PaperlayerView exact revision on arXiv

Source-linked AI summary

Marginal Fidelity Does Not Establish User Simulation in Demographic Synthetic Survey Panels: Response Contracts, Support Collapse and Conditioning Failure

Alexander Doudkin

arXiv:2609.07305v1cs.CLcs.HC

TL;DR

Aggregate agreement with published survey marginals does not establish individual simulation, and the measured fidelity depends on the response contract. The paper compares committed-set panels, per-option probability vectors, and a no-persona population query; the query can outperform committed panels on the aligned marginal endpoint.

  • Problem

    Whether matching published survey marginals validates demographic synthetic respondents remains an open question, because the same persona can return a set or a probability vector under different response contracts.

    Source: §1 · p. 1, §6 · p. 6

  • Method

    The paper compares committed sets, per-option probability elicitation, and a no-persona population-prevalence query on aligned demographic multiselect instruments.

    Source: §1 · pp. 1–2

  • Results

    A no-persona query averaged 6.27 MAE versus 12.39 for committed panels and won 9 of 9 aligned pairs; constraint-aware vectors averaged 5.34 and beat the query on 4 of 9.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Population-marginal agreement is evidence about an elicitation contract and an estimand obtainable without simulated respondents, not evidence of individual simulation.

    Source: §8 · pp. 7–8

    The readouts are separate calls and compare realised human check-all answers with latent probability propensities, so better marginal recovery need not mean more faithful respondent-level behaviour.

    Source: §9 · p. 8

Abstract

from arXiv · show

Demographic synthetic survey panels are often validated by matching aggregate answers to published surveys. We test what that certificate establishes across six multiselect batteries from four survey organisations in three countries. The headline analysis is restricted to three instruments whose synthetic cohort and human target share the stated population frame; three other batteries remain sensitivity analyses. The response contract dominates measured fidelity. In the aligned instruments, committed sets leave 66 of 128 model-battery option slots empty in panels of up to 500 respondents, versus 0 of 128 under per-option probability elicitation. Across eight uncapped model-instrument comparisons, probabilities reduce option-marginal MAE by 4.53 to 7.30 points. The capped instrument reverses on two models until the vectors are projected onto its stated maximum. These are measurement effects: human targets are realised check-all responses, whereas the vectors are latent inclusion propensities. Published marginal agreement also fails to discriminate respondent simulation from direct population estimation. On nine aligned model-battery pairs, a no-persona population-prevalence query averages 6.27 MAE versus 12.39 for committed panels and wins all nine comparisons. Constraint-aware probability vectors average 5.34 and beat the query on four of nine, so the baseline challenges the validation criterion rather than proving direct estimation uniformly best. On three unpublished demographic cells, neither approach beats reciting the national distribution. Population-marginal agreement is therefore evidence about an elicitation contract and an estimand obtainable without simulated respondents, not evidence of individual simulation.

1. Introduction

This paper examines whether aggregate agreement validates demographic synthetic survey respondents, showing that fidelity depends on how answers are elicited and that a population query can outperform committed panels. It frames marginal agreement as evidence about a response contract, not proof of individual simulation.

1. Introduction

  • A no-persona population-prevalence query outperforms committed panels on the published-marginal endpoint, so aggregate agreement cannot by itself validate individual simulation.The paper treats this as a limitation of the validation criterion, not proof that direct estimation is uniformly best.Source: §1 · p. 2
  • In aligned instruments, committed sets leave 66 of 128 model-battery option slots empty, versus zero under probabilities; the vectors are latent propensities, unlike realised human check-all answers.The authors call the observed zeros support collapse and describe it as an empirical, sample-size-dependent diagnostic, not proof of incapacity.Source: §1 · p. 2
  • The contribution combines a multiselect support-collapse diagnostic, a no-persona baseline for testing marginal validation, and an exploratory decomposition of subgroup error into ordering and level components.The paper also includes a descriptive sparseanchor sensitivity analysis.Source: §1 · p. 2

