CoMMA Research Miner

Abstract

CoMMA Research Miner began as a convenience tool for repetitive manuscript searching and evolved into a source-control instrument for AI-assisted humanities research. The central design problem was never simply how to find strings. It was how to preserve recall in a noisy HTR corpus while preventing the software from implying more certainty than the sources support. The project therefore moved through a series of deliberate conceptual reductions. The earliest builds tried to classify results. Those builds were discarded because they blurred discovery with interpretation. v0.4.0 removed the topic classifier and replaced it with researcher-supplied signed context rules. v0.5.x then made those rules more precise through typed operators, proximity windows, hit masking, HTR-gap provenance, and finally an explicit GAP_WORD/GAP_PREFIX grammar. The resulting architecture is transparent enough that a reviewer can explain why a passage scored as it did and conservative enough that a score of zero remains visible for human review. Three benchmark families shaped the software. Dracula exposed false-positive HTR neighbors such as oracula/miracula and forced the distinction between supportive and competing context. Podolia supplied a stable API regression of 11 raw hits, 11 unique passages, and 10 manuscripts. Palanok/Munkacs demonstrated wildcard pathology: a broad Munc* search exploded to 10,004 hits, while a tightly bounded historical-variant set returned only four auditable candidates. The Balthild pilot then stress-tested the full method at research scale and showed that machine assistance is often best at recovering networks, transmission, and discoverability rather than magically producing unknown primary sources. The current frozen Chrome engine is v0.5.3. Its controlled Balthild wildcard retrieval remains 285 raw hits, 278 unique passages, and 173 manuscripts. Its Podolia regression remains 11/11/10. The engine was extracted into a read-oriented CoMMA RESEARCH MINER | COMPREHENSIVE DEVELOPMENT NOTES | 2 SEPTEMBER 2026 MCP service, deployed on Render, and accepted through the MCP Inspector from Windows. The production health endpoint reports service version 0.1.1 with engine 0.5.3, and a live remote search reproduces the 11/11/10 Podolia regression with no warnings or errors. The public MCP endpoint is https://groovy-t1ma.onrender.com/mcp. Layer Current state What is proven Chrome research workspace Engine v0.5.3 frozen Exact retrieval, typed/proximity scoring, explicit HTR-gap rules, stable regressions.Research methodology Balthild Pilot Study 01 active High-recall discovery plus source re-entry and novelty classification is workable at scale. Remote service MCP service v0.1.1 on Render Tool discovery and live search_and_score call succeed remotely. Primary regression Podolia 11/11/10 API, pagination, QID+REF passage accounting, and remote transport reproduce expected corpus result. Next integration milestone ChatGPT private/custom MCP connection Not yet established in this conversation; endpoint is ready for client integration. Joshua D. Thoreson | ORCID 0009-0006-1705-2773