VIGIL™ by CAT Sentinel
Multimodal Voice
Design Framework
How VIGIL™ generates, formats, and delivers proactive CAT briefings across text, structured data, and US broadcast-quality audio — grounded in NPR and Bloomberg Radio standards.
Version 1.0
For Design & Engineering
May 2026
vigil.pionedata.com
01
The US audio models VIGIL™ is designed after
Three reference points, each contributing something distinct to the VIGIL™ voice format. The target is the intersection — not the average.
NPR — Sentence discipline

NPR's training guide for broadcast journalists establishes the foundational rule: write the way you speak, not the way you write. One idea per sentence. Subject before verb before object. Numbers kept simple — listeners cannot rewind. The test: speak it before you type it. If it sounds like a document being read aloud, rewrite it. NPR's average sentence in a news spot runs 12–16 words. This is the structural discipline VIGIL™ voice scripts inherit.

Bloomberg Radio — Financial data delivery

Bloomberg Daybreak anchors Nathan Hager and Karen Moskow deliver market-moving data in a format that is precise, direct, and action-oriented. The structure is consistent: state the number, state what it means, state what moves next. No warm-up, no preamble. The Bloomberg briefing assumes a sophisticated audience that wants the fact and the implication — nothing more. VIGIL™ voice scripts inherit Bloomberg's data-forward cadence and professional register.


02
Two texts from one intelligence object
Every VIGIL™ briefing cycle generates two parallel text outputs per role — the written brief and the voice script. These are not the same document. They are generated by different Claude API prompt templates from the same verified CAT data.
Written brief — designed for eyes
Structured, data-dense, precise. Tables are appropriate. Figures are exact. Sentences are complete. Designed for filing, referencing, and forwarding. Read at the reader's own pace.
Voice script — designed for ears
Written from scratch by Claude with US broadcast constraints. One idea per sentence. Numbers rounded. Active voice throughout. No embedded clauses. Designed to be absorbed once, at normal speech pace, by someone who cannot rewind. Approximately 85–100 words for a CRITICAL brief. Run time: 45–60 seconds.
Same data — UW voice script

Texas is in an active SCS accumulation. CAT Sentinel has recorded one-hundred-and-sixty-eight high-severity alerts there in the past nine months. That's one event every day and a half. Three separate systems are active right now across Harris, Dallas, and Travis counties. Combined portfolio exposure: around sixty-eight billion dollars. The pricing gap is significant. The observed storm frequency in Texas is running at eighteen percent per year. The model assumption is five. That's a thirteen-point gap — and it's been compounding for three renewal cycles. Action is required before the next renewal batch.

✗ Do not do this — document read aloud

"As of this briefing dated October 15th, 2024, CAT Sentinel has identified a total of 168 HIGH-severity Severe Convective Storm alerts within the state of Texas over the preceding 9-month period, representing a frequency of approximately one event per 1.6 days, with three simultaneous active events currently located across Harris, Dallas, and Travis counties."

✓ Correct — broadcast voice script

"Texas is in an active SCS accumulation. One-hundred-and-sixty-eight high alerts in nine months. That's one storm event every day and a half. Three systems are active right now across Harris, Dallas, and Travis. Portfolio exposure: sixty-eight billion dollars."