1.1 Relation to prior work

  • Prior categorical work contrasts distribution description with sampling; multiselect instead involves inclusion marginals, set size, and co-selection rather than a single simplex [Meister, Guestrin, and Hashimoto (2025); Zhang et al. (2025)].SATA-Bench documents count biases on multiselect questions with known answers [Xu et al. (2025)].Source: §1.1 · p. 2
  • Forced-choice formats elicit more endorsements than check-all, and stated probabilities can diverge from realised choices, making readout a measurement decision [Smyth et al. (2006); Yamin et al. (2026)].The probability readout is described as the graded limit of forced-choice.Source: §1.1 · p. 3

2. Readouts, estimands, and what coverage adds

The section distinguishes population marginals from respondent-level set structure and treats coverage as a prerequisite to calibration. Its coverage diagnostic is computable from simulator output, but probability vectors’ positive values do not establish joint behavioural support.

2. Readouts, estimands, and what coverage adds

  • A set induces distinct population estimands, including inclusion marginals, cardinality, pairwise dependence, constraint compliance, and the distribution over sets; better marginals do not establish better sets.Option-marginal MAE says nothing about which options co-occur or how many an individual selects.Source: §2 · p. 3

2.1 Coverage as a prior question

  • Coverage asks whether any respondent expresses an option before assessing its calibration; zero coverage and zero between-respondent variance are computable from simulator output alone.The human cost of zero coverage is summed human prevalence across those options, not a population share.Source: §2.1 · p. 3
  • Zero coverage is empirical and sample-size-dependent, while a continuous readout avoids it by staying positive; that indicates marginal expressivity, not joint behavioural support.Coverage is necessary, not sufficient, and a persona with positive option probabilities may still realize the same set every time.Source: §2.1 · p. 3
  • The comparison uses direct multiselect sets against raw per-option probability forecasts, averaged without projection or calibration; the paper does not identify a psychological mechanism for the contrast.The psychological motivation is given separately and is not an identified mechanism.Source: §2.1 · p. 3

3. Design

The design compares two separate readouts for four-attribute synthetic cohorts across six multiselect batteries. Its headline analysis is limited to instruments whose synthetic cohort and human target share the stated population frame.

3. Design

  • Each cohort contains 500 demographic-only personas, and the retained allocation audits show one-way marginals within 0.1 points; the first-500 convenience slices deviate from audit targets by up to 5.4 points.Because source-safe relationships were unavailable, the generator used seeded independent assignment rather than official four-way joint cells.Source: §3 · p. 3
  • Six batteries span four survey organisations in three countries, but only CC24_300, CC24_305, and German QA3 enter the headline subset because the other three have frame or eligibility mismatches.The three excluded batteries remain sensitivity analyses; QA3 has a maximum of two selections.Source: §3 · pp. 3–4
  • Each persona receives independent calls for a JSON check-all set and a 0–100 forecast per option, using their conventional temperatures; capped-instrument answers exceeding the limit are rejected.Invalid outputs are re-asked at most twice, and probability forecasts are averaged without projection, normalization, calibration, or repair.Source: §3 · p. 4
  • Scoring uses option-marginal MAE against released shares, with paired bootstrap intervals resampling distinct prompts rather than respondents.The intervals condition on one cohort, prompt set, and collection window; distinct prompts are resampled with replacement for both readouts.Source: §3 · p. 4

4. Support collapse under committed-set elicitation

Committed-set panels exhibit many observed zero-coverage options in the aligned headline subset, including options with substantial human prevalence. This is a sample-dependent diagnostic, and probability readouts’ zero holes do not establish joint behavioural support.

4. Support collapse under committed-set elicitation

  • 66 of 128 aligned model-battery option slots draw zero selections under committed sets, versus 0 of 128 under probabilities, which stay positive by construction.The zeros are descriptive observations, not hypothesis-test rejections.Source: §4 · p. 4
  • Zero-coverage options include emergency-room visits at 22.1 percent human prevalence, German immigration at 23.0 percent, and life events with human prevalence from 2.5 to 3.9 percent.One German run leaves 108.0 points of summed human prevalence unexpressed; repeated demographic prompts are dependent.Source: §4 · p. 4
  • The statistic is empirical and sample-size-dependent, and positive probability vectors show marginal expressivity rather than joint behavioural support.Coverage is necessary, not sufficient; on one battery the probability readout has substantially worse error despite perfect coverage.Source: §4 · p. 5

5. The readout effect and its boundary

Across aligned uncapped instruments, probability elicitation improves marginal agreement over committed sets, while the capped instrument reverses on two models until its vectors are projected onto the stated maximum. The evidence is bounded by sample-dependent diagnostics and does not establish population-level effects.

