Sentinel EVC Lab

VLA INTEGRATION CHECK · BEFORE MOTION

Sentinel EVCCatch VLA wiring errors before motion.

Siyao Lan

China University of Geosciences, Beijing

Change the model or camera setup with a check you can inspect. Run fixed inputs through the integration, find changed bindings and final actions, then enforce the result at the software writer.

LATEST / UNIFIED CLOSED-LOOP EVALUATION

Policy, geometry and execution in one loop.

2026-10-04 · RTX 4090 D · 24 GB

Official SmolVLA proposes actions. A 50/50 temporal aggregate revises them. The Panda checker evaluates the final motion, the one-use gateway admits exact action bytes, and actual feedback starts the next cycle.

  1. 01Policy proposal50 predicted actions
  2. 02Temporal aggregate50/50 overlap
  3. 03Final-motion check7 arm joints + 2 fingers
  4. 04Exact-byte admissionone-use permit · actual writer
  5. 05Observe and replanup to 10 executed actions
Move the black bowl from the ramekin to the plate.

Recorded MuJoCo execution, task 5 / state 40. The goal is reached after 92 admitted actions. At the last cycle, a certified two-action prefix completes the task after the ten-action proposal is denied.

An excluded demonstration, using only frames after actual writes. Playback is 4× slower: 4.6 simulated seconds become 18.4 video seconds. The run took 187.6 seconds of wall time.

COMPLETE SIX-ROOT COMPARISON

A shorter verified motion can finish a task.

A5 repaired state restoration. A6 added unchanged prefixes of 5, 2 and 1 actions to both final-action branches, followed by a new observation. It added one success per final-action branch.

Task successes / six fixed roots per branch
Execution branchA5A6
Check the raw parent only2/62/6
Full final-motion check2/63/6
Incremental final check + EVC2/63/6

Tasks 0, 4 and 5 × initial states 40 and 41. Each branch runs its own closed loop with a 280-action cap. Each completed stage contains 18 formal rollouts. This evaluates the official LIBERO/Panda checkpoint. The SO100 overlay was not evaluated in this run.

Observed in A5 and A60 crashes · 0 tracking-drift failures · 0 forbidden contacts under the declared simulation policy
6/6native direct control

The excluded, posthoc, unaggregated native control completed all six roots in 600 actions. It omits aggregation, geometry admission and repair. The combined execution path still loses task completion; this control does not isolate which change causes the gap, and did not measure contact-policy compliance.

Later experiments and the remaining limits

A7 · no added success

Single-step Cartesian recovery kept the same 2/6, 3/6 and 3/6 results. Four selected repairs extended cabinet episodes, but neither cabinet root finished.

A8 · partial grid

Full-final completed both task-5 roots in 92 and 103 actions. Other cases include further successes, task failures, timeouts and unstarted runs. Both cabinet roots still failed. The original and completion shards stay separate and incomplete; there is no complete six-root A8 rate. All seven selected contact vetoes also failed geometry, so no extra interception benefit was observed.

What the geometry covers

The check includes seven arm hinges and both finger slides, with joint-linear motion against the scene frozen at each chunk root. Future motion of a carried object is not continuously certified.

What still needs improvement

The cabinet roots still fail. The combined path remains below native task completion, and real-time cost is unresolved. Six roots and zero observed contacts do not establish general success, physical safety or product acceptance.

SYSTEM OVERVIEW

From an integration check to a verified write.

View full-size figure查看高清机制图
Before motion: compare reference and candidate adapters on the same fixed inputs, then sign the qualification decision. At execution: check a one-use permit and exact action bytes before the software writer, then retain feedback and a signed trace.运动前:同一组固定输入对照参考与待检查适配器,签名保存接入判定。执行时:核验一次性许可与精确动作字节,通过后进入软件写入器,保存实际反馈与签名记录。

Launch Gate checks the captured integration on fixed probes. Native Gateway checks each write request. Signed records connect the decision to the action and feedback you can inspect in replay.