03
Voice script writing rules — US broadcast standard
These rules are enforced via the Claude API system prompt for voice script generation. Violations cause a script to fail quality review and regenerate before TTS rendering.
RuleStandardNPR/Bloomberg rationale
Sentence length 15 words maximum per sentence. Target 10–12. One idea per sentence — always. NPR training: listeners cannot rewind. Information lost when sentence exceeds working memory span.
Sentence structure Subject → verb → object. Never reverse. No embedded clauses. No "which" or "that" relative clauses. NPR rule: embedded clauses force the listener to hold two ideas simultaneously — comprehension drops.
Voice Active voice throughout. No passive constructions. "Texas recorded" not "was recorded in Texas." Bloomberg Radio standard: passive voice dilutes urgency and obscures the actor of the action.
Numbers — rounding Round to the nearest significant figure that preserves meaning. $68.4B → "around sixty-eight billion." 18.3% → "eighteen percent." 13.3pp → "thirteen points." NPR: "Numbers can require your listener to do math, which distracts from your story." Precise figures belong in the written brief only.
Numbers — large figures Spell out fully spoken: "sixty-eight billion dollars" not "$68B" or "68 billion." The listener hears words, not symbols. Bloomberg Radio convention: numbers are always spoken in full at first reference. Abbreviations are visual tools only.
Numbers — count limit Maximum 3 distinct numbers per 60-second brief. More than 3 creates a math problem for the listener. Additional figures go in the written brief. NPR guidance: "There isn't enough time in radio to provide context for multiple statistics."
Acronyms First reference: always expand. "Severe Convective Storm — SCS." After that: spell out each letter, do not try to pronounce as a word. "S-C-S" not "sic-s." Pronunciation dictionary handles this. NPR: acronyms that don't work spoken aloud break listener flow. The Pronunciation Dictionary is mandatory.
Tense Present tense for active conditions ("Texas is in an active accumulation"). Simple past for concluded events ("the March outbreak tested six states"). Future simple for forecasts ("the renewal window closes in ninety days"). Bloomberg Radio: present tense creates immediacy and currency. Avoids "as of today" constructions.
Quotation No direct quotation. Paraphrase all Event Book™ data findings. "The Event Book shows" not "According to the CAT Event Book™ Peril Trend Atlas, Chapter 3, the observed frequency..." NPR: listeners cannot see quotation marks. Paraphrase with attribution word ("shows," "finds," "records") is cleaner for audio.
Jargon Insurance jargon is permitted — VIGIL™ audience is professional. But each term must work spoken aloud. "TIV" → "T-I-V." "FNOL" → "F-N-O-L." "AOP" → "A-O-P." "PML" → "P-M-L." "Per-occurrence" → as written, it works spoken. NPR: "The more non-conversational language in a script, the harder it is to deliver naturally." Domain terms stay; document syntax goes.
Opening sentence Always starts with the situation: peril + geography. Never starts with "This is a briefing about..." or "CAT Sentinel has detected..." Open with the fact, not the meta. Bloomberg Daybreak format: lead with the fact that matters. The anchor context comes second, not first.
Word economy No filler phrases: "it's worth noting," "it's important to mention," "as previously stated," "in terms of." Cut every word that does not carry information. NPR: "Words we would never say out loud creep into scripts." Marketplace host Ryssdal's rule: if you'd cut it in speech, cut it in script.
Script length CRITICAL brief: 85–105 words / 45–60 seconds at broadcast pace. HIGH brief: 65–85 words / 35–50 seconds. WATCH brief: 50–65 words / 30–40 seconds. Bloomberg Daybreak market reports: 45–90 seconds. NPR news spot: 45–60 seconds. These are proven listener attention windows for professional audio.

04
Voice script structure — the VIGIL™ five-beat format
Every VIGIL™ voice script follows a five-beat structure. This is not optional — it is what makes the briefing consistently absorbable. Each beat has a word budget and a function.
1
The situation — 10–15 words
Peril. State. Active status. One sentence. No numbers yet. Opens with the fact that demands the listener's attention. Example: "Texas is in an active Severe Convective Storm accumulation."
2
The number that matters — 10–20 words, max 2 figures
The single most important data point from the signal — rounded, spoken in full. May include the comparison figure if the gap IS the story. "One-hundred-and-sixty-eight high alerts in nine months. That's one storm every day and a half." The "that's" construction is a Bloomberg Daybreak convention — it makes the number human.
3
The portfolio implication — 15–25 words
What this means for the carrier's book. TIV at risk. Coverage gap. Concentration signal. One or two sentences. This is the Marketplace "here's why it matters" beat. "Combined portfolio exposure is around sixty-eight billion dollars. The pricing gap has been compounding for three renewal cycles."
4
The primary action — 15–20 words, one action only
The single most urgent action item, stated as a directive. Owner named. Deadline stated in plain language — not ISO timestamps, but "within forty-eight hours" or "before end of business Thursday." Additional actions are in the written brief. "Sarah Chen in Texas underwriting needs to apply a surcharge before the next renewal batch. Deadline: forty-eight hours."
5
The VIGIL™ sign-off — 10–15 words
Timestamp, full brief reference, next scheduled update. Consistent across every brief — like a Bloomberg market report sign-off. "This is your VIGIL briefing for October fifteenth, six-fourteen Central. The full written brief is available now."
Full sample voice scripts — CRITICAL SCS Texas event
Four simultaneous scripts generated in parallel by Claude API — same event, four roles, four registers.
CRO Executive Brief · Voice Script
Critical ~52 sec Auto-generated 06:14 CST

