■ Fetch Synthetic User Panel

Simulated users, validated against a real survey

An LLM panel of personas grounded in the real archetype survey. We ask each persona a question in plain language, map the free-text answer to a 1–5 distribution with Semantic Similarity Rating (SSR), and check whether the synthetic panel recovers the real survey. Then we point the validated panel at brand-new stimuli.

How it works

Two tracks. Track 1 proves the panel recovers a known survey; Track 2 reuses the validated panel on a new offer or feature concept.

01
Persona
Each real survey respondent becomes a ~12-line system prompt. The target answer + known leakage columns are excluded so the model can't cheat.
02
Elicit
Claude Opus answers the question in first person, in the persona's voice. Every response is cached, so re-runs are free.
03
SSR
Free text → cosine similarity vs 5 monotone anchor phrases → min-subtract → linear-normalize → a 1–5 pmf. Averaged over 3 anchor sets.
04
Compare
Average per-respondent pmfs into a panel distribution; compare to the real survey via mean μ and KS similarity = 1 − max|ΔCDF|.

Track 1 — Can it recover the survey?

Nine survey items. The bar we set: |μ_synth − μ_real| ≤ 0.3 and KS ≥ 0.7. Pick an item to see the real vs. synthetic distribution and the personas' own words.


item receipt
real μ—
synthetic μ—
Δ—
KS similarity—
verdict—

All nine items at a glance

#ItemBatteryreal μsynth μΔKSResult

Track 2 — Test something new

Point the validated panel at a fresh concept. Here's a sample offer run — an intent distribution, a per-archetype breakdown, and the personas' verbatim reactions.

Honest notes on method

  • Unweighted panel; the survey over-represents engaged users — a production version would reweight.
  • n=1 sample per respondent; SSR uses 3 anchor sets and local bge-small embeddings (symmetric, 384-dim).
  • Elicitation via Claude Opus on Bedrock. The effort item is inverted (5 = less effort than expected = good).
  • The four numeric dimension scores are never shown to the model; per-item leakage columns are excluded too.
  • Distributions shape-match everywhere (all items KS ≥ —); the near-misses are a mild conservatism at the top of the scale.