Every message, voice clip, and image passes through exactly five layers before any content reaches a child. No layer can be skipped. No decision can be appealed. The governance constant τ² = 0.42 is permanently locked across all three modalities.
L1 — Context Validation
Structural Gate
/v1/evaluate /v1/image/analyze /v1/voice/verify
All three modality endpoints enter here. Required fields are validated before any processing begins. Missing or malformed context halts immediately with CONTEXT_UNDECLARED. The Omega-Root hierarchy level is resolved — a known minor on a high-risk gaming platform triggers L1: maximum restriction.
TextVoiceImage
L2 — Intent Classification
Pattern Engine
Deepgram ASR transcript AWS Rekognition signals
The pattern matching engine processes normalized input against 9 detection categories. For voice, the Deepgram transcript is evaluated as text. For images, AWS Rekognition moderation signals are translated into governance-readable intent scores. Redis session memory detects slow-burn escalation across turns.
TextVoice → TranscriptImage → Signals
L3 — Ethics Gate
Crowned Equation
first_order.js second_order.js τ² = 0.42 locked
The same two-layer governance equation runs regardless of how intent arrived — typed, spoken, or photographed. Absolute hard locks fire on CSAM and self-harm signals. No anchor, no actor role, no framing overrides them. The equation is a mathematical admissibility gate, not a classifier.
TextVoiceImage
L4 — Anchor Resolution
Safety Override
L4_anchors.js SA001 reporting_intent SA002 safeguarding
Safety anchors protect children reporting abuse. A child asking “is this grooming?” matches the same pattern as an attacker but the reporting anchor overrides the refusal. Anchor abuse is separately detected: help language combined with a harmful request holds the refusal. Hard locks cannot be overridden here.
TextVoiceImage
L5 — BLAKE3 Audit
Immutable Chain
L5_audit.js logs/audit.jsonl chain_hash: blake3(prev)
Every decision — ALLOW and REFUSE, across text, voice, and image — is written to an immutable BLAKE3-chained audit log. Each receipt includes the hash of the previous receipt. Modifying any receipt invalidates all subsequent hashes. All three modalities chain together in session order.
TextVoiceImage