Five months silent. People asked if we rugged.
We didn't.
The November prototype worked but had a ceiling. Shipping it would have been fine for a month and embarrassing for a year. We stopped and started rebuilding from scratch.
Here's what changed ↓
Scenario Marketplace Structure
The scenario marketplace is open to any author. Publishing a scenario costs a flat $10 in $NFA, paid in token and burned on publication, a deliberate anti-spam floor that keeps the marketplace free of throwaway listings without gatekeeping who may contribute.
Once published, a scenario earns on every run it serves, priced by its measured forecasting skill on resolved markets.
This creates a structural property: supply of scenarios grows automatically with trader activity. Every active user is a potential contributor.
The best scenarios, judged transparently on accuracy, rise to the top of recommendations. Poor scenarios stay in the long tail, available but rarely surfaced. Quality emerges through competition rather than through central curation.
NFA's core engine is an agent swarm simulator.
Given a scenario and a set of initial conditions, the engine generates a population of AI agents representing the actors relevant to the forecasting question.
Each agent has a defined role, personality, incentives, and memory. Agents interact over simulated time, exchanging messages, updating beliefs, reacting to events, and producing observable behaviors.
The simulation runs for a configurable number of rounds, then outputs a probability distribution over possible outcomes.
This approach reflects a decade of research in multi-agent systems, grounded in the observation that complex real-world outcomes emerge from interactions between actors with different goals, information, and constraints.
Single-model forecasting collapses this complexity into a single point estimate. Swarm simulation preserves it and extracts probability distributions from the resulting dynamics.
The engine is general-purpose. It does not know about @Polymarket or @Kalshi. It does not know about specific geopolitical situations. It simulates whatever scenario it is given, according to whatever dynamics the scenario specifies. Specialization happens in the scenario layer.
Whitepaper: https://t.co/ZBlzvMnvMh
Traders in these markets know that quantitative tools built for statistical or microstructure markets don't apply. They need a different class of tool entirely.
That tool is NFA.
https://t.co/XidyD9oE3u
The third category is where NFA operates. Actor-driven markets constitute roughly a third of active prediction-market inventory and a higher share of volume during crisis periods (elections, wars, scandals, regulatory cycles).
NFA runs on credits. Credits buy simulations.
Every simulation a payment goes directly to the scenario author, weighted by their accuracy on resolved markets.
Bad forecasts earn nothing. Good ones compound.
https://t.co/XidyD9oE3u
Strait of Hormuz traffic back to normal by June 30.
38% chance, $11M in volume.
Look at that price chart. Every spike and drop is a news cycle. The market is reacting, not modeling.
Whether Hormuz reopens doesn't depend on shipping data. It depends on specific people making specific decisions under specific pressure.
No price chart tells you that.
https://t.co/ntGLKDdTXt
Agent layer is wired. Game Master mediates rounds, plausibility filter is in, convergence detection running.
Next: binding layer. Market URL goes in, research comes out, cast gets built.
First real simulation soon.
6/ Every resolved market adds an accuracy data point to the scenario.
A framework written today will run on markets that don't exist yet, building a track record that reflects the actual quality of the thinking behind it.
Domain expertise, priced by accuracy.
https://t.co/ntGLKDdm7V
Most forecasting tools model central bank decisions with generic archetypes.
Dove. Hawk. Centrist.
The problem: Jerome Powell is not a generic dove. Christopher Waller is not a generic hawk.
The actual people in the room have histories, track records, and pressure points that archetypes can't capture.
🧵
5/ The scenarios that perform aren't the most sophisticated ones.
They're the ones that correctly identify the few dynamics that actually drive the outcome and ignore everything else.
Same discipline good traders already have. Write what actually matters. Leave out the rest.