Not another panel. A market simulator.

1. The problem with synthetic panels

Most synthetic-user tools prompt a language model to roleplay a persona, then ask it questions one at a time. It is a chat window wearing a costume — fine for a first gut-check, useless for the thing that actually decides a launch: how an opinion moves through a room. Adoption is not an average of isolated answers. It is a process on a network.

The 2026 buyer guides for synthetic research are explicit on the gap: the category is weakest precisely where it is most often sold — modelling social contagion, negative word-of-mouth, and the cascades that determine whether a feature lands or a price change rebounds ^[9]. Holon was built against that gap.

2. The four-state model

Each of n agents occupies one of four states at discrete time t:

• S — Susceptible. Has not adopted. Receives both positive and negative pressure.
• I — Adopter (Infectious). Has adopted. Emits a persuasion signal η to its out-neighbours.
• R — Recovered. Has churned silently. Emits nothing.
• D — Detractor. Has churned loudly. Emits a negative signal weighted ν ≈ 2.5·η.

Transitions S→I, I→R, I→D, and D→R are stochastic and depend on the agent's role, budget, and the pressure arriving from its in-neighbours in a directed weighted graph G = (V, W). The split between R and D is governed by a role-specific loud-exit probability β (Power Users churn loudly; Lurkers vanish silent — §3 of Note 02 ^[11]).

3. Adoption hazard

We use a continuous-time hazard formulation discretised at Δt = 1 day. The probability that susceptible agent i adopts during a tick composes an exogenous innovation term p (advertising, discovery) with endogenous imitation q·φ⁺, divided by a role-based friction coefficient f, and dampened by negative pressure φ⁻:

The exponential form keeps λ in [0,1] for any pressure, avoids double-counting when multiple neighbours push simultaneously, and matches the standard continuous-time hazard convention used in epidemic and diffusion models ^[2]. The novelty is f (friction, see Note 02 ^[11]) and the explicit negative term r · φ⁻ that lets aversion suppress adoption even when social proof is high.

Prior values for (p, q) are taken from the Sultan-Farley-Lehmann meta-analysis ^[2], rescaled from annual category penetration to community-response hazards. Calibrated cities (Scale tier) fit per-agent (p, q, μ) to client historical curves and report a per-account MAPE on a held-out window.

4. Social pressure on the graph

Pressure is computed from the graph, not from a prompt. Let W ∈ ℝⁿˣⁿ₊ be the column-normalised weighted adjacency matrix where Wⱼᵢ > 0 means j influences i. Active adopters emit a signal u(t) = η ⊙ i(t); detractors emit a negative signal u⁻(t) = ν ⊙ d(t) with ν ≈ 2.5 η. The pressure received by every node is a single matrix product:

Because each step is a sparse matrix-vector multiply, the whole city updates simultaneously. There is no per-agent loop, no per-agent API call. a is the receiver's affinity; ⊙ denotes element-wise product. Cluster-level ("tribal") pressure can be added with a membership matrix C ∈ {0,1}ⁿˣᴷ to model community contagion ^[4].

5. Contagious churn

Naïve models treat churn as an independent coin flip. Holon makes it social: an adopter's churn hazard rises with personal aversion av and with the negative pressure from neighbours who already defected — but only if the agent is itself exposed (Nitzan & Libai's neighbour-defection effect ^[3]):

The (0.15 + 1.3 av) factor captures Nitzan & Libai's finding that the social-churn effect is near-zero for satisfied customers. When churn fires, a role-specific β decides whether the agent exits quietly (R) or loudly (D). Detractors decay back to quiet at rate δ ≈ 0.07/tick — a ~14-day half-life calibrated on online-firestorm lifecycle studies ^[5].

6. Why it's fast

Every quantity above is a vector or a sparse matrix. A full 90-tick run on 1,000 agents is ~90 sparse matrix-vector products — milliseconds on a single core. Below is the inner loop in NumPy, the same code that powers the reference engine; the TypeScript port in /engine is a direct transliteration.

def step(state, W, p, q, r, mu, f, eta, nu, a, dt=1.0, rng=None):
    rng = rng or np.random.default_rng()
    I  = (state == 1).astype(float)
    D  = (state == 3).astype(float)
    u_pos =  eta * I
    u_neg =  nu  * D
    phi_pos = a * (W.T @ u_pos)          # positive social pressure
    phi_neg = a * (W.T @ u_neg)          # negative pressure (negativity bias)
    # adoption hazard (Bass-style with friction and aversion)
    lam = 1 - np.exp(-dt * np.maximum(0, p + q*phi_pos/f - r*phi_neg))
    # churn hazard, gated by personal aversion (Nitzan & Libai)
    gam = 1 - np.exp(-dt * (mu + 0.04*aversion + 0.14*phi_neg*(0.15+1.3*aversion)))
    return advance(state, lam, gam, beta_role, delta=0.07, rng=rng)

Zero LLM calls per tick means: no rate limits, no latency, no per-run bill. You can sweep 500 random seeds and read a distribution rather than a single hopeful point.

7. Reproducibility

A run is a pure function of (scenario, seed). The engine uses a seeded Mulberry32 PRNG; Date.now() and Math.random() are forbidden inside it (linted at CI). We hash the final state, so any shared URL replays byte-for-byte — the precondition for publishing backtests that other people can independently verify.

Below is the measured battery — every row is the live output of the harness in scripts/engine-bench.ts, not a target we hope to hit.

8. What the model misses

High-stakes decisions dominated by emotion or identity (politics, sensitive medical, security). Novel categories with no historical precedent. Rare-event tail risk. We log these limits per backtest in a mandatory "what the model misses" section. The model is a pre-filter, not an oracle — designed to kill bad ideas quickly so research budget can be spent on the survivors.

Built on research published in

Univ. of Chicago Press

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