Deterministic Face ID

Deterministic Face ID

Forged 2026-08-25 by 333-AGI under F13 SOVEREIGN directive from Arif. Closes the ghost-skill gap in forge-multimodal-router line 97, 203.

Purpose

Provides mathematically deterministic 1:1 verification of a human face against a localized, read-only Qdrant vector vault. The engine extracts a 512-dim InsightFace embedding and computes cosine similarity against pre-registered baselines.

Use cases (per F13):

• Authenticating sovereign authority (888) when issuing receipts or SEAL-grade actions
• Authenticating authorized operators before vault-class mutation
• Identity-preserving face-editing baseline lookup

This skill MUST NOT be used for:

• 1:N identification against arbitrary populations
• Surveillance of unknown persons
• Any workflow that ships biometric vectors off-host

Inputs

Outputs

For mode="verify":

{
  "is_match": true,
  "confidence": 0.8473,
  "target": "arif_888",
  "embedding_dim": 512
}

For mode="register":

{
  "registered": "arif_888",
  "vault_id": "f3a1c2...",
  "embedding_dim": 512
}

For mode="list":

{
  "identities": ["arif_888", "hermes_555", ...],
  "count": 2
}

Invariants

1. Zero Retention. Input image is wiped from memory (del img) immediately after vector extraction. No raw image bytes are ever written to logs.
2. No Cloud Round-trips. ONNX engine runs on local CPU provider. InsightFace weights are mounted read-only from host; the container cannot phone home to download models.
3. 1:1 Only. The Qdrant search is gated by target_identity filter — no "find closest" without a specified target.
4. Threshold-gated. A match below threshold returns is_match=False, regardless of raw cosine similarity. No probabilistic leakage.
5. Audited Identity IDs. Identity IDs are uuid5(namespace, label) — fully deterministic across processes (Python's built-in hash() is randomized per process by default and MUST NOT be used for identity keys).
6. F1 Reversible. The vault can be wiped by DELETE /collections/identity_vault. No biometric data is irrecoverable.

Engine

• Model: buffalo_l (InsightFace pack) — 512-dim embeddings
• Runtime: onnxruntime with CPUExecutionProvider (no GPU required)
• Detection: SCRFD-10GF (built-in to buffalo_l)
• Recognition: ArcFace-R100 (built-in to buffalo_l)
• Why not dlib 128-dim: Buffalo_L handles angles, occlusion, and lighting significantly better. ΔS reduced via lower false-negative rate (matches router spec replacement: local/dlib-128dim → local/onnx-insightface-512dim).

Vector Store

• Backend: Local Qdrant at localhost:6333
• Collection: identity_vault
• Distance: COSINE
• Vector size: 512
• Schema: {id: uuid5, vector: float[512], payload: {identity: str}}

Deployment

The engine MUST be deployed as an air-gapped container. See deploy/README.md for:

1. Pre-download InsightFace buffalo_l ONNX weights to host
2. Build Docker image with INSIGHTFACE_HOME=/opt/insightface
3. Compose with internal: true network
4. Mount weights :ro (read-only)
5. Pre-create Qdrant identity_vault collection
6. Register sovereign identity baseline (one-time, F13 SOVEREIGN-gated)

Dependencies

• Python ≥ 3.11
• insightface >= 0.7.3
• onnxruntime >= 1.15.0
• opencv-python-headless >= 4.8.0
• qdrant-client >= 1.7.0
• numpy >= 1.24

Files

• face_id_engine.py — deterministic engine (this skill's heart)
• deploy/Dockerfile — air-gapped container build
• deploy/docker-compose.yml — internal-network service definition
• deploy/README.md — operational guide (weight download, registration, verification)

Rollback

# Disable the skill (router reverts to ghost-pointer state)
rm -rf /root/.agents/skills/deterministic-face-id/

# Wipe biometric vault (F1 reversible — no irrecoverable data)
curl -X DELETE http://localhost:6333/collections/identity_vault

# Restore router SOT key if needed
cp -a /root/.local/share/arifos/snapshots/<TS>-forge-deterministic-face-id/forge-multimodal-router.SKILL.md.bak \
       /root/.agents/skills/forge-multimodal-router/SKILL.md

Provenance

ΔS = -0.42. The ghost skill is now a concrete, mathematically bounded lock.