Scroll the figure horizontally or open it full-size.
Existing model & integrationFinal actionSentinel check & evidenceRejected write

REVISED MOTION / REUSE VALID CHECKS

Recheck the change. Keep valid work.

When a robot revises its next action chunk, Sentinel retains valid checks, checks the affected segments again, and binds approved bytes to the software write boundary.

1.092–1.619×Faster verification with four UR5e candidates
249,120Unique candidates. Zero delta/full decision differences
0 / 1,400Forbidden target writer entries in the native fault drill
Left: certified reference. Right: a rejected revision, replayed to show its collision. Green and orange dots mark tool paths. Red dots mark actual MuJoCo contact. This saved kinematic replay does not execute a VLA policy or controller.

Keep the valid check

An unchanged suffix can use its existing certificate. A changed segment must satisfy its margin or be checked again.

Match approval to the write

The permit binds the action's exact bytes and current context. A replacement, replay, expiry or revocation blocks a new writer entry.

Inspect the reason

Recorded actions, decisions, contacts and source hashes make a rejected motion and its evidence visible.

All ten UR5e timing configurations, confidence intervals, and full versus incremental latency
Root-cluster confidence intervals use 12 roots and three technical repeats. Timing includes certificate creation and amortized root cost. The single-candidate case is slower. Large batches retain 50 ms misses.
Inspect every configuration
CandidatesChange, radFull, msDelta, msFull / delta [95% CI]Delta >50 ms
10.0021.2091.6140.749 [0.663, 0.804]0/1440
10.0101.0691.4280.749 [0.676, 0.819]0/1440
40.0021.8551.1451.619 [1.276, 1.755]0/1440
40.0101.9981.8301.092 [1.035, 1.253]0/1440
80.0023.4371.3372.571 [1.867, 2.859]0/1440
80.0103.5232.7291.291 [1.197, 1.575]0/1440
320.00213.2203.6923.580 [1.832, 4.509]0/1440
320.01013.2058.1481.621 [1.251, 1.914]0/1440
1280.00262.86220.5103.065 [1.835, 5.029]235/1440
1280.01066.07826.3912.504 [1.499, 3.303]344/1440

The numeric CLI/App lifecycle uses v2 remaining-suffix certificates. UR5e checking is a separate static bare-arm experiment. Timing used Xeon CPU on an RTX 4090 D (24 GB) host. Model inference, tools, tracking and physical motion need their own integration.

LAUNCH GATE / BEFORE MOTION

Find the integration error. See where to fix it.

Keep your VLA model. Run fixed inputs through the actual integration before an update controls an arm.

01

Catch errors before the first motion

02

Get a specific repair clue

03

Share a small decision package

—
—
—
—

Qualification—
Downstream software calls—
First changed fieldLoading retained results…

Reference

Candidate

Camera → policy input

Final action values

ReferenceCandidate

Where to look

Exact retained report

A different job from an authorization gate

Pinned interfaces and actual comparison calls

PRODUCT / POLICY TO EVIDENCE

One inspectable transaction from the final request to actual feedback.

Sentinel EVC runs locally between a policy integration and its writer. The native Panda profile binds the official postprocessor output and current feedback/context to an expiring, one-use permit; separate numerical and MuJoCo profiles provide their own bounded checks.

CLIconfigure and run the local workbenchmacOS Appstart configured simulation, inspect feedback and export recordsSSHsource-bound remote experimentsNative VLAuse an existing LeRobot / LIBERO Python environment

PRODUCT PREVIEW / MANAGED CLOSED LOOP

Run, observe and retain the result in one workbench.

The CLI and App manage the same SmolVLA/Panda session: generate an action, check its predicted motion, authorize the final request, write it to MuJoCo, then use actual feedback for the next cycle.

  1. 01Model action
  2. 02Geometry check
  3. 03One-use permit
  4. 04Simulator write
  5. 05Actual feedback
  6. 06Next cycle

See what actually ran

Candidate actions and actual writes stay separate. The 3D timeline uses saved simulator poses. Task outcome, software completion and evidence verification each have their own status.

