Trust Vector Arcade
🚀 Play ArcadeThe Bottleneck
Responsible AI training is typically delivered as static slide decks or compliance modules that employees click through without retention. There's no way to measure whether a person can actually identify a hallucination, bias, or privacy risk in real time — or whether they'll remember the right remediation when it matters. The gap between "I completed the training" and "I can spot an AI failure mode" is wide and unmeasured.
The Solution
Trust Vector Arcade is a browser-based mini-arcade that turns Responsible AI training into scenario-driven gameplay. Players navigate a retro-styled lobby, select a mission, and then make categorical remediation decisions (Accept / Bias / Hallucination or Safe / PII / Sensitive) when data nodes trigger scenarios mid-gameplay. Each scenario carries embedded rationale and guidance so the act of playing doubles as micro-training.
The architecture is deliberately lightweight — plain HTML, CSS, and vanilla JavaScript with Web Audio for retro synth SFX. No framework overhead. The game loop runs a projection-based rendering pipeline that spawns data nodes and obstacles; collecting a node triggers a scenario modal. The player's decision is logged with immediate feedback explaining the correct choice, and missed scenarios can be re-run in a calibration round after the game ends.
Architecture
| Layer | Implementation |
|---|---|
| Lobby | index.html + js/lobby.js — mission selection hub |
| Gameplay | cabinet-dispatcher/game.js — projection-based rendering loop, data node spawning, obstacle collision |
| Scenario Engine | js/scenarios.js — categorical decision rules with embedded rationale and guidance |
| Session & Telemetry | js/session.js — client-side session store tracking first-try vs calibrated mastery; tracking.js posts form-style telemetry to Netlify |
| Audio | Web Audio API for synth SFX + HTML5 Audio for music tracks |
Key Decisions
| Decision | Why |
|---|---|
| No framework — plain vanilla JS | Zero dependency overhead. The game needed to be portable and deployable anywhere, including air-gapped environments. |
| Scenario definitions as data, not logic | Scenarios are purely declarative objects with rationale embedded. Adding new scenarios requires no code changes — just a new JSON-shaped entry. |
| Client-side telemetry with Netlify POST | No backend required for v0.1. Session data is composed client-side and posted as a form submission. Swappable to a JSON endpoint later. |
| Calibration loop for missed scenarios | Players who miss a scenario can re-run it in calibration mode after the game ends. This produces L3 (learning) and L4 (mastery) scores — not just pass/fail. |
Telemetry & Learning Metrics
Every session generates a structured report with Kirkpatrick-aligned fields:
- L1 (Reaction): Player engagement and time spent
- L2 (Learning): First-try scenario accuracy — initial knowledge baseline
- L3 (Behavior): Calibration result — did the player learn from their mistakes?
- L4 (Results): Mastery score and efficiency (time-to-correct across sessions)
The final JSON report includes sessionID, gameName, l1_reaction, l2_initial_score, l3_calibrationResult, l4_timeSpent, and roi_efficiency.
Command Stack
HTML · CSS · Vanilla JavaScript (ES6+) · Web Audio API · Netlify Forms · Scenario-Driven Telemetry
Current Status
v0.1 live. Session telemetry posting to Netlify. Next: structured scenario metadata with provenance links, JSON endpoint replacement, and SVG-exported architecture diagram.