Twelve citizens. Built to say no.

1. Why a cast, not personas

The dominant way to build a synthetic respondent is to hand a language model a personality — often an OCEAN trait vector — tell it to roleplay, and ask it questions one at a time. The failure mode is documented and consistent: the model is kind to you. The Nielsen Norman Group is blunt that synthetic users please the moderator rather than mirror real skepticism ^[1]; the 2026 peer-reviewed numbers are unforgiving — a 79% surface match that over-generalises and under-varies ^[2], a correlation of just r ≈ 0.20 with real people across nineteen pre-registered studies ^[3], and an effect whose sign flips in 32% of runs ^[4].

Holon takes the opposite route. A citizen is not a personality you converse with; it is a published stat block — a single point in an eight-axis parameter space. You cannot flatter a hazard equation. A Skeptical CFO refuses a price increase not because we prompted it to "be skeptical" but because its price-friction coefficient divides incoming social pressure by four before that pressure ever reaches the adoption hazard. Resistance is structural, not theatrical — and because every parameter is public, every refusal is inspectable.

2. The twelve

Each citizen occupies a distinct, necessary position in the parameter space. Before the cast shipped it had to pass two tests. Redundancy: no two citizens are duplicates on two or more axes — if a role collapsed onto another, it was cut, and twelve survived. Coverage: every stimulus type (a feature, a price change, a story) has both its natural igniters and its natural brakes, so no event can glide through unopposed and none is dead on arrival.

The cast reads in three behavioural bands — the Movers who lift the curve, the Gates who stop a cascade, and the Quiet weight that decides off-screen:

None of this is hidden. The parameters are public on purpose — that is the whole point of an instrument rather than a black box. A closer look at five of the underlying stat blocks, where the design choices are sharpest:

3. Why this number

Not five (Rogers)

Rogers' innovators / early adopters / majority / laggards ^[5] are temporal quantiles — they sort people by when they adopt over a category's lifetime, not by the role they play in a single room. Useful for a product's decade; useless for one buying decision, where what matters is why each seat resists. Our innovation-leaning citizens (Early Adopter, Visionary, Champion, Influencer) sum to 23% of the city — broadly consistent with Rogers' ~16% innovators-plus-early-adopters, but split by motivation instead of collapsed into a single bucket. A simulator needs to tell a CFO's "no" from a Burned-out PM's "no" — Rogers cannot.

Not forty-five (the 90-9-1 rule)

An adversarial review pushed for ~45% Lurkers, citing the 90-9-1 participation-inequality rule ^[6]. We rejected it as a category error: 90-9-1 describes who posts in an online community, not who sits in a market. A private B2B buying conversation is far flatter than a public forum. Our Lurker is 17% — the silent-but-present majority of an active decision ecosystem, not a census of everyone who could ever buy.

And three we are holding back

We designed fifteen and ship twelve. The three on the bench — a Compliance Officer, a Procurement lead, an IT Admin — are gatekeepers that only bind in formal enterprise deals. Industry research puts the typical B2B buying group at 6–10 stakeholders, rising to 9–11 in 2025 from 5–7 in 2017 ^[7] — but most self-serve and mid-market motions never convene that committee. The +3 arrive with the Enterprise template at the Scale tier; until then, the Skeptical CFO carries the security-review veto.

4. Built to say no

Every citizen carries a friction coefficient f ≥ 1 per event type. It divides social pressure inside the adoption hazard itself, so resistance is a causal property of the agent rather than a cosmetic haircut on the final probability:

Friction touches only positive pressure. Negative signal is weighted ν ≈ 2.5η and bypasses the dampener entirely — Baumeister's "bad is stronger than good" puts the negativity ratio at 2–4× across affect, learning and persuasion ^[14], and Anderson found dissatisfied customers carry their experience to more people than satisfied ones do ^[15].

Resisting enthusiasm is human; resisting fear is not. That single asymmetry is what lets the cast produce the bad-news cascades real markets show — and that politely averaged panels miss.

5. Risky decisions need many voices

One enthusiastic neighbour is enough to forward a meme. It is not enough to make a CFO sign. Centola's complex-contagion experiments ^[8] — and the book that consolidates them ^[9] — show that costly or risky behaviours need social affirmation from multiple distinct sources. Holon's deliberative roles (CFO, Cautious Buyer, Department Head) only count social proof once they have ≥ 2 distinct adopter contacts; below that threshold their positive pressure is attenuated to 15%:

# distinct-source gate (vectorised approximation)
n_distinct = (W_in > 0).multiply(I).getnnz(axis=1)   # per-receiver count
under_gate = n_distinct < k_role                     # boolean mask
phi_pos[under_gate] *= 0.15                          # attenuate

The measurable effect: the adoption S-curve rises more slowly (24→29 ticks in our city) and clusters can block a cascade — closer to how enterprise software actually clears a room. The buying data agrees: 40–60% of B2B deals end not in a loss to a competitor but in no decision, and ~41% stall because internal stakeholders fail to align ^[10]; legal, compliance and IT act as silent vetoes a lone champion cannot override. A simulator that lets one excited voice carry the deal is telling you a comfortable lie.

6. Loud exits and silent ones

Not everyone who leaves tells you. TARP's customer-complaint studies ^[11] find only ~10% of dissatisfied customers ever complain publicly while ~70% never say a word. Each role gets a loud-exit probability β deciding whether it churns into a loud Detractor (D) or a silent exit (R):

“Your dashboard reports 11% churn. The city shows another 20% of silent disengagement the dashboard can't see.”

— the insight the loud/silent (D/R) split is engineered to surface

7. Network positions, and why hubs aren't enough

Three citizens earn special positions on the graph: the Influencer is a hub (out-degree ×10), the Department Head carries a team tree (10 edges weighted ×4 — the top-down motion), and the Champion is an amplified leaf (persuasion triples once it adopts — the bottom-up motion). Reach, structure and amplification, kept distinct on purpose.

But reach is not influence — and here we deliberately encode a finding that cuts against our own Influencer. Watts & Dodds showed that large cascades are driven not by a handful of influentials but by a critical mass of easily-influenced ordinary people ^[12]. So Holon does not treat the Influencer as a magic spreader. Its ×10 reach only converts if its neighbours are themselves susceptible and the complex-contagion threshold (§5) is met; fire a hub into an unreceptive, single-source neighbourhood and the cascade fizzles — exactly as Watts predicts, and consistent with the influence-vs-susceptibility split Aral & Walker measured on a real network ^[16].

8. Calibratable, not clairvoyant

We say "archetypes calibrated to data," never "digital twins" — the distinction is the product's spine. Priors for (p, q, μ) come from the Sultan-Farley-Lehmann diffusion meta-analysis ^[13], rescaled from annual category penetration to community-response hazards. At the Scale tier we fit those parameters to a client's own historical adoption and churn curves and report a per-account MAPE on a held-out window — never a single global accuracy number.

What the cast cannot see: decisions dominated by identity or emotion (politics, sensitive medical), genuinely novel categories with no behavioural precedent, and rare-event tail risk. Every backtest ships with a mandatory "what the model misses" section. The cast is a pre-filter for bad ideas — designed to kill the weak ones cheaply — not an oracle for good ones.

Grounded in research published in

Univ. of Chicago Press

Princeton University Press

Methodology builds on peer-reviewed research from the venues above. See references for exact papers.