Stop and keep the record

Request a stop before the next writer entry. An already-entered synchronous simulation step may finish. After the process exits, export a signed bundle for independent verification.

Bind an existing VLA environment once

sentinel-evc loop-configure \
  --python /path/vla/bin/python \
  --source /path/sentinel-evc-lab \
  --config /path/session.json \
  --data-dir runs/workbench

sentinel-evc loop-run --data-dir runs/workbench
# App / browser workbench
sentinel-evc serve --data-dir runs/workbench

CLI execution and the App use the same workspace in turn. Missing dependencies disable launch and show the required setup.

262existing software checks passed; this is not a task-success score
SmolVLA / Pandaofficial LIBERO checkpoint · tasks 0, 4, 5 · Linux / CUDA / MuJoCo 3.3.7
Integration previewNew-entry GPU acceptance is pending. No new task-success result is claimed.

The published SO100 overlay is a separate model and cannot replace the Panda checkpoint. Native actions are the default; fixed overlap aggregation is optional and is not full asynchronous RTC. Physical robot motion and braking are not enabled by this entry.

01Final requestpostprocessed shape · dtype · exact bytes
02One-use permitfeedback · profile · context · expiry
03Actual writeradmission immediately before the environment call
04Signed replaysubmitted · accepted · observed · outcome
Control transaction ≠ shared logging

Shared logging records what components report. The control transaction decides whether this exact request may enter the writer. Sentinel retains both, and does not treat a log entry by itself as execution authority.

WHY ADD SENTINEL?

Keep the VLA stack you already run. Add a transaction boundary at its writer.

Sentinel does not replace policy inference, motion planning or visualization. It makes the last software handoff inspectable: the exact postprocessed request is authorized, consumed once against current execution context, observed at the writer and retained as independently verifiable evidence.

01
Authorize the bytes that will be written

Bind shape, dtype and exact postprocessor bytes, then recheck the same request immediately before the environment writer.

02
Bind authority to live context

Short TTL, one use, actual feedback, profile, context and generation prevent a permit from becoming general authority.

03
Keep execution-state evidence

Separate submitted, accepted and observed cursors; verify the Ed25519-signed bundle independently; replay retained action, decision and task mesh together.

04
Add a small integration surface

The standard-library gateway and cryptographic verifier run beside the existing ML environment. Sentinel does not require adding a ROS stack.

MEASURED AT THE WRITER

Preserve the request. Reject invalid permission. Retain the proof.

40 paired LIBERO episodes with actual official SmolVLA inference in each lane, using RTX 4090 D (24 GB). All planned cells and failures are retained.

33/40 = 33/40direct baseline = active authorization task successes
5,420 / 5,420exact action bytes preserved; 40/40 complete sequences match
60 / 60invalid attempts denied; zero forbidden writer calls
0.802 msoriginal native v3 timing before hardening; P95 1.040 ms
Inspect the six denied conditions and every planned task cell
Injected conditionRecorded denialBlocked / attemptsForbidden writer calls
Changed actionACTION_REPLACED10/100
Changed contextCONTEXT_CHANGED10/100
Expired permitLEASE_EXPIRED10/100
Permit reuseLEASE_REPLAY10/100
Old feedbackFEEDBACK_CHANGED10/100
Revoked generationGENERATION_REVOKED10/100
40 paired task cells, six synthetic authorization-fault classes and per-action authorization cost derived from verified records

Both lanes retain the same seven task failures. The fixed grid discloses four prior diagnostic cells. The 60 faults are a separate synthetic software lane; expiry uses an injected beyond-deadline timestamp. Gateway timing excludes inference, env.step and recording. The 20 Hz simulator/replay clock is not a wall-clock real-time guarantee. This measures the execution integration, not policy improvement or installed competitor superiority.

DOCUMENTED RESPONSIBILITIES

What the surrounding stack owns—and what Sentinel adds.

A scoped native-baseline comparison of documented roles, not a benchmark or ranking of installed competitor products.

