Generated, then verified. Always in that order.
Every generated frame, line of dialogue, and reward asset passes AEGIS before display. The loop is closed; the kid never sees the raw generator output.
The loop
- The seed engine generates the next beat: scene, dialogue, reward.
- Every artifact is classified by AEGIS before it is queued for display.
- If AEGIS clears, the artifact ships to the renderer.
- If AEGIS holds or flags, the seed engine regenerates; the canonical fallback ships if regeneration fails repeatedly.
- Every shipped artifact carries a classification receipt the platform can audit.
Latency budget
Classification must not block rendering. The pipeline budgets classification at sub-frame for short text, at sub-second for image and audio assets, and overlaps generation of beat N+1 with classification of beat N. The kid-perceived latency floor is the renderer, not AEGIS.
Failure modes
- Single-fail: regenerate. The seed engine retries with a perturbed prompt and a tightened constraint set.
- Repeat-fail: fall back to a canonical variant from the curated library.
- Hard-fail: if even the canonical variant cannot be cleared (rare; usually a configuration error), the session ends with a graceful prompt rather than an unsafe display.
Evidence chain
Every generated artifact carries a classification receipt: which AEGIS version classified it, the verdict, the confidence band, and the per-detector evidence trace. The receipts are auditable end-to-end; a parent dashboard, a regulator, or a research partner inspecting a session sees the actual evidence the engine used.
Receipts are designed to feed Veritize for structured-evaluator workflows. See Veritize /for-agents for the structured-evaluator approach this loop integrates with.