5. The readout effect and its boundary

  • Twenty battery-model runs completed, but intervals are conditional on retained cohorts and prompts rather than uncertainty over populations, models, or instruments.The headline subset contains eleven model-battery runs.Source: §5 · p. 5
  • Across five models on CC24_305, probability readout effects run +5.18 to +7.30, with every clustered interval excluding zero and the effect surviving a matched readout-by-temperature comparison.The battery ran across three providers.Source: §5 · p. 5
  • All eight uncapped comparisons favour probabilities by 4.53 to 7.30 points, while projection changes the two capped-instrument reversals to +12.76 and +9.79.The post hoc projection suggests unconstrained mass explains the observed reversal but does not isolate a causal cap effect.Source: §5 · p. 5
  • The committed-set binary-sampling floor is 1.9 to 4.3 points, and heuristic zero-count calculations are diagnostics rather than calibrated tests because prompts are dependent and human prevalence is fixed.Recall probes do not establish absence of contamination; retrieved-and-adjusted priors remain possible.Source: §5 · p. 6
  • The defensible pattern is a same-direction effect on two aligned uncapped instruments across eight model-instrument runs, plus a constraint-sensitive boundary on one capped instrument.The authors do not meta-analyse batteries because model spread is not a sampling variance and instruments are not a random sample from a defined population.Source: §5 · p. 6
  • Probability outputs are closer to the human mean set size in 6 of 8 aligned uncapped runs, while raw capped-battery vectors imply 1.2 to 3.9 times the human set size.Marginal MAE and respondent-level structure are different estimands.Source: §5 · p. 6

6. What marginal agreement does not establish

A no-persona population-prevalence query can match published marginals without simulated respondents and outperforms committed panels in the aligned comparison. On unpublished demographic cells, neither readout beats national-table recitation, so marginal agreement does not certify subgroup conditioning or individual simulation.

6. What marginal agreement does not establish

  • The cell targets are unpublished weighted marginals for three four-way demographic cells, with intervals measuring decoding variation rather than independent respondents or population uncertainty.The target file and selection rule were fixed before calls according to the retained study record; the digest cannot independently prove that chronology.Source: §6 · p. 6
  • On unpublished cells, neither readout beats national-table recitation: the averages are 7.65 for the better panel readout, 7.47 for the query, and 7.00 for the national table.The probability readout favours five of six cell-model comparisons, from +1.88 to +12.05, but can also return exact zeros on 8 of 13 options.Source: §6 · p. 6
  • The no-persona query averages 6.27 MAE versus 12.39 for committed panels and is better on all 9 matched pairs; constraint-aware probabilities average 5.34 and beat the query on 4 of 9.The query does not establish that direct estimation is uniformly best.Source: §6 · p. 6
  • On held-out cells, the query wins three of six comparisons against the better panel readout, and both average worse than national-table recitation; the query yields no respondent sets, cardinality, or co-selection.The authors make no training-cutoff claim from the gpt-4o-mini identifier because its exact dated resolution was recorded after collection.Source: §6 · p. 7

7. Exploratory decomposition of subgroup calibration

Exploratory analyses on additional CES demographic cells describe subgroup ordering and level bias, then test option-specific offsets on held-out cells. These analyses are hypothesis-generating, not confirmatory comparisons with survey estimators.

