Presentation mock v2 · same data, reordered · Jul 10, 2026

Ten professionals.
Three positions.

A language model gives you one answer. A recruited population argues. This is the same expert-networks study, rebuilt so the disagreement — the thing ChatGPT cannot produce — is the first thing a buyer sees.

7 yes as co-pilot — with hard conditions 2 refuse AI as interviewer of record 1 comfortable delegating

Q4 of 7: "Imagine that instead of a human researcher, an AI conducted the expert interview for you…" counts illustrative — wire to response data

Divergence device 01 — the spectrum

Every respondent is a position, not a data point

Ten people placed where they actually stand, reject → embrace. Hover any dot for their words. The scatter is the argument: a model that was "just ChatGPT" would put all ten in the same place.

◄ Rejects the premiseEmbraces with conditions ►
KS
CS
HP
MT
SD
CC
PT
DG
JT
AS
positions illustrative  Placement needs wiring to per-response stance classification — the quotes are real. Dot color = position group from the hero split.
Divergence device 02 — the confrontation

Let the panel argue in its own words

Two real positions from the same study, set against each other. One screen like this kills "it all sounds the same" faster than any methodology paragraph — and it's built from quotes the report already contains.

The adopter case
"Co-pilot" — AI runs scheduling, capture, extraction and contradiction-hunting. The human probes.
Jason Talbot · investment analyst, Cincinnati — with Dale Gomez, Paula Thomas, Anthony Strand and three others alongside
VS
The refusal
An interview is a trust relationship. Guardrails don't change that — she won't delegate it, full stop.
Cleotilde Seymour · investment analyst, rural California — with Kimberly Strait rejecting the premise entirely
Divergence device 03 — the demand matrix

Who demands what before they'll trust it

The panel's adoption conditions as a scannable grid — one row per person, one column per demand. Ten different condition fingerprints is what a real population looks like; a single LLM has exactly one. Column totals are the product roadmap, ranked by the market. cells partly inferred — quotes real, wire to response data

Respondent Verbatim +
timestamps
Raw audio /
attestations
MNPI / PHI
guardrails
Human
sign-off
Procurement-safe
pricing
Kimberly Straitpublic-sector · rural PA
Cleotilde Seymourbuy-side · rural CA
Michelle Thaohealthcare · East LA
Sonia Danielhealthcare · DC
Dale Gomezbuy-side · Fort Worth
Jason Talbotbuy-side · Cincinnati
Paula Thomascorporate ops · Charlotte
Chelsea Chavezcorporate ops · Milwaukee
Herbert Proctorpublic-sector · rural GA
Anthony Strandcorporate ops · Denver
demanded by 9/102/103/1010/104/10
  volunteered this condition   blocker — condition can't be met by features   not surfaced
Divergence device 04 — the contested index

Which questions split the panel

Consensus is a finding; contest is a bigger one. Ranking the seven questions by disagreement gives the reader a map — and tells the client exactly where a follow-up study earns its money. splits illustrative — computable from stance classification

Q7
On-demand synthetic experts & personas — acceptable for scoping and prep; contested at decisions
Q4
AI conducts the expert interview — the co-pilot / pilot line
Q6
AI "expert" trained on historical transcripts — split on trust at the regulated edges
Q5
Reusing historical transcript libraries — welcomed, if recency- and role-filtered with citations
Q3
Value vs friction of expert calls — shared: "scar-tissue" value, heavy friction
Q2
Usage trend over 2–3 years — near-consensus: repriced and repurposed, not cut
Q1
How you get up to speed — consensus: the disciplined, transcript-first playbook

And the split isn't noise — it's structured by sector. Small multiples make that scannable in two seconds: illustrative

Buy-side
3 respondents
Split on delegation. Adopt for transcript mining; Cleotilde holds the human-interview line.
Healthcare finance
2 respondents
Conditional yes. Highest compliance bar on the panel: raw audio, attestations, MNPI/PHI.
Corporate ops
3 respondents
Most adoptable. Pilots and pricing decide; houses the panel's one enthusiast.
Public-sector K-12
2 respondents
Premise contested. Procurement optics block paid expertise before AI is even discussed.
The receipts — who answered

Recruited, not generated

The skeptic's first question gets a polling-house answer, one scroll after the argument. Every number here is read from the study record. live data

Respondents
10 recruited
drawn from a census-grounded synthetic population of 340,000 U.S. residents
Fielded
Jun 19
2026 · 7 questions · 70 individual responses
Panel
6F · 4M
ages 30–55 · median household income ≈ $180K
Sectors
4
buy-side · corporate ops · healthcare finance · public-sector K-12

Panel income vs. U.S. households

Deliberately senior — and here is exactly how the panel differs from the U.S. baseline. Benchmark: Census ACS 2022 (B19001), as on the live page.
Under $50K
0%
35%
$50K – $100K
0%
29%
$100K – $150K
30%
17%
$150K – $200K
30%
9%
$200K+
40%
12%
This panel (n=10) U.S. households

