getminds/marginal-fidelity-survey-eval

Synthetic Survey Evaluation: Marginal Fidelity and Response Contracts

Matching survey averages does not establish that an AI persona simulates an individual. This small, reproducible evaluation release accompanies Alexander Doudkin's arXiv:2609.07305v1, Marginal Fidelity Does Not Establish User Simulation in Demographic Synthetic Survey Panels: Response Contracts, Support Collapse and Conditioning Failure.

Read the practical walkthrough: Do AI Personas Simulate People—or Just Match Survey Averages?

Part of Minds Research on synthetic audiences and evaluation. The models in this paper were called directly through provider APIs. These are not Minds product-performance results.

Main comparison

Method	Mean absolute error, percentage points
Committed persona answers	12.39
Direct population query, no personas	6.27
Persona probabilities, raw	11.44
Persona probabilities, selection cap applied where relevant	5.34

These are descriptive, unweighted means over nine population-aligned model–instrument pairs. The direct query beats committed answers on 9/9 pairs. Constraint-aware probabilities beat the query on 4/9. Lower aggregate error does not establish accurate individual responses or joint behavior. The cap correction is a sensitivity analysis; raw probabilities reverse the benefit on two models on the capped instrument.

Dataset contents and provenance

This is a derived evaluation-results dataset, not a human respondent dataset or a training corpus. Each of its 18 rows is one model–instrument comparison from the public supplement. Nine rows belong to the headline comparison; nine remain explicitly labeled sensitivity comparisons. All original model aliases are retained to avoid overstating exact model routing.

The archive in source/supplementary.tar.zst is byte-identical to the public arXiv supplement. Its SHA-256 is cda972f58a98cde541fe01fccb109c1639b66830598bca3345b4d6f078319bba. See provenance.json for the source URL and individual input hashes. The source files remain unchanged; derived outputs live in data/.

Column	Meaning
battery	Original survey-instrument identifier
model_alias	Model label used in the published analysis; not a guaranteed resolved version
headline_pair	Whether the pair belongs to the population-aligned headline subset
exclusion_reason	Reason a pair is sensitivity-only; empty for headline pairs
committed_mae_pp	Error from realized option selections
population_query_mae_pp	Error from direct population-prevalence queries without personas
probability_raw_mae_pp	Error from raw persona inclusion probabilities
probability_constraint_aware_mae_pp	Probability error after applying the QA3 selection cap; unchanged elsewhere
cap_applied	Whether the instrument's maximum-two constraint was applied
valid_population_query_repetitions	Number of valid repeated population queries contributing to the estimate

MAE averages absolute gaps from human option prevalence within each instrument. Units are percentage points throughout. The evaluation split is for analysis, not held-out training evaluation: these published results should not become a model-selection target for claims of fresh validation.

Reproduction

Download this repository and run:

python reproduce.py

This verifies the source JSON hashes, rebuilds the CSV, and checks the headline means against the unchanged public results. It uses the standard library and requires no credentials, network, or model calls. reproduce.ipynb provides an executed notebook; Matplotlib is optional for its chart cell.

To audit the larger paper, unpack the original supplement:

mkdir original-supplement
tar --zstd -xf source/supplementary.tar.zst -C original-supplement
cd original-supplement
python scripts/verify_numbers.py
python scripts/audit_numbers.py

The original audits require NumPy. Follow the supplement README for extended analyses and separately acquired microdata. The compact reproduction here recomputes summaries from released intermediate results; it does not rerun the raw-response analysis or recollect any model outputs.

Scope, exclusions, and limitations

• Headline instruments: CES 2024 CC24_300, CES 2024 CC24_305, and German Eurobarometer QA3. Other batteries are sensitivity-only because of endogenous eligibility/empty-array ambiguity or population-frame mismatch.
• The paper studies static four-attribute demographic prompts. It does not establish the behavior of richer interview-grounded or interactive simulators.
• Persona rows can repeat the same prompt. Counts of synthetic responses are not counts of independent demographic specifications.
• Continuous positive probabilities avoid zero coverage by construction; this does not demonstrate behavioral support or correct combinations of choices.
• Human outcomes are realized multiselect choices; probability outputs are stated latent propensities. Their error comparison is partly a measurement comparison.
• Public-target contamination cannot be ruled out. Held-out demographic-cell targets come from public microdata; no claim of pristine pretraining exclusion is made.
• Raw provider envelopes and exact resolved model versions are not available for every call. Consult the original provenance notes before attempting a new collection.
• Subgroup anchoring and cap projection remain exploratory/sensitivity analyses. This release is not evidence of universal model or product superiority.

No new participants were recruited for this release. The table contains aggregate evaluation scores, not human free text or individual human records. The original supplement's privacy and third-party data notices remain applicable.

License and citation

The article and derived evaluation table are released under CC BY 4.0, consistent with the public paper. New reproduction code is MIT licensed; see LICENSE-CODE. The unmodified source supplement preserves its upstream notices, and source survey material remains subject to the respective providers' terms. The Hub license tag does not relicense third-party material.

@misc{doudkin2026marginalfidelity,
  title={Marginal Fidelity Does Not Establish User Simulation in Demographic Synthetic Survey Panels: Response Contracts, Support Collapse and Conditioning Failure},
  author={Alexander Doudkin},
  year={2026},
  eprint={2609.07305},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2609.07305}
}

This is a companion to an arXiv preprint. No conference acceptance or completed peer-review status is claimed.