Saud Rifat

Research

ORCID iD icon ORCID Profile · 0009-0001-0822-5293 ↗

Research Papers and Frameworks

Cognitive Singularity Theory

A safety-first measurement framework for studying recursive cognitive transitions and state-aware alignment.

Cognitive Singularity Theory (CST) proposes a measurement framework for detecting when an AI system shifts from reactive tool behavior into recursive self-modification. It introduces the Cognitive Autonomy Index (CAI) as an exponentially weighted moving average (EWMA) composite that tracks calibration accuracy, belief revision magnitude, and Higher-Order Thought activation over time.

To reduce false positives, CST includes a fail-closed Hard Gate with anti-sandbagging that requires measurable calibration improvement before attributing autonomy. The transition threshold $T_{\mathrm{CST}}$ is defined via non-circular change-point detection on an independent regime indicator $R(t)$. Safety is framed as state-aware alignment with a Trust Model requiring externalized monitoring.

Key contributions
  • Defines the Cognitive Autonomy Index ($\mathrm{CAI}$) as an EWMA composite with measurable proxy components
  • Introduces a fail-closed Hard Gate with anti-sandbagging to reduce false positives
  • Defines $T_{\mathrm{CST}}$ via non-circular change-point detection on independent regime indicator $R(t)$
  • Frames safety as state-aware alignment with a Trust Model requiring externalized monitoring
  • Includes RQAA proxy module for recurrence-guided attention as a candidate workspace proxy
Core equation
$$ \mathrm{CAI}(t) = (1 - \lambda)\,\mathrm{CAI}(t\!-\!1) + \lambda \left( k_{\mathrm{CAI}} \cdot e^{-|P_{\mathrm{pred}}(S_{t+1}) - P_{\mathrm{act}}(S_{t+1})|} \cdot \frac{D_{\mathrm{KL}}(P_t \| P_{t+1})}{\max D_{\mathrm{KL}}} \cdot P_{\mathrm{eff}}(\mathrm{HOT} \mid S_t) \right) $$
$$ \mathrm{CAI}(t) \geq T_{\mathrm{CST}} \quad\text{where}\quad T_{\mathrm{CST}} = \text{median CAI at } R(t) \text{ change-point} $$

The Cognitive Autonomy Index

An operational framework for detecting recursive cognitive transitions in AI systems.

This paper introduces the Cognitive Autonomy Index (CAI), a composite operational metric for studying recursive cognitive transitions (RCTs) in AI systems. CAI combines belief revision magnitude (KL divergence), workspace activation coherence, and CalibrationGain to evaluate whether a system is entering a higher-risk recursive update regime.

The Hard Gate is a fail-closed review condition: if calibration improvement is not verified, or if anti-sandbagging checks fail, the system receives no metacognitive credit and recursive escalation is routed into review. Three controlled simulations illustrate the framework's expected behavior and failure modes under toy conditions.

Key contributions
  • Defines the CAI as a composite metric integrating KL divergence, workspace activation, and CalibrationGain
  • Specifies a fail-closed Hard Gate requiring verified calibration improvement before metacognitive credit is granted
  • Includes three proof-of-concept simulations illustrating expected behavior and failure modes
  • Introduces the RQAA candidate proxy module with preliminary divergence signals near putative regime boundaries
Hard Gate - fail-closed review condition
$$ P_{\mathrm{eff}}(\mathrm{HOT} \mid S_t) = P(\mathrm{HOT} \mid S_t) \cdot \mathbf{1}[\mathrm{CalibrationGain} > 0] \cdot \mathbf{1}[\mathrm{Gap}_{\mathrm{before}} \le \mathrm{Gap}_{\max}] $$

Recasting Lewin's Field Theory

A practical model for tracking behavior shifts as emotional filtering and meta-awareness change over time.

This work extends Kurt Lewin's classic formula $B = f(P, E)$ into a dynamic, time-indexed model that accounts for real moment-to-moment changes in behavior. The updated framework introduces two explicit state variables: $\mathrm{EFilter}_t$ (emotional filtering) and $\mathrm{HOT}_t$ (meta-awareness), forming $B_t = f(P_t, E_t, \mathrm{EFilter}_t, \mathrm{HOT}_t)$.

Version 3.2 incorporates construct validity criteria from a multi-auditor review, including a framework failure statement, falsifiable hypotheses tied to drift-diffusion models, and explicit psychometric validation criteria with candidate EEG markers.

Key contributions
  • Extends Lewin's $B = f(P, E)$ into time-indexed behavior dynamics
  • Introduces $\mathrm{EFilter}_t$ and $\mathrm{HOT}_t$ as explicit state variables
  • Defines operational proxies and within-subject falsifiable hypotheses
  • Includes construct validity shield from five-auditor red-team review
  • Proposes four falsifiable within-subject hypotheses (H1-H4)
Core equation
$$ B_t = f(P_t, E_t, \mathrm{EFilter}_t, \mathrm{HOT}_t) $$