Real trainable feed-forward networks: 7 ray sensors plus speed/yaw and learned previous-action feedback. Selection, crossover, Gaussian mutation, behavioral novelty, and genetic niche fitness sharing, and mirrored evolution strategies evolve independent steering/throttle policies. No prescribed line.
GENETIC DIVERSITY RESCUE · MEASURED NEURAL WEIGHTS
A deterministic 24-racer sample measures real neural-weight distance against the first random generation. Optional rescue boosts mutation below 45% of the initial distance and adds a few fresh genomes below 25%. Elite winners, niche representatives, archive replay and ES pairs remain protected. No scripted racing line, steering or fitness modification.
Waiting for completed training · rescue defaults OFF.
| Gen | RMS weights | Initial % | Mutation × | Fresh |
|---|
Audit not run.
VERIFIED LAP-FIRST PARENT SELECTION
When enabled, a genuinely completed six-gate lap outranks any incomplete candidate; among finishers, the fastest actual lap wins regardless of behavioral novelty. Before any finisher, progress fitness and novelty retain their original roles. This changes parent selection, never steering or physics. Compare OFF / ON with matched seeds and frozen clean/blind evaluation in BENCH.
Lap-first selection enabled · awaiting completed generations.
Lap-first selection audit not run.
Sector curriculum off · full-lap training.
Practice chains: one sector · no measured completions.
Chain audit: not run.
Start: neutral
Start audit: not run.
SENSOR ROBUSTNESS / DOMAIN RANDOMIZATION
Optional teacher-free ray noise and missed returns during training only. Clean blind evaluation remains unchanged. Mirrored evolution-strategy pairs receive identical seeded sensor disturbances for fair comparison. Adjust the two training-ray sliders above to restart learning with the new conditions.
Training sensor augmentation: off.
Robustness audit: not run.
SECTOR MASTERY
Skill archive: empty.
QUALITY–DIVERSITY POLICY MAP
Full-lap policies are stored in a 6×6 MAP-Elites archive: rows = verified track progress (low → high); columns = mean absolute neural steering (low → high). Tap a cell to inspect the best recorded neural policy's measured progress, steering and lap time in that behavior niche. Practice-only runs cannot populate this map.
No full-lap policies archived yet.
Select a filled cell to inspect its policy.
Archive audit: not run.
VERIFIED NEURAL LEARNING EVIDENCE
Real completed training generations only: lap completion rate, fastest actual lap, verified progress, and separate frozen-policy clean three-start validation. Neutral starts and augmented starts are not pooled; sector practice is excluded from full-lap rates. Observational only, not a steering or fitness input.
Waiting for completed neural generations.
| Gen | Start | Train laps | Fastest | Best % | Clean laps |
|---|
Evidence audit not run.
Actual practice completion and verified progress identify difficult sectors. Adaptive targeting trains the weakest while preserving full-lap generations.
No sector-practice measurements yet.
Exploration archive: 0 · adaptive mutation: 1.00×
OFFICIAL-LAP MUTATION CONTROLLER · VERIFIED TRAINING
Optional neural exploration feedback from real six-gate neutral-start lap times. Practice, augmented starts, physics mismatches, unverified finishes, and study episodes cannot trigger it. Existing raw-fitness fallback remains active until valid evidence exists. No steering is scripted.
Awaiting completed training generations.
Controller audit not run.
VERIFIED SECTOR PACE · ADAPTIVE PRACTICE
Opt-in practice targeting from actual six-gate neutral-start neural laps. Recent median sector times are compared against each sector’s best observed time with matching circuit and physics. Requires five finishing generations and a measurable gap. No prescribed racing line, steering or reward changes; isolated from other studies.
Awaiting verified sector times.
Audit not run.
Genetic niches: clustering neural weights…
Mirrored ES: initializing…
Mirrored ES objective audit not run.
Boundary-aware gradient audit not run.
Pairwise ES signal audit not run.
Exact spatial collision/sensors · frame budget active
Novelty neighbor search: awaiting independent parity check.
Temporal feedback and legacy policy migration: not checked.
Fast sensor/neural engine: ready for independent parity and throughput audit.
Sensor throughput: awaiting simulation.
Arc-length and checkpoint audit: not run.