Texas is in a critical SCS accumulation. [300ms] Portfolio exposure in the active alert zone is around sixty-eight billion dollars. [200ms] The base-case loss estimate is in the range of nine to fourteen billion. [200ms] Reserve adequacy is currently within band — but the base scenario is within twenty percent of the upper threshold. [200ms] Three actions are in progress. [200ms] The most time-sensitive is a reinsurer notification — the treaty attachment probability is above seventy percent. [200ms] David Park owns that notification. Deadline: end of business today. [300ms] This is your VIGIL critical briefing for October fifteenth, six-fourteen Central. The full written brief is ready now. // 92 words · NPR sentence structure · Bloomberg data format · Marketplace implication beat

Underwriting Action Brief · Voice Script
Critical ~58 sec Auto-generated 06:14 CST

Texas SCS — critical level. [300ms] One-hundred-and-sixty-eight high alerts in the past nine months. That's one storm event every day and a half. [200ms] Three systems are active right now across Harris, Dallas, and Travis counties. [200ms] The pricing gap is significant. Observed frequency is running at eighteen percent per year. The model assumption is five. That's a thirteen-point gap — compounding for three renewal cycles. [300ms] Sarah Chen needs to apply a surcharge of eighteen to twenty-two points to all Texas renewals due in the next ninety days. Do not wait for the scheduled pricing review. [200ms] Deadline: forty-eight hours. [300ms] Full exposure table and analogue data are in the written brief. // 102 words · strong action directive · Bloomberg data delivery · Marketplace context beat

Claims Pre-Position Brief · Voice Script
Critical ~55 sec Auto-generated 06:14 CST

Active SCS event — Texas, three counties. [300ms] First-notice volume forecast: between three-hundred-forty and five-hundred-twenty claims. [200ms] Peak is expected within forty-eight hours of the system clearing Harris County. [200ms] You need eight adjusters staged in Lubbock before that window. [200ms] Marcus Webb owns the staging call. Deadline: six hours from now. [200ms] Also flag your preferred roofing network. Demand surge in the Amarillo corridor is running around twenty-two percent above normal rates right now. [300ms] Full timing curve and vendor contact list are in the written brief. // 96 words · operational directives · concrete numbers · adjuster-centric register

Management Portfolio Brief · Voice Script
Critical ~48 sec Auto-generated 06:14 CST

Active critical event — SCS, Texas. [300ms] Sixty-eight billion dollars of portfolio exposure is in the alert zone. [200ms] Four action items have been dispatched across underwriting, claims, and risk management. [200ms] Two are complete. One is in progress. One is approaching its deadline in the next six hours. [200ms] There are no pricing model flags beyond the active SCS event. [300ms] The T-plus-fourteen debrief is scheduled for October twenty-ninth. [200ms] The full action tracker is live on the dashboard. // 79 words · portfolio-wide view · completion status · no role-specific detail


05
Role-specific tone calibration
Each briefing role receives a different voice register — same data, different delivery style. This is encoded in the Claude API system prompt per role and matched to a distinct ElevenLabs voice profile.
Executive
CRO Voice
Bloomberg Surveillance anchor register
Measured, authoritative pace
Slightly slower than other roles
Financial precision with gravitas
No hedging — states fact then implication
Pauses are longer — 400–500ms
Underwriting
UW Voice
Bloomberg Market report register
Direct, data-forward pace
Slightly faster than CRO
Emphasis on the gap number
Action directive delivered crisply
Pauses: 200–300ms between beats
Claims
Claims Voice
NOAA Weather Radio operational register
Operational, immediate pace
Fastest of the four roles
Logistics-forward — who, where, when
Numbers emphasized (count, hours, %)
Shortest pauses — 150–200ms
Management
Mgmt Voice
NPR Business Desk anchor register
Calm, summary pace
Measured — no urgency tone
Status-forward: complete vs pending
No role-specific jargon
Pauses: 300ms — reflective rhythm

