CM Workshop · Beliefs Analysis

Technical sub-questions and expert distribution questions. Only respondents who answered each question are shown.

Process mode (E5) — % of production by mode in 2036

Only Elliot Swartz provided E5 values (fed-batch 20%, perfusion 50%, continuous 30%).

Discussion responses (open text)

Submission timeline

When did responses arrive, relative to the workshop on May 8, 2026?

Individual responses

All 15 non-test submissions. Sorted by CM_01 estimate (low to high; no-estimate at end). Click to expand.

Methods, notes and annotations

Statistical approach

Log-normal distribution: Each respondent's estimate is modelled as log-normal — the standard choice for cost data, which is bounded at zero and typically right-skewed. Given median m and 80% CI [lo, hi], we set μ = ln(m) and σ = (ln(hi) − ln(lo)) / 2.564 (since the 80% interval spans ±1.282σ on the log scale). Respondents without a CI are assigned a user-selectable default σ (ranging from 0.3 = narrow to 1.8 = very wide).

Three aggregation methods are available:

• Linear mixture (equal CDF weight): F_agg(x) = (1/n)·ΣFᵢ(x). Not log-normal — preserves multimodality. Each respondent contributes equal weight at every CDF percentile, but a respondent with a wider CI (more diffuse) adds less density at any specific point, so they shift the mode of the aggregate less than a confident respondent. Conceptually: "I give equal credence to each respondent's full distribution."
• Geometric pool (log opinion pool): log f_agg = (1/n)·Σlog fᵢ + C. Under log-normality collapses to a single log-normal: location = precision-weighted mean of log-medians, width = √(n/Σ(1/σᵢ²)). Externally Bayesian if experts are independent. Confident respondents pull the location more; diffuse ones widen the result.
• Bayesian posterior: Treats each respondent as an independent noisy measurement of the shared true θ. Location = precision-weighted mean (same as geometric), but width = 1/√Σ(1/σᵢ²) — dramatically tighter, since n independent signals combine multiplicatively. Appropriate if respondents truly have independent information; likely over-confident if their knowledge overlaps.

Note: all three methods agree on the location of the aggregate (precision-weighted mean of log-medians) when using precision weights. They differ primarily in the width of the aggregate: linear (widest, not log-normal) → geometric (intermediate) → Bayesian (narrowest).

Geometric mean of point estimates: exp(mean(ln(mᵢ))). Appropriate for log-symmetric distributions; less sensitive to the highest estimates than the arithmetic mean. Shown alongside the precision-weighted geometric mean (where respondents with tight CIs get more weight).

Exclusions: Three submissions have all slider-controlled probability fields at exactly 50%. The beliefs form uses a touchedSliders mechanism (added in beliefs.js) that prevents untouched sliders from being submitted — but these three submissions predate that fix and were submitted via an earlier form version. They represent probable no-interaction responses, not deliberate 50% choices, and are excluded by default. One submission (TEST_submission) is excluded entirely. The "FN" pseudonym submission ($1/kg, no CI) appears genuine based on substantive CM_10 reasoning and "other thoughts" content.

Respondent categories

Categories are assigned based on name + affiliation. "Researcher" includes both academic researchers and nonprofit analysts (GFI). "Industry" includes company practitioners. Several anonymous respondents cannot be categorised and are labelled "unknown."

The two anonymous May 8 all-50% submissions are the only ones that appear to be "default" (sliders not moved). The May 4 anonymous submission has a consumer-perspective reasoning ("current chicken price in my country") and a very wide CI — it is flagged as non-expert and can be excluded.

2027 near-term question

The beliefs form did include a supplementary CM_01 question for December 31, 2027 (fields: cm01_2027_median, cm01_2027_ci_lower, cm01_2027_ci_upper, cm01_2027_reasoning). However, zero respondents filled in any 2027 values — the question appears after a long scroll and was skipped by everyone. The 2027 section in the CM_01 tab notes this and suggests prompting respondents explicitly in the follow-up round. The dashboard is structured to support a 2027 CI-bar chart once data exists.

Update analysis (pre/post)

Two post-workshop responses have arrived (Stout, PersonABC). No participant has submitted twice, so pre/post updating within a person cannot yet be measured. When paired responses exist, the dashboard will compute and display the shift in μ (log-scale location) and σ (uncertainty) for each respondent.

Hypothes.is annotations

Public annotations: Install the Hypothes.is browser extension and annotate freely. Public annotations on this URL are monitored by the workshop Slack bot (checks every 2 hours on the Linode server at 45.79.160.157 via hypothesis_to_slack.py).

Private group annotations:

1. Go to hypothes.is/groups/new and create a group (e.g. "CM Workshop Internal")
2. Copy the join link and share with collaborators (Zhuoran, Tommy, etc.)
3. In the Hypothes.is sidebar, switch from "Public" to your group name before annotating
4. To monitor this group's annotations automatically, update hypothesis_to_slack.py on Linode with the group's API tag

Data completeness