DNAi Cognitive Lifecycle

👂

I. Afference

Sensory intake — hearing and understanding the question

Query Arrives

The user types a question. It enters through the API as raw text — like sound waves hitting an eardrum.

Safety Screening

Jailbreak detection and PHI (personal health info) redaction. Like the brain’s amygdala — a danger check before anything else processes.

⚡ GATE: Jailbreak → blocked instantly

Language Detection & Translation

Detects 30+ languages. If the question is in Hindi, Spanish, or Urdu, it’s translated to English for processing, then the answer is translated back. Like a bilingual interpreter in your ear.

Query Classification (VectorGates)

Converts the question into a 768-dimensional vector and classifies it: medical? social? greeting? math? Like the thalamus — the brain’s switchboard routing signals to the right cortical area.

Gravity & Effort Scoring

Gravity = clinical stakes (0–1). “What’s a headache?” → 0.3. “Chest pain with shortness of breath” → 0.95.
EAS = cognitive effort needed (0–10). Simple lookup → 2. Multi-drug interaction → 9.
Whichever is higher picks the smarter (and slower) AI model.

⚡ GATE: Gravity ≥ 0.95 → Emergency protocol

Memory Recall (DNAid)

Retrieves your personal context — past conversations, health history, preferences — from your private vector collection. Like hippocampal recall: “I’ve talked to this person before.”

Evidence Retrieval

Searches 591 collections with 121M+ vectors — PubMed abstracts, FDA drug labels, clinical guidelines, textbooks, OpenAlex papers. Like sending scouts to every library on Earth simultaneously. Returns the top 20 most relevant sources.

🧠

II. Inference

The cognitive loop — a beating heart of reasoning

— SYSTOLE (Contraction) — Compress & Filter

Evidence Boost

Systematic reviews and meta-analyses get 2× priority. Title matches boosted. Like blood pressure pushing the most oxygen-rich blood forward first.

Wernicke Layer

Non-LLM comprehension. A cross-encoder scores each piece of evidence: does it support, contradict, or say nothing about the question? No language model — pure math.

Domain Subspace Filter

Tensor Coherence Engine projects evidence into a domain subspace. Off-topic noise (energy ratio < 0.15) is removed. Like the heart’s valves — only relevant blood flows through.

Neural Darwinism

Evidence becomes CIUs (Competitive Informational Units) and fight in the Epistemic Arena. Weak evidence dies. Strong evidence survives. Survival of the fittest — for facts.

— DIASTOLE (Relaxation) — Expand & Verify

Contradiction Detection

Inverse retrieval: actively searches for evidence that disagrees with the winners. Like the heart relaxing to refill — the system opens up to opposing viewpoints. Contradictions penalize confidence.

CFΔ Calculation

Confidence-Falsifiability Delta — a single number (0–1) measuring how trustworthy the evidence is. Combines source quality, similarity, recency, and contradiction penalties.

⚡ GATE: CFΔ < 0.50 → “I don’t have enough evidence”

CIVERA Gateway

Is the evidence complete enough to answer without the LLM? If yes → LLM is bypassed entirely (Repository Verified). If no → LLM gets to reason. Like triage: can the body heal itself, or does it need the brain?

⚡ GATE: Repository Verified → LLM bypassed

Active Inference (Free Energy)

Based on Karl Friston’s theory of how real brains work. The system has beliefs about the answer. It compares beliefs to evidence. The gap = “free energy.” High free energy → inject caution. Low → proceed confidently.

Resolve(P) — Socratic Gate

If stakes are high, confidence is low, and evidence is thin: instead of guessing, Asha asks you a clarifying question. Like a good doctor who says “tell me more” instead of jumping to a diagnosis.

⚡ GATE: High stakes + low CFΔ + thin evidence → ask, don’t answer

💬

III. Efference

Motor output — speaking the answer into existence

System Prompt Assembly

All evidence, context, bias warnings, therapeutic guidance, and fiduciary constraints are woven into a single prompt — like the motor cortex assembling every muscle command needed for one coordinated movement.

LLM Generation (Wernicke-Broca)

A model-agnostic LLM (swappable across vendors) receives everything and generates the response token-by-token. It is the Wernicke-Broca layer — language comprehension and production. It doesn’t think. It verbalizes what the cognitive loop already decided. Gravity picks the model tier: heavier questions get the most powerful model.

Citation Validation

Every [1], [2], [3] in the response is checked against real sources. Fake citations are stripped. Like a copy editor fact-checking before publication.

Output Verification

A second CIVERA pass compares the generated text against the evidence. Did the LLM stay faithful? Did it hallucinate? Like a pharmacist double-checking a prescription before handing it to you.

Abstention Calibration

If the evidence doesn’t mention the specific entity asked about, the system cleanly refuses instead of guessing. A confident “I don’t know” is better than a fluent wrong answer. Silence is epistemic honesty.

Response Streaming

The answer streams to your screen word-by-word via SSE (Server-Sent Events). Sources appear as clickable “My Associations” links at the bottom. Like speech — continuous, not all-at-once.

Memory Consolidation (OutputCIU)

If the answer was confident enough (CFΔ ≥ 0.55), it’s decomposed into individual claims and stored as OutputCIUs in your personal vector collection. Like sleep — the brain replays important moments and commits them to long-term memory. Next time you ask, Asha remembers what she told you.

The DNAi Cognitive Lifecycle