JEVfire browser demos

Start with the visual field guide → The new learn.html page explains token scoring, caching, constrained decoding, calibration, and open training recipes with interactive diagrams. It does not load a model. It also links to the measurement receipts behind 71 ms per action on M4 Max: 71.26 ms mean worker inference in the final-build Mario run, with 511 accepted choices. The hardware name is owner-supplied; saved browser telemetry records Apple / Metal 3. See full methodology.

World 1-1 adds a third game: original Canvas2D art, playable keyboard/touch controls, a scripted baseline and local Qwen choosing finite maneuvers with persistent instruction caching and an explicit physics guard. The world keeps moving during inference; raw three-field controls remain an advanced comparison. Game notes.

Slipstream: four racing strategies → · Last Hearth: six villagers →

All three demos run Qwen 3.5 0.8B locally through WebLLM and WebGPU. The same pinned model files can be reused from the browser cache; each page creates its own worker and engine. Neither requires an inference server or API key.

Slipstream

Open driving.html, load Qwen, and start a three-lap race. Four cars receive separate editable prompts. Click a numbered car or its roster card to inspect its speed, race position, resources, current maneuver, last model choice and saved input. Apply a prompt edit to use it for that car's next decision.

Nine finite actions cover acceleration, braking, holding pace, left/right lane changes, boost, risky left/right overtakes, and pit service. Only currently available actions are scored. Worn tyres and wet corners lower grip; excessive speed causes spins and damage, with retirement at 100 damage. Pitting involves driving to the entry and a six-second stopped service. Three laps decide the winner; unfinished cars retire at the 180-second simulation cutoff.

Following assistance brakes for close traffic, never for corners. A risky pass briefly disables it. Lane tracking and pit routing are simulation rules. Try scripted drive uses a separately labeled controller with no inference. Boost, spins, contact, overtakes, pit visits and finishes appear in Race control.

Observations include actual/target speed, position and rival gaps, nearby traffic in all three lanes, closing speeds and time to contact, the next corner's safe speed and braking distance, tyres, damage, boost, and pit distance. Read the full design and context contract.

The model scores all eligible drivers in one fleet request, sequentially on one engine, and immediately starts the next request. Stable policies are cached. It cannot invent keys or actions, but can choose badly. Physics runs independently of inference; a higher simulation speed gives the model less time to react per simulated second. Reset, pause, policy edits and unavailable actions invalidate stale results. Forced pit routing and paused probes add no AI decisions.

See driving QA and policy experiments for measured results and limitations. The CUDA benchmark headline is not a browser speedup claim.

Last Hearth — six villagers, one local model

Play Last Hearth →

Keep a village alive against hunger and increasingly strong orcs. Two collectors bring home food, two fighters train and protect them, and two builders repair the settlement, construct defenses, and heal wounded allies. Edit the objective and role policies; Qwen chooses each person's actions. 3D art credits and licenses.

Play

1. Click Load Qwen 3.5 in a recent desktop browser with WebGPU and shader-f16. The 4-bit model is about 450 MB, plus runtime files. Allow roughly 2 GB of free GPU/unified memory; requirements vary by device.
2. Start the village. Click a person, nameplate, or roster card to open its inspector. See level/XP, health, hunger, current job, and recent decisions. On smaller screens, details open in a compact panel with a close button.
3. Edit the village order and role policies. Balance food reserves against early defenses, or risk faster foraging. Policies persist in this browser; Restore returns a role to the supplied policy.
4. Try again with the same seed. The separately labeled scripted baseline runs without a model download. It does not follow prompt edits and is never substituted for failed model inference.

Drag to orbit, scroll to zoom, and use the crosshair to restore the camera. The game pauses when its tab is hidden. Simulation speed does not increase model throughput. Seeds fix random spawning/roaming, but model arrival times and update cadence also affect play; identical seeds do not guarantee identical AI runs.

