TWO GAME AGENTS · BUILT WITH OVERLIFT

Neon Drift

Meet the hunter and the intelligence above the arena.

VANTA-9 pursues the player while Zarvox comments and proposes directives. The current work binds learning to the choices the player actually had and records how accepted advice affects legal moves. Circling and overall playing strength remain under evaluation.

Play Neon Drift See project screenshots
Neon Drift OverLift NPC Agent Fabric introduction showing Zarvox and VANTA-9 roles, game rules, controls, and Enter the Grid action
OverLift NPC Agent Fabric introduction The new Neon Drift introduction makes the two-agent product story visible immediately: Zarvox observes and advises, VANTA-9 responds and hunts, and the deterministic arena remains authoritative.

Recorded project scope

Implemented

ND-OL01 through ND-OL06 governed Zarvox semantics, compiler, receipts, replay, and Agent Lens

More in this category (4)
  • ND-NPC01 through ND-NPC04 VANTA-9 capsule, world authority, multi-agent coordination, and product proof
  • HUNTER, PREDATOR, SENTINEL, and ADAPTIVE VANTA-9 profiles over one unchanged legal action set
  • Bounded capsule recovery with world-epoch, checkpoint-digest, clock, generation, and outcome-binding checks
  • Shared Kokoro WebGPU model with separate character voices, required first-page loading and warm-up

Demonstrated

Controlled recovery fixtures compare restored memory, outcome settlement, and subsequent decisions with uninterrupted execution

More in this category (2)
  • The current voice source package records a successful .NET build and real HTTP delivery checks
  • Controlled voice tests cover shared loading, permissions, stale Worker replies, cancellation, and retry; they do not execute browser GPU synthesis

Planned

Automatic browser lifecycle recovery and the broader recovery fault campaign

More in this category (2)
  • Geometry-aware VANTA tactics, multi-horizon Zarvox strategy, and cooperation comparisons that measure harmful advice
  • Independent opponents, owner play, actual WebGPU listening, and full-path latency qualification

Research

Broader engine-neutral adapters and external studio pilots

More in this category (2)
  • Larger crowds and faction-level agent capsules after the two-agent proof
  • Additional measured path evaluators that must earn authority through comparison
PROJECT SCREENSHOTS

Explore the interface, workflow, and results shown in the project screenshots.

Neon Drift gameplay with Zarvox voice ticker, VANTA-9 pursuing the player, power cores, and the neon arena
Two-agent gameplay proof Live gameplay shows the product claim in physical behavior: Zarvox communicates through the arena while VANTA-9 pursues, adapts, and acts inside the same deterministic world rules. View full-size image
Earlier Neon Drift game start screen with the neon arena, rules, controls, and Enter the Grid action
Original product view The earlier product view remains as supporting evidence of the interface evolution from a game introduction into a complete two-agent NPC Agent Fabric story. View full-size image
MARKETING

Who it helps and why.

Traditional NPCs repeat bark lines and fixed branches until the player can see the machinery. Unbounded generative AI creates the opposite problem: stale dialogue, illegal actions, inconsistent identity, latency, and decisions that cannot be reproduced.

What makes it useful: Neon Drift connects character behavior to inspectable game evidence. Zarvox can propose a directive; VANTA can accept, modify, delay, reject, or abstain. The game validates the resulting action, while tactical quality and cooperation are evaluated separately.

BUSINESS / EXECUTIVE

Business value.

Neon Drift makes agent proposals, permitted actions, and recorded outcomes inspectable within a game. The inspected recovery and voice work extends that design. Whether the characters feel convincing, varied, and fair still requires owner play and independent evaluation.

Explore the business outcomes
  • Inspectable agent decisions and outcomes
  • Bounded memory and recovery evidence
  • Separate character roles under one game authority
ARCHITECTURE

How the system fits together.

The Semantic Bridge turns authoritative game events and bounded perception into meaning that agents can use. MiniLM can rerank admitted candidates and detect conceptual repetition, while Q-Lens evaluates complete response and action paths. Neither can invent legal actions or directly mutate the game world.

A representative workflow

  1. 01

    Inspect the game event or perception packet available to a character, together with its own memory context.

