The ids are the same — a twin is named expected_freeze##<solution>##<field id>; the engine strips the prefix and the collaboration suffix, so the field node and its twin share one id. A node present in one world only was added or deleted there.

The baseline tells who changed — when the two worlds differ, the baseline (the expected world right after the clone, when both were equal) tells which side moved: the field (something happened) or the plan (the process did something). The clone is progressive, so the baseline is only taken once the size of the expected world is stable between two computations; it is stored in the SupervisionBaseline collection and survives a restart.

Matched changes cancel out — when the process does what the field also shows (a task resolved the gas leak and the sensor confirms it), both worlds changed the same way: nothing is weighted. Only a change found in one world counts.

Involved instances — an instance is involved when a task of the solution which is not over yet uses it (actor through Assigned_To, consumed, produced or required resource), or when an objective of the solution treats or prevents it (Solves → Plans → Treats | Prevents, and the objectives its tasks were deduced for through Created_By, such as the corrective objectives of an Event_Flow). The ACTIVE clone raised Near a treated FREEZE or REFERENCE problem counts as treated too. Changes on involved instances weigh more than the others.

Backend plugin — a module gind-detection-engine-<name> extends AbstractSupervisionStrategy (name, categories it feeds, extension name, checked by default or not, default configuration of a project, compute(SupervisionContext, configuration)) and declares itself in META-INF/services/fr.emac.gind.monitoring.supervision.AbstractSupervisionStrategy. The r-iota webapp depends on the plugins at runtime; ServiceLoader finds them at startup.

Frontend extension — a project gind-npm-strategy-supervision-<name> is an Angular bundle loaded by r-iota from supervision-strategy-extensions.json. It registers itself in window.gindSupervisionStrategyExtensions and mounts its configuration in the box: mount(host, {strategy, configuration, onChange, readonly}) returns update, setReadonly and destroy.

Sessions — the SupervisionEngine is started with the r-iota webapp. A knowledge space is watched once r-iota asks for its state, or a few seconds after a solution is supervised. A one-second tick computes each session whose period is over; computations never overlap.

Period — 10 s by default, changed in the box even while a solution is supervised (it does not change what is measured). It cannot be shorter than 5 s, has no maximum, and is never shorter than five times the last computation, so a large use case slows itself down (about 0.5 s per computation on IMT Albi Crisis).

Business time — the now of a computation, the start of the supervision when the process instance gives none, and the dates of the history points are read on the clock of the use case; only the scheduling of the computations keeps the server time.

History — a point (date, value of each category, value per strategy, up to 200 contributions) is kept in memory and in the SupervisionHistory collection each time the values change, so the curves survive a reload. Every computation is also pushed to the r-iota pages as a supervisionStatusEvent.

Configuration — the checked strategies and their configurations are stored in the hidden project property supervision configuration (JSON). While a solution is supervised the configuration is frozen; the server refuses to save it until the solution is cancelled. The period and the alert definition, kept in the same property, can still be changed.

Alert definition — what r-iota counts as an alert (pulsars, alert counters, status light, Find solution(s)): any active actuality, and an active risk whose probability reaches a threshold, 50 % by default (Reset to 50 %, then Save). Below it a risk is only watched: no pulsar, not counted; the “No alert detected” view lists the watched risks with their probability. The definition is kept in the project and can be changed while a solution is supervised.

Forecast certainty — a forecast of a sensor is a set of curves (low, central, high quantiles). A forecast actuality counts once it appears on a share of the curves, 50 % by default; a forecast risk takes the probability it reaches on the central curve. The timeline of r-iota and the deduction with the forecasts use this share; the “No alert detected” view lists the alerts foreseen by the forecasts and offers “Find solution(s) on the forecast”.

Alert box — when an alert category newly reaches 1 (and the supervision box is not open), a single box opens on the most severe crossed category, with the contributions that explain it and the adaptations. The radar button turns red and reopens this box.

Adaptations — Runtime · Modify the running process: deduce again from what was done, then change only what is not started yet in the running process; the tasks added and removed are shown on it and applied together, or not at all. Design time · From what was done keeps the tasks done and in progress with their actors; From scratch deduces the whole treatment again. Both reuse the strategy and options of the last Find (forecasts included); the new solution is reviewed (graph of its process, editable) and replaces the supervised one only when it is deployed. Continue keeps the current solution. Before a new deduction the box lists the problems to treat (actualities on the field, risks at or above the alert probability, actualities projected by the forecasts), all checked: the objectives of an unchecked problem are left out.

Improvable — never opens the box: the radar button gets a blue badge and a light banner tells that a better solution may exist; the operator can hide it.

Cancel the current solution — two choices. Memorize the tasks already done writes the done tasks, their actors, dates and outputs on the solution (memorized done tasks) before cancelling, so that the next deduction can keep those actors; memorize nothing starts again from scratch, the deduction being free to choose better actors. After cancelling, the tasks of the To-Do Lists show as cancelled, r-iota is no longer supervised, the alerts are visible again and a new Find solution(s) covers every current problem.

The deduction does not read memorized done tasks: the adaptation From what was done starts from the supervised solution itself.