Roles and rules

Role	Choices	Tradeoff
Collectors	forage_safe, forage_bold, relax	Safer routing versus faster, riskier food runs
Fighters	train, defend, relax	Future strength versus intercepting current attacks
Builders	repair, build, heal, relax	Preserve buildings, invest in towers, or spend food treating allies

Foraging includes selecting a patch, walking, harvesting, and returning a full basket. Defending selects a threatening orc and taunts it away from villagers. Building and repair select legal sites. The model chooses high-level behavior; navigation and combat are game code, not a hidden second model.

Hunger declines continuously. Zero hunger means death. At 28 hunger, a visible automatic-needs rule interrupts work, finds actual food at the hall, in a carried basket, or at a stocked patch, and resumes the chosen job after eating. It does not create food or count as an AI decision. Relaxing also walks home to eat and recover. Builders travel to wounded allies and spend one stored food per treatment to restore up to 20 HP. Training permanently raises fighter damage. Orcs roam, acquire nearby villagers/buildings, and attack. New waves become stronger. Finished towers shoot orcs. The run ends when everyone dies or the hall falls. An allowed but unwise decision can lose the game.

Each villager has a distinct prompt: Mira is cautious, Bram likes short trips and leisure, Aldric protects others, Sable prioritizes training, Tomas builds, and Nell prioritizes treatment. Personality influences scores; it does not hard-code a job. Stamina falls while working and recovers at home. Rest becomes available when tired, hungry, wounded, already on a break, or when no useful job remains. Bram takes earlier breaks (70 stamina), Tomas later ones (30); other thresholds are shown in context. Unavailable work is removed from the candidate set. When only relaxation is available, a labeled game rule applies it without invoking the model or increasing AI ticks.

Character levels and decisions

Every character starts at level 1 and earns one level per 100 XP from productive work. Collectors earn 10 XP per harvested food; fighters earn 2 XP per productive training second and 0.5 per actual damage dealt to an orc. Builders earn 0.1 XP per building HP repaired, 60 per tower constructed (shared according to actual progress), and 0.5 per ally HP healed. Walking, resting, model calls, and tower attacks do not award XP. Reset starts everyone at level 1 again. Levels are experience milestones, with no added combat or work-speed multiplier; fighters' existing training strength still determines damage.

The inspector distinguishes current job and activity from the latest Qwen decision. The six-entry history identifies Qwen decisions, scripted commands, forced single-option jobs, and automatic meal breaks with simulation timestamps. Repeated identical scripted/rule assignments are collapsed; actual repeated model decisions remain visible. Scores and the saved input snapshot belong to the shown Qwen decision, not to a newer world state or automatic action. A scorer provides preferences, not a generated explanation of its reasoning.

What the model knows

Each villager gets compact, role-specific observations rather than a dump of every world coordinate:

• Health, hunger, estimated time until starvation, walk time home, current action, carried food, and number of attackers.
• Shared food, living population, wave, hall health, and threatened people or buildings.
• The nearest two orcs: distance, bearing, health, level, current attack target, and estimated direct approach time.
• Collectors see food patches, stock, route danger, and trip estimates. Fighters see allies/buildings under attack. Builders see damage, unfinished defenses, and wounded allies. Each actor also sees its current available jobs.

These are observations computed from simulation state, not model-generated facts. There is no screenshot interpretation in this browser build. A route's estimated risk can become stale as orcs move. Context design.

Read the counters correctly

One AI tick = one accepted model decision for one living villager. One WebLLM engine scores actors in a fair round-robin, taking fresh observations before each request. It scores the available labels (up to four) at one final output position; JavaScript builds a result such as {"mira":"forage_safe"}. Generated text is ignored. Missing or nonfinite scores fail explicitly.

Display	Measurement
AI ticks/sec	Applied NPC decisions per wall-clock second over a rolling 10-second window; shorter during warmup
AI rounds/sec	Completed rounds covering actors currently eligible for AI (multiple available jobs, no meal break)
Last inference	One actor's scoring time, including prefill
Model decisions	Total applied AI decisions in this run
Render FPS	Actual draws per elapsed wall time over 5 seconds; capped near 60
Frame p99	99th percentile of animation-frame gaps over 5 seconds; lower is smoother

Scripted mode shows no AI rate. Paused rates are zero. Pauses, resets, changed orders, dead actors, automatic meal breaks, and newly unavailable jobs discard pending results. Round counts use the currently eligible actor set; rule-only jobs and meal breaks do not add model decisions. The browser SDK reuses exact instruction prefixes but does not implement vLLM's parallel GPU batching. The repo's 10.3× CUDA result is a separate benchmark, not a claim for this game.