06
TTS engine — ElevenLabs configuration
ElevenLabs v3 (eleven_v3) is specified for VIGIL™ voice delivery. Three reasons: native audio tag support for pacing control, the 10,000+ voice library for role differentiation, and 75ms Flash v2.5 latency for push notification delivery.
Recommended voice profiles
These are starting points for selection — carriers may configure their own ElevenLabs voice via the VIGIL™ configuration dashboard. Professional Voice Cloning (carrier's own named voice) is available at Tier 3+.
🎙
CRO Brief — "Marcus" profile
Male, 45–55 range, American English, measured cadence. Closest reference: Bloomberg Surveillance anchor. Stability: 0.75. Clarity: 0.85. Style: 0.2.
CRO · Management eleven_v3 speed 0.92
🎙
UW Brief — "Claire" profile
Female, 35–45 range, American English, direct and precise. Closest reference: NPR business correspondent. Stability: 0.70. Clarity: 0.88. Style: 0.25.
Underwriting eleven_v3 speed 0.97
🎙
Claims Brief — "Jordan" profile
Male or female, 30–40 range, American English, operational and clear. Closest reference: NOAA Weather Radio anchor. Stability: 0.80. Clarity: 0.90. Style: 0.15.
Claims eleven_v3 speed 1.02
🎙
Management Brief — "Alex" profile
Gender-neutral, 40–50 range, American English, calm summary register. Closest reference: Marketplace host cadence. Stability: 0.78. Clarity: 0.85. Style: 0.2.
Management eleven_v3 speed 0.95
Pronunciation dictionary — mandatory at deployment
All insurance CAT domain abbreviations must be pre-loaded into the ElevenLabs Pronunciation Dictionary before VIGIL™ goes live. Without this, "SCS" renders as "sic-s" and "TIV" renders as "tiv." The dictionary is maintained as a versioned file alongside the VIGIL™ system prompt.
SCS→ "S-C-S" (spell each letter) WLF→ "W-L-F" HUR→ "hurricane" (expand, not spell) FLD→ "flood" (expand) WIN→ "winter storm" (expand) TIV→ "T-I-V" FNOL→ "F-N-O-L" AOP→ "A-O-P" PML→ "P-M-L" WUI→ "W-U-I" UW→ "underwriting" (expand in full sentence) CRO→ "C-R-O" NWS→ "N-W-S" CAT→ "catastrophe" (in full phrase "CAT event") NOAA→ "Noah" (standard US pronunciation) pp→ "percentage points" (always expand)
Audio tag encoding in voice scripts
ElevenLabs v3 supports audio tags embedded in the script text. These are stripped before TTS input — the pause markers visible in the script samples above are converted to native ElevenLabs audio tag format by the VIGIL™ TTS adapter layer.
Pause encoding — VIGIL™ script → ElevenLabs format

Script marker: [300ms]
ElevenLabs audio tag: <break time="0.3s"/>

Script marker: [CRITICAL] severity announcement
ElevenLabs audio tag: <emphasis level="strong">critical</emphasis>

Named owner: spoken with slight emphasis
ElevenLabs audio tag: <emphasis level="moderate">Sarah Chen</emphasis>


07
Channel delivery — multimodal contract per channel
Each delivery channel has a defined multimodal contract — what text appears, whether audio is present, and how the user interacts with it. Configuration lives in CAT Sentinel, not in VIGIL™.
Channel Voice / Audio Text format User interaction Severity scope
VIGIL™ Dashboard Full audio player. Auto-play on CRITICAL (configurable). Word-by-word transcript sync during playback. Script visible below player. Full structured written brief. Tables, figures, source citations. Play / pause / seek. Playback speed 0.75–1.5×. Download MP3. Read transcript. Mark action complete. All tiers
Email "Listen to this briefing — 1m 02s" link as first line. Hosted MP3 URL. Opens in browser or mail client audio player. Written brief as HTML email body. PDF attached on CRITICAL/HIGH. Click to listen. PDF download. "Read full brief" CTA links to dashboard. All tiers
Slack / Teams Slack: Audio file preview card (native Slack audio player). Teams: Adaptive card with play button linking to dashboard audio. Severity badge + scroll-stop number + one-sentence situation + action count. Not the full brief. Play in-thread (Slack). Open dashboard (Teams). @mention owner. React with ✓ to acknowledge. CRITICAL + HIGH
Push notification (iOS/Android) Tap notification → VIGIL™ app opens → voice brief begins playing automatically. No tap-to-play step — immediate audio on open. Title: [CRITICAL] SCS Texas. Body: $68B TIV · 3 actions dispatched. No more text. Tap to hear brief immediately. Swipe to see action items. Force-touch for quick complete. CRITICAL always. HIGH configurable. WATCH off by default.
VIGIL™ Mobile App Full audio player. Same as dashboard. Continues playing in background (audio session category: playback). CarPlay / Android Auto compatible. Full brief readable. Action tracker swipe-to-complete. Dark mode. Background audio play. Lock screen controls. Siri / Google Assistant: "Play my VIGIL briefing." All tiers
Webhook / API audio_url field in JSON payload — hosted MP3, 7-day expiry. audio_duration_seconds field. audio_word_count field. Full briefing JSON — written brief + voice script text + structured data. Carrier system can play audio in meeting room, pipe to Teams call, embed in carrier UI. All tiers (configurable by severity in carrier webhook config)
CarPlay and Android Auto — the commute use case