What sits behind every dot

The evidence trail a ChatGPT answer doesn't have.
Each respondent opens into a grounded biography, an OCEAN personality profile, an ingested media diet — this panel reads The Economist, NPR, MarketWatch, the Star Tribune, the Chicago Sun-Times, PJ Media and more — the recent news they actually read, and a voice profile. The roster below leads with professional authority; the lifestyle detail lives one click down, where it stops reading as a tell and starts reading as depth.
Cleotilde SeymourRural CA · 55 · F
Buy-side · Investment analyst
Base case from filings and regulator data before any call; won't delegate interviews end-to-end.
Dale GomezFort Worth, TX · 49 · M
Buy-side · Investment analyst
Reads 8–12 transcripts before booking the first call; one GLG call killed an OEM thesis.
Jason TalbotCincinnati, OH · 53 · M
Buy-side · Investment analyst
Evidence-first; coined "co-pilot" after being burned by a mis-screened "operator."
Sonia DanielWashington, DC · 44 · F
Healthcare · Management analyst
"MNPI/PHI paranoia" — calls only at regulatory or staffing pivots; named the "Tuesday-at-2 p.m. reality."
Anthony StrandDenver, CO · 31 · M
Corporate ops · Financial manager
Most AI-open of the panel — still insists on human sign-off for final interpretation.
Michelle ThaoEast LA, CA · 30 · F
Healthcare · Finance manager
Raw recordings and expert attestations required before acting on any summary.
Chelsea ChavezMilwaukee, WI · 36 · F
Corporate ops · Management analyst
Pilots and operator references decide; surfaced bilingual / code-switching capture needs.
Kimberly StraitRural PA · 38 · F
Public-sector · Management analyst
One-pager framing with kill criteria; "no summary-only deliverables."
Paula ThomasCharlotte, NC · 52 · F
Corporate ops · Management analyst
"No bundles, no annuals" — procurement-safe, predictable pricing or no deal.
Herbert ProctorRural GA · 39 · M
Public-sector · Financial manager
Paid networks untenable on procurement optics and open-records risk; peers and pilots instead.
The receipts — what they said

No uncited sentence

Every summary claim carries its count, its named respondents, and a verbatim quote — with a route to the full responses. The narrative survives as connective tissue between receipts.

4 OF 10 NAMED ◆
Transcript-first is the default workflow; live calls are the last mile
"I read 8–12 transcripts before the first call."
Dale Gomez — investment analyst, Fort Worth
Echoed by Jason Talbot, Paula Thomas, Cleotilde Seymour
→ read the four full responses
3 OF 10 NAMED ◆
Spend was repriced and repurposed — not cut
"No bundles, no annuals."
Paula Thomas — management analyst, Charlotte
Sonia Daniel: calls only at regulatory pivots · Dale Gomez: transcripts up, calls flat-to-down
→ read the three full responses
4 OF 10 NAMED ◆
Calls deliver "scar-tissue" value — and heavy friction
One GLG call killed an entire OEM thesis.
Dale Gomez — the value case in one sentence
Sonia Daniel's "Tuesday-at-2 p.m. reality" · Jason Talbot's mis-screened "operator" · Kimberly Strait on rate bloat
→ read the four full responses
5 OF 10 NAMED ◆
AI is welcome as co-pilot — never as the interviewer of record
"Co-pilot."
Jason Talbot — the panel's own word for the acceptable role
Volunteered conditions: verbatim audio + timestamps, tenant-isolated custody, MNPI guardrails, human sign-off
→ read the five full responses
Fix regardless — on the live shared page today

Four tells that undercut everything above

01An internal test reference is client-visible. The live Recommendations read "For Claude's Ditto API test context, the winning wedge is…" — internal harness note, dead brand name, shareable artifact. This one line out-damages any formatting choice.
02The footer says "Digital Twins of the World's Population." Banned phrase — everything else says synthetic populations. It's on every shared study page.
03Unfinished states ship to the reader. "Summary pending" and "Analyzing correlations… taking longer than usual" are visible on the share link.
04"Across 70 respondents." The correlations overview multiplies 10 people × 7 questions into a respondent count — a skeptic reads it as a fabricated n. Say "10 respondents, 70 responses."

Mock v3 · Jul 10, 2026 · built from the live shared study (fielded Jun 19). Divergence devices: the position spectrum, the head-to-head confrontation, the demand matrix, and the contested index with sector small multiples — plus the ledger and traceable insight cards from v1. All demographics, quotes, segment contrasts and outlier positions are read from the study record; ◆ and "illustrative" mark stance allocations pending wiring to per-question response data. No data output altered — only order, hierarchy, and citation surface.