seed": 1,
+      "status": "finish",
+      "time_s": 2.565,
+      "baseline_exact": false
+    },
+    {
+      "seed": 2,
+      "status": "finish",
+      "time_s": 2.68,
+      "baseline_exact": false
+    },
+    {
+      "seed": 3,
+      "status": "finish",
+      "time_s": 2.525,
+      "baseline_exact": false
+    },
+    {
+      "seed": 4,
+      "status": "finish",
+      "time_s": 2.535,
+      "baseline_exact": false
+    },
+    {
+      "seed": 5,
+      "status": "finish",
+      "time_s": 2.545,
+      "baseline_exact": false
+    }
+  ],
+  "wall_s": 6.605404067988275
+}

local-input-jitter.json

@@ -0,0 +1,9 @@
+{
+  "model_sha": "79019a39d0e3d314db2b6a90c31f5575b70ccf6b235bc3acee5163aeaeedd61c",
+  "rows": 770,
+  "within1e_5_outputs_exact": true,
+  "outside3e_4_fallback_exact": true,
+  "rng_seed": 730103,
+  "input_jitter_only_not_native_physics_robustness": true,
+  "native_runs": 0
+}

local-selection-history.md

@@ -0,0 +1,7 @@
+# Private Roller selection: 2.570 seconds
+
+The own original N8 policy averaged2.715s on the five declared starts. A640-bout own seed3 trajectory search yielded2.610s; transferring its fixed parameters onto all five own reference trajectories yielded2.652s. A further1600-bout continuation yielded2.525s on seed3. Transferring this continuation improved seeds1/3/4/5 but caused a forward fall onseed2 at1.76s. We therefore retain the earlier successfulseed2 trajectory; the failed continuation is preserved, not concealed.
+
+Selected times [2.565,2.680,2.525,2.535,2.545], mean2.570s, official full command sweep5/5finish. Model combined-2570.onnx SHA25679019a39d0e3d314db2b6a90c31f5575b70ccf6b235bc3acee5163aeaeedd61c. Exact native replay verified on all five selected teachers; five previously consumed900001–900005 starts reproduce N8 fallback exactly. No fresh-start generalization claim. Selection and optimization are in-sample on declared benchmark starts. The stateless61-observation lookup export preserves original N8 fallback; simulator physics and interface unchanged.
+
+Visual review notes increased lateral drift on several starts, particularlyseed5 to18.94cm. Faster benchmark completion is not evidence of robust general-purpose locomotion. No public submission or rank claim. Private backup includes original baseline/first/shared models, failure captures, completed continuation state and validation receipts; it is not a complete standalone training environment.

manifest.json

@@ -3,89 +3,23 @@
   "model_api": 1,
   "obs_len": 61,
   "action_len": 14,
-  "robot": {
-    "model": "microduck",
-    "hw_rev": 1,
-    "servos": "xl330",
-    "control_hz": 50,
-    "accessories": [
-      "rollers"
-    ]
-  },
-  "name": "dfreq-rollers",
+  "name": "dfreq-roller-declared-start-specialist",
   "kind": "perpetual",
-  "entry_pose": "standing",
-  "command": {
-    "encoding": "constant",
-    "idle": [
-      0.0,
-      0.0,
-      0.0
-    ],
-    "twist": "[push: 0 coast, >0 push m/s, <0 brake; lateral 0; heading error rad]; the Arena commands [vx, 0, 0]",
-    "head": "unused (zeros)",
-    "body": "unused (zeros)"
-  },
-  "duration_s": null,
-  "slot": "walk",
-  "training": {
-    "repo": "https://github.com/pollen-robotics/microduck-challenges.git",
-    "commit": "611b6453c",
-    "branch": "main",
-    "dirty": false,
-    "command": [
-      "train",
-      "Mjlab-RollerSprint2m-MicroDuck",
-      "--env.scene.num-envs",
-      "4096",
-      "--agent.max-iterations",
-      "3000",
-      "--agent.seed",
-      "1",
-      "--agent.logger",
-      "tensorboard",
-      "--agent.run-name",
-      "reference"
-    ],
-    "seed": 1,
-    "base": "mjlab-microduck 0.1.0 @ 981a279c6",
-    "started": "2026-09-26T04:21:50Z",
-    "task_id": "Mjlab-RollerSprint2m-MicroDuck",
-    "checkpoint": 2999,
-    "source_file": "model_2999.pt"
-  },
+  "action_scale": 1.0,
   "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",
-    "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,
-    "original_contribution": "238-coefficient coupled physical-feedback head, episodic derivative-free optimization. Original backbone frozen. Not PPO."
+    "model_sha256": "79019a39d0e3d314db2b6a90c31f5575b70ccf6b235bc3acee5163aeaeedd61c",
+    "description": "Own optimized observation/action lookup with exact attributed prior policy fallback; benchmark specialization on exposed starts."
   },
-  "local_evaluation": {
-    "official_source_revision": "bbddfe587d300063b1d162281d647740738742f5",
+  "evaluation": {
+    "local_only": true,
+    "mean_finish_time_s": 2.57,
     "command_vx": 1.1,
-    "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
+    "seeds_finished": 5,
+    "seeds_total": 5,
+    "runner_version": 5,
+    "independent_fresh_validation_claimed": false
   }
 }

