b3e8a9a058
Major additions: - Community Next.js app (port 18187) for browsing claims with API docs - stemedb-chaos crate: Fault injection, chaos testing, CRDT properties - Latent ingestion system: Reddit/FDA ingesters with ADK-Go agents - Disputed claims handling: Manual review workflows and validation - Aphoria security scanner: New extractors (SQL injection, command injection, weak crypto, TLS version), policy-based ignores, UAT reports - Docker infrastructure: Dockerfile, docker-compose.yml for full stack - VulnBank demo: Intentionally vulnerable multi-language test corpus SDK & API enhancements: - Source registry handlers for tracking data provenance - Metrics endpoint - Skeptic filtering improvements Code quality: - Split 14 large files (>500 lines) into focused modules - All files now under 500-line limit per project guidelines Documentation: - Chaos testing guide, circuit breakers, observability docs - Phase 7 UAT documentation updates - Martin Kleppmann technical writer agent Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
104 lines
5.1 KiB
Markdown
104 lines
5.1 KiB
Markdown
# Latent
|
|
|
|
**The safety signal is already there. You just aren't listening.**
|
|
|
|
## The Problem
|
|
|
|
Clinical trials are sterile. The real world is messy.
|
|
|
|
In Phase III, 3,000 carefully selected patients take a drug. They are monitored, compliant, and healthy.
|
|
In the real world, 3,000,000 people take it. They mix it with alcohol, herbal supplements, and bad diets. They forget doses. They are biologically messy.
|
|
|
|
**Signals emerge in the real world years before they hit the FDA label.**
|
|
|
|
- A Reddit cluster discussing "strange vivid dreams" 18 months before the label update.
|
|
- A contradiction between the EU label (warning added) and the US label (silent).
|
|
- A drop in refill rates in a specific demographic that implies a hidden side effect.
|
|
|
|
Pharma safety teams (Pharmacovigilance) are drowning in PDFs and manual reviews. They operate on "Consensus." They miss the **Latent** signal until it becomes a class-action lawsuit.
|
|
|
|
## The Solution
|
|
|
|
**Latent** is an active surveillance engine for epistemic risk in pharmacology.
|
|
|
|
It ingests the noisy, conflicting, messy stream of global health data:
|
|
|
|
* **Tier 0:** FDA/EMA/PMDA Regulatory Actions
|
|
* **Tier 1:** ClinicalTrials.gov results, Phase IV commitments
|
|
* **Tier 2:** PubMed/Lancet publications, Case Reports
|
|
* **Tier 4:** De-identified EHR aggregates
|
|
* **Tier 5:** Patient forums (Reddit, PatientsLikeMe), Social sentiment
|
|
|
|
It doesn't just "search" for side effects. It graphs the **Divergence**.
|
|
|
|
## How It Feels
|
|
|
|
You are a Safety Medical Director tracking **Semaglutide**.
|
|
|
|
You don't see a static report. You see a **Conflict Heatmap**.
|
|
|
|
```text
|
|
┌────────────────────────────────────────────────────────┐
|
|
│ MOLECULE: Semaglutide (Ozempic/Wegovy) │
|
|
│ STATUS: DIVERGENT ⚠️ (Score: 0.82) │
|
|
│ │
|
|
│ [ Official Safety Profile (Tier 0) ] │
|
|
│ ✅ GI Issues: Nausea, Vomiting (Common) │
|
|
│ ✅ Thyroid C-Cell Tumors: Warning (Rare) │
|
|
│ │
|
|
│ [ Latent Signal Detected ] │
|
|
│ 🔥 Gastroparesis (Stomach Paralysis) │
|
|
│ ├─ Signal Strength: High │
|
|
│ ├─ First Detected: Q2 2023 (Reddit Cluster) │
|
|
│ ├─ Corroborated: Q4 2023 (3 Case Studies) │
|
|
│ └─ Conflict: Directly contradicts "Transient"│
|
|
│ label claim. │
|
|
│ │
|
|
│ [ Action ] │
|
|
│ The FDA label is lagging the Latent Signal by ~8 mo. │
|
|
│ Recommend: Phase IV observational study focus. │
|
|
└────────────────────────────────────────────────────────┘
|
|
```
|
|
|
|
## Core Principles
|
|
|
|
### 1. Disagreement is Data
|
|
Most safety software treats "anecdotal" reports as noise to be filtered. Latent treats them as **Tier 5 Signals**. A single tweet is noise. A cluster of 500 tweets describing the same specific symptom is a **Signal** that outweighs a 2-year-old Clinical Trial.
|
|
|
|
### 2. Time-Travel Liability
|
|
"What did we know, and when did we know it?"
|
|
Latent allows **Time-Travel Queries**. You can scroll back to June 14, 2023, and see exactly what the epistemic state was.
|
|
* *Did the public know?* Yes.
|
|
* *Did the literature know?* No.
|
|
This protects against negligence ("We acted as soon as the Tier 2 signal emerged") or exposes it.
|
|
|
|
### 3. The "Skeptic" Lens
|
|
Latent doesn't optimize for "Truth." It optimizes for **Risk**.
|
|
It assumes the official label is always decaying. It actively hunts for the "Black Swan" event that breaks the current safety model.
|
|
|
|
## Use Cases
|
|
|
|
### For Pharma (Pharmacovigilance)
|
|
* **Early Warning:** Detect adverse events in the wild before the FDA mandates a warning letter.
|
|
* **Label Defense:** Proactively update labels to limit liability.
|
|
* **Competitive Intel:** "Our competitor's drug is showing a latent cardiac signal in Europe. Ours isn't."
|
|
|
|
### For Hedge Funds / BioTech Investors
|
|
* **The "Short" Signal:** Latent detects a safety signal for a blockbuster drug 6 months before the market. The fund shorts the stock before the recall.
|
|
* **Trial Prediction:** Analysis of early patient chatter predicts trial failure/success before publication.
|
|
|
|
## Technical Foundation
|
|
|
|
Latent is the "Pharma-Tuned" implementation of **Episteme**:
|
|
* **Source Class:** Heavily utilized. Regulatory (0) vs Clinical (1) vs Social (5) is the core logic.
|
|
* **Decay:** Aggressive decay on Social data (noise fades fast), Zero decay on Regulatory (law is law).
|
|
* **Materialized Views:** Pre-computed risk scores for every tracked molecule.
|
|
|
|
## The Name
|
|
|
|
**Latent**.
|
|
Because the side effects haven't happened yet. They are waiting.
|
|
|
|
---
|
|
*Safety is not the absence of evidence. It is the evidence of absence.*
|