The Functional Cognition Framework - Realised as A Par-c-Driven, Self-Auditing Agentic Architecture Sustaining Long-Horizon Coherence on a Single Commodity GPU

Chain-of-Thought prompting and its variants have made large language models appear to reason in steps, yet those steps are often independent rounds that cohere only on the surface; a confident model can present the shape of a synthesis while, in fact, confabulating. The Functional Cognition Framework (FCF), realised here as a running system internally named "Dave", is built to measure and audit multi-agent reasoning rather than assume it. Where the dominant approach scales capability spatially - through larger models and ever-longer context windows - the FCF scales it temporally. A persistent orchestrator and a small fixed core-swarm drive up to twenty-four parallel thought-lineages over a three-tier memory hierarchy, on a single commodity GPU, and hold coherence across long horizons by consolidating and pruning memory between short cycles rather than by enlarging the context window. For reasoning tasks, every fused step receives a Shapley contribution score and a geometric novelty verdict that separates genuine synthesis from confabulation. For build tasks, work is decomposed to the method level and assured by a fix loop that ships only the best passing candidate behind a held-out gate, so that nothing incorrect ships past that gate - zero false ships, subject to the gate's coverage; successful procedures are distilled into a calibration-gated ledger and reused - experience-based learning that replaced an earlier, slower genetic approach. On a pre-registered corpus of requirements for fifty-two software modules and 1,131 methods, the system shipped every operator-tested method and fifty of the fifty-two modules under a hidden held-out gate, with no false ships and roughly nine in ten methods correct on the first attempt, in about an hour and a half of build time and without a single manual edit to the generated code. A concurrency study on the same GPU shows aggregate throughput rising roughly elevenfold as parallel width grows, and the dependency structure of the work exposes a further 3.7x reduction in build time that parallel scheduling could realise. We offer the FCF as a reproducible, self-auditing route to long-horizon agentic work on modest hardware; its self-monitoring signals are reported strictly as functional, measurable engineering metrics for explainability and audit, not as claims about machine experience.