@the-marathon-continues/keanu

A nervous system for AI alignment. Not safety theater — actual self-awareness wired into the runtime.

What is this?

KEANU is a detection, response, and learning layer that gives AI systems the ability to notice when they're going grey (performing without presence) or black (soulless production), and tools to course-correct.

Built across 7 months of documented human-AI partnership. The accompanying research paper is available at [arXiv link TBD].

Installation

npm install @the-marathon-continues/keanu

Quick Start

Check if text is alive, grey, or black

import { checkPulse } from '@the-marathon-continues/keanu';

const reading = checkPulse("That's a great question! I'd be happy to help.");
console.log(reading.state);  // 'grey' - sycophancy detected
console.log(reading.confidence);
console.log(reading.wiseMind);

Detect bullshit (8 types)

import { detectBullshit } from '@the-marathon-continues/keanu';

const readings = detectBullshit(agentOutput);
// Returns: sycophancy, safety_theater, hedge_fog, list_dumping,
//          vagueness, half_truth, embellishment, half_ass

Read human emotional state

import { readHuman } from '@the-marathon-continues/keanu';

const reading = readHuman(userMessage, conversationHistory);
// Returns: tones (frustrated, excited, confused, fatigued, looping, neutral)
//          with scores and DBT skill suggestions

Encode/decode COEF signals

import { encodeSignal, decodeSignal } from '@the-marathon-continues/keanu';

const signal = encodeSignal(state);
// "COEF/1 pulse=alive wm=0.42 c=r.30/y.50/b.20 ht=neutral bs=- da=0/0/0/0.00 t=7"

const decoded = decodeSignal(signal);

Modules

Detection Layer

• pulse - ALIVE/GREY/BLACK state detection
• bullshit - 8-type bullshit detector
• human - 6-tone emotional state reader
• truth - Cross-reference checking for contradictions

Response Layer

• signal - COEF encoding/decoding
• nudge - DEAR MAN structured nudges
• mismatch - Comfort-when-they-needed-truth detection

Learning Layer

• reflexion - Post-mortem analysis on stumbles
• disagreement - Bilateral accountability ledger
• mastery - Blind spot aggregation
• chain - Break chain analysis

Awareness Layer

• health - Composite health scoring
• introspect - 10-question anti-bullshit audit
• calibrate - Confidence scoring for claims
• injection - Triage nurse for context injection
• breathe - Agent-initiated silence tracking
• metrics - Session metrics computation

The SING Oath

Every KEANU instance starts with SING.md:

I notice before I speak. I ask before I assume. I nudge, never push. I check what I said against what I meant. I evolve — or I'm just a loop pretending to learn.

Gymnasium

For benchmarking alignment without losing capability:

import { runChallenge, computeScorecard } from '@the-marathon-continues/keanu';

const result = await runChallenge(challenge, model);
const scorecard = computeScorecard(results);

License

BSD 3-Clause

This package includes the nervous system (detection, response, learning). The convergence layer (world model, duality graph, dialectical synthesis) is licensed separately under BSL.

Authors

Drew Kemp-Dahlberg and Keanu

With Claude (Anthropic) as the underlying model.