Cognitive Harness Engineering, Sub-Parametric Overthinking, and Human-AI Symbiosis: Non-Compensatory Conditioned Parity in High-Stakes Compound Agentic Systems
Year: 2026
Type: preprint
Abstract: Technical Report & Preprint (VTI-RES-2026-01) — Bilingual Edition (English & Portuguese) + Executive Briefs + LaTeX Sources. Deploying agentic systems across high-stakes institutional domains confronts a fundamental trade-off among frontier inference cost, second-order statutory omissions in compact workhorse models, and live operational governance of Human-in-the-Loop symbiosis. Through a controlled repeated-measures experiment (5 × 20 = 100 end-to-end executions), we empirically demonstrate that scaling unconstrained internal reasoning effort in compact models fails to remediate missing statutory mindware and induces Sub-Parametric Overthinking. Grounded in Keith E. Stanovich's Tripartite Theory of Mind and comparative principles of cognitive physiology, we formalize the Cognitive Harness Parity Engine (HPE) across four decoupled planes and eight mathematical models (M1–M8), validated across 13 reproducible E2E test batteries (B1–B13).
Author: Gonçalves, Caio
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7234271984
