WRECK VALLEY · 0.61.4 · IMPLEMENTATION & VERIFICATION
Jev sees the whole map.
The human minimap showed every opponent. Jev only received rivals within 180 metres. That mismatch is removed: the model now receives global opponent markers, map geometry, supplies, objectives and HUD state.
Inspect what the model received
These are controlled observation fixtures sent to the real Jev provider. Select a case to inspect the recorded map and response. They are separate from gameplay footage.
Loading observations…
Cyan: Jev. Coral: rival. Gold: next waypoint. North points upward. A route can initially lead away from the rival to get around a building. Positions update even when the opponent is on the other side of the world.
A map and a usable route
- Map information: self and rival positions, direction and velocity; city roads, buildings, freight lanes, park paths, coast and island routes; ramps, garages, water, supplies and event markers. HUD information includes health, nitro, grip, speed, score, time, protection and city heat.
- Navigation: paths connect the city, freight district, park, beach and island. Ground and car-width clearance checks reject deep-sea shortcuts and paths through buildings. Existing ramp navigation still handles ramps locally.
- Fresh pursuit: the driver replans moved destinations without requiring a different model tactic. Stationary destinations keep their route; shortcut prefixes are discarded so the car advances instead of revisiting old junctions.
- Visible decisions: the control room shows the map stored with the selected request, including the rival and the next route waypoint. Past decisions retain their own map.
Control room recording
UI integration recording: deterministic model stub, real server, ledger and physics. Checks cover recorded maps, live controls, history, session changes, disconnect recovery and mobile layout.
Real-model gameplay
The final browser run applied 20/20 real model responses with no provider or JavaScript errors. Tactics: pursue: 17, stage: 3. Confirmed cost: $0.004863. The isolated $0.006 test cap correctly switched to local fallback.
Authoritative counters across the entire run: 63 nitro steps, 0 drift steps and 358 braking steps at 60Hz. The closest sampled separation was 3.5m and nitro fell to 81.1%. Reviewed frames show contact with the player. The clip is an excerpt. Keyboard input used the production client without car-position or physics overrides; Chromium/SwiftShader rendering and automated inputs are not a human playtest or hardware FPS benchmark.
Early browser runs exposed junction oscillation and poor local ramp detours. Stable routes, shortcut advancement and graph-level ramp bypasses address those cases. Jev also holds its lane while the rival has a spawn shield instead of heading toward the city centre. Each fix has regression coverage.
Regression coverage and cost tradeoff
- 469 tests, lint and production build passed.
- Global markers update across the map; inactive actors and private identity data are excluded.
- Shared-physics tests close a 309m chase, turn back toward a rival behind the car, handle the spawn detour and drive from beach to island through shallow water.
- Graph tests verify segment clearance and connectivity through all districts, including the wider beach.
- Request size remains bounded at 12KiB with the existing conservative 32K-token reservation.
Production caps stay at $0.05 per match and $1 per UTC day. Full map requests contain more tokens: these fixtures cost about $0.00024 each. At that observed rate, the per-match cap can switch to local fallback before a six-minute round ends. The monitor shows actual charges, remaining budget and fallback state. Navigation and map drawing make no additional model requests.