Built to say no.

1. Why panels say yes

Ask an LLM persona whether it likes your idea and, more often than not, it is kind to you. The Nielsen Norman Group is blunt: synthetic users tend to please the moderator rather than mirror real skepticism, which makes them unreliable precisely for the high-stakes decisions teams most want to de-risk ^[1]. The 2026 synthetic-research meta-review reaches the same conclusion: directional confidence on simple questions, systematic over-acceptance on hard ones [9 · Note 01].

If your simulator always says yes, it is theatre. Holon encodes resistance at four levels of the hazard model.

2. Friction as a parameter

Every role carries a friction coefficient f ≥ 1 per event type. It divides the social pressure that reaches the adoption hazard, so resistance acts at the causal level rather than as a cosmetic multiplier on the final probability:

Factorising f by role, policy and memory keeps the system explainable: every refusal traces back to which factor dominated. A sample of the deployed table:

3. Fear is not dampened

Friction applies only to positive pressure. Negative signal is weighted ν ≈ 2.5η and bypasses the dampener entirely. The constant is not arbitrary:

• Baumeister et al. (2001) ^[3] — meta-analysis: "bad is stronger than good" at ratios 2.0–4.0 across affect, learning, social interaction, persuasion.
• Anderson (1998) ^[4] — negative WOM 2–4× the impact of equivalent positive WOM in customer satisfaction studies.
• Goldenberg, Libai & Muller — empirical NWOM impact on consumer brand evaluation consistent with ν ≈ 2.5×.

Resisting enthusiasm is human; resisting fear is not. That asymmetry is what produces the bad-news cascades that real markets exhibit — and that politely averaged synthetic panels typically miss.

4. Risky decisions need many voices

Simple contagion (one enthusiastic neighbour is enough) overstates B2B adoption. Centola's complex-contagion experiments ^[5] show costly or risky behaviours require reinforcement from multiple distinct sources. Deliberative roles — CFO, Cautious Buyer, Department Head — only count social proof once they have ≥ 2 distinct adopter contacts in their neighbourhood; otherwise their positive pressure is attenuated:

# 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: adoption S-curves rise more slowly and clusters can block cascades — closer to how enterprise software actually spreads (Aral & Walker [8 · Note 01]).

5. Loud exits and silent ones

Not everyone who leaves tells you. TARP's customer-complaint studies ^[6] find that only ~10% of dissatisfied customers complain publicly, while ~70% never complain at all. Holon gives each role a loud-exit probability β:

Your dashboard sees the loud exits. The marketplace feels both. Holon surfaces both in the verdict.

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

— the insight the D/R split is engineered to surface

6. Intervals, never oracles

Because friction, negativity and gating are stochastic, a single run is meaningless. We sweep 500 seeds and report a p10–p90 fan per metric. Calibrated cities (Scale tier) fit (p, q, μ) to a client's historical adoption and churn curves and report a per-account MAPE on a held-out window — never a global accuracy claim.

Built on research published in

Univ. of Chicago Press

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