Bone science answers stop relying on model memory: give the right passage and one model goes from 42% to 78%

Bone-science question answering is split into five tabs over one shared substrate: search, cited chat, self-correcting reasoning, image identification and image mechanics. General models answer from parametric memory, cannot point to specific peer-reviewed evidence, and have no durable way to accept an expert's correction. On a 30-question benchmark spanning seven bone-science domains, retrieval reached a mean reciprocal rank of 0.928 at k = 10, and every question not returned at rank 1 had a relevant passage at rank 2 or 3.

Reasoning runs as draft, deterministic physics check and evidence-grounded critic, and a user's correction is compiled into an editable rule applied on every later request. Letting a model judge its own output on confidence tends to reward fluent, confident text over correct text. Eight built-in rules encode established relationships such as cortical modulus, trabecular BV/TV and the WHO T-score, and the critic receives only one-hop knowledge-graph edges plus retrieved literature.

The imaging tab trains a small head on frozen BiomedCLIP features to hint a vision-language model about bone region, and withholds that hint when an input falls outside the training scope. Strong medical vision-language models typically depend on large curated instruction sets and substantial GPU training. On the patient-disjoint MURA held-out split, region classification reached 92.6% accuracy against a seven-class chance rate of 14%, and the out-of-distribution guard placed a spine micro-CT below the 0.827 nearest-neighbour threshold.

The demonstrated scope is bone-science literature retrieval, cited question answering, self-correcting reasoning with user rules, upper-limb radiograph region identification, and displacement and axial strain prediction from a single undeformed micro-CT slice. The system runs as a public beta at bonegraph.org, with all inference local on a single NVIDIA Jetson AGX Orin and no third-party API calls. The authors position it as a research and educational tool rather than a clinical diagnostic device.

The full-text corpus covers only open-access PDFs reachable through OpenAlex and the resolution pipeline, so paywalled articles have metadata but no body text in the index. The knowledge graph is machine-extracted, not verified edge by edge and fragmented, so it serves only as one-hop evidence. The retrieval benchmark is 30 questions run once, and prompt wording moved every condition by 6 to 12 percentage points on the 50-question reasoning benchmark. The vision classifier is trained only on upper-limb radiographs and its abnormality head recalls 0.61, which the authors call preliminary. The mechanics model was trained on a particular vertebral micro-CT dataset and generalises best to comparable slices.