Verification

What it is

Signal AI’s shipped transcript summary (GET /call/transcript_analysis/<call_record_id>, transcript_summary) already turns a call recording into prose a manager reads instead of listening to the call. Prose like “the caller asked about pricing three times and mentioned a competitor by name” is convenient exactly because it saves the manager from listening to the whole call — and exactly that convenience is what makes a wrong claim expensive: nothing about reading it invites the manager to go check. Verification attaches a citation to a specific, checkable claim in that output — a timestamp, a transcript line, a call ID — so the manager can jump straight to the passage the claim is based on with one click, instead of either blindly trusting the summary or re-listening to the entire call to spot-check it. The citation is not a decoration; it is the mechanism that makes “trust but verify” cheap enough that people actually do the second half. This is the natural partner to a summarization or search feature built on Signal AI: any time the output makes a specific claim about what happened on a call, the claim should carry a way to check it against the source. Not every claim can carry one — see the boundary with Caveat below — but every claim that can should.

Choose this when / choose something else when

Situation	Do this instead	Why
The claim is a broad, qualitative read with no discrete source passage (“the caller seemed frustrated”)	Caveat	Hedging the claim in words is honest; a citation link pointing at nothing specific is worse than no link
The user lacks permission to view the underlying transcript or recording	Name the source in text with no link	A citation that resolves to a permissions error teaches users that citations are decorative
The claim is trivial and its source is already visible on the same screen	Tooltip	Full navigation is unnecessary friction when the source is already in view
The system took an action rather than made a statement about existing data	Stream of thought, or an audit log	Verification checks a claim about data that already exists; an action needs a record of what happened, not a citation

Agency tier

Suggests. Verification composes with output the user still has to weigh and act on; it is a way to check a claim before trusting it, not an action of its own. The tier to watch here is not this pattern’s own — it’s whatever surface a verified citation feeds. If a citation’s presence ever becomes the sole justification for an unsupervised action (for example, auto-flagging a call as fraudulent because a citation happened to exist), that action has escalated to Acts-unsupervised without anyone deciding it, and a citation a human never actually opened is not the review that tier requires. See Agency tier.

Anatomy

┌───────────────────────────────────────────────┐
│  Call summary                                   │
│                                                  │
│  Caller asked about pricing three times¹        │
│  and mentioned a competitor by name².           │
│                                                  │
│  ¹ 0:42, 3:15, 6:02      [ View in transcript ]  │
│  ² 4:58                  [ View in transcript ]  │
└───────────────────────────────────────────────┘

Part	Role
Card	Container for the summary
Inline citation marker	Superscript numeral attached to one discrete claim
Citation list	Maps each marker to its timestamp(s) or transcript line(s)
Link	Navigates to the exact cited position — never a Button, since it resolves to a URL
Tooltip	Optional hover preview of the cited excerpt before navigating

Outcome states

State	Treatment
Working	No citations exist yet; the summary area shows Skeleton, not a claim with a citation pending
Streaming/partial	Citations attach per completed claim as the summary streams in; a claim rendered without its citation attached yet reads visually as unfinished, not as a bare, unverifiable assertion
Confident and right	The easy case — the citation, once opened, supports the claim exactly as stated
Confident and wrong	The primary case, and the hardest one for this pattern specifically. A claim can be stated confidently and carry a citation that looks legitimate while pointing at the wrong timestamp, the wrong call, or a passage that says something different. Nothing in the interface distinguishes a correct citation from an incorrect one without the user actually opening it — that is a named limitation of this pattern, not an oversight. The design response is to make opening a citation as close to zero-friction as possible (one click, deep-linked to the exact position, no separate page load), because the pattern’s entire defense depends on people actually doing that
Uncertain	If the system cannot attach a confident citation to a specific claim, the claim is dropped or rewritten as a hedge via Caveat — never given an approximate or invented citation
Refused	The summary states which claims could not be verified and omits them, rather than presenting an uncited claim with the same visual weight as a cited one
Empty	No claims worth citing — the summary says so (“nothing notable in this call”) rather than showing a citation list with no citations beside a confident-sounding summary
Interrupted	If generation is canceled mid-stream, already-attached citations stay valid and clickable; claims left without their citation are removed rather than shown dangling
Degraded	If source linking falls back to a lower-fidelity source (transcript text only, no audio playback), that is stated next to the link (“View transcript text — recording unavailable”), not discovered only after clicking
Rate-limited/over budget	Not applicable to this pattern directly. Verification is a display and navigation behavior on already-generated output, not a metered operation itself; a source-fetch throttle, if one exists, is a technical concern outside what this page documents
Stale	Applicable. A cited source can change after the citation was generated — a transcript correction, a redaction. If the source has changed since the citation was made, that is flagged; the link is never shown as current when it might not be

Disclosure & recourse

Does the user know this is AI? Yes — the citation markers are themselves part of the disclosure, visible inline with the claim they support, not confined to a settings toggle.
What did it use? Exactly what the citation shows — the specific timestamp or transcript excerpt the claim is grounded in.
How sure is it? This pattern substitutes a check for a score: instead of a confidence number, the user can verify the claim directly. A claim that cannot carry a citation is the signal to distrust it more — see the boundary with Caveat.
How does the user check it? That is this pattern’s entire mechanism — click the citation, land on the exact cited passage.
How does the user correct it? Undecided beyond disregarding or editing the summary text itself on a Drafts-tier surface. Whether a user can flag “this citation doesn’t support this claim,” and whether that flag reaches anyone, is an open gap below.
How does the user get out? The raw transcript and recording remain reachable directly, without going through the summary — the summary is additive to the source data, never a gate in front of it — and AI-generated summaries can be turned off per view.

