The same toy model in a Blender-modelled room: the rapport beam between the chairs brightens with trust, the client leans in, and the tower fills with readiness. Drag to orbit, pinch or wheel to zoom.
The science
What this is. A two-variable toy model of the dynamics of one conversation about change. It is inspired by the vocabulary of motivational interviewing (Miller and Rollnick, Motivational Interviewing, 3rd ed. 2013): open questions, affirmations, reflections and summaries ("OARS"), the "righting reflex" (pushing advice makes people argue the other side) and the idea that change talk grows when the client feels understood. It does not model cognitive behavioural therapy, solution-focused therapy or any other named technique, and it is not a validated model of any real person.
Equations. Time in minutes. With move shares pi = xi / Σx (style, not volume):
dW/dt = kW (E - W) with E = clamp(0.15 + 0.8 Σ wi pi), weights reflection 1.0, affirmation 0.85, open question 0.7, summary 0.5, unsolicited advice -0.6; kW = 0.12 /min.
dR/dt = kR [ u (1 - R) - d R ], u = 0.6 W (0.9 pf + 0.8 po + 0.5 ps + 0.45 pa), d = pg (1.3 - 0.9 W) + 0.12; kR = 0.08 /min. The constant 0.12 is a built-in pull of ambivalence; advice hurts more when rapport is low.
Turns. Every half minute a counsellor move is drawn from the mix shares and the client replies with change talk with probability 0.9 R, sustain talk with 0.9 (1-R) or neutral 10 %. This is a random illustration of the averages, not a transcript.
Checks. (1) With constant input E, RK4 equals the closed form W(t) = E + (W0 - E) e^(-kW t). (2) With W frozen, R equals R* + (R0 - R*) e^(-kR(u+d)t), R* = u/(u+d). (3) 40,000 sampled moves reproduce the mix shares within a few standard errors.
Limits. All coefficients are illustrative choices, not measurements. Real conversations depend on culture, context, timing and the individual. This tool cannot assess anybody and gives no advice.
Models and 3D assets: the room, chairs, pawns, readiness tower and wall clock were modelled and texture-baked in Blender (named parts such as Rapport_Beam, Readiness_Liquid, Pawn_Client) and are driven by the equations above in Three.js r160.