Controls

Exemplar page — first pass, entirely proposal. No Invoca product ships this pattern today. Nothing below is code fact or measured behavior — it is a proposal offered for review. See Coverage, stated honestly.

What it is

A manager reviewing a Signal AI summary wants more than the two-sentence version — they want every objection the caller raised, not just the outcome. A Control lets them ask for that directly: a “Brief / Full” toggle next to the summary, scoped to this one run, that changes what gets generated the next time without the manager having to type a request describing the change they want. The same job applies to a marketer running a natural-language search who wants to restrict it to a specific date range or team before running it, rather than including that constraint in the words they type. A Control is a fixed, named parameter — detail level, date range, tone, team scope — that shapes an AI run without being the request itself. Controls are scoped to the feature they tune. They are not durable, product-wide preferences (those are ordinary Switch or Select settings with no AI dimension) and they are not the request itself (that is Open input). A Control changes how the next run behaves; it never is the run.

Choose this when / choose something else when

Situation	Do this instead	Why
The user is expressing a new request in their own words, not adjusting a parameter of an existing one	Open input	A typed request carries intent a fixed control can’t capture; a Control only narrows or reshapes an intent already established
The interface is proposing what to do next, rather than letting the user change how the current thing is done	Suggestions	A Suggestion offers a next action; a Control reshapes the current one
The setting is a durable, product-wide preference with no AI dimension (e.g., default page size)	Switch or Select as an ordinary settings control	Controls governs parameters scoped to one AI feature’s run, not general application preferences
The user is reviewing what the AI already did, rather than adjusting what it will do next	Governors: Stream of thought	Oversight of a completed or in-progress run is a different job than tuning the next one

Agency tier

Suggests. A Control changes what will be proposed or generated, but the user still triggers the run and reviews its output — changing “Brief” to “Full” does not itself act on anything; it changes the input to the next Suggests-tier result. If a Control is ever wired to change an already-running, unsupervised action’s scope in real time — narrowing what an autonomous process is currently doing rather than configuring the next discrete run — that is Acts-tier behavior wearing a Tuner’s interface, and the escalation must be recorded explicitly, not left implicit in the control’s own simplicity. Per TITAN-AI overview, escalating a feature’s tier is a decision, never a default.

Anatomy

┌─────────────────────────────────────────────┐
│  Summary detail:   ( Brief )  [ Full ]       │
│  Date range:        [ Last 30 days ▾ ]       │
│                                               │
│                              [ Regenerate ]  │
└─────────────────────────────────────────────┘

#	Part	Component	Required
1	Control group container	Card	Yes
2	A discrete, mutually-exclusive control	Button group or Switch	Yes — one or more
3	An enumerable-list control	Select	No
4	Apply / regenerate trigger	Button	Yes, unless a control applies immediately on change

Outcome states

State	Treatment
Working	Regenerate button enters `loading`; the previous output stays visible, dimmed, until the new one replaces it
Streaming / partial	If the run itself streams, changing a control mid-stream cancels the in-flight stream rather than blending old and new output
Confident and right	The output reflects the control values the user actually set
Confident and wrong	Most detail below.
Uncertain	Not applicable to the control itself. Uncertainty belongs to the generated output, not to the setting that shaped it — see the invoked Action’s own page.
Refused	A combination of control values is invalid or unsupported (e.g., a date range with no available data) — refused before running, with the specific incompatible values named
Empty	The combination is valid but returns nothing (e.g., “Full” detail requested on a call with too short a transcript to elaborate) — distinguished from a refusal by stating that this combination legitimately has nothing to return
Interrupted	User changes a control while a previous run is still in flight; the previous run is cancelled outright, not left to finish and silently discarded
Degraded	A control’s higher-cost option (e.g., “Full” detail) falls back to a cheaper generation path under load; the output states this rather than silently returning a shorter result under the “Full” label
Rate-limited / over budget	Controls stay adjustable, but the regenerate action is blocked with the reason stated
Stale	The applied control state was captured before the underlying data changed; the output does not silently claim to reflect the current data if it doesn’t

Confident and wrong. The distinctive failure here is a control silently not doing what its label says — a date-range control that defaults to “last 7 days” instead of the “last 30” the user actually selected, or a “Full” detail toggle that has no measurable effect on the output. The output that results reads exactly as confidently as one generated from the correct settings, because nothing about its tone changes based on which parameters actually produced it. The user notices only by checking the applied control state against the result — for example, counting whether the number of calls referenced is consistent with a 30-day window — which requires them to already suspect something is wrong before they’d think to check.

Disclosure & recourse

Does the user know this is AI, at the moment it matters? The controls themselves are settings, not AI output — disclosure at this stage is about the run they configure, and that run’s own page carries the identifier requirement. This page requires only that the controls are visibly attached to the AI feature they tune, not floating as generic settings.
What did it use? Not applicable to the control itself — this question belongs to the run’s output, which states what data it drew on given the applied settings.
How sure is it, and does that change what the user should do? Not applicable. Controls carry no confidence of their own; only the output they shape does.
How does the user check it? By the applied control state remaining visible alongside the output it produced — a result with no visible record of which settings generated it cannot be checked against them at all.
How does the user correct it? Change the control and regenerate. Whether a correction here also updates the AI’s stated confidence or caveats in the output, or requires a separate resubmission to reflect, is undecided — see Gaps.
How does the user get out? A reset to default values is available, and running the feature at all remains optional regardless of how the controls are set.

Reference

No model, prompt, tool schema, latency budget, or cost has been defined for how any control value maps to an actual change in model behavior.