7. Exploratory decomposition of subgroup calibration

  • Thirty additional CES cells on two models are used to describe subgroup error and explore offset correction after the initial three-cell result, not as confirmatory estimator comparisons.The later analyses were specified after the initial result.Source: §7 · p. 7
  • Against cell truth, probability MAE is 6.08 and 5.09 and committed-set MAE is 11.36 and 12.14, all above the national-table score of 4.85.Simulated and human cell deviations have per-option median correlations of 0.79 and 0.86, while mean absolute per-option level bias is 4.62 and 4.48 points; selected cells make the correlations descriptive upper bounds.Source: §7 · p. 7
  • Ten anchors reduce haiku’s probability MAE from 6.08 to 4.42, and one reduces gemini’s from 5.09 to 3.52, with empirical draw-win rates against the unadjusted national table of 95 and 96 percent.These resampling summaries are conditional on selected cells, models, and a post hoc correction, not success probabilities for new instruments or populations.Source: §7 · p. 7
  • The second CES battery shows the same descriptive pattern across four models: probability outputs preserve cell ordering more than committed sets, and offsets reduce MAE.Anchor counts reaching a 95-percent empirical draw-win rate range from two to ten, not a preregistered budget guarantee.Source: §7 · p. 7
  • The exploratory analyses reuse CES microdata, compare against a deliberately weak table baseline, and omit small-area, regression, and multilevel estimators.Their role is hypothesis generation.Source: §7 · p. 7

8. Conclusion

For static demographic synthetic survey panels, the response contract changes measured fidelity, and matching population marginals cannot establish individual simulation. The conclusion is limited to the reported panels and estimands.

8. Conclusion

  • Committed sets leave 66 of 128 model-option slots empty, while probabilities leave none; probabilities reduce marginal MAE in all eight uncapped comparisons, subject to the capped-instrument boundary.The aligned subset is the basis for these comparisons.Source: §8 · p. 7
  • A no-persona query beats committed panels on all nine aligned matched pairs, though constraint-aware probability vectors do not lose to the query on average.Subgroup analyses suggest ordering signal with biased levels but establish neither an anchor budget nor superiority to survey estimators.Source: §8 · p. 8
  • Evaluate a synthetic user on the individual and joint estimands it is claimed to simulate; marginal agreement is evidence about a readout, not a person.Source: §8 · p. 8

9. Limitations and ethics

The paper limits its claims to three aligned instruments and cautions that mode, date, prompt, and elicitation choices constrain how fidelity comparisons should be interpreted. Its interval estimates, grounding evidence, and target provenance have additional limits; the appendix also qualifies the endpoint specification and option-level analyses.

9. Limitations and ethics

  • Absolute MAE levels are not comparable across batteries because mode and format can shift check-all endorsement by several points.The paired, within-instrument contract comparison is not threatened by this difference, but no single MAE is a fidelity score for a survey population.Source: §9 · p. 8
  • Intervals are conditional on their resampling unit, and all batteries were collected in one window, leaving time effects inseparable from model effects.National-battery intervals resample prompts, while fixed-cell held-out intervals capture decoding noise; absolute levels are less stable than paired contrasts.Source: §9 · p. 8
  • Grounding evidence is limited to a one-time narrative variation: written profiles lowered error on one model and battery but supplied no confirmatory evidence.Effects of interviews, histories, goals, or persistent interaction remain untested.Source: §9 · p. 8
  • The endpoint compares realized human check-all answers with latent per-option propensities, so better marginal recovery need not mean more faithful respondent-level behavior.The separate-call design also leaves within-call consistency outside the study, and target contamination cannot be excluded.Source: §9 · p. 8
  • The headline claims cover only CC24_300, CC24_305, and QA3; the other batteries are sensitivity evidence, not pooled panel-versus-target comparisons.Prompt variants also shrink the paired readout difference on one model, and the German labels are unofficial translations.Source: §9 · p. 8

Appendix A. Option-level and sensitivity analyses

  • The appendix identifies option-marginal MAE with paired prompt-cluster bootstrap intervals as the endpoint, but shipped hashes cannot establish that it preceded collection.Its option-level file contains 206 evaluated slots over 60 distinct options; pooled option analyses use all 206 rows without clustering.Source: Appendix A · p. 11, §A.1 · p. 11

A.2 Where the advantage comes from

Across these analyses, elicitation contract and instrument constraints shape marginal fidelity, while aggregate agreement can also be achieved without simulated respondents. Exploratory results qualify what probability vectors reveal about set size and respondent-level structure.

