What an assistant can do, operation by operation

How an assistant reads what is there. CHECK answers a question about a place at a moment: how much of something is in an area, what value a measured quantity takes across it, which way things are moving. It reports; it does not interpret.

1. Activity densityhow much of something is in an area: people, vehicles, events
2. Scalar fieldthe value of a measured quantity across an area, such as water depth, temperature or noise
3. Flow vectorsthe direction and rate of movement across an area
4. Place resolutionturn a place name into coordinates and a boundary

Simulation output read by CHECK is reported as simulation output, never as a field observation.

How an assistant traces movement through a place. SEEK follows paths: between two points, outward from one point, or between whole areas.

1. Origin–destination flowshow much movement goes from each area to each other area
2. Reachability isochroneseverywhere that can be reached from a point within a given time or distance
3. Routingthe path between two points and how long it takes

The order in which runs finished is reported as an order, never as a trend or a cause.

How an assistant watches a run or a place over time. MONITOR keeps standing watches and reports what has changed.

1. Cancel a watchstop a standing watch
2. Context summarythe setting those events happened in
3. Drift detection, CUSUMwhether a model’s behaviour has moved away from what it was checked against
4. Recent eventswhat has happened since the last look
5. Active watchessee the standing watches that are running
6. Current statewhat a run or a watched place looks like right now
7. Threshold subscriptiona standing watch that reports when a value crosses a line
8. Trend analysis, linear slopethe direction a value is moving and when it would cross a threshold

A status read is a snapshot. It does not create a watch or stream state; a watch is a separate, named subscription.

How an assistant sets up a hypothetical. IMAGINE describes a scenario that differs from the present in stated ways, and records the conditions it assumes, so that whatever is later computed carries them.

1. Counterfactual scenarioa scenario that differs from the present in stated ways, ready to be proposed
2. Evidence emissionrecord the conditions the scenario assumes, so the result cannot be read without them

IMAGINE never starts a run. A scenario becomes a run only after the person approves it.

How an assistant runs the simulation. EXPLORE advances an approved scenario through time and records everything needed to reproduce the run exactly.

1. Forward runadvance an approved scenario through time, deterministically, and record its results

A completed run establishes what the model did under stated conditions. It does not by itself establish how accurate the model is or how uncertain the result is; those are GAUGE’s to establish.

How an assistant puts results side by side. COMPARE lines up scenarios or states in time, place and units, and says whether they differ.

1. Pareto frontierthe options that cannot be improved on one objective without losing on another
2. Multi-scenario comparisonthe same measurements across several scenarios, aligned so they can be read together
3. Two-state test, Welch’s t-test and Mann–Whitney Uwhether two sets of results differ by more than chance would explain

Simulation output and real observations are never compared as if they were the same kind of evidence. A numerical difference is reported as a difference, not as a cause, a preference or a prediction.

How an assistant says why. EXPLAIN assigns a difference to a stated cause, and only when a stated cause map shows that the cause can be separated from everything else that could have produced the difference.

1. Causal attribution, backdoor identification and adjustment setshow much of a difference is assigned to the stated cause, given the variables the cause map says must be held fixed; if the map cannot separate the cause, EXPLAIN says so and stops
2. Evidence regimes
3. Simulationthe evidence is simulation output
4. Observationalthe evidence is real observations
5. Hybridboth, kept distinct

Identification is checked before any effect is estimated. An explanation that the cause map cannot support is refused rather than weakened.

How an assistant says how sure. GAUGE attaches an uncertainty statement to a result and enforces how strong that statement may be. The method is chosen from the kind of result; the analysis mode sets a ceiling of its own; the achieved strength is the minimum of every limit that applies.

1. Uncertainty assessmentattach an uncertainty statement to a result by one of the methods below, bounded by the rule above
2. Distributional
3. Parametric intervalan uncertainty range that assumes the result follows a known distribution shape
4. Resampling
5. Block bootstrapresampling in blocks so that patterns over time are preserved
6. Bootstrap intervala range from resampling the data itself, with no distribution assumed
7. Re-running the simulation
8. Monte Carlo ensemblere-running with different random seeds and reading the spread
9. Paired delta ensemblerunning two scenarios under identical seeds so that only the change between them shows
10. Parameter perturbationre-running with the inputs nudged to see how far the answer moves
11. Sensitivity analysischanging one input at a time to find which inputs matter most
12. Place by place
13. Spatial bootstrapresampling, reported place by place
14. Spatial ensemblethe ensemble spread, reported place by place
15. From real observations
16. Observational bootstrapresampling real observations rather than simulation output
17. Observational ensemblean ensemble anchored to real observations; the ceiling rises with how much of the area the observations cover
18. Against real observations
19. Calibrationchecking the model against real observations and reporting how well it fits
20. Inline calibrationthe same check, run inside another analysis
21. Analysis modes and their ceilings
22. Audit only, CS0report what was computed and claim nothing about uncertainty
23. Calibrated predictive, CS3the model has been checked against observations for this use
24. Decision grade, CS4the strongest ceiling; every other limit still applies
25. Ensemble active, CS3an ensemble of runs is available to draw on
26. Internal uncertainty, CS1describe the spread inside the model only
27. Validation active, CS3observations are available to check the model against

Descriptive spread across runs is reported as spread, never as a confidence interval, unless a method that supports an interval was run.

How an assistant looks for the best option. CHOOSE searches the allowed values of a decision against stated objectives and hard limits, and ranks what it finds. It is the only operation that can reach decision grade, and it cannot recommend at all until GAUGE has attached an uncertainty to the candidates.

1. Decision searchsearch the allowed values of a decision against the objectives and hard limits, by one of the methods below
2. Search
3. Bayesian optimizationa search that learns from each trial where to try next
4. Grid searchtry every combination on a fixed grid
5. NSGA-IIan evolutionary search for the set of best trade-offs among several objectives
6. Random searchtry random combinations within the allowed ranges
7. Robust optimizationprefer options that stay good when the inputs are uncertain
8. Weighted-sum enumerationscore every option by a weighted total of the objectives
9. Before a recommendation is made
10. Constraint evaluationthe hard limits an option must satisfy; options that fail are dropped, not ranked lower
11. Grade gatea recommendation is withheld unless the attached uncertainty supports the claim strength asked for
12. Recommendation sensitivitywhether the ranking survives small changes to the inputs; an unstable ranking is not recommended

A ranking is a ranking of the submitted candidates under the stated objectives. It does not establish what the person prefers.

How an assistant proposes to learn. ADAPT derives a proposed change from saved evidence and hands it to the person. Nothing is learned until a person accepts the proposal.

1. Belief-revision proposala proposed change to a value, derived from one exact piece of saved evidence, for the person to accept or not
2. Model recalibration proposala proposed change to a model’s parameters based on observations, for the person to accept or not

A proposal applies no change and does not estimate its own relevance. The person’s acceptance is a separate, recorded act.