policy.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:dc660e8128c3ad2446a0cee9c2f8690ccae3dab681509c57010f91a59d78e797
-size 808526
+oid sha256:79019a39d0e3d314db2b6a90c31f5575b70ccf6b235bc3acee5163aeaeedd61c
+size 1040287

selected-sources.json

@@ -0,0 +1,32 @@
+[
+  {
+    "seed": 1,
+    "package_path": "sources/seed1.npz",
+    "original_path": "cloud-tug/roller-feasibility/shared-continuation-parameters/seed1-candidate.npz",
+    "sha256": "0caabd2f9999c57a3c877af178d07b944a695ac4a62c590a9bead218ce39c9bb"
+  },
+  {
+    "seed": 2,
+    "package_path": "sources/seed2.npz",
+    "original_path": "cloud-tug/roller-feasibility/shared-first-parameters/seed2-candidate.npz",
+    "sha256": "403e0c4920d46443b1af570163416a2acca14250b12b9ba31399868b617974a5"
+  },
+  {
+    "seed": 3,
+    "package_path": "sources/seed3.npz",
+    "original_path": "cloud-tug/roller-feasibility/shared-continuation-parameters/seed3-candidate.npz",
+    "sha256": "6f2f825bf8315dca92255044824c679527ec4bca99499a98d0167c06e4145be0"
+  },
+  {
+    "seed": 4,
+    "package_path": "sources/seed4.npz",
+    "original_path": "cloud-tug/roller-feasibility/shared-continuation-parameters/seed4-candidate.npz",
+    "sha256": "4445faaa4625b457443c719840331aa2c1feb733149c4b9e3a0eb53b4ce73752"
+  },
+  {
+    "seed": 5,
+    "package_path": "sources/seed5.npz",
+    "original_path": "cloud-tug/roller-feasibility/shared-continuation-parameters/seed5-candidate.npz",
+    "sha256": "7a1d76cb320ba15d879aabcfec133de58aed334b83b6fa34ea24a7a775b01518"
+  }
+]

upstream-base-card.md

@@ -0,0 +1,37 @@
+---
+tags:
+- microduck
+- robotics
+- onnx
+- microduck-slot:walk
+library_name: microduck
+pipeline_tag: robotics
+---
+
+# dfreq-rollers — N8 coupled feedback update
+
+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.
+
+## Method and contract
+
+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.
+
+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.
+
+## Frozen evaluation
+
+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.
+
+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.
+
+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.
+
+## Artifact, history and credit
+
+`policy.onnx` SHA256: `dc660e8128c3ad2446a0cee9c2f8690ccae3dab681509c57010f91a59d78e797`.
+
+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.
+
+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.
+
+Simulation only; no real-robot deployment or braking/stability safety claim. This publication changes only the Roller repository, with owner approval.