The model cannot invent a field or choose outside the role enum. It can still choose badly, misunderstand observations, or follow an injected instruction that changes its preference. Scores are not calibrated confidence. Exact structural guarantee.

The supplied policies remain experimental. The previous release matched only 7/12 and 4/12 development scenes; those tests predate automatic meals, healing, personalities, and available-action filtering. They are historical evidence, not scores for this version. Results and limitations. Prompt tuning here has not established an optimal survival policy.

Smooth rendering and inference

A Web Worker keeps model orchestration off the main thread, but WebLLM and the browser compositor still share a GPU. The new SDK checkpoints unchanged prompt prefixes, including attention and recurrent state. The optimized adapter avoids sampling and intermediate vocabulary readback; an unsupported runtime falls back to independent prefills. See the SDK implementation and limits.

Last Hearth retains its responsiveness controls: Balanced uses 32-token chunks with 16 ms yields and a 200 ms inter-decision pause; Maximum uses 128-token chunks with no artificial in-worker yields and a 30 ms UI pause. Slipstream instead scores an entire fleet snapshot with 128-token submissions and immediately starts the next fleet request: no inter-request pause. Its per-driver policies are cached. World 1-1 defaults to continuous maneuver selection with persistent policy caching, 256-token submissions, no artificial yields or post-answer delay, and local physics forecasts. Advanced raw buttons uses three fields and two cache levels. Only its optional Decision steps mode waits at action boundaries. Measured design.

This browser runtime scores independent suffixes sequentially. It does not expose simultaneous multi-sequence GPU execution. Higher throughput can compete with rendering, and changed submission shapes can alter close floating-point scores.

The scene batches static foliage, caches shadows, limits pixel ratio, caps draws near 60 FPS, and updates label content at 10 Hz. Low graphics reduces resolution further. Frame measurements and reproducible scripts include cold and warm observations; hardware and shader compilation affect the results.

Develop and verify

Requires Node.js 22.12+ or a newer version supported by Vite:

cd web
npm ci
npm run dev

WebGPU requires HTTPS or localhost. Explicitly load the model in the opened page. Inference runs in a Web Worker through WebLLM; Three.js renders through WebGL. There is no inference API key or remote model server.

npm test
npm run format:check
npm run build
npm run preview

CPU tests cover hunger, meals, death, training, construction, repair, aggro, taunts, tower combat, observations, seeded balance, role contracts, stale results, fair scheduling, and measured tick rates. GPU checks require a supported browser and downloaded model. QA scripts, fixtures, and results.

Pinned runtime

Component	Pin
WebLLM	@mlc-ai/web-llm 0.2.85
Tokenizer runtime	@mlc-ai/web-tokenizers 0.1.6
Weights	Qwen3.5-0.8B-q4f16_1-MLC
Weight revision	0ec138972555613c1d7812a821778ad0398c8790
Library	MLC v0_2_84/base/Qwen3.5-0.8B-q4f16_1_cs1k-webgpu.wasm
Library revision	025bcaf3780fa8254f5e5efd3bfea0a5397248f4
Context	2,048 tokens; input capped at 1,800
Labels	Verified distinct single tokens mapped to finite actions; up to four village choices or nine driving choices

Models/tokenizers download from Hugging Face/CDNs, the compiled library from GitHub, and fonts from Google Fonts. Mission text and game state stay on-device. WebLLM caches completed files in IndexedDB; browsers may evict them. This is not an offline-installable PWA, though inference works without a network after the page, assets, and model load. Upstream models and dependencies retain their licenses. The application's MIT license does not relicense them.

The earlier Signal Run demo remains in commit c75ea086fc5aeee2eb8971b24631b827bd510045.