Neural Governance Engine

Run local ONNX inference for PII, injection, and intent detection — sub-50 ms, 85-95% LLM cost reduction.

The Neural Governance Engine (NGE) is Palveron's local inference layer. Five ONNX models run inside the gateway, turning most LLM-assist round-trips into sub-50 ms local calls.

Prompt → NGE pipeline (5 stages) → Decision
            1. Regex            fast, deterministic
            2. Aho-Corasick     multi-keyword matching
            3. ONNX NER         entity extraction
            4. NLI Contextual   semantic intent scoring
            5. LLM-Assist       escalate borderline only

What NGE detects

Signal	Source stage	Use case
Structured PII (SSN, IBAN, credit card, email, phone)	Regex	Strict, low-false-positive masking
Brand-specific keywords, custom denylists	Aho-Corasick	Internal product names, blocked phrases
Named entities — `PERSON`, `ORG`, `LOC`, `EMAIL`, `SECRET`, `INJECTION`	ONNX NER × 2	Context-aware PII detection beyond regex
Intent classification — `injection`, `toxicity`, `off_topic`, `pii_density`	NLI Contextual	Catches paraphrased malicious intent
Edge cases	LLM-Assist	Last-resort cloud escalation for low-confidence local results

Engine modes

Configure under Settings → Security → Neural Governance Engine or via env var NGE_MODE:

Mode	Behavior	When to use
`disabled`	Regex + LLM-assist only (pre-Sprint-54 behavior)	Benchmarking or fall-back
`nge_local`	Local models only — never calls the cloud LLM	EU data residency, zero-egress; foundation for air-gapped deployment (on the roadmap)
`nge_fallback` (default)	Local-first; borderline cases escalate to LLM	Most customers — best balance
`llm_only`	Skip local, always call cloud LLM	Benchmarking accuracy upper bound

Palveron's internal benchmarks (nge_fallback vs. llm_only) show:

85-95% reduction in LLM-assist calls
Under 1 percentage point accuracy drop on a 50k-prompt evaluation set
6-10× latency improvement at the p50 (median)

Sensitivity presets

A single slider maps to internal thresholds across all five stages:

Preset	Thresholds	Typical industry
Strict	Low — more BLOCKs and MODIFYs	Healthcare, finance, government
Balanced (default)	Vendor-tuned	General enterprise
Tolerant	Higher — fewer interventions	Internal R&D, developer tooling

Don't tune sensitivity per-policy — use presets at the project level and rely on per-policy enforcement actions (BLOCK, APPROVAL, ANONYMIZE, FLAG) to handle granularity.

Shadow Mode

When Shadow Mode is on, the engine evaluates every request but does not enforce. Traces show what would have happened; the request always passes through unchanged.

Use Shadow Mode to:

A/B test a stricter sensitivity preset before rolling out
Onboard a new agent type without false-positive blocks
Validate NGE accuracy against a known-good corpus

Switch enforcement back on under Settings → Security → Neural Governance Engine → Shadow mode when satisfied with the trace history.

Reading NGE results in traces

Open any trace in the Trace Explorer. NGE-driven traces include:

Engine badge — nge_local, nge_fallback (local), nge_fallback (LLM), llm_only, disabled
Confidence bars — per NLI dimension (injection, toxicity, off_topic, pii_density) with the active threshold line
Entity highlights — color-coded inline annotations on the prompt: PII, SECRET, INJECTION, TOXICITY
Stage timing — milliseconds spent in each pipeline stage

Stage              ms     Decision contribution
Regex              0.3    ssn match → MASK
Aho-Corasick       0.4    no hit
ONNX NER           18.2   PERSON × 1, EMAIL × 1
NLI Contextual     22.1   injection: 0.04 (under 0.5)
LLM-Assist         —      not invoked
Total              40.9   ✓ MODIFIED

NGE in the API response

POST /api/v1/verify returns NGE fields alongside the decision:

{
  "decision": "MODIFIED",
  "engine": "nge_local",
  "nge_policy_scores": {
    "injection": 0.04, "toxicity": 0.02,
    "off_topic": 0.18, "pii_density": 0.91
  },
  "detected_entities": [
    { "type": "SSN", "start": 26, "end": 37, "confidence": 0.99 }
  ],
  "deterministic_results": { "regex_matches": ["ssn_us"] }
}

See api/verify for the full response schema.

Language packs

Default packs: English + German. Each additional pack (Spanish, French, Italian, Portuguese, Dutch, Polish) adds one ONNX model (~250 MB) and loads on first use.

Enable extra packs under Settings → Security → Neural Governance Engine → Language packs. Multi-lingual prompts auto-route to the closest pack via the gateway's language detector.

Self-hosted considerations

On self-hosted deployments, set:

NGE_MODE=nge_fallback
NGE_MODELS_DIR=/app/models/nge

Models are distributed separately (they're ~4 GB total). Download:

python scripts/download-nge-models.py --models-dir ./models/nge

Mount the directory into the gateway container (the Helm chart does this for you via a PVC). NGE warms up on boot — the gateway's /ready endpoint returns 200 only after every model has finished loading.

NGE scores are assistance signals, not legal verdicts. They inform the policy engine; the policy engine still owns the final decision. Treat them like a smoke alarm — informative, but the firefighter (your policy) decides what to do.

Neural Governance Engine

On this page