Reproducible runs

To compare two runs, keep the population, design, respondent instructions, model, analysis, and artifact lineage explicit. A fixed seed controls a particular draw. It does not establish that the whole pipeline or an external model provider will return identical bytes.

The current request schema already defaults llm_temperature to 0.0, resample_population to false, and population_seed to 100. Design and latent-variable generation need additional controls for a planned comparison.

Field	Comparable value	Default	What it pins
llm_temperature	0.0	0.0	respondent sampling
resample_population + population_seed	false + a fixed seed	false, 100	which respondents answer
design_seed	any fixed integer	null (fresh design)	the choice tasks and their order
deterministic_stages	true	false	the pre-survey stages (latent variables, persona bios, mood statements) run at llm_temperature instead of a fixed 0.7
latent_variables_from_run	the first run's name	null	reuses the first run's latent-variable battery and scores, so respondent prompts repeat

Apply the settings to a complete request

Merge this settings fragment into the study request from Run an experiment, after reviewing its population, attributes, levels, privacy, and cost:

{
  "llm_temperature": 0.0,
  "resample_population": false,
  "population_seed": 100,
  "design_seed": 42,
  "deterministic_stages": true
}

Retain the returned wandb_run_name with the request and artifacts. For a later run of the same definition, set latent_variables_from_run to that actual run name. This reuses the latent-variable bundle instead of generating a new one. It requires the same population; inspect the reused artifact and its respondent mapping rather than assuming that a matching count proves matching identities.

Measure the variation that remains

Keep the model and provider version, complete settings, realized design, population artifact, latent-variable bundle, and analysis code with each run. Compare the same estimand on the same scale and report the repeat-run variation for your own study. Do not apply a correlation from another benchmark as a universal noise threshold.

Temperature zero does not guarantee identical model responses. A repeated synthetic result is also not validation against human behavior. Use the research validity checklist when deciding what the comparison supports.

What the controls do not replace

Seeds preserve repeatability of specific random draws when the surrounding inputs and implementation remain fixed. They do not remove randomization within a design, freeze external services, repair a population mismatch, or make two different study definitions comparable.