Abstract
Field Computer-implemented systems for monitoring whether previously remediated employee knowledge deficits recur, using cumulative sum control charts applied to operational performance metrics rather than semantic embedding similarity. Background After remediation, organizations need to detect whether the same deficit returns. One approach stores vector embeddings and computes cosine similarity against incoming communications. The approach described here treats the employee's performance as a continuous signal and applies statistical process control methods borrowed from industrial quality management to detect shifts. Technical Description When a knowledge gap is marked as remediated, a snapshot service records the employee's relevant KPI baselines into a remediation_snapshot record containing: employee_id, knowledge_id, kpi_ids (a list of the KPIs mapped to the remediated domain), baseline_mean, baseline_stddev, baseline_control_limits (upper and lower), and remediation_date. These baselines are computed from the employee's KPI readings during the 30 days immediately following remediation. A recurrence watch manager creates a kpi_watch_window entry containing: watch_id, employee_id, knowledge_id, start_date, end_date (default: 90 days post-remediation), kpi_ids, and status (active, recurrence, absorbed). During the watch period, a monitoring process runs at configurable intervals (default: daily). For each KPI in the watch window, it retrieves recent readings from the kpi_readings table and computes a CUSUM statistic. The CUSUM accumulates the log-likelihood ratio of each observation coming from a deficit distribution (characterized by the pre-remediation KPI pattern) versus a proficient distribution (characterized by the post-remediation baseline). Formally: C_t = max(0, C_(t-1) + (x_t minus k)), where x_t is the standardized deviation and k is a configurable allowance parameter (default: 0.5 sigma). If the CUSUM statistic exceeds a decision threshold h (default: 4 sigma, derived from the Wald sequential probability ratio), the watch status transitions to recurrence. The system stores evidence including the CUSUM trajectory, the specific readings that pushed it over threshold, and timestamps. If the watch window expires without the CUSUM exceeding h, the status transitions to absorbed, and a non_recurrence_certificate record is created referencing the full KPI history and CUSUM trajectory. Distinguishing Characteristics This system does not use vector embeddings. No cosine similarity is computed. There is no gap fingerprint or remediation fingerprint stored as a dense vector. The detection mechanism is a CUSUM control chart operating on scalar KPI time-series data, computing cumulative log-likelihood ratios. There is no forgetting-curve decay function modulating any threshold. The Wald decision threshold is fixed. There is no monitoring lease object that self-terminates. The watch window is a simple time-bounded record in a database, not a computational entity that registers with event streams and deallocates resources.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Davis, Kenneth, "Statistical Process Control with CUSUM Charts for Post-Remediation Knowledge Deficit Recurrence Detection", Technical Disclosure Commons, (March 25, 2026)
https://www.tdcommons.org/dpubs_series/9633