  2. 02

    Follow the character's proposal or coordination request through the legal-action compiler.

  3. 03

    Compare the accepted or rejected action with the recorded outcome in Agent Lens and replay evidence.

Explore the architecture details

I built Neon Drift as an OverLift NPC Agent Fabric reference implementation. Zarvox and VANTA-9 have separate bounded contexts, typed directives and dispositions, an authoritative action compiler, and decision receipts. Recent work restores bounded capsule memory and pending outcomes, and shares one Kokoro WebGPU model between the separately configured VANTA and Zarvox voices. MiniLM remains a separate semantic reranker.

ENGINEERING

Engineering choices.

I separated each character's observations and proposals from the game state. A deterministic compiler checks whether a proposed action is legal, fresh, fair, and physically possible before the world accepts it. Bounded choices make decisions inspectable, while the quality of the resulting opponent still needs play evaluation.

Explore the technology and engineering details
Technology stack.NET 10, MVC/Razor, deterministic game loop, Workers, local MiniLM, Kokoro WebGPU, Semantic Bridge, Q-Lens, action compiler, capsule recovery, receipts and replay
  • Zarvox and VANTA-9 maintain separate identity, memory, goals, permissions, and decision evidence
  • Semantic Bridge, local MiniLM, Q-Lens, compiler, Agent Lens, and Replay Lab form one inspectable path
  • Silence, delay, refusal, cancellation, and abstention are valid actions that prevent repetitive interaction
OVERLIFT METHOD

How this product applies the OverLift method.

The OverLift Method preserves authoritative events, translates them through the Semantic Bridge, admits only legal candidates, evaluates consequences with Q-Lens, compiles one bounded action, and seals the outcome in a receipt. The field may reshape the search landscape; it may not rewrite the ground beneath it.

Explore the shared OverLift architecture

The method is reusable. Each product supplies its own facts, rules, execution, and validation. These tiers describe the shared architecture; the project scope above records which capabilities are implemented here.

Why determinism matters here: A player must encounter the same physical rules, legal actions, collision outcomes, and fairness boundary even when the characters adapt. Determinism keeps intelligence from becoming an invisible cheat.

  1. T0
    Exact baseline

    Define the authoritative result before adding acceleration, semantics, or agents.

  2. T1
    Measured evidence

    Instrument state, latency, quality, and failure so improvement is observable.

  3. T2
    Bounded ownership

    Move expensive or high-frequency work behind clear worker and memory boundaries.

  4. T3
    Deterministic transport

    Version messages, snapshots, identities, deadlines, and replay inputs.

  5. T4
    Semantic assistance

    Let local retrieval and models propose meaning only inside admitted evidence.

  6. T5
    Candidate paths

    Compare several allowed next actions with explicit risk, cost, fit, and uncertainty.

  7. T6
    Shadow verification

    Challenge the production route without giving the challenger authority.

  8. T7
    Governed production

    Promote only measured, replayable paths with fallback and operational receipts.

  9. Tier-S
    Experimental challenger

    Run advanced evaluators in shadow, compare them with the control, and graduate only repeatable gains.

CATS

Context-Aware Tier Selection

CATS chooses a qualified execution tier for the evidence, device, latency, permissions, and failure conditions of the task.

SEMANTIC BRIDGE

Meaning becomes typed product state

The Semantic Bridge maps a goal to explicit entities, evidence, uncertainty, tools, and permissions. Missing or contradictory information remains visible.

DETERMINISM

Validate the action and retain the result

Models may retrieve, compare, or propose. Exact software checks facts, permissions, state changes, and execution, then records what happened.

OverLift securityAgentic AI Glossary

Read the OverLift MethodOverLift.ai See other products
PROOF

Evidence and limits.

The inspected game source package records controlled recovery tests, restored-versus-uninterrupted decision comparisons, and .NET build/HTTP checks for Kokoro voice integration. It does not establish browser GPU synthesis, listening quality, automatic browser recovery, sub-50 ms fresh speech, or stronger tactical scores. This portfolio update adds no new game qualification.

View project screenshots

Back to the project overview ↑

Discuss Neon Drift View resume All projects