Spaces:

ConvergentIntelligence
/

README

Configuration error

App Files Files Community

README / README.md

reaperdoesntknow

Update README.md

ce0e742 verified 20 days ago

preview code

raw

history blame contribute delete

7.34 kB

metadata

title: Convergent Intelligence LLC
emoji: 🔬
colorFrom: blue
colorTo: indigo
license: apache-2.0

Convergent Intelligence LLC

AI Governance · Algorithmic Bias Detection · Intelligence Analysis

Where classical analysis fails to see, we begin.

Who We Are

Convergent Intelligence is a research-driven consultancy specializing in AI governance, algorithmic bias detection, and applied intelligence analysis. We build mathematical frameworks that work where standard methods break — and we publish the models, papers, and code to prove it.

Founded September 11, 2025. DUNS: 144950019. SAM.gov registered (UEI: HC76F13L4KS8). Federal contract-ready.

Principal: Roy S. Colca Jr. — B.S. Pure Mathematics (CCNY), M.S. Applied Intelligence (Mercyhurst University, 3.89 GPA). Background spanning pure mathematics, intelligence analysis, penetration testing, and field operations.

Three Divisions

Research Division

The mathematical and empirical engine. We develop Discrepancy Calculus (DISC) — a measure-theoretic framework that treats singularities as primary structure rather than pathology — and deploy it across model architectures, training methodologies, and intelligence analysis pipelines.

Published Papers:

Discrepancy Calculus: Foundations and Core Theory — The eight axioms, the Mesh Fundamental Identity, the Meta-Discrepancy impossibility theorem. DOI: 10.57967/hf/8194
Structure Over Scale: CPU-Native Training of Sparse Cognitive Architectures at $1.60 Per Model — Seven methodological pillars, 15 models trained on CPU at FP32, total compute cost $24. DOI: 10.57967/hf/8165
From Three Teachers to Dual Cognition — Topology-aware multi-teacher distillation, role-conditioned self-critique at 1.7B scale. DOI: 10.57967/hf/8184

Companion Monograph: "On the Formal Analysis of Discrepancy Calculus" — 203 pages, 41 chapters, four parts (Analytical Foundations, Structures/Geometry/Time, Quantum Discrepancies, Theory of Other). The complete proof apparatus.

Consulting Division

Algorithmic risk assessment, bias auditing, and AI governance for regulated industries. We don't just detect bias — we mathematically characterize why standard detection methods fail, using the Meta-Discrepancy Theorem to identify regimes where classical statistical testing is provably insufficient.

Development Division

Infrastructure, tooling, and operational systems. JARVIS intelligence analysis platform, FRACTURE real-time threat assessment pipeline, OSINT fusion capabilities, and the cix-gateway edge computing stack on Cloudflare.

The Portfolio

50 models · 22,500+ downloads · 8 collections · 3 published papers

Architecture Families

Family	Models	Downloads	Key Innovation
DistilQwen	14	8,892	Topology-aware knowledge distillation via BV decomposition
MoA / DiscoverLM	7	3,171	Metric-native attention with triangle inequality enforcement
Qemma	5	2,204	Cross-architecture fusion via Gap Envelope Integral
SAGI / Swarm	3	1,347	Swarm intelligence with discrepancy mechanics routing
Symbiotic	3	1,304	Hybrid symbolic-transformer with persistent memory
DNA-AI	2	984	Depth-native architectures
DualMind	6	862	Dual cognition — explore, examine, respond on shared weights

Collections

DistilQwen — BF16 proof-weighted distillation from Qwen3-30B-A3B → 1.7B/0.6B. Three teacher variants, nine models.
DualMind — Single architecture, dual cognition. Five models including the Opus 4.6 reasoning variant.

Flagship Models

Model	Downloads	What It Proves
Qwen3-1.7B-Thinking-Distil	1,188	TKD preserves reasoning structure standard KD destroys
TopologicalQwen	1,134	Full BV decomposition in the distillation pipeline
DiStil-Qwen3-1.7B-uncensored	1,030	Alignment removal preserves capability
LFM2.5-1.2B-Distilled-SFT	1,024	Cross-architecture TKD (LFM → Qwen)
DiscoverLM-70M	784	Metric attention with proper geometry beats dot-product at 1/1000th the parameters

The Mathematics: Discrepancy Calculus (DISC)

Every model in this portfolio is built on Discrepancy Calculus — a measure-theoretic framework that quantifies the mismatch between integration and differentiation via the discrepancy operator:

$Df(x) = \lim_{\varepsilon \downarrow 0} \frac{1}{\varepsilon} \int_x^{x+\varepsilon} \frac{|f(t) - f(x)|}{|t - x|}\,dt$

For smooth $f$: $Df(x) = |f'(x)|$ (classical recovery). For rough $f$: $D$ localizes irregularity to null sets while preserving integral structure.

The Mesh Fundamental Identity — the DISC replacement for the Fundamental Theorem of Calculus:

$f(b) - f(a) = \underbrace{\int_a^b f'(x)\,dx}_{\text{smooth}} + \underbrace{\sum_{x \in J_f} \Delta f(x)}_{\text{jumps}} + \underbrace{D^c f(I)}_{\text{Cantor drift}}$

Standard methods see only the first term. DISC preserves all three.

The Meta-Discrepancy Theorem (Theorem 11.15) proves: when gap measure and discrepancy energy are both positive, the classical derivative/FTC/MVT package is impossible on positive measure. This is why standard knowledge distillation, standard bias detection, and standard statistical testing fail at structural boundaries — and why DISC-informed methods work where they don't.

Full theory: Discrepancy Calculus: Foundations and Core Theory (DOI: 10.57967/hf/8194)

Core Thesis

Structure beats scale. A 69M parameter model with proper geometry outperforms architectures 100x its size on structural reasoning tasks. A $24 training budget on CPU at FP32 produces models with organic community adoption. The transformer's dot-product attention is a hardware-constrained design choice, not a mathematical optimality — and we have 50 models proving the alternative works.

The research is the product is the marketing is the credibility. Every model card documents the mathematics. Every paper links to the models. Every download validates the methodology. The portfolio compounds.