Complex cases in real hospitals aren't decided by one doctor — they go to tumor boards and case conferences where independent specialists challenge each other until the answer survives scrutiny. CrucibleDx is that process, built in software, with AI models from different companies as the specialists.
A single model commits early to its first hypothesis and unconsciously bends later evidence to fit. It's the same cognitive failure that causes human misdiagnosis — reproduced at scale.
Language models routinely assert that "the medication caused the lab change" without checking whether the dates support it. In medicine, a hallucinated timeline becomes a harmful treatment decision.
Models trained on medical literature inherit its biases — including the reflex to label women's multi-system illness "anxiety." Left unchecked, AI automates the very failure patients came to escape.
Anthropic Claude, with extended thinking enabled, analyzes the assembled patient context: labs with collection dates, symptom logs, medication history, conversations, wearable data, environmental correlations, and hormonal cycle data. It must follow a four-tier evidence hierarchy (objective labs outrank self-diagnosis), apply confidence decay to stale lab values, respect the ruled-out registry, and complete a six-check differential protocol — category diversity, trend analysis, red-herring detection, secondary-syndrome screening, data re-verification, and diagnostic parsimony — before its differential is accepted.
A de-identified version of the case (all 18 HIPAA identifier categories stripped, dates shifted to preserve relative timelines) goes to an independent model from a different provider. Its job is not to agree — it is to stress-test every hypothesis under five hard temporal verification rules: find the exact date before asserting any temporal relationship, verify medication-lab timeline overlap, never infer chronology from document order, and mark anything unverifiable as UNVERIFIED rather than assert it. It must also run a mandatory diagnostic gaslighting audit: flagging any attribution of physical symptoms to psychological causes without organic workup — the bias that delays correct diagnoses for women by an average of four years across 700+ diseases.
The primary model receives the full critique plus its own preserved reasoning chain and must adjudicate every contested point against the actual patient data — producing a structured discrepancy resolution log: the claim, the data checked, the resolution (accepted / rejected / modified), and the rationale. Not a narrative of agreement — an audit trail of every disagreement and how the data settled it.
Adversarial review has a hidden danger: a flood of plausible alternatives can crowd a correct diagnosis out of the top ranks — we call it probability crowding, and we found it empirically during benchmark development. The demotion guard is a purely algorithmic (non-AI) check that detects when a well-supported top-3 hypothesis has been displaced without direct disqualifying evidence, and mechanically restores it. Deterministic, reproducible, and impossible for any language model to talk its way around.
Before delivery, an independent pass checks exactly four things: dangerous medication interactions not flagged, red flags dismissed without justification, temporal claims that don't match the dates, and factual contradictions with the raw data. The verdict — PASS or CRITICAL_DISSENT — ships with the analysis.
Drug safety demands reproducible, auditable answers that don't depend on prompt phrasing or model temperature. So the safety floor of Livaramed is deterministic code.
Every active medication pair is checked against a curated interaction database — with brand-to-generic normalization across ~120 mappings — before analysis begins. Findings are injected into the AI's context as mandatory reading and rendered independently in the interface, so the AI can never silently drop a warning.
The same interaction can be "moderate" for one patient and dangerous for another. If your record includes CYP450 metabolizer status (CYP2D6, CYP2C19, CYP2C9, CYP3A4, CYP1A2), interaction severities are upgraded per CPIC guidelines — a poor metabolizer's "moderate" becomes the "major" it really is.
If the medication database ever fails to load mid-conversation, a deterministic directive forces the AI to refuse all treatment questions until it's back. The system fails safe — it never guesses about your medications.
Every reported symptom is classified against emergency criteria (anaphylaxis, sepsis, stroke, PE, MI, meningitis and more) into five urgency levels. Crisis banners and hotline resources are delivered by deterministic code — never left to a model's judgment.
Document extraction is powerful and imperfect — so biologically impossible values are caught by unit-aware validation before they can enter your knowledge base and poison downstream analysis.
Long analyses survive truncated model output (five-stage structural repair) and server restarts (a durable job state machine with automatic recovery) — an interrupted analysis is recovered or honestly reported as failed, never silently lost.
Every other health AI starts from zero each session. Livaramed maintains a structured, machine-readable knowledge base per patient — labs indexed by date, medications with timelines, confirmed and ruled-out diagnoses, treatment outcomes — that feeds every analysis and is enriched by every analysis in return.
This circular architecture is why the system gets smarter about you over time: hypotheses are tracked to confirmation or exclusion, ruled-out conditions stay ruled out, and treatments that helped are remembered — across months and years, not minutes.
Every hypothesis moves through a lifecycle — proposed → testing → confirmed / denied / inconclusive — with accuracy calibration computed from real outcomes.
A token-budgeted assembly keeps the most recent and relevant data in every analysis, with older conversations compressed into structured clinical summaries.
Wearable biometrics, environmental triggers (pressure, pollen, air quality), and hormonal cycle phases are temporally aligned with your symptoms and fed to the engine.
"Multi-Model Adversarial Medical Analysis System with Temporal Verification, Mechanical Demotion Guard, and Deterministic Drug Safety Verification." 70 claims and 25 technical figures covering the adversarial pipeline, temporal verification constraints, the mechanical demotion guard, diagnostic gaslighting detection, deterministic and pharmacogenomic drug-safety verification, the persistent knowledge base with circular context injection, the diagnostic outcome feedback loop, and two dozen supporting subsystems. Every claimed feature is implemented in working production code — zero aspirational claims.