Each string is one rule-based signal raised on one employee. Detection and Triage are fully automated and always shown; Evaluation is where an analyst decides false positive vs. true positive; Closure only unlocks after that decision, and its content depends on which way the case was decided.
Confirmed (can reach a definite True in this export): Off-Shift Access & Shift Pattern Drift,
Off-Day Badge Use with Shift Drift, Excessive Hours Request & Departmental Trend, Coordinated Cage Access (Group).
Candidate (the export can only ever partially confirm these): Unexplained Shift Switch & Coverage Check,
Falsified Tour / Drill Records.
Every string's Detection tab ends with a metric row and a set of drill-down buttons. Four are available on every rule: the monthly signal timeline (how this rule and every other rule evaluated each month, with the anomaly score alongside), the anomaly stream detail (the unsupervised stream stated separately from the rule hit, since the two are independent), rule condition coverage (which of the rule's catalog conditions this export can actually answer, and which have no data source — this is what decides confirmed vs candidate), and raw shift data where the employee has overlapping shift records that would undermine any hours or badge figure. The group cluster adds three more covering its members, the maintenance-window join and the mantrap reads.
Every string here is one rule on one employee, with a single exception: Coordinated Cage Access is a cluster detection, so it surfaces as one string covering four employee IDs rather than four separate strings. It is pinned to the top of the list in every sort order — a confirmed multi-person pattern has no meaningful individual anomaly rank to sort on, and the mitigation spans three rosters plus a supervisor's swap-approval permission rather than one person. Each member's own profile is reachable from the evidence tables inside Detection. Two data limits bound how far this rule can be taken: the maintenance-window link is a temporal join against the tenant's published CMMS window (it records opportunity, not what work was done), and the cage signal is a badge read at the mantrap, not what happened inside — CCTV and the VMS are not connected sources in this deployment.
Tailgating / Piggybacking has no real data source in this export — it needs VMS analytics or a second-factor reader — and is shown as "cannot evaluate" for nearly every employee. That isn't a signal, it's an absence of data, so turning it into an individual case per employee would create noise rather than insight. It remains visible only as a caveat inside each employee's full profile.
Restricted-Zone Access Without Current Certification has partial month-level signal in the underlying data but isn't yet rolled up into a per-employee case in this export, so it is not shown as a string here. Buddy-Punching, Visitor Management Bypass, Privileged Engineer Cage Access, Pre-Resignation IP Exfiltration, After-Hours Privileged System Access, Infrastructure Sabotage and Fibre / Network Tap have no data source at all in this export.
There are two separate feedback signals in this console, shown in the counter top right. Marking a string "false positive" or "true positive" records the ground-truth outcome of the case — the mechanism by which Iveron's rule thresholds and anomaly baselines get retuned over time. The 👍/👎 buttons inside Evaluation and Closure are a separate, lighter-weight signal: they rate whether the automatically generated guidance text at that stage was accurate and useful, independent of how the case itself turned out — that's what would be used to improve the guidance-generation model itself. This preview does not persist either signal between sessions.
Full pipeline detail: analysis/README.md
and CLAUDE.md in the project repository.