Ray audit: run to compare exact raycast against fine brute-force sampling.
Swept collision audit: not run.
Continuous collision contacts this generation: 0.
NEURAL FAILURE ATLAS · WHERE LEARNING STALLS
Real neutral-start training outcomes, binned by checkpoint-verified forward progress. Segment risk = failures / racers reaching that segment. Practice, augmented starts, benchmarks and blind evaluation are excluded. Observational only: no steering script, fitness changes or neural inputs. Turn on the track overlay under VIEWS.
No neutral-start training outcomes yet.
Failure-risk audit not run.
FAILURE-AWARE SECTOR CURRICULUM
Optional, evidence-calibrated practice selection. Neutral-start failures only; at least 12 entrants and 3 failures are required per sector. Wilson 95% lower confidence bounds prevent a few random crashes from dominating. Default OFF; full-lap generation frequency, learned steering and official time trials are unchanged. Benchmark and transfer studies never use this adaptive signal.
Curriculum risk targeting OFF · awaiting neutral-start evidence.
| Sector | Entered | Failures | Risk | 95% lower |
|---|
Risk-targeting audit not run.
No post-upgrade practice decisions recorded yet.
| Track segment | Reached | Failures | Risk |
|---|
INDEPENDENT POLICY VALIDATION
The best real neural policy from each completed full-lap generation is tested on three clean starts. The best independently validated candidate is ranked by completed laps, then fastest verified lap, then minimum and mean progress. Its frozen weights can be saved to the existing policy vault without altering evolution or official lap records.
Validation winner: awaiting completed clean trials.
Validation isolation: not checked.
Waiting for a completed full-lap generation.
VERIFIED PACE / OFFICIAL LAP SPLITS
Pace reward uses actual checkpoint-verified progress and elapsed time, not a prescribed path. Official sector splits are measured at exact forward gate crossings on neutral-start completed laps.
No official lap splits yet.
Pace/split audit not run.
BEST VERIFIED LAP · LEARNED DRIVING PROFILE
Actual forward-progress samples from the fastest neutral-start completed lap. Blue is observed neural driving speed; amber is the approximate centerline physics envelope. Gaps mean no observation, not zero speed. The envelope does not steer or train the policy.
No completed neutral-start lap recorded.
Lap telemetry audit not run.
VERIFIED TIME-TRIAL HALL OF FAME
The eight fastest independently completed neutral-start laps on this exact circuit. Each entry preserves actual neural weights, official gate splits, candidate ID and physics. No synthetic laps or steering instructions; records never affect selection.
| Rank | Official lap | Generation | Candidate |
|---|
Awaiting a completed neutral-start lap.
Leaderboard audit not run.
Replay recorded neural positions without controlling training.
Replay audit not run.
VERIFIED SECTOR PACE LAB · LIVE GATE DELTAS
Compare actual official six-sector lap times from independently finished neural racers. Only records with matching current vehicle physics and exact track geometry qualify. A best-sector sum is theoretical, not a physically achieved lap or a steering instruction.
Awaiting compatible verified completed laps.
| Gate | Reference | Comparison | Best observed | Cumulative Δ |
|---|
Live selected candidate: waiting for a verified sector crossing.
Live delta uses the selected candidate's most recent physically verified gate time, not predicted finish time or geometric proximity. Practice, augmented starts and study episodes are excluded. Negative delta means ahead of the reference at that gate.
Sector comparison audit not run.
GENERATION OFFICIAL SECTOR PACE · LEARNING OVER TIME
The fastest physically completed six-gate neutral-start lap in each generation, including finishers outside the Hall of Fame. Dots show actual finishes; a stepped line shows the best observed time to date. Missing finishes are gaps, not zero or predicted lap times. Observational only, never a neural steering or selection input.
Awaiting verified neutral-start generations.
| Gen | Verified laps | Official lap | Metric |
|---|
Trend audit not run.
GENERATION HISTORY
Blue = full-lap verified progress · amber = sector practice · green = completed lap. Novelty uses 36 track/lane occupancy bins; selection favors both speed and distinct behavior.
COMMON TRAINING / VEHICLE
Overdrive uses more active-tab CPU with the same physics step and model. Switching apps normally suspends mobile browser computation.
