Matrix · Corp · Intelligence · Substrate

ECHO

It remembers every wound

Qwen3.5-27B · Opus 4.6 Distilled · Rust
A living model that crystallizes its mistakes into Scars — structured memories that grow stronger with every correction. The more it fails, the harder it is to fool.
descend
Interactive Demo

Watch a Scar form

Correction Interface
Active Scar Feed
Factual
"Paris is the capital of Germany"
Paris is the capital of France
Hallucination
"Einstein won the Nobel Prize in 1925"
Einstein won in 1921 for photoelectric effect
Logical
"All primes are odd therefore 2 is not prime"
2 is the only even prime number
Scar Lattice — Live
Pre-scan Output
// Awaiting query...
struct Scar {
  id: Uuid,
  domain: String, // "math.calculus"
  scar_type: ScarType,
  original_claim: String,
  correction: String,
  confidence_at_time: f32, // was 0.94
  times_triggered: u32,
  weight: f32, // 0.0–1.0
  fingerprint: Vec<f32>, // 128-dim
  created_at: DateTime<Utc>,
  last_triggered: Option<DateTime>,
}

enum ScarType {
  Factual | Logical | Contextual
  | Hallucination | Overconfidence
}
The Primitive

A Scar is a structured wound

Every correction ECHO receives becomes a Scar — a typed, weighted memory object that lives in the Scar Lattice. Unlike RAG or fine-tuning, Scars form instantly, cost no compute, and grow stronger every time they fire.

The fingerprint is a 128-dim semantic vector computed in pure Rust — no ML model, just math. Similarity search across 10,000 Scars completes in microseconds via Rayon parallel scan.

weight Influence strength. Increases every time this Scar fires in a pre-scan. Decays slowly over time.
confidence_at_time How certain the model was when it was wrong. High confidence + wrong = overconfidence Scar.
times_triggered How many future queries activated this Scar. A proxy for how common this mistake pattern is.
fingerprint 128-dim semantic vector. Used for cosine similarity matching against new queries in O(n) via rayon.
Domain Weakness Map

Where it bleeds most

Risk score = Scar volume + average weight + confidence at mistake. High risk domains get automatic confidence suppression.

math
0.82
history
0.71
science.physics
0.64
coding
0.48
geography
0.31
technology
0.22
The Loop

How ECHO heals itself

💬
Input User prompt arrives

Raw query enters ECHO. Before a single token is generated, the Scar system intercepts.

🔬
Pre-Scan Lattice searched for wounds

128-dim fingerprint computed in Rust. Cosine similarity scan across all active Scars via Rayon. Top-k matches retrieved in microseconds.

⚠️
Caution Injection Past wounds prepended

If matching Scars found — a caution prompt is prepended: what was wrong before, what the correction was. Confidence suppressed up to 30% in risky domains.

🧠
Generation Qwen3.5-27B runs in Rust

Full inference via candle. The base model generates with scar context already in its window — no architecture change needed, just smarter input.

🩸
Correction User corrects → Scar forms

POST /v1/echo/correct. Scar crystallizes instantly. Duplicate check — if a similar Scar exists, its weight increases instead of creating a duplicate. Saved to disk atomically.

Architecture

By the numbers

0 Billion parameters
0 Scar fingerprint dims
0 Max active Scars
0 Scar types