# AI Disclosure, Bias Monitoring, and Human Oversight

Version: 2026-07-22.3  
Status: readiness evidence, not an independent fairness audit

## What the mentor is

The mentor is a local structured teaching model with an optional allowlisted textbook metadata lookup. It is not a general-purpose conversational LLM. Every generated teaching display begins with `AI MENTOR`. Locally validated rules—not online prose—determine concepts, formulas, structures, and assessment goals.

## Signals used

The adaptive model can use completed game actions, task criteria, unsuccessful attempts, prior mentor assessment outcomes, independent performance, concept variety, and later retrieval evidence.

## Signals excluded

It does not collect or infer age, actual school grade, intelligence, race, ethnicity, gender, disability, income, identity, or diagnosis. Input method, mouse skill, keyboard skill, movement speed, and edit speed do not raise or lower the instructional band.

The automatically selected middle-school, high-school, or college profile is an instructional challenge band only. It must never be shown as the player's actual educational level or as psychometric mastery.

## Monitoring measures

For each released version, evaluate plan validity, concept distribution, support recommendations, overrides, concern reports, false advancement, false downshifting, completion, and accessibility outcomes. Compare results across voluntarily recruited test groups only when a qualified study has consent, safeguarding, privacy, adequate sample design, and a legitimate purpose. Do not add protected attributes to normal gameplay telemetry.

## Documented remediation examples

1. Risk: a fast builder could be labeled as understanding after one motor sequence. Change: mastery language was removed; advancement now requires multiple completed assessments, independent evidence, different concepts, hysteresis, and conservative performance-only records. Speed and input method are excluded.
2. Risk: pointer-only editing and visual-only teaching disadvantage keyboard and assistive-technology users. Change: the canvas is focusable and labeled, mentor content has a live text equivalent, reduced-motion styling is included, and dedicated touch controls now cover movement, looking, mining, placing, and selection. Keyboard terrain editing and validated assistive-technology action parity remain unresolved accessibility gaps.

These examples document design changes; they do not prove the absence of bias.

## Player and educator agency

- Shift+H cycles future mentor support among automatic, more guidance, and more challenge.
- H shows or hides tutorial hints.
- P cycles future lesson pace.
- V skips the active AI recommendation without changing evidence, ability estimates, or competency status.
- B flags the current/recent lesson and downloads a non-identifying report.

## Reporting gap

The B control is an in-product flag and local export, but the included static build has no developer-operated receiver. Before applying for certification, add a secure receipt, acknowledgement, triage, correction, appeal, and response workflow. Publish who reviews reports and expected response times. Never ask a child to include unnecessary identifying information.

## Release gate

No version should claim responsible-AI certification until the official reviewer accepts the public documents, in-product disclosure, reporting workflow, override, remediation evidence, and external fairness evaluation.