Model catalog

Event-count likelihoods

For cell i, BARRACUDA observes a non-negative event count n_i over a positive observation time T. The models distinguish sampling variation, a structural non-engaging fraction, and continuous cell-to-cell rate heterogeneity.

Key	Structure	Parameters reported by donor-ignorant fits
homo	One shared Poisson rate	lambda
z2p	Structural zero component plus one shared engaging-cell rate	lambda, p_zero
dis2p	Engaging-cell rates follow a Gamma distribution	mu_lambda, sigma_lambda
hetero3	Gamma rate heterogeneity plus structural zeros	mu_lambda, sigma_lambda, p_zero

p_zero is a model component, not a generic explanation for every observed zero. A zero can still arise from the count likelihood for an engaging cell.

Donor hierarchy

Donor-aware versions use the same four scientific mechanisms but introduce a population distribution over donor-level parameters. They separate:

• population-level location and dispersion;
• between-donor parameter variation; and
• within-donor cell/count variation.

The public donor utilities canonicalize the historical backend name phi_0 to p_zero. Variance decomposition depends on the supplied donor weights. A leave-one-donor-out moment calculation recomputes weighted posterior mixtures; it is not a refit and must not be described as leave-one-out predictive cross-validation.

Ordered contact-kill trajectories

Trajectory models jointly represent contact opportunity and the binary lethal decision. The decision log odds may vary between cells (sigma_eta) and may change with previous non-lethal (beta_f) or lethal (beta_s) contacts.

Key	Heterogeneous baseline decision propensity	History effects
homogeneous_history_independent	No	No
homogeneous_history_dependent	No	beta_f, beta_s
heterogeneous_history_independent	sigma_eta	No
heterogeneous_history_dependent	sigma_eta	beta_f, beta_s

Public notation uses beta_f and beta_s; research backends may store those variables as beta_x and beta_y. Public helpers translate the names.

Nested boundaries

Several simpler mechanisms sit on boundaries of larger models—for example sigma_lambda = 0, p_zero = 0, sigma_eta = 0, or a history coefficient of zero. Recovery near a boundary can be asymmetric and weakly identified. Report boundary-aware summaries and inspect posterior mass rather than relying only on a posterior mean.

Priors are part of the model

Bayes factors compare prior-predictive models, not likelihood labels alone. Changing prior bounds or scales changes the scientific model and its marginal likelihood. Preserve the complete settings object and do not combine evidence from fits that used incompatible model definitions or priors.