SIMULATION-FIRST RENDER SCHEDULER
Separate neural physics from canvas drawing. Reducing visual refresh frees CPU without skipping physics ticks, sensors, or neural decisions. Editing and paused views remain fully responsive.
Adaptive graphics scheduling ready. Neural physics remains exact.
Rendering audit not run.
Essential controls here; specialized evolution settings under Advanced.
Resets all training/motion parameters; retains current track and saved policies.
Particle mode · neural steering and throttle.
Car dynamics: not audited.
Four-stage car contact: not audited.
MOBILE PHYSICS PERFORMANCE LAB
Profile real neural sensor, network, road-contact and vehicle physics across three separated circuit sectors at 16, 64, 128 and 192 agents. Frame-sliced samples report average and 95th-percentile CPU cost; safe pacing uses the slower estimate. Training state is isolated. This measures throughput, not learning success.
Ready to profile real neural physics without changing training.
| Agents | Agent steps/s | Mean ms / step | P95 ms / step | Crashes | Finite |
|---|
Adaptive pacing audit not run.
LIVE EVOLUTION GENERATION PROFILER
Measures actual completed training generations, including neural/physics simulation, selection and offspring creation, archive maintenance and autosave. CPU timings exclude drawing and browser idle time. This observes real evolution without changing genomes, rewards or RNG.
Waiting for a complete normal training generation.
| Gen | Agent ticks | Physics ms | Evolution ms | Save ms | Lap |
|---|
Generation timing audit not run.
MORE MOTION SETTINGS
Car mode: four-wheel chassis, planar tire-slip/yaw model, combined braking/cornering grip, aerodynamic drag and downforce. Reference power 520 kW, mass 1030 kg. Other coefficients are estimates, not homologation data.
FILTER & COLOR
Lineage counts: waiting for population.
OVERLAYS
TRACK EDITOR
Choose Freehand, Polyline or Grid snap above. DRAW starts a new track; EDIT POINTS opens the current one. Tap to add a point, drag to move it, or tap near a segment to insert. APPLY rebuilds the track; CANCEL preserves training.
Choose a drawing method and press Draw or Edit Points.
SEEDED CIRCUIT DESIGN LAB
Generate a reproducible, collision-safe closed circuit without touching learned weights until you press APPLY. The exact generated centerline opens in the existing point editor: drag handles to refine it, APPLY to start fresh training, or CANCEL to restore your current race. Geometry uses a separate random stream, never the evolution seed.
Choose a seed and complexity to preview a new drivable circuit.
PHYSICS-AWARE CORNER PACE MAP
Inspect a geometric speed envelope for the current track or live point-editor preview. Uses measured centerline curvature, particle yaw response or hypercar tire-grip/steering limits, plus forward acceleration and backward braking constraints. The colored overlay is an engineering diagnostic, NOT a racing line, learned policy, predicted lap time, or training reward. Actual neural cars may need to drive slower.
Enable Corner pace map above to visualize constrained sections.
Blue = relatively unconstrained · amber = slowing required · red = severe bend. The car steering warning concerns the centerline only; a different line within the road may be feasible.
Physics envelope audit not run.
LANE SEPARATION / TOPOLOGY
Checks exact closest distances between nonlocal centerline segments against the physical racer-center corridor. Narrow hairpins or overlapping lanes can create ambiguous progress and unintended shortcuts. Risky edits require an explicit override; existing saved tracks remain compatible.
Open the point editor to inspect drivable-lane separation.
Topology audit not run.
LIVE NETWORK ACTIVATIONS
Input rays → hidden layers → steer / throttle. Blue = positive, coral = negative; links are learned weights.
NEURAL DECISION FLIGHT RECORDER
Capture the selected live candidate's real pre-step sensor readings, network outputs and hidden activations alongside post-step vehicle physics, sector crossings and terminal state. No simulated or scripted decisions. Recording follows the candidate's unique ID, not the selection index; it stops on death, generation replacement or the 2,400-sample limit.
Select a living candidate to record its neural decisions.
Blue: steering (−1…1) · Amber: throttle (−1…1) · Green: speed (0…1). Tap or drag the timeline to inspect a recorded decision.
No samples yet.