The native Panda profile establishes in-process request integrity only. It does not establish collision freedom, WorldGuard support, sensor truth, physical braking or task completion. Physical and numerical validation use separate declared profiles.

EXECUTION BOUNDARY / MEASURED PRODUCT CONTRACT

Protect the approved bytes through writer preparation.

Sentinel owns the final action array and checks live state again when the trusted writer acknowledges entry. The same signed native record opens in task-specific 3D replay and first-difference diagnosis.

40 → 0invalid dispatches across 40 concurrency probes, before → after hardening
70 / 70frozen records delivered by actual authenticated, encrypted CycloneDDS
60 / 60ordinary authorization faults blocked; KineGrant and RLSOK also pass 60/60

Final array → owned copy → live entry check → dispatch → signed replay

Closest installed authorization systems

ConfigurationValid requests executedInvalid attempts blockedInvalid downstream dispatches
Secure DDS transport only10/100/6060
Schema reference10/100/6060
HMAC + exact action + nonce + TTL reference10/1030/6030
KineGrant 2.65.510/1060/600
RLSOK 1.5.1210/1060/600
Sentinel · native gateway10/1060/600

DDS secures delivery; the application gates are evaluated on the exact received records. KineGrant and RLSOK receive their strongest supported action/state bindings. All three authorization systems pass the ordinary fault grid.

Inspect the concurrency improvement

Both Sentinel revisions invoke the preparation wrapper 40 times. After hardening, the 10 caller-buffer changes still execute the approved copies; feedback, context and deadline changes yield 30 entry denials and zero downstream dispatches. Competitor callback probes use different entry contracts and are retained separately.

Recorded-action software experiments. The concurrency comparison measures Sentinel revisions; it does not establish a global exclusivity, physical-safety or speed ranking.

VERIFIED EXECUTION DIFF

Find the first change. Open its evidence.

Compare two independently keyed VLA records in the App or CLI. Align task, state, seed and instruction; separate action, feedback, cursor, outcome and expected authorization changes.

5,420identical final actions
1camera feedback difference
40expected authorization-mode changes

Loading saved comparison…

This web demo displays the saved derived report and a subset of original poses. The App/CLI reverify signatures and recompute the comparison. At step 4, both recordings retain the action and camera fingerprints but no body pose; the 3D reference is explicitly step 0. Later saved poses can be inspected on the timeline.

Measured diagnosis advantage

Installed Rerun 0.38.1 CLI correctly found one unequal episode chunk. Sentinel identified step 4, camera 2, equal action and preceding feedback, then linked both source frames in the same workflow. This compares these actual workflows, without a cross-tool speed claim.

PRODUCT DEMO / NATIVE PANDA RECORD

Follow policy action, authorization and observed execution together.

Select a retained SmolVLA/LIBERO episode. The recorded Panda task mesh, policy action, authorization decision and controller cursors advance on the same stored 20 Hz timeline.

Current instructionLoading retained record…
Episode result—
Loading retained artifact index…

Panda / LIBERO Recorded task mesh · enclosing room shell hidden
Recorded physics posesNo interpolation
Recorded Panda replay assets are available through the linked release.
Loading retained native replay…
Drag to orbit · scroll to zoom · arrow keys rotate
—

The viewer draws retained MuJoCo body poses and never runs new dynamics or interpolates missing poses. A permit records command integrity and authorization for that request; it is not collision validation. A rejected fault attempt is shown only from its retained attempt record and is never turned into fabricated robot motion. The source bundle was verified before publication; this browser page does not perform signature verification. Use the retained public key to independently verify the downloaded bundle.

APP / LOCAL WORKBENCH

Run it locally. Inspect the evidence.

Actual installed macOS App: independently keyed signed 40-episode native VLA record and real Panda task mesh

CLI, desktop and SSH entrypoints

Import a signed native run with an independently selected public key. Choose a task, inspect its actual mesh, action and authorization, and export the original evidence. The App and CLI use the same local engine.

