Inspiration

While experimenting with Large Language Models for complex policy and infrastructure questions, I noticed a recurring failure mode: the model always gives an answer, even when the evidence is incomplete, contradictory, or deeply uncertain.

In high-stakes domains, a confident but fragile answer is often worse than no answer at all. Most AI systems collapse uncertainty into a single plausible narrative, creating the illusion of certainty and enabling decision laundering.

PRDE was inspired by a simple question:
What if an AI system was allowed to refuse to decide unless the decision survived structured, adversarial disagreement across incompatible futures?


What it does

PRDE evaluates high-stakes decisions by simulating multiple mutually incompatible future worlds and forcing adversarial reasoning between independent agents.

Instead of producing a single answer, the system:

  • Generates five parallel world models
  • Assigns five reality-bound agents with conflicting perspectives
  • Runs structured adversarial debate using Quantitative Bipolar Argumentation Frameworks (QBAF)
  • Monitors statistical convergence using Wald’s Sequential Probability Ratio Test (SPRT)
  • Outputs one of three outcomes:
    • CONVERGED_ROBUST
    • CONVERGED_FRAGILE
    • FAILED_UNCERTAIN (epistemically correct refusal)

If evidence does not converge mathematically, the system refuses to decide.


How we built it

PRDE is implemented as a strict multi-stage pipeline:

  1. Decision Compilation — validates and structures the user query
  2. World Generation — produces five incompatible future scenarios
  3. Agent Binding — permanently binds one agent to each world
  4. Adversarial Debate — agents generate SUPPORT/ATTACK arguments in a QBAF graph
  5. Convergence Detection — SPRT determines whether evidence supports robustness, fragility, or uncertainty

All reasoning outputs are:

  • Schema-validated JSON
  • Logged with tamper-evident audit trails
  • Guarded by preflight validation and rate-limited Gemini API usage

A mock-first architecture ensures zero API calls during development and testing.


Challenges we ran into

The hardest problems were not model quality, but system safety and correctness:

  • Preventing API quota waste during iterative debugging
  • Avoiding silent failures caused by schema mismatches
  • Ensuring probabilistic outputs did not force convergence
  • Designing refusal paths that cannot be overridden
  • Managing rate limits and retries without hiding failures

These issues required architectural solutions rather than prompt-level fixes.


Accomplishments that we're proud of

  • A decision engine that refuses to answer when uncertainty is irreducible
  • Deterministic refusal enforced outside the LLM
  • Full mock-first, quota-safe Gemini integration
  • Mathematical convergence guarantees using SPRT
  • Tamper-evident audit logs for every decision
  • Over 1000 automated tests validating safety and correctness

Most importantly, PRDE treats uncertainty as a first-class output, not a failure.


What we learned

This project fundamentally reshaped our understanding of AI decision systems:

  • Refusal is a feature, not a weakness
  • Averaging opinions is unsafe in fat-tailed risk domains
  • Human-in-the-loop alone does not prevent automation bias
  • Safety must be enforced at the architecture level, not through instructions
  • Statistical decision theory integrates naturally with LLM pipelines

Building PRDE required thinking less like a chatbot designer and more like a safety engineer.


What's next for PRDE — Parallel Reality Decision Engine

Planned next steps include:

  • Adaptive world generation for tail-risk discovery
  • Formal verification of refusal thresholds
  • Improved visualization of evidence corridors
  • Regulatory-friendly audit exports
  • Applying PRDE to real-world policy, infrastructure, and safety datasets

The long-term goal is to make epistemic humility a standard capability of AI systems used in high-stakes decision-making.

Built With

Share this project:

Updates