NCLUDE: Edge-Native Neuro-Symbolic AI for Inclusive Education in Low-Resource Settings

INCLUDE is an offline-first, neuro-symbolic architecture for visually impaired (VI) learners in low-resource, post-conflict settings, optimized for ultra-constrained hardware (<2 GB RAM). Grounded in an IRB-approved 2025 empirical baseline (N=9,385 census; N=250 survey, Cronbach's α=0.87; N=60 longitudinal pilot), the artifact contains:  Source code for the four-node architecture: (1) Edge AI Layer with 4-bit QLoRA INT8 quantization (4.5 GB → 720 MB) behind a deterministic finite-state machine; (2) Asynchronous CRDT Sync Engine (24 KB local delta; 102.4 KB curriculum micro-modules; 15 MB/day rural budget); (3) Cloud Backend with Local Differential Privacy (Gaussian mechanism, ε=0.1, δ=10⁻⁵) and k-anonymity (k≥5); (4) Decolonial Knowledge Vault (9 historical pillars, 247 proverbs, 1,482 semantic relations). 166 deterministic test oracles achieving 100% line coverage over 598 source statements (Peng's Level 4 reproducibility).  Machine-readable manifests (data/reproducibility_manifest.json, data/prototypes_manifest.json) locking all empirical baselines, privacy parameters, and hardware constraints. Sole-author manuscript (PDF/LaTeX) and graphical abstract. Verified claims: deterministic architectural simulation and test oracles validate a 128 ms inference-latency bound, 99.8% cross-modal packet integrity within the 150 ms ISO 9241-11 window, and a 287 ms simulated end-to-end pipeline; expert-validated knowledge-graph curation (Fleiss' κ=0.94) reduces cultural misalignments by 67.8% (p<0.001); the pilot shows a +18.5% absolute learning gain (r=0.94) with a novel 25-minute cognitive-threshold effect in Afro-Asiatic decoding.Integrity note: all latency figures are deterministic simulation bounds enforced by test oracles, not physical on-device measurements. Raw census microdata remain restricted by the Tigray Education Bureau; only anonymized aggregates are deposited. IRB Ref. TEB/2025/IRB-047.Licenses: code MIT; documentation CC-BY 4.0. Funding: DAAD (Ref. 91627144).  G1 (Graphical Abstract). INCLUDE replaces cloud-dependent foundation models with a four-node offline-first fog architecture. Deterministic test oracles enforce a 128 ms simulated inference bound, 99.8% cross-modal integrity (<150 ms ISO 9241-11), and a 287 ms simulated pipeline; expert-validated curation reduces cultural misalignments by 67.8%; the pilot (N=60) shows +18.5% gain with a 25-minute cognitive threshold. All latency values denote simulation bounds (Peng Level 4, 166 oracles, 100% coverage)