LEARNED SENSOR → ACTION SENSITIVITY
Exact local derivatives of the frozen recorded neural policy, including hidden layers. Shows which normalized inputs most affect steering and throttle at the selected decision. Memory inputs stay fixed; these are NOT simulated steering corrections, causal attributions or a future driving prediction.
Inputs 1–7: rays · 8: speed · 9: yaw · 10: previous steering · 11: previous throttle.
No neural sensitivity sample selected.
Gradient audit not run.
Recorder isolation not audited.
RESUMABLE TRAINING
Checkpoint initializing…
Checking saved generations…
Study launch protection: ready. Training is durably checkpointed before a study replaces it.
Recovery audit not run.
Checkpoint audit not run.
SAVE / TRANSFER
Saves neural weights, exact particle physics, RNG, track and settings during active generations and on backgrounding. Exported JSON includes a resumable checkpoint. Older project files still load.
SAVED TRACKS
Store exact circuit geometry and editable control points separately from training. Loading a track starts new training; saved neural policies remain available.
FAMOUS CIRCUIT-INSPIRED TRAINING LAYOUTS
Schematic interpretations, not scaled race-track maps. Preview and move control points; APPLY restarts training, CANCEL preserves your run.
Circuit-inspired training geometry, not exact GPS/survey alignment.
SAVED NEURAL CANDIDATES
Each generated particle or car receives a unique ID. Saved policies contain neural weights—not recorded trajectories, brake points, or a prescribed racing line.
Saved candidate architecture will be detected automatically; other settings remain adjustable.
Audit not run.
NEURAL TRANSFER / ADAPT ON THIS TRACK
Warm-start a NEW evolutionary population from a saved neural policy. The original weights stay exact in slot 1; the selected share receives mutated descendants, and the remainder starts from independent random networks. All candidates learn from real sensors and lap/progress fitness. This restarts the current training; save a checkpoint first if needed.
Select a saved policy above to warm-start training. The first generation compares warm-start and fresh cohorts; this is not a controlled multi-seed A/B study.
Storage: checking durable browser database…
Read/write/delete test not run. Does not modify neural training.
Transferred neural weights can be tested on unseen circuits. They are not guaranteed to beat newly initialized candidates.
CROSS-TRACK POLICY TRANSFER GAUNTLET
Evaluate the selected saved neural candidate against freshly initialized neural networks on matched circuits, seeds and neutral/offset starts. This runs real sensors, learned actions and collision physics. No training or prescribed driving line; current training is paused and restored afterward. Fresh policies are untrained, not freshly evolved competitors.
SAVED-POLICY HEAD-TO-HEAD DUEL
Choose a second saved neural policy. Both drive identical circuits with identical neutral/left/right starts and the SAME sensor and physics engine. This compares actual saved policies, not a scripted driver or untrained random baseline. Blind mode excludes the UNION of both policies' recorded training geometry; older unknown training exposure cannot be ruled out. Fixed policies are deterministic, so each start is evaluated once.
Select a primary candidate above and an opponent here. Use the Evaluation circuits selector to choose a blind holdout.
Isolation audit not run.
Select a saved candidate above. Three matched starts per circuit and seed.
No transfer study recorded.
| Track | Saved % | Fresh % | Laps S/F | Fastest S/F |
|---|
Progress is checkpoint-verified, and laps require ordered sector crossings. These tests measure transfer, not guaranteed superiority. Circuit width and vehicle mode must be compatible. Results are stored separately from the training checkpoint.
MULTI-CIRCUIT NEUROEVOLUTION LAB
Compare novelty search, sector curriculum, adaptive targeting, sector replay, quality-diversity replay, or multi-sector chain practice OFF / ON from identical initial neural weights and seeds, on your circuit, an oval, a technical circuit, or all three. Practice generations never count as completed laps or full-circuit progress.
For sensor robustness, lap-first selection, failure-atlas curriculum, official-lap mutation, sector-pace practice, or genetic diversity rescue: matched seeded training, followed by frozen-policy clean tests on the training circuit AND geometry-excluded blind circuits. Nine starts add deterministic symmetric offset, yaw and launch-speed stress tests to each frozen policy. No holdout episode selects parents or weights. Unseen means excluded from this study’s training geometries, not necessarily from earlier historical exposure.
Ready. Experiments temporarily replace the visible circuit and population; both are restored afterward.