The highest-value voice delivery moment for VIGIL™ is the commute. A CRO or UW lead who receives a CRITICAL push notification at 6:14 AM, gets in their car, and asks CarPlay "Play my VIGIL briefing" should hear the full 52-second CRO voice brief begin immediately — without unlocking the phone, without opening an app, without reading anything. This is the "VIGIL™ keeps the watch so the general can act at dawn" use case made literal. CarPlay audio session configuration and background audio mode must be implemented in the VIGIL™ iOS app from Phase 1.


08
Proactive delivery — configuration in CAT Sentinel
VIGIL™ has no independent delivery configuration UI. Every voice preference is configured in CAT Sentinel and read by VIGIL™ at trigger time. This keeps carrier preferences in one place and prevents configuration drift.
Quiet hours — voice suppression rules

During configured quiet hours (default: 10 PM – 6 AM carrier-local time):

CRITICAL severity: voice brief generated and pushed immediately. Overrides quiet hours always.
HIGH severity: voice brief generated, held until quiet hours end, then pushed with a "held since [time]" annotation in the audio sign-off.
WATCH severity: voice brief generated but not pushed until quiet hours end. No annotation.

The written brief is always delivered immediately regardless of quiet hours. Only voice push delivery is suppressed.


09
Claude API — voice script generation prompt architecture
The voice script system prompt is separate from the written brief system prompt. Both use the same verified event data as input. The voice script prompt enforces all US broadcast rules structurally — not as guidelines but as hard constraints.

10
Reference sources — US broadcast standards applied
SourceWhat VIGIL™ takes from itWhat VIGIL™ does not take
NPR Sound Reporting (Kern, 2008 / Socolovsky, 2025 ed.) One thought per sentence discipline. Write-the-way-you-speak rule. Numbers test: if it requires the listener to do math, remove it. Speak-before-you-type script development process. Conversational warmth and storytelling arcs. NPR's "why this matters to you personally" register is too soft for a professional briefing audience.
Bloomberg Daybreak (Hager / Moskow format) Data-forward opening — state the number first. "That's [X]" construction to humanize a statistic. Direct action implication after the data. Financial precision without hedging language. Market commentary and analyst opinion. VIGIL™ briefs are factual — no commentary on what the data "suggests" or "might mean."
NPR Marketplace (Ryssdal format) The closing context sentence that connects data to decision. "Here's why this matters for your book" beat at position 3 in the five-beat structure. Consumer economics framing. "What this means for your household" register is wrong for an institutional professional audience.
NOAA Weather Radio Operational urgency without panic. Logistics-forward: who, where, when, how many. Consistent structure that listeners learn to expect. Used specifically for the Claims voice register. The full NOAA script formula is too terse for a VIGIL™ brief — it strips context entirely in favor of pure instruction.
WSJ What's News podcast Professional register for financial news. The WSJ "three things you need to know" structure is close to the VIGIL™ five-beat format. The WSJ podcast pace is slower than Bloomberg. VIGIL™ is closer to Bloomberg's speed for its professional audience.

VIGIL™ — Multimodal Voice Design Framework
Version 1.0 · May 2026 · vigil.pionedata.com · For design & engineering review