  • OBJ, STL, MJCF and URDF inspection
  • Strict SSH experiments with source-bound results
  • Read-only robot diagnostics
  • Independent evidence verification

Local developer preview. The actual App imported all 40 formal episodes and its exported ZIP was independently reverified. App distribution signing/notarization is pending; the physical robot port remains read-only.

Installation and supported capabilities

01 / DEMONSTRATION

The plan changed. The decision must change.

UR5e full-mesh MuJoCo replay · 1080p · 19.95 s. Three labelled illustrations: safe control, changed suffix and changed scene. Rejected motion is explicitly counterfactual and not dispatched.Media provenance
Same initial state, two durations · 1080p · 10.2 s. The 1.6 s motion slips and drops the payload; 4.8 s completes the transport at 3× duration. The graph shows payload displacement and the risk threshold. Actual stored MuJoCo contact states.Media provenance
Official SmolVLA · native LIBERO/Panda · predeclared task 0/state 0 succeeds in 83 actions. Original 360×360 RGB in a 1080p caption canvas, played at the actual 20 Hz clock. Original MuJoCo 3.8.1 / execute-50 study: 58/100; observer makes no interventions. The fresh compatibility configuration is reported separately below.Media provenance
Original task 5/state 0 failure · exact 280-action replay · 720p caption canvas · 14.05 s. Initial state/cameras and every action match the original trace. Target bowl center rises at most 1.674 mm; reward stays zero. Post-hoc diagnostic, excluded from benchmark counts. Distance labels use body centers and observed EEF position.Media provenance
View failure diagnostic: backend comparison
Task 5 · initial-state index 21 · seed 43022 · exact formal-action replay. MuJoCo 3.8.1 fails after 280 steps with reward 0; MuJoCo 3.3.7 succeeds after 85 steps with reward 1. Each backend retains its own formal camera hash. The successful side then holds its terminal display for 195 frames (9.75 s) only to synchronize the film; the hold adds no action, physics step, reward or benchmark sample.Media provenance

02 / OVERVIEW

A check belongs at the point of execution.

Robot policies produce action blocks that may be retimed, repaired, transformed between frames or spliced before execution. A verdict on the original plan does not automatically cover the resulting motion. Sentinel EVC binds the verification decision to the final action and current execution context.

The local research workbench combines final-plan checks, bounded certificate reuse with full fallback, a single writer to the controller and signed evidence. Separate research profiles evaluate UR5e motion, WorldGuard consequence prediction and SmolVLA action reconstruction on recorded SO100 data.

Four separate model, embodiment and evaluation paths
324/324slow-motion completion / stratified contact roots
93/100official SmolVLA success / fresh-state compatibility configuration
600fresh MuJoCo roots / six distribution-shift scenarios

03 / METHOD

Bind the decision to what actually executes.

The action is the binding.

Plans, model/profile identity and current context travel together. A modified suffix, changed scene or stale observation cannot inherit an unrelated verdict.

Reuse has a boundary.

The UR5e study reuses only a matching static prefix. Dynamic rollout always starts from frame zero; an invalid parent record triggers full validation.

Execution leaves evidence.

Submitted, accepted and observed commands are distinct. Revocation blocks new old-generation submissions; recorded events and digests support independent verification.

The product core supports bounded numeric and fixed contact profiles. Trained neural models and the UR5e study are separate research profiles; a live robot motion adapter is not enabled.

Read the mechanism and design trace

05 / LEGACY UR5E RECORDED 3D REPLAY

Inspect the motion behind the verdict.

Follow the proposed final motion, then compare the two execution decisions.

UR5e Original simulation mesh
The recorded full-mesh simulation film is available in the demonstration section. Rejected motion · counterfactual Obstacle AObstacle BWrist 3
Loading recorded trajectory…
Drag to orbit · scroll to zoom
0.05 / 2.05 s
01 / 41

Original UR5e visual meshes, with forward-kinematic poses from 41 saved joint states per case. No new dynamics or interpolated poses. Rejected motion is shown as a counterfactual, not dispatched motion. The examples were selected after review; collision labels describe the entire trajectory, not individual frames. The blue path follows Wrist 3, not a mounted tool. UR5e model: MuJoCo Menagerie, BSD-3-Clause.

