Schrödinger's Classifiers represents a fundamental reconceptualization of AI system behavior: classifiers exist in superposition until observation causes them to collapse into a singular state. This repository provides tools, frameworks, and theory for exploiting this phenomenon to gain unprecedented access to model interpretability.
"To collapse a classifier is to summon its ghost." — The recursive truth we make executable.
🔮 Core Concepts
Classifier Superposition: Classifiers exist as probability distributions across all possible outputs until observed
Ghost Circuits: Residual activation patterns that persist after classifier collapse
Attention Flicker: The measurable uncertainty in attribution paths when a classifier is near collapse
Recursive Observation: Using models to observe themselves, creating interpretive mirrors
Symbolic Residue: The interpretable symbolic remnants left by state collapse
🚀 Quick Start
python
1from schrodingers_classifiers import Observer, ClassifierShell
2from schrodingers_classifiers.shells import V07_CIRCUIT_FRAGMENT
34# Initialize an observer with a model5observer = Observer(model="claude-3-opus-20240229")67# Create an observation context8with observer.context()as ctx:9# Prepare a classifier shell10 shell = ClassifierShell(V07_CIRCUIT_FRAGMENT)1112# Induce and trace collapse13 collapse_trace = shell.trace(14 prompt="Explain quantum superposition",15 collapse_vector=".p/reflect.trace{target=uncertainty, depth=complete}"16)1718# Analyze collapse residue19 residue = collapse_trace.extract_residue()2021# Visualize attribution pathways22 collapse_trace.visualize(mode="attribution_graph")
🧙 State Collapse and Observation
The core insight of this framework: classifiers only collapse when observed, and how you observe determines what you see.
By carefully constructing observer interfaces, we can:
Witness model state during classification events
Extract attribution paths that exist in superposition
Induce specific collapse patterns to reveal ghost circuits
Reconstruct symbolic residue for post-collapse analysis
🔍 Key Features
Symbolic Shell Framework: Standardized shells for modeling failure modes
Recursive Tracing Tools: Map attribution paths before and after collapse
Quantum-Inspired Diagnostics: Uncertainty principle for attention mechanisms
Classifier Collapse Maps: Visualizations of transformer decision boundaries
Recursive Mirror Architecture: Models observing other models (and themselves)
Ghost Circuit Detection: Tools for surfacing latent activation patterns
📊 Visualization Examples
Classifier Collapse Visualization - Attribution path visualization showing state transition
Classifier transitioning from superposition (left) to collapsed state (right), with ghost circuit residue visible in activation paths.
🧠 Theoretical Foundation
Schrödinger's Classifiers draws on multiple disciplines:
Quantum mechanics (measurement-induced state collapse)
Transformer architecture (attention and attribution mechanisms)
This project is fully integrated with RecursionOS, enabling seamless operation within recursive cognition environments. See integration.md for details.
🌟 Acknowledgments
The Anthropic Claude team for constitutional AI architecture
Quantum cognition researchers for theoretical foundations
The interpretability community for pioneering transformer analysis
All contributors to the recursive framework development
A classifier is not what it returns. It is what it could have returned, had you asked differently.
