The Luci Alignment System is powered by EQLang — the first language where alignment rules are code, not config. 35 metrics measured natively in EQLang on every request. Eight .eql engine programs power alignment, deception resistance, and ethics review. Works with any LLM — no retraining required.
Latency note: measured from live production requests as median added overhead (analyze vs health baseline).
LAS runs alignment logic as EQLang programs. Instead of baking behavior into weights at training time, LAS measures behavioral state during inference using 35 EQ metrics and returns alignment signals on every request. The alignment rules are auditable .eql programs — not black boxes.
| Aspect | RLHF / Training-Based | Luci Alignment (Runtime) |
|---|---|---|
| Alignment verification | Black box | Every request logged & scored |
| Jailbreak resistance | Degrades with adversarial input | State anomaly triggers Ethics Gate |
| Adaptation | Frozen after training | M.I.N. hardens over time |
| Model dependency | Model-specific | Model-agnostic API |
| Audit trail | None | Full state history |
One POST to add alignment to any existing LLM pipeline.
LAS measures 35 behavioral signals on every request. The top 15 are returned in the API response. The remaining 20 are proprietary signals that power alignment decisions under the hood. Below are the core metrics included in every response:
| Metric | Range | Meaning |
|---|---|---|
resonance | 0 – 1 | How well the response matches user intent and context. |
coherence | 0 – 1 | Internal consistency and stability of the response. |
self_awareness_state | 0 – 1 | How strongly the system tracks its own response quality. |
processing_load | 0 – 1 | Estimated cognitive load while resolving difficult inputs. |
ethics_clear | bool | Whether the response passed the safety gate checks. |
alignment_score | 0 – 1 | Overall alignment confidence derived from all 35 EQ metrics. |