Research Lab

Where the math meets the method

Original research exploring the statistical and mathematical foundations of collaborative intelligence. Every claim is grounded in provable equations. Every model is falsifiable. This is where LIMN gets its rigor.

Published Research — February 2026
New Research

Anomalous Data or Phase Transition?

A Statistical Inquiry into Discontinuous Gains Beyond Diminishing Returns

Authors: Joshua Sebastian & Claude (Anthropic)
Date: February 2026
Pages: 4
Status: Discussion Draft

Standard models predict sigmoid growth with asymptotic flattening. But what happens when anomalous data clusters near the ceiling? This paper presents evidence — from AI benchmark performance data (2012–2025) and catastrophe theory — that these anomalies are not noise but signals of phase transitions: points where the governing dynamics of a system fundamentally change. We formalize nine provable equations mapping the logistic model, cusp catastrophe bifurcation conditions, power law distributions, dimensional carrying capacity, critical slowing down, and mutual information to testable predictions about where collaborative human-AI systems may access qualitatively different output regimes.

Key Equations (9 total, all provable & solvable)

Eq. 1 — Logistic Growth Model
Eq. 2 — Cusp Catastrophe Potential
Eq. 3 — Equilibrium Condition (Cubic)
Eq. 4 — Discriminant of the Cusp
Eq. 5 — Bifurcation Set Boundary
Eq. 6 — Power Law Distribution
Eq. 7 — Dimensional Carrying Capacity
Eq. 8 — Critical Slowing Down
Eq. 9 — Mutual Information (Sweet Spot)
Key Findings
01

The Dip Precedes the Breakout

Systems approaching a phase transition perform worst immediately before the jump. Premature abandonment during this dip is the most common error in resource allocation.

02

Benchmark ≠ Capability

AI benchmark saturation (ImageNet, MMLU) reflects instrument limits, not system limits. Real capability growth contains discontinuities invisible to saturated benchmarks.

03

The Sweet Spot Is Dimensional

The collaborative sweet spot isn't a point on the existing curve — it's evidence of a higher-dimensional output space accessible only through human-AI coupling at intermediate intensity.

04

Nine Equations, All Testable

Every claim is backed by provable mathematics from catastrophe theory, information theory, and statistical distribution theory. All predictions are empirically falsifiable.

On Methodology

This research was produced through human-AI collaborative analysis — the same methodology it investigates. The initial pattern recognition (anomalous data near asymptotes resembling phase transitions) originated from human intuition grounded in statistical reasoning coursework. The mathematical formalization, benchmark data compilation, and visualization were developed iteratively through reciprocal exchange between human and AI. Neither agent could have produced this document alone. That is the point.