Evaluation

Not evaluated. No eval set exists confirming that a given control setting (e.g., “Full” detail) produces a measurably different output from its alternative.

Content

Element	✅	❌
Control label	Summary detail	Options
Control values	Brief, Full	Short, Long
Apply trigger	Regenerate	Go
Refused combination	No calls in the selected date range have transcripts to summarize.	Invalid selection.

Control labels name the dimension being adjusted, not the mechanism — “Summary detail,” not “AI settings” — matching Field label’s own rule that a label states what the value means, not how it’s implemented.

Accessibility

Changing a control that triggers regeneration announces the new loading state via aria-live="polite", distinct from the initial run’s announcement, so a screen reader user knows a change is in progress rather than assuming the original result still stands.
Regeneration’s indeterminate wait needs a non-visual equivalent — an accessible status text, not only a visual spinner change.
No control’s effect is conveyed by color alone; a selected discrete control (e.g., “Full” in the button group) is marked with aria-pressed or equivalent state, not color contrast alone, per TITAN-COLOR-03.
When regeneration completes, focus stays on the Regenerate control rather than jumping to the new output automatically — the user changed a setting deliberately and is not disoriented by staying where they are, unlike a fresh Open input submission where focus does move to the new response.

Constraints

ID	Constraint	Rationale
TITAN-TUNE-01	The control values that produced a given output remain visible alongside it, not only at the moment of setting them.	Disclosure & recourse, question 4 — a result cannot be checked against settings the interface has already hidden.
TITAN-TUNE-02	A control that applies immediately (no separate regenerate step) states so visually — the absence of an apply button is not sufficient signal on its own.	Otherwise identical-looking control groups behave differently with no way to tell which is which before touching one.
TITAN-TUNE-03	Changing a control while a previous run is in flight cancels that run outright; it is never left to complete and silently discarded.	A discarded-but-completed run wastes the cost of generating it and risks a race where the stale result renders after the new one.
TITAN-TUNE-04	An invalid or unsupported combination of control values is refused before running, naming the specific incompatible values.	”Invalid selection” tells the user something is wrong and nothing about what to fix — same reasoning as TITAN-ERR-03.
TITAN-TUNE-05	A degraded fallback for a higher-cost control option states that it fell back, rather than returning a lesser result silently labeled as the option requested.	An output that doesn’t match the label the user selected, with no disclosure, is the control-level version of confident-and-wrong.
TITAN-TUNE-06	Controls provide a reset to default values.	Recourse requires an easy way back to a known-good state, not just forward adjustment.
TITAN-TUNE-07	A Control never itself performs an unsupervised action; changing one only ever configures a future Suggests-tier run unless that escalation is explicitly stated.	Per TITAN-AI overview — escalating tier is a decision, never a default, and a Control’s simplicity makes it an easy place for that line to blur unnoticed.

Divergences

Not applicable — nothing is shipped yet to diverge from.

Gaps

Whether a control change updates the AI’s stated confidence or caveats in an already-visible output, or requires a full regeneration to reflect, is undecided.
How many controls a single feature may expose before the group should be considered too complex for this pattern (and better served by a dedicated settings surface) is undecided.
Whether control state persists across sessions per feature, per record, or not at all is undecided.
What “measurably different” means for validating that a control actually changes model behavior — see Evaluation — has no defined threshold.

Volatility

This page assumes that a named control value (detail level, tone, scope) reliably produces a distinguishable change in model output — that “Full” is actually fuller than “Brief” in a way a user would notice. If a given control turns out to have no measurable effect, that control is decoration wearing the shape of a setting, and this page’s guidance for it specifically needs reverification, not the whole pattern. Dated 2026-09-02; revisit on the first real implementation or the first eval run comparing control values against output.

AI Experience overview — vocabulary, agency tiers, and the six disclosure questions this page answers
Open input — expressing a new request, as distinct from adjusting a parameter of an existing one
Suggestions — proposing a next action, as distinct from reshaping the current one
Governors: Stream of thought — reviewing what the AI already did, the complementary oversight job to tuning what it does next
Actions: Summarize — Signal AI’s shipped job, the concrete case this page’s detail-level example is drawn from
Button group, Switch, Select, Card — the components this pattern composes

Why it works this way

A Control’s whole value is that it’s cheaper than rephrasing a request in words. Once a parameter is named and fixed — detail level, date range, tone — a toggle is faster than typing a sentence that means the same thing, and the interface can validate it before running rather than discovering an unsupported combination mid-generation. That speed only holds if the control actually does what its label claims; a control whose setting doesn’t measurably change the output is worse than no control, because it teaches the user a lever exists where none does.

Overview

Wayfinders

Inputs

Tuners

Governors

Trust builders

Identifiers

Actions

What it is

Choose this when / choose something else when

Agency tier

Anatomy

Outcome states

Disclosure & recourse

Reference

Evaluation

Content

Accessibility

Constraints

Divergences

Gaps

Volatility

Why it works this way

​What it is

​Choose this when / choose something else when

​Agency tier

​Anatomy

​Outcome states

​Disclosure & recourse

​Reference

​Evaluation

​Content

​Accessibility

​Constraints

​Divergences

​Gaps

​Volatility

​Related

​Why it works this way

What it is

Choose this when / choose something else when

Agency tier

Anatomy

Outcome states

Disclosure & recourse

Reference

Evaluation

Content

Accessibility

Constraints

Divergences

Gaps

Volatility

Related

Why it works this way