Two Approaches to AI Safety
The AI industry has converged on a safety model: build the most powerful model you can, then add guardrails after the fact. RLHF (Reinforcement Learning from Human Feedback), constitutional AI, red-teaming, content filters. All of these operate on the same principle: constrain the output.
This is like building a car that goes 300 mph, then adding a speed limiter. The car still wants to go 300 mph. You’re just hoping the limiter holds.
We built the other kind of car.
What “Fiduciary” Actually Means
In law and medicine, a fiduciary is someone who is legally obligated to act in your interest, not their own. Your doctor, your lawyer, your financial advisor — they have a duty that goes beyond “don’t be evil.” They must actively be good. They must prove their work. They must tell you when they don’t know.
This is the standard we hold our AI to. Not alignment — fiduciary duty.
Post-Hoc Alignment
- Generate first, filter after
- Policy-based guardrails
- Can be jailbroken by clever prompting
- Safety team reviews outputs
- “We try our best”
Fiduciary Architecture
- Evidence verified before generation
- Axiom-enforced constraints
- Harmful paths are unreachable, not blocked
- Every response is auditable to source
- “Enforced by code, not by hope”
How It Works in Practice
In our architecture, the language model doesn’t decide what to say. The language model is the voicebox — the Wernicke-Broca layers that turn structured evidence into natural language. The thinking happens upstream.
Before the model generates a single token, the system:
1. Retrieves evidence from 100 million+ verified knowledge atoms, each cryptographically sealed with a Name (what it is), a Form (how it’s structured), and a Dharma (its purpose and constraints).
2. Validates the evidence using the buddy system — no single atom can be used alone. A second, independent atom must confirm. If confirmation fails, confidence drops.
3. Computes a confidence score (the Confidence-Falsifiability Delta) that determines whether the system should speak, hedge, or stay silent.
4. Applies 46 governing axioms that constrain the inference path. These axioms cannot be overridden by the model, the user, or even us. They’re compiled into the pipeline.
Only then does the language model receive the evidence and the behavioral constraints. It verbalizes what the architecture has already validated. It doesn’t decide. It reports.
Why This Matters for Real People
When Asha tells a patient about a drug interaction, that answer traces through peer-reviewed literature, verified by a second independent source, scored for confidence, and auditable back to the original papers. If the evidence is insufficient, Asha says so.
When Harley builds a workout for a client with a herniated disc, the exercise selection traces through anatomy, rehabilitation science, and contraindication databases. The system knows what to avoid before the trainer even asks.
When Artha (Harley’s built-in AI CFO) projects a trainer’s revenue, it draws on financial models grounded in the trainer’s actual business data.
The Standard We Hold Ourselves To
We didn’t choose this standard because it’s marketable. We chose it because we’re physicians. We take care of patients. We know what it costs when trust fails. And we built the AI we wished existed — for our patients first, and then for everyone.
Deepan Singh, MD, FAPA & Paridhi Anand, MD
Co-Founders, DNAi Systems