GENERALIST MULTI-TRACK EVOLUTION
Evolve the SAME neural genomes on every selected circuit before selecting parents. Score = 70% mean + 30% worst-circuit verified result, with optional neutral/offset-start robustness inside each circuit. Completed laps and faster times receive priority. No scripted steering or memorized racing lines. Original training is restored after the study; save the resulting neural policy to Library.
Seeded mode varies the two off-center starts by generation and circuit using bounded reproducible offsets, yaw and launch speed. Neutral starts, blind holdouts, neural steering and episode budgets remain unchanged; both training arms use the same perturbation schedule.
REPLICATED NEURAL EVOLUTION · INDEPENDENT SEEDS
Train matched cold-start generalist/specialist networks for each independent seed, then automatically duel their learned policies on the SAME frozen, geometry-excluded holdout circuits. No scripted steering. Unseen means excluded from this study’s recorded training suite, not guaranteed absent from all historical exposure.
Generated blind packs are geometry-excluded BEFORE training, then frozen across all independent seeds and both arms. Pack seed and evolution seed are independent. Blind outcomes never select neural weights. Earlier historical exposure cannot be guaranteed absent.
Generated-pack audit not run.
Nine-start blind evaluation retains the three fixed starts and adds three seeded symmetric left/right pairs per unseen circuit. Both frozen policies face identical starting offsets, heading and launch speed; these tests never train or select a policy. Nine starts triple holdout compute.
Blind-start audit not run.
Ready for independent-seed replication.
Seed-level uncertainty appears after three completed blind-tested seeds.
| Seed | Training G/S laps | Training G/S progress | Blind G/S laps | Blind G/S fastest | Blind G−S progress | Blind W/T/L |
|---|
BLIND GENERALIZATION MATRIX · CIRCUIT × INDEPENDENT SEED
Compare actual unseen-circuit laps, fastest times and verified progress. Each seed contributes one paired result per circuit, regardless of the number of correlated starts. Inspect neutral, lateral or stress launches separately; no result feeds back into training.
| Circuit | Seeds | Laps A/B | Fastest A/B | Δ lap rate | Δ progress (pp) | Seed W/T/L |
|---|
Blind circuit results appear after a completed paired seed.
Matrix audit not run.
MEASURED NEURAL LEARNING CURVES
Real in-sample training fitness by generation (best neural genome, 0–3000), averaged across complete matched seeds. A learning curve measures training, not unseen-track generalization or fastest lap.
Learning curves appear after matched training completes.
Curve audit not run.
Cold-start experiments create independent, reproducible neural weights from the chosen seed. Both training arms receive identical initial weights and equal episode budgets. Uncheck to continue evolving your existing population instead.
Paired mode starts both arms from identical neural weights and RNG seed. Generalist trains on each selected circuit; specialist trains on the CURRENT circuit from equally many rotated start lines. Each arm receives the same number of full physics episodes per genome and generation, including the selected lateral starts. No scripted driving. Holdout duels use circuits excluded from both training suites.
CONTROLLED TRANSFER LEARNING · WARM VS FRESH
Select a saved neural policy in LIBRARY. Compare warm-start descendants against an entirely fresh population over the SAME circuits, generations, physics and seeded start conditions. Non-warm genomes are identical in both arms. The source parent is frozen, and both populations evolve independently. BLIND PAIRED DUEL evaluates frozen winners on geometry excluded from recorded training; a single seed is exploratory.
Uses the LIBRARY policy and warm percentage, plus BENCH seed, circuit suite and generation count. Original training is restored afterward.
REPLICATED TRANSFER · INDEPENDENT SEEDS
Select a compatible LIBRARY policy, then replicate WARM versus FRESH across the independent-seed count, frozen holdout suite, and 3/9 blind starts configured above. Both arms share fresh initial genomes outside the warm cohort, physics, training episodes, and start schedules. Frozen source weights and recorded training geometry survive seed-boundary resume; unknown historical exposure cannot be inferred. STOP, RESUME, CSV, and JSON use the shared replicated-study controls.
Compare the generalist with matched fresh policies on geometrically unseen circuits. Geometry used in this recorded generalist study is excluded even if a circuit is renamed or its starting point changes. Earlier training exposure cannot be reconstructed.