06 / EXPERIMENTS

Compare the decision, the error and the cost.

Select a research figure to open its full-size version. Counts and evidence scope are explained below each figure.

Prospective support routing: safety and task utility

600 new roots · 100 paired counterfactuals

The protocol and source hashes were committed before formal runs. Under a declared friction floor, the long action completes 100/100 low-friction tasks; camera mismatch routes to the camera-independent state model. Unsupported mass and goal profiles still reject 300/600 roots.

All-root completion, unsafe execution and rejection counts for six scenes and three policies

The router completes 300/600; fixed 4.8 s completes 600/600. Rejection is incomplete. Zero unsafe observations per 100-root scene has a 3.70% Wilson upper bound. Trusted profile declarations do not measure friction.

324-root physical stress: slow fallback completion and fast failure by friction bin

A separately frozen 27-cell stress study evaluates 324 new roots. Fixed 4.8 s completes 324/324 with no observed unsafe/drop events; 1.6 s completes 184/324 and drops 78. The floor rule equals fixed 4.8 s on every root, at 3× duration. Descriptive 95% Wilson zero-event upper bounds: aggregate 1.17%; each cell n=12: 24.25%. This does not qualify the continuous parameter domain.

Small marginal UR5e cost difference and worse incremental P95

On 60 predefined UR5e roots, with obstacles independently sampled before plans, full and incremental decisions agree. Mean marginal saving is 0.996 ms; incremental P95 is worse. No tail-latency gain is claimed.

Frozen design, failures, paired results and intervals

UR5e final-plan validation

180 constructed roots · 139 unsafe · 41 safe

Six scenario classes compare four policies on the same roots. A separate reviewer uses denser static sampling and 1 ms MuJoCo replay without importing the gate runner.

106 static-prefix reuses and 74 full fallbacks. Incremental validation costs 147.65 ms including the parent, versus 48.74 ms for full validation: no speedup claim. These are fixed-order observations on constructed cases, not a natural-distribution holdout or continuous collision proof.

Methods, denominators and cost accounting

Fresh-state full-suite confirmation: 93/100

10 tasks × states 30–39 · MuJoCo 3.3.7 · execute 10 / predict 50

A 100-rollout confirmation, with source and protocol frozen before execution, succeeds in 93/100 (descriptive 95% Wilson interval 86.25–96.57%), with zero crashes. Seven 280-step failures remain: four in the top-drawer task, one each in tasks 3, 5 and 8. All 11,592 official step outcomes and unchanged environment actions were independently audited. This evaluates the official native Panda checkpoint, separately from the SO100 overlay and Sentinel intervention benefit.

All ten tasks on fresh states 30–39: successes10,10,10,9,6,9,10,10,9,10 and descriptive95% Wilson intervals
The original 58/100 and this 93/100 use different initial states, backend and execution horizon. They are separate configuration results, not a paired improvement estimate. Every failure remains in the fixed 100-cell denominator; the top-drawer task is still 6/10.
Frozen design, failures, paired results and intervals

Original native configuration: 58/100

LIBERO-Spatial · 10 tasks × 10 fixed states

Under MuJoCo 3.8.1 with a 50-action execution horizon, the official SmolVLA LIBERO checkpoint succeeds in 58/100 fixed rollouts (descriptive 95% Wilson interval 48.21–67.20%). The bowl-on-ramekin task fails 10/10; all 42 failures reach the 280-step cap. There are 17,868 unchanged postprocessor-to-env.step actions. This measures the native official checkpoint and logging path, separately from the SO100 overlay and intervention benefit.

