CCI-MRC-C

Computable Research Pipeline for Closed Causal Information in Neural Dynamical Systems

Closed Causal Information (CCI) is a diagnostic metric for internal temporal dependence in sequential systems.

$$ I(H_t; H_{t+1}\mid E_t) $$

It measures how much future internal state depends on current internal state after conditioning on external input.

CCI measures dependence, not causation or consciousness.

Overview

We consider a system with internal state H_t, external input E_t, and stochastic noise epsilon_t:

$$ H_{t+1}=F(H_t,E_t,\epsilon_t) $$

The core objective is to isolate the contribution of internal dynamics from input-driven effects.

$$ E(X)=\sum_{t=1}^{T} I(H_t;H_{t+1}\mid E_t) $$

Conceptual Diagram (Vision Layer)

Conceptual interpretation only. Metric claims are defined in CLAIMS.md.

The concise vision synthesis from the external blueprint is documented in theory/master_blueprint.md and explicitly tagged as Interpretive Layer (Vision).

Visual Pipeline Panels

Panel breakdown:

Visuals are explanatory overlays for communication and do not replace metric evidence in results/ and claim boundaries in CLAIMS.md.

Interactive CCI Calculator (Experimental)

The repository includes an experimental CCI calculator workflow (Gaussian CMI):

• Input tensors: H_t, H_t1, E_t
• Covariance estimation with numerical stabilization
• Cholesky-based log-det computation
• Adaptive jitter policy for near-singular covariance
• Outputs: CCI (bits), rolling CCI trend, optional delta-CCI, used jitter diagnostics

Expected input shapes

• H_t: [N, d_h]
• H_t1: [N, d_h]
• E_t: [N, d_e]

Gaussian Estimation

$$ H(X)=\frac{1}{2}\log\left((2\pi e)^d|\Sigma_X|\right) $$

$$ I(X;Y\mid Z)=\frac{1}{2}\log\frac{|\Sigma_{XZ}||\Sigma_{YZ}|}{|\Sigma_Z||\Sigma_{XYZ}|} $$

The implementation uses covariance factorization with Cholesky stabilization.

Benchmark Reporting Policy

• Primary benchmark comparison uses bias_corrected CCI.
• raw CCI is retained for diagnostics but can be inflated in high-dimensional finite-sample settings.
• Transformer non-zero CCI is conditional dependence evidence only and not proof of autonomy.
• Methods note: docs/methods.md.
• Claims contract: docs/claims_contract.md.

Why Frontier Labs May Care

• Hidden-state diagnostics beyond outputs.
• Regime-shift monitoring (rolling CCI / delta-CCI).
• Architecture-level comparison (bias-corrected CCI).
• Safety-relevant exploratory signal (latent loop persistence).

Reproducibility Card

• Primary metric: bias_corrected = max(raw_cmi - permutation_null_mean, 0).
• Raw retained for diagnostics.
• Seeds + per-run in results/cci_values.json.
• Cholesky + adaptive jitter (1e-7..1e-4).
• pytest sanity + shape checks.
• Interpretation boundaries and non-claims are governed by CLAIMS.md.

What CCI Captures

• Internal temporal dependence
• Hidden-state persistence
• Regime shifts in latent dynamics (via CCI trend / delta-CCI)

Open Collaboration Requests

• Non-Gaussian CMI estimators.
• Cross-model benchmark suite.
• Online regime monitoring.
• External validation datasets.

Open an issue with a reproducible benchmark proposal.

See docs/applications.md for evidence-tiered application scope. Documentation updates should preserve boundaries defined in CLAIMS.md.

How to contribute benchmarks

• Use issue templates in .github/ISSUE_TEMPLATE/:
  • benchmark_proposal.yml
  • repro_mismatch.yml
  • bug_report.yml
• Use .github/pull_request_template.md for PR checklist and reproducibility notes.
• Compare against results/reference_v0_1_1.json when reporting mismatches.

Non-Claims (Important)

• No causal proof.
• No consciousness detection claim.
• 1.0 bit is operational in this setup only, not universal.
• See docs/claims_contract.md and CLAIMS.md for required qualifiers.

Applications

Current evidence-backed use cases

• AI model diagnostics: reactive vs recurrent separation
• Training regime shift monitoring
• Hidden-state persistence tracking
• Architecture comparison (FF/RNN/Transformer)
• Anomaly/regime change signal in agent loops (exploratory)

Hypotheses / future validation

• Medical, legal, and policy contexts are future validation hypotheses and not current evidence-backed claims in this repository.
• Consciousness-related interpretation remains speculative and cannot be inferred from CCI alone.
• See docs/applications.md for full evidence tiers and CLAIMS.md for scope.

Benchmarks in This Repository

• Feedforward vs recurrent comparison
• Training dynamics tracking
• Input invariance test (noise vs structured)
• Transformer baseline evaluation

Comparative benchmark reporting uses bias-corrected CCI as primary value.

EEG Experimental Extension

This repository now includes an EEG-oriented experimental pipeline as an extension layer on top of the current CCI benchmark core.

New Modules

• app.py
• core/models.py
• core/engine.py
• data/sample_eeg.py
• research/cci_validation.py

How to Run EEG Demo

pip install -r requirements.txt
streamlit run app.py

EEG Validation Script

python research/cci_validation.py

This writes an exploratory JSON summary to results/eeg_validation.json.

EEG Test (Live)

streamlit run app.py
python research/cci_validation.py

Output:

• results/eeg_validation.json

Scientific Scope Note

This EEG extension is exploratory and provides a diagnostic dependence signal. It does not constitute proof of consciousness or causal agency.

Note on thresholds

1.0 bit is an operational threshold in this setup. It is not a universal constant and must be interpreted with assumptions from CLAIMS.md.

External Research Note

• A concise application-oriented technical note is available at paper/Paper.pdf.
• It presents CCI application domains in a diagnostic, input-conditioned, non-causal framing.
• It is intended as a citation/readability anchor for reviewers.

Quick Start

pip install -r requirements.txt
python run_all.py