Reference

No model, citation-extraction method, latency budget, or cost has been defined for this pattern yet. This is fact register — nothing below is invented to fill the gap.

Evaluation

Not evaluated. No eval set exists for citation accuracy — whether a citation actually supports the claim it’s attached to.

Content

Moment	✅	❌
Citation marker	”¹ 0:42 — “Do you have anything cheaper?""	"[source]“
Citation link	”View in transcript"	"Learn more”
Degraded source	”Recording unavailable — showing transcript text only”	(no mention of the shortfall)
Stale source	”This call’s transcript has changed since this summary was generated”	(link shown as if nothing changed)

Accessibility

A citation marker needs accessible text beyond a bare superscript numeral — for example aria-label="View source at 0 minutes 42 seconds" — not a number alone.
Streaming citations follow the same live-region rule as Stream of thought: announce on claim completion, never per token.
Verifiability is never color-only. A claim that could not be cited says so in text; it does not simply omit a colored marker and rely on the absence being noticed.
Activating “View in transcript” moves focus to the destination heading or timestamp marker in the transcript view, and that move is announced.

Constraints

ID	Constraint	Rationale
TITAN-VERIFY-01	Every discrete, checkable claim about call data carries a citation to the specific source passage, not to the call as a whole.	”See the call” is a redirect, not a citation on a specific claim — passage-level discipline is what makes the citation useful.
TITAN-VERIFY-02	A claim the system cannot cite to a specific passage is dropped or rewritten as a hedge, never given a fabricated or approximate citation.	A wrong-looking-right citation is worse than no citation — it is the exact failure this category of pattern exists to prevent.
TITAN-VERIFY-03	Citation links are Link, never Button.	Navigation to a resolvable location follows Button’s own navigation rule.
TITAN-VERIFY-04	Citation navigation lands on the specific timestamp or transcript line, not the top of the call.	Forcing the user to re-search inside the source defeats the point of citing a passage.
TITAN-VERIFY-05	When the underlying source is degraded (text-only, unavailable, redacted), that is stated next to the citation, not only discoverable by clicking through.	A disclosure that arrives after the fact is not disclosure.
TITAN-VERIFY-06	A citation is never styled or worded to imply verification has already happened (“confirmed,” a checkmark icon).	That language claims a review step nothing performed. See the overview’s own rule against unsubstantiated claims.
TITAN-VERIFY-07	If a cited source changes after the citation was generated, the pattern flags the citation as stale rather than showing it as current.	A citation is a promise about the state of the source at generation time, and it can go stale exactly like the answer it supports.

Divergences

Not applicable. Nothing is shipped yet to diverge from.

Gaps

Whether a user can flag a broken or incorrect citation, and whether that flag ever reaches a model owner, is undecided.
Whether citation-to-passage mapping is produced by the model itself or a separate retrieval or alignment step is undecided — this affects how reliably this page’s no-fabricated-citation rule can actually hold.
No decision on how long a generated citation stays valid once the underlying recording or transcript is deleted under data retention policy.
Whether citations render for every role, including a user without permission to view the underlying recording, or whether the claim shows with no citation at all in that case, is undecided.

Volatility

This guidance depends on the eventual model or pipeline being able to produce passage-level grounding at all — some retrieval architectures can cite a document but not a line or a timestamp within it, which would make this page’s passage-level citation requirement unbuildable as written. It also depends on Invoca’s data-retention rules for recordings and transcripts, which determine how long a citation stays resolvable before this page’s staleness-flagging rule has to apply. Reverify when a model or pipeline is selected for the first feature that cites call data, and whenever retention policy changes. Dated 2026-09-02. Stream of thought and Caveat, the other two pages in this batch. AI Experience overview, for the vocabulary and the agency-tier and outcome-state definitions this page uses. Actions: Summarize, Signal AI’s shipped transcript summary, which this pattern is written directly against. Link, Tooltip, Skeleton, and Card, the components this pattern composes.

Why it works this way

A citation is cheap to build and expensive to trust blindly — the entire value of this pattern is in shrinking the distance between reading a claim and checking it to a single click, because anything longer than that gets skipped under real workload, and a citation nobody opens verifies nothing. Pointing at a specific passage rather than the whole call is what makes checking fast enough to actually happen, which is the only thing that makes “trust but verify” different from “trust.”

Status

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

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

Status

​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

​Status

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

Status