WRECK VALLEY · 0.61.5 · IMPLEMENTATION & VERIFICATION

Jev fights to survive.

Health pickups now trigger a committed retreat. Nitro helps it escape and close the gap. Controlled drifts help turns, and a fresh run-up replaces ineffective bumper pushing.

479Automated tests passed

3 / 3City repair routes completed

20% → 99–100%Health restored through actual pickups

48 healthRival damage in the contact-reset fixture

What was going wrong

A read-only audit of 133 recent applied production decisions found 21 requests below 40% health: 20 chose pursuit and only one chose repair. The driver also disallowed nitro during repair trips, particularly at low health. Arrivals could miss small pickups, and prematurely skipped route corners could trap the car beside ramps.

What changed

Survival first. At 40% health or less, commit to repairs until at least 75%. A pickup restores the normal 40 health, so a badly damaged car seeks a second one. This immediately supersedes an older attack request. The model receives the same constraint and an explicit escape policy.
Choose a useful destination. Ready supplies near the rival receive a distance penalty. Jev retains its chosen pickup until collected or unavailable. If all pickups are cooling down, it seeks a clear escape corridor and switches to a refill when available.
Nitro and drift that reach the controls. Escape and attack policies time safe nitro and brief drift pulses between model calls. Grip, obstacles, airborne state, heading and speed still constrain execution. Jev brakes for the final pickup approach.
Set up another hit. A fresh attack policy can reverse out of sustained slow frontal contact when the rear corridor is clear, then accelerate into a new ram. Protection, low health, obstacles and stale decisions cancel this manoeuvre.
Follow the road accurately. Nearby route corners advance at 2.5m instead of 8m; blocked retained segments trigger replanning. Slow, tight turns reduce overshooting.
Explain local actions. The live monitor shows survival and run-up reasons beside actual inputs. Historical decisions retain their own information.

Jev still chooses high-level tactics and driving policies. Survival constraints, route following and split-second controls run locally. Vehicle health, damage, traction, nitro supply and speed physics remain the same as the existing game. No extra health or teleporting was added.

Shared-physics verification

Controlled starting positions, a stationary rival and a held attack policy; no model calls. These scenarios run the actual city simulation with the authored roads, ramps and pickups. All three repair trips collect two supplies. The previously failing ramp route now finishes in 24.5 seconds.

South road · Around ramps · North road. Vertical jumps are real pickup collections.

Scenario	Duration	Jev health	Refills	Nitro / drift inputs
South road repair	15.2s	20% → 100.0%	2	82 / 78
Repair around ramps	24.5s	20% → 99.5%	2	137 / 78
North road repair	22.9s	20% → 98.9%	2	291 / 80
Break contact and ram again	14.0s	100% → 98.6%	0	146 / 0

The contact scenario created 20.3m of separation and reduced the rival to 51.7% health. Input counts are physics substeps, not separate boosts or drifts: these simulations run two 120Hz substeps per 60Hz tick.

What the real model chose

Four paid observation fixtures, separate from gameplay. Both low-health cases chose repair + escape; both healthy cases chose pursue + attack. Total confirmed cost: $0.000997.

Scenario

Loading…

Recorded gameplay

Real Jev provider, production client, automated keyboard inputs, Chromium/SwiftShader at low quality. This 30-second excerpt skips loading and shows Jev closing in and making contact. Its HUD also shows a reverse setup; the latter part reaches the deliberately small test cap and clearly labels local fallback.

Across the entire run: 19/19 applied responses, 154 nitro input substeps, 5 drift input substeps, 391 braking input substeps, and 3.5m closest sampled separation. Tactics: pursue 14, stage 3, strike 2. Cost: $0.004797. Jev stayed above 91.2% health, so this run did not exercise a critical-health retreat.

The gameplay run showed only a brief drift signal. Longer drift and refill execution is demonstrated by the controlled physics scenarios, not this clip. Reviewed frames show contact, but this is not a human playtest, a win-rate comparison or a hardware performance benchmark. The scripted human later drove into a roadside barrier; that is a limitation of the test route.

Control room recording

Deterministic model stub with real sessions, sockets and ledger. Browser checks cover session selection, live controls, historical maps, mobile layout, disconnects and the new survival reason. The reason display has a separate rendering fixture.

Checks and costs

479 automated tests, lint and production build passed. New regression coverage includes repair commitment, rival camping, unavailable pickups, real pickup collection, the failing ramp route, safe reverse setup, damaging re-engagement, escape nitro and pickup braking. Existing map, vehicle, ordinary AI, provider, stale-response and budget tests also pass.

Production limits remain $0.05 per match and $1 per UTC day. No extra model calls are made for these local driving changes. The real-model browser test used an isolated $0.006 match cap and stopped paid control correctly. Reports, clips, screenshots and databases remain outside source control.