Fresh seeded networks by default. Disable Fresh to begin from your existing trained networks. Original training is protected and restored.
| Generation / circuit | Best verified | Laps | Fastest |
|---|
Audit not run.
PAIRED LEARNING RESULTS
No experiment yet. Compare matched verified progress, lap completions, and actual fastest laps across circuits. A short run is exploratory, not proof of superiority.
| Circuit / trial / arm | Best % | Laps | Fastest |
|---|
OFFICIAL-LAP MATCHED INFERENCE · NO FABRICATED TIMES
Run paired training trials to compare verified finishes.
Audit not run.
CLEAN + BLIND FROZEN-POLICY VALIDATION
Matched frozen-policy evaluation appears here after sensor robustness, lap-first selection, or failure-atlas curriculum studies. It never selects parents.
Run at least three matched trials for trial-level uncertainty on blind circuits.
| Training → evaluation | Trial | OFF / ON % | Laps O/N | W/T/L |
|---|
Audit not run.
FAILURE-ATLAS CURRICULUM A/B · VERIFIED INTERVENTION
OFF and ON start from the same seeded neural weights, with identical physics and sector-practice share. ON can choose practice sectors using ONLY its own neutral-start training failures. Each arm begins with a separate empty atlas; frozen clean/blind results cannot influence selection. Three or more paired trials are needed for exploratory uncertainty.
Choose Failure-atlas curriculum OFF / ON above. Risk influence requires sufficient actual training failures.
Study isolation audit not run.
OFFICIAL-LAP MUTATION A/B · VERIFIED INTERVENTION
OFF/ON use identical cold-start weights, seeds, physics and generation budgets. Each arm builds its OWN neutral-lap evidence ledger; OFF records evidence without using it. ON can change mutation exploration only after comparable verified finishes. Frozen clean and blind evaluations never feed back into training.
Choose Official-lap mutation OFF / ON above. No study yet.
Mutation study audit not run.
VERIFIED SECTOR-PACE PRACTICE A/B · ISOLATED NEURAL EXPERIMENT
OFF and ON evolve independent populations from identical seeded initial neural weights and episode budgets. Both accumulate separate official six-gate neutral-start lap records. ON may choose a different practice sector from its own verified timing gaps after at least five compatible finishing generations. No lap means no evidence; practice/augmented starts and frozen validation are excluded. An altered sector choice is measured, never assumed.
Select Verified sector-pace practice OFF / ON above. The intervention may remain inactive if neural racers do not finish laps.
Paired sector-pace audit not run.
GENETIC DIVERSITY RESCUE A/B · FROZEN POLICY VALIDATION
Both arms start from identical seeded neural weights, circuit geometry, vehicle physics and generation budgets. OFF leaves diversity rescue disabled; ON measures actual neural-weight collapse and may boost mutation or inject a bounded number of fresh genomes. Each arm establishes its own initial genetic reference. Intervention is measured from changed mutation multipliers or actual immigrant insertions; if diversity remains healthy, no benefit can be attributed to rescue. Frozen clean and geometry-excluded blind tests never train either arm.
Select Genetic diversity rescue OFF / ON above. A study without an observed genetic intervention is inconclusive.
Genetic rescue study audit not run.
Original training, neural weights, mid-episode state, custom circuit, and local autosave are protected. Suite experiments can take several minutes on mobile. Closing the tab leaves your last autosaved training unchanged.
2D FLIP / PIC FLUID LAB
Real particle-to-grid velocity transfer, staggered MAC-grid pressure projection, and FLIP/PIC advection. Track edges are solid walls. A sealed start/finish divider makes the inlet and outlet opposite ends of one channel. Sustained channel pressure head drives liquid around all bends toward the outlet, even after pumping stops. Values use screen pixels and seconds, not calibrated 3D CFD units.
Hold FLOW to inject; releasing stops new fluid immediately. Already-released fluid keeps moving down the channel under its adjustable pressure head. CONTINUOUS is a latching alternative for hands-free flow. Set sustained drive to zero to test inertia/gravity alone.
Switch to FLIP mode to initialize the current track channel.
Fluid checks not run.
The neural simulation is suspended, not deleted. Switch back to neural racing to resume; fluid settings persist locally, while liquid particles are session-only.