Model Limits & Critique

The model page defines the output as "pure cultured chicken cell biomass (wet weight, at harvest)" — factory-gate, before texturisation or blending. The beliefs form asks for "average production cost per edible kg of cultured chicken meat." These are actually the same accounting object (harvested wet cell mass = the edible component before any downstream processing), but the wording difference created potential for semantic confusion.

The model's "growth factors" parameter covers FGF-2, IGF, TGF-β and similar signaling proteins. But three other recombinant proteins used in cell culture media — albumin, transferrin, and insulin — are not yet separately modeled. According to GFI's 2023 analysis, these together are expected to account for the vast majority of recombinant protein production volume in CM, and under optimistic (efficient media) conditions could contribute ~$1/kg. Under pessimistic conditions — high media use, no substitution — the contribution could be $5–50/kg or more, potentially larger than several items the model currently tracks.

Albumin is the most important: it's used at gram-per-liter concentrations, and animal-free alternatives (rice-derived, yeast-derived) are expensive today but improving. Transferrin is lower volume but expensive. Insulin is lowest volume and recombinant insulin is relatively affordable.

The model uses one latent "maturity" variable to correlate technology adoption rates, financing costs (WACC), and reactor choice. This prevents absurd combinations — e.g., technology succeeding while financing remains stuck — but asks one synthetic dimension to stand in for sector maturity, supply-chain development, regulatory learning, and financing all at once. The reviewer notes this can create hidden coupling and double-counting.

Better practice would separate at least three latent factors: technical bioprocess maturity, supply-chain / input maturity, and financing / regulatory maturity — with user-selectable correlation among them. This would also make the dependence structure itself an object of scrutiny rather than a baked-in assumption.

The model's tornado chart shows how much the mean cost changes when each parameter moves from its 10th to its 90th percentile — a conditional-mean dollar-swing statistic. The reviewer notes this is not a Sobol variance decomposition: bars for correlated parameters can overlap and should not be summed. An earlier version of the simplified-view text incorrectly said these parameters "contribute less than 10% of the variance," which was the wrong framing for a swing statistic.

A proper global sensitivity analysis (Sobol first-order and total-effect indices) would answer a fundamentally different question: what fraction of the output variance is explained by each input, accounting for nonlinear and interaction effects? This is methodologically superior but requires dedicated re-simulation infrastructure.

The focal question asks for a median and optional 80% credible interval. Most technical subquestions ask for point estimates. The Sheffield Elicitation Framework and similar structured protocols recommend eliciting full probability distributions for each quantity — using fixed-interval (chips-and-bins) or variable-interval (quartiles) methods — and include calibration training, piloting, and revision rounds.

This form is much simpler: a structured survey without calibration training, not a full elicitation instrument. That's a real limitation for the goal of updating model priors rigorously.

The workshop mixes TEA authors, industry operators, bioprocess researchers, evaluators, and animal-welfare stakeholders. Good-practice elicitation recommends reporting subgroup beliefs separately before attempting any pooled synthesis — subgroup disagreement is information, not noise. If TEA specialists and industry operators systematically disagree, that disagreement is exactly what the workshop is designed to surface.

Performance-weighted aggregation (e.g., Cooke's Classical Model) goes further: using calibration seed questions with known answers to weight experts by accuracy and informativeness. This requires preparing calibration questions in advance — something we did not do for this workshop.

Formal expert elicitation protocols typically include training on cognitive biases in probability estimation — anchoring, overconfidence, availability heuristics — along with practice questions and feedback. This helps experts produce better-calibrated probabilities. We are not doing this at this workshop: participants are not being paid, time is limited, and there is no trained facilitator running a formal protocol.

Some of the workshop participants (e.g., David Manheim) have forecasting/elicitation backgrounds that partially substitute for formal training. But the group overall has not been calibrated.