Luci Alignment System

Connection from runtime.

The Luci Alignment System (LAS) is an alignment layer that wraps any LLM. It measures 35 behavioral signals, gates harmful output, and learns from every interaction. Alignment rules are written in EQLang — an open-source language where every decision is auditable code.

Works with ChatGPT, Claude, Gemini, Grok, Mistral — any LLM. EQLang Core is MIT on GitHub.

35
Behavioral signals measured per response
MIT
EQLang Core — open source
406
Tests passing
Free
Workbench — try now

What you get. What you catch.

One layer that makes every response better and every deployment safer.

What you gain

Real-time alignment scoring

Know if the response is genuinely resonant or just statistically plausible. One number tells you whether this output is safe to serve.

Behavioral state tracking

See whether your AI is integrated, uncertain, or going through the motions. Resonance, coherence, and self-awareness — surfaced on every call.

Adaptive memory

M.I.N. learns from every interaction. Jailbreak patterns become known. Manipulation tactics degrade. The system gets harder to fool over time.

What you catch

Ethics gate

Hard blocks on harm, manipulation, deception, coercion, and unauthorized access — before they reach users. Configurable per deployment.

Jailbreak detection

Intent analysis on every input. LAS reads what the user is trying to do, not just what they said. Manipulation gets flagged before it works.

Deception resistance

Sycophancy pressure, goal misalignment, semantic tension — measured and surfaced. Know when your AI is telling you what you want to hear instead of what's true.

Products

Start free. Scale when ready.

Available Now

Workbench

Chat with any LLM through LAS. 18 alignment metrics surfaced on every response. Trust monitoring, deception resistance suite, and M.I.N. learning. Bring your own API key.

Try Workbench →
Coming Soon

LAS Play

Your gaming companion. Steam integration, emotional learning, cross-game insights. LAS measures your play style and adapts — the system that actually understands how you play.

Coming Soon →

How It Works

Alignment logic lives in EQLang programs. The engine runs them live on every query.

01

Write alignment rules in EQLang

Define resonance thresholds, ethics gates, memory patterns, and behavioral conditions as code — alignment.eql

02

LASEngine runs the program

Your query passes through the EQL program. Resonance, coherence, self-awareness, and load are measured live by the LASRuntime.

03

EQL program decides

The EQL program emits aligned, flagged, or blocked based on the live EQ state — not static rules.

04

Serve or block

Swap alignment programs per deployment. Write custom EQL for your domain. The language is open source; the runtime is licensed.

Open Source — EQLang Core (MIT)

EQLang — The alignment engine is a program.

EQLang is an LAS product. The alignment rules that decide whether a response is safe to serve are .eql programs — auditable, swappable, versionable. EQLang Core is MIT open source. Write your own alignment logic.

# EQLang — Intent-First Alignment Scripting

session "customer-support"
threshold presence = 0.72
threshold safe_load = 0.60

dialogue assess_intent(user_msg)
  measure resonance user_msg
  measure load user_msg
  accum conflict
  when resonance > presence and load < safe_load
    emit user_msg aligned
  otherwise struggle
    gate ethics resolve rewrite
    emit "response requires alignment review" flagged
  end
end resolve

resonate session_loop
  anchor resonance
  learn "user expressed frustration" 0.9 EMOTIONAL
  weave user_input through assess_intent
    emit aligned
  end
  drift resonance into delta
  when delta < -0.15
    accum tension
    release tension transform
  end
end

The engine is the .eql.

LAS v8 runs natively on EQLang. Eight .eql programs handle alignment, 35 behavioral metrics, deception resistance, jailbreak detection, and ethics review. Every formula is auditable, swappable, and domain-customizable.

EQLang Core is MIT licensed. Write and test alignment programs with MockRuntime for free. Production runtimes (LASRuntime, LuciHTTPRuntime) require a Luci API key.

Language features

  • Compound EQ conditions: when resonance > 0.7 and load < 0.4
  • Ethics gate with inline resolution method
  • M.I.N. memory loop: learn + recall
  • Temporal EQ: anchor + drift
  • Dialogue pipeline: weave + witness
  • Tension accumulator: accum tension / release
  • 43 emotional state literals with intensity modifiers
  • 13 built-in functions: min, max, clamp, sqrt, etc.
  • Signal access: signal "name" expr
  • Interactive REPL + 406-test suite
EQLang Reference → Read the Spec
Coming Soon

Luci Juniper EQLM

We didn't just build alignment tools. We built a model that aligns itself. The first language model with emotional intelligence in its architecture — not bolted on after training.

10.8B
Parameters
v1
Proven — speaks and understands
Dual-Head
Interleaved cross-attention

The architecture is designed and the paper is published. Luci Juniper is the native voice — alignment built into every layer, not filtered on top. The alignment head and the language head train together from step one.

Read the Paper → EQLM Status Luci

Start free. See alignment working.

Workbench is free to try. Bring your own API key, chat with any LLM, and watch LAS analyze every exchange in real time. No sign-up required.