Appearance
Composer Live
Composer Live connects selected original observations to an interactive comparison and a saved decision brief. Change your priorities or segment, explore Decision Terrain, and use Impact Replay to see how changed inputs affect the result.
Where to find it
Open Composer Live from the panel picker. When a map of the open space is on screen, the Composer sidecar also shows a Live button in its input footer, next to Attach files. The point-action menu offers Explore in Composer Live for the points you have selected.
Build an analysis
- Select between 1 and 100 original observations on a map.
- Choose Use selected sources.
- Choose the field that groups observations into options and, optionally, a segment field.
- Choose two or three distinct numeric criteria. For each, specify whether higher or lower values are preferred, whether to use the mean or sum per option, and fixed scale bounds. Declare units and an optional absolute measurement tolerance per observation.
- State why the observations are comparable and decide whether incomplete observations should stop the computation or be explicitly omitted.
- Open the live preview.
The field picker suggests up to 32 fields from the selected originals. An analysis supports up to 20 options. Numeric strings and booleans are not treated as measurements. Values outside the declared scales remain visible; they are not silently clamped before scoring.
New records elsewhere on the map are not automatically added. The selected source identities define the analysis.
Explore Decision Terrain
Move the priority sliders, choose a segment, or click the terrain. The brief and option comparison recompute from the declared inputs.
With three criteria, each point in the triangle represents a different allocation of priorities. With two, the terrain is a horizontal band. Regions are calculated from the linear weighted scores; they show where each option leads. Explore nearest tie moves to the nearest feasible equal-score boundary.
Region area is a share of possible priorities, not a probability that an option is correct. Tied regions can overlap. A Pareto-frontier label means that no other included option is at least as good on every criterion and better on one. It does not establish causality or statistical confidence.
With nonzero measurement tolerances, stronger fill shows where an option stays at least tied for best even at the worst permitted measurement values. Lighter regions are sensitive to those bounds. The brief distinguishes a current leader from one that stays strictly ahead throughout the declared tolerances. Means retain the per-observation bound; sums accumulate it. These are user-declared worst-case bounds, not inferred statistical confidence intervals.
Scores use the declared fixed bounds and weights. A score can fall outside 0–1 if current measurements exceed those bounds. The comparison bars clip their lengths; the numeric score remains visible.
Save and review
Save workspace retains an immutable candidate with its definition, controls, original input values, computation identity, and result. Reopen it from Saved workspaces.
Inspect the evidence and assumptions, enter a review note, and confirm before choosing Accept snapshot. Acceptance establishes the baseline. A new candidate cannot replace it if the source data or accepted baseline changed during review.
Explore with Live preview, or use Saved snapshot to inspect frozen historical inputs and results. Reset to saved controls restores that snapshot's priorities and segment. Saving new controls creates another review candidate; it does not modify the earlier snapshot.
If another review changes the accepted baseline, refresh and choose Save review candidate to compare your current priorities against that latest baseline. The older candidate remains in history. A save whose response is lost can be retried safely without creating a duplicate snapshot.
Follow Impact Replay
The open panel checks for updates after 30 seconds without an active read while the page is visible, when the window gains focus, or when you refresh it. Background checks wait for active reads to finish. If a request fails or times out, Refresh retries it, including source-field discovery. The panel does not run a background monitor when closed.
Replay shows changed declared source fields, controls, or computation definitions and the stages whose outputs changed: original evidence, selected segment, measurements, decision, and brief. The original brief stays beside the proposed brief. A changed input can leave the leading option unchanged.
Changes to unrelated metadata do not invalidate the analysis. A deleted, moved, or reassigned source causes its saved contents to be withheld. Current space and map permissions are checked on every request.
The panel lists the 50 most recently updated workspaces and the latest 50 snapshots in a workspace. The currently open workspace and snapshot remain selectable even when a list refresh is pending or the selected baseline is older than the latest 50 snapshots. Saved receipts remain in the backend; these lists are bounded navigation views.
Use Composer
Composer can inspect workspaces with get_composer_live, compute a read-only preview with preview_composer_live, and save an explicitly requested candidate with save_composer_live_candidate. Ask mode permits inspection and previews. Adoption stays in the human review interface.
The initial computation pipeline is deliberately bounded: original observations → segment → grouped numeric measurements → linear decision → brief. It does not execute generated code, infer missing measurements from prose, or claim that a recommendation is scientifically verified.