SCI: Surgical Cognitive Interpreter

A Metacognitive Control Layer for Signal Dynamics

This repository contains the reference implementation of SCI, a closed-loop metacognitive controller that wraps existing models and turns prediction into a regulated process rather than a one-shot function evaluation.

SCI is introduced in:

Vishal Joshua Meesala
SCI: A Metacognitive Control for Signal Dynamics
arXiv:2511.12240, 2025
https://arxiv.org/abs/2511.12240

The paper formalizes interpretability as a feedback-regulated state: SCI monitors a scalar interpretive signal ( SP(t) ), defined over reliability-weighted, multi-scale features, and adaptively adjusts an interpreter’s parameters to reduce interpretive error
[ \Delta SP(t) = SP^*(t) - SP(t), ]
under Lyapunov-style stability constraints.

---

1. Motivation

Most neural networks are deployed as open-loop function approximators: they map inputs to outputs in a single forward pass, with no explicit mechanism to regulate computation, explanation quality, or clarification depth.
In safety–critical domains (medicine, industrial monitoring, environmental sensing), this is brittle:

• Easy and ambiguous inputs receive the same computational budget.
• Explanations are static and post hoc, with no adaptation under drift.
• There is no explicit notion of “interpretive error” that can be monitored or controlled.

SCI addresses this by introducing a closed-loop metacognitive layer that:

• Monitors a scalar interpretive state ( SP(t) \in [0,1] ).
• Computes interpretive error ( \Delta SP = SP^* - SP ) relative to a target clarity level.
• Updates interpreter parameters ( \Theta ) according to a Lyapunov-inspired rule with safeguards.
• Allocates more inference steps and adaptation to ambiguous or unstable inputs.
• Exposes ( \Delta SP ) as a safety signal for abstention, escalation, or human-in-the-loop review.

Empirically, SCI:

• Allocates 3.6–3.8× more computation to misclassified inputs than to correct ones.
• Produces an effective scalar safety signal ( \Delta SP ) with AUROC ≈ 0.70–0.86 for error detection across vision, medical, and industrial benchmarks.

---

2. Conceptual Overview

SCI is a modular architecture with four main components.

2.1 Decomposition ( \Pi )

A multi-scale, multimodal feature bank ( P(t, s) ) that organizes raw signals ( X(t) ) into interpretable components:

• Rhythmic components (frequency bands, oscillations)
• Trend components (baselines, drifts)
• Spatial / structural components (sensor topology, modes)
• Cross-modal interactions (coherence, correlation, causal couplings)
• Latent composites ( \Pi^* )

Each feature is weighted by a reliability score ( w_f(t) ) derived from:

• Signal-to-noise ratio (SNR)
• Temporal persistence
• Cross-sensor coherence

These weights ensure degraded or untrustworthy features are down-weighted.

---

2.2 Interpreter ( \psi_\Theta )

A knowledge-guided interpreter that maps the reliability-weighted feature bank into:

• Markers ( m_k ): human-meaningful states or concepts
• Confidences ( p_k(t) ): calibrated probabilities
• Rationales ( r_k(t) ): sparse feature-level attributions and/or templated text

This component can be instantiated as a linear or shallow neural head on top of ( P(t, s) ), optionally constrained by domain rules or ontologies.

---

2.3 Surgical Precision (SP)

( SP(t) \in [0,1] ) aggregates calibrated components such as:

• Clarity / selectivity
• Pattern strength
• Domain consistency
• Predictive alignment

In the minimal implementation, ( SP ) is normalized entropy of a marker or predictive distribution:
high SP corresponds to focused, confident internal usage of markers;
low SP corresponds to diffuse or ambiguous internal state.

---

2.4 Closed-Loop Controller

The controller monitors ( \Delta SP(t) ) and updates ( \Theta ) when interpretive clarity is insufficient.

[ \Theta_{t+1} = \text{Proj}_{\mathcal{C}}\left[\Theta_t + \eta_t\left(\Delta SP(t)\nabla_\Theta SP(t) + \lambda_h u_h(t)\right)\right], ]

where:

• ( \eta_t ): step-size schedule
• ( \lambda_h ): human-gain budget
• ( u_h(t) ): bounded human feedback signal (optional)
• ( \text{Proj}_{\mathcal{C}} ): projection enforcing constraints (trust region, sparsity, parameter bounds)

Lyapunov-style analysis shows that, under suitable conditions on ( \eta_t ) and ( \lambda_h ), the “interpretive energy”

[ V(t) = \tfrac{1}{2}(\Delta SP(t))^2 ]

decreases monotonically up to bounded noise, so explanations become more stable and consistent over time.

This yields a reactive interpretability layer that not only explains but also stabilizes explanations under drift, feedback, and evolving conditions.

---

3. Repository Structure

The repository is organized as follows (file names may evolve slightly as the framework matures):

sci/                  # Core SCI library
  __init__.py
  config.py
  controller.py       # SCIController: closed-loop update over Θ using ΔSP
  decomposition.py    # Decomposition Π and reliability-weighted feature bank
  interpreter.py      # Interpreter / marker head and SP computation
  reliability.py      # Reliability scores (SNR, persistence, coherence)
  sp.py               # SP scalar and related metrics
  utils.py            # Shared utilities and helper functions

configs/              # Example configuration files
  mnist.yaml
  mitbih.yaml
  bearings.yaml

examples/             # Jupyter notebooks (to be populated)
  mnist_sci_demo.ipynb
  ecg_sci_demo.ipynb
  bearings_sci_demo.ipynb

experiments/          # Experiment scripts, logs, and analysis

scripts/              # Training utilities, Hub utilities, etc.
  push_to_hub.py

run_sci_mitbih_fixed_k.py
run_sci_bearings.py
run_sci_signal_v2.py  # Signal-domain SCI experiments

plot_metacognition_hero.py  # Plotting script for metacognitive behavior
sc_arxiv.pdf                # Paper PDF (for convenience)
sci_latex.tex               # LaTeX source of the paper

pyproject.toml
setup.cfg
LICENSE
README.md