DCA at DocVQA 2026

DCA at DocVQA 2026 is Planet AI's submission in the >35B-parameter category at the ICDAR 2026 DocVQA challenge — multimodal reasoning over documents across eight heterogeneous domains (business reports, scientific papers, slides, posters, maps, comics, infographics, and engineering drawings). No single Vision-Language Model excels across all of them, and confident hallucinations on numbers and table structures are a primary source of error. The system combines three components: (i) IDA, a layout-aware OCR engine producing structured Markdown that anchors text-heavy domains; (ii) multi-perspective page reading by independent frontier VLMs (Gemini 3.1 Pro, Gemini 2.5 Pro, Sonnet 4, Qwen3.5) guided by model-adapted question reformulations; and (iii) agentic reasoning (Claude Opus 4.6) that synthesizes the perspectives through domain-aware trust hierarchies and cross-perspective hallucination detection. It is built on Luna, Planet AI's cognitive AI platform implementing the Distributed Cognitive Architecture (DCA), which extends frozen foundation models with hierarchical memory, executive control, and convergent dynamics. The submission scored 60.00% — a +20-percentage-point improvement over the strongest frontier-model baseline (~40%), of which roughly +7 pp is attributable to IDA's deterministic text extraction and +13 pp to DCA's orchestration — placing first in the >35B-parameter category of the official, externally juried leaderboard. This technical report is the empirical anchor of the DCA paper family, companion to DCA — Foundations (biological motivation; DOI 10.5281/zenodo.20738104) and DCA — Theory I (formal convergence theory; DOI 10.5281/zenodo.20732538).