Ali-Lazem/DBPM: Dynamic Bidirectional Pattern Memory

First public release of the DBPM gating module, accompanying the paper "Dynamic Bidirectional Pattern Memory: A Production-Scale Empirical Characterisation of Inference-Time Gating in Clinical NLP" (Lazem & Teahan, 2026).

What's included

• dbpm.py — the standalone DBPM gating module: the memory update operator (Algorithm 1), three-tier gating (BLOCK / DOWNGRADE / ALLOW), source-reliability weighting, real-time wall-clock decay, cross-task propagation, and the two verifier-independent predicates (the NER-coverage QA signal and the ontology-violation RE predicate).
• graders/ — scripts that reproduce the paper's reported lift, confidence interval, and per-category numbers from released aggregate outputs.
• cohorts/ — held-out and development cohort UID lists indexing the public PMC-Patients corpus.
• results/ — aggregate gating-outcome counts (counts and rates only; no clinical text).

Key results reproduced by this release

• The verifier-fed relation channel persists 0 patterns despite 785,797 logged rejections at production scale (167,034 patients).
• Of five question-answering gate signals, the NER-coverage signal is selective: path lift 1.84 (95% CI [1.75, 1.93], p < 10⁻¹⁵), within a 1.52–1.84 band across four replications.

Scope

The gating module only. The surrounding extraction pipeline (orchestration, model serving, the extraction loop, and the intelligence layer) is a separate system and is not included. The cohort lists and aggregate counts permit reproduction of the reported gating numbers without it.

Citation

See CITATION.cff. Archived on Zenodo: [DOI to be added on first archive].