Case file No. 221B

Consulting engineer · 221B Labs

221B
02 Research

What should an agent do when the reward signal is a trap? My work sits where reinforcement learning, quantum-inspired search and the theory of memory meet.

Deceptive reward landscape
  • Decoy peak
  • Agent
  • True goal behind barrier
Featured study2026 – Present

Tunnelling Through Deception

A quantum-inspired exploration operator for agentic AI

Reinforcement-learning agents get lured into deceptive reward traps long before they find the true goal. The Tunnelling Exploration Operator (TEO) lets a DQN agent ‘tunnel’ through the barrier instead of climbing over it — in a classical exponential-decay form and as a simulated 4-qubit variational circuit in PennyLane.

  • 15×15 deep-deception grid with a 3-cell barrier ring
  • 40-seed pre-registered study, no interim analysis
  • Reproducible, tested Python codebase with a live comparative demo
  • Manuscript in preparation for IEEE submission
Normalised return · 40 seedsp < 0.0001
Baseline
0.494
TEO · classical
0.875
TEO · quantum
0.877
Normalised return across 40 seeds, p < 0.0001
MethodScore
Baseline0.494
TEO · classical0.875
TEO · quantum0.877

Undergraduate Researcher · Dept. of AI & ML, Malla Reddy University · Guide: Manasa Chandupatla

Code & comparative demo on GitHub ↗
Theory · 2026 preprintwith S. M. R. Prakash & S. M. Asad

Predictive state is not explanatory state

A system tuned to forecast can quietly lose the ability to explain itself. For threshold processes over binary event streams, prediction needs only log₂(t+1) bits — but answering every counterfactual about the past needs the full t. An exponential separation, verified with two independent implementations and 572 adversarial trials up to t = 4096.

Paper & verification code ↗
Forecasting · log₂(t+1) Counterfactual audit · t
016324864history length t →64 bits≈6 bits