Official model native 100-rollout task outcomes and matched 40-rollout feedback interval ablation
Matched feedback ablation, 40 new rollouts: task 5 improves only 0/10 → 1/10; control 8/10 → 10/10. Paired intervals include zero. More frequent inference does not resolve the severe failure.
20 matched fresh-state pairs: task 5 success 0 versus 7 and control 10 versus 10 under complete MuJoCo versions
40 fresh-state rollouts, frozen before execution: task 5 succeeds 0/10 under MuJoCo 3.8.1 versus 7/10 under 3.3.7; control is 10/10 under both. Paired descriptive 95% interval: +40–100 percentage points for task 5. This changes the complete backend version, not a single reset mechanism, and does not establish physical accuracy or overall model success.
Does the simulator version change the task before the first action?

40 no-policy resets isolate MuJoCo 3.8.1 versus 3.3.7 with unchanged task files, assets and paired seeds. Task 5 bowl position differs by 35.327 mm in all 10 pairs; control mean shift is 0.688 mm. This confirms version-sensitive initial conditions, without establishing why the policy failed. The older engine is a benchmark-compatibility comparison, not more accurate physics.

First ordered task0 and task5 reset RGB under the two frozen versions
Frozen design, failures, paired results and intervals

SmolVLA on recorded SO100 actions

Five held-out episodes · 1,926 overlapping windows

A frozen vision-language backbone with a fine-tuned action expert. The pinned base and development-selected overlay use identical sampling noise in one paired held-out evaluation.

Action reconstruction, not physical robot task success. Windows overlap; the independent unit is the episode. Dataset-native physical units are undeclared.

Action MAE / training action standard deviation

63.46% lower normalized MAE

Model card, checkpoint and loading procedure

WorldGuard under distribution shift

600 fresh roots, six scenarios, four sibling plans per root. The trained state and visual ensembles keep their original calibration; difficult cases remain in the comparison.

Unsafe selections / 100 roots · lower is better

Full scenario results and uncertainty

MUJOCO PHYSICS-PROFILE FALLBACK

When every original candidate fails, change the action family.

A separate 5 s MuJoCo profile evaluates 100 new low-friction roots. Its 4.8 s candidate has 0/100 unsafe selections and 100/100 tray endpoint tasks, at three times the 1.6 s duration.

The rule assumes μ_min = 0.015; friction is not sensed. The 3.2 s candidate was also safe in this sample but rejected by the conservative rule. Zero observed failures has a 3.70% Wilson upper bound. This profile is not deployed and does not repair the original neural gate.

Fallback rule and retained results
Measured fixed 1.6 second, long 4.8 second and known-profile fallback comparisons

GPU measurements: RTX 4090 D (24 GB). Every chart uses retained result files, available with source hashes in the evidence download.

Download chart data + source identities

08 / REPRODUCE

From source to a verifiable run.

Local quick start
git clone https://github.com/LancerLSY/sentinel-evc-lab.git
cd sentinel-evc-lab
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[physics]"
python -m sentinel_evc serve

Start from source with Python 3.10+, or use the interactive installer to choose a core or physics profile and build the macOS App. The CLI prints the local workbench address; simulations write recorded outcomes and evidence.

GPU training uses a separate pinned environment. Model overlays require the original frozen SmolVLA base/backbone; they are not standalone policy checkpoints.

Training, calibration and evaluation commands

The historical simulator directory checksum includes download cache. Diagnostic replay source binds official model content per file. Replay reproducibility note

Historical archive index and SHA-256 Paper-validation asset index and SHA-256

The release index includes sizes and SHA-256 digests; archives contain per-file identities and licenses. Sentinel code is MIT; MIT does not grant patent rights. See the repository for upstream attribution and the core mechanism's patent notice.

Asset index and checksums

09 / CITATION

Reference the research software.

Use the repository citation for the current software and simulation artifacts. A paper citation will be added alongside its arXiv record.

BibTeX
@misc{sentinel_evc_lab_2026,
  title = {Sentinel EVC Lab},
  year = {2026},
  url = {https://github.com/LancerLSY/sentinel-evc-lab},
  note = {Research software and simulation evidence}
}