A.2 Where the advantage comes from

  • On the rare-option tercile, direct multiselect leads probabilities by 4.29 points; pooled summaries show a prevalence gradient but do not identify its mechanism.Common options often saturate or collapse under committed sets, while a positive probability floor can be costly for rare options.Source: §A.2 · p. 11

A.4 Battery-level effects, plotted

  • Figure 2 shows the capped German battery as the only reversal across the six batteries; projecting vectors onto its cap removes it.Bars show means across three to five models per battery, and ticks show individual models.Source: Figure 2 · p. 12

A.3 Why MAE plus paired bootstrap remains the primary analysis

  • Prompt-clustered paired option-marginal MAE tests the contract comparison; matching temperatures shifts uncapped differences by +0.18 points on average.The temperature analysis reports that all four uncapped batteries still favor probabilities.Source: §A.3 · p. 12, §B.2 · p. 16

A.5 Constraint projection sensitivity (exploratory)

  • Projection changes the capped battery from favoring committed sets on two of three models to favoring probabilities on three of three; the worst case improves from −20.46 to +9.79.The analysis says this is consistent with, but does not establish, the cap as the cause because projection imposes the mass and cap together.Source: §A.5 · pp. 12–13

A.6 Post-stratifying the cohorts (exploratory)

  • Raking shifts the all-instrument descriptive mean by 0.15 points for committed sets and 0.04 for probabilities, but does not validate the cohorts or their population frames.The cohorts are raked to their recorded one-way margins; this cannot repair absent joint-distribution evidence or frame mismatches.Source: §A.6 · p. 13

A.7 Implied set size

  • In the aligned uncapped subset, probabilities are closer to human set size in 6 of 8 runs, while committed sets are closer in 2.On the capped battery, raw vectors imply 1.2 to 3.9 times the human mean; expected size says nothing about realized-set distributions or co-selection.Source: §A.7 · p. 14

A.8 Realized sets from the vectors (exploratory)

  • Realized vector-based sets beat committed sets on set-size distribution in 8 of 11 combinations and on co-selection in 9 of 11, but persona-specific joint structure is small and battery-dependent.The shuffled control is within 0.04 on set size for every combination, and on one battery a random rearrangement performs better.Source: §A.8 · p. 14

B.3 Population-prevalence baseline

  • The population-prevalence query beats committed sets on 9 of 9 aligned pairs; constraint-aware probability vectors average 5.34 and the query beats them on 5 of 9.The query diagnoses the weakness of committed-panel marginal validation, but is not the best estimator on average.Source: §B.3 · p. 17

B.4 Contamination probe

The appendix probes whether population-prevalence estimates reflect retrieval, tests prompt and interval choices, and qualifies the interpretation of the contract comparison. These analyses leave approximate or learned priors possible and treat psychological explanations as unproven.

B.4 Contamination probe

  • Forcing recall improves prevalence-query accuracy by 0.54 and 1.57 points, but the probes cannot rule out a retrieved-and-adjusted prior.Named-source prompts reproduce the neutral prompt’s systematic errors rather than correcting them; literal lookup is excluded, but a hybrid account is not.Source: §B.4 · p. 18

B.5 Validity and pairing

  • Of 19,510 collected responses, 19,410 passed validation; the capped-battery probability run with 406 valid vectors retains the same comparison direction on paired cases.On the 406 personas valid under both contracts, the difference is +4.47 [3.94, 6.03] (clustered).Source: §B.5 · p. 18

B.6 Prompt ablation: is collapse induced by the persona instruction?

  • Removing the typicality clause or asking for a randomly drawn person leaves zero coverage at 8 options, while the readout effect shrinks from +4.76 to +3.31 and +3.15.The random-draw wording does not restore missing options, though it lowers committed-set error; the prompt comparison is limited to one model and battery.Source: §B.6 · p. 19

Appendix C. Clustering the bootstrap by distinct prompt

  • All twenty prompt-clustered intervals exclude zero, and the median is 2.27 times wider than the respondent-bootstrap interval.Clustering widens intervals threefold on the UK and German batteries but 1.65 to 1.95 times on CC24_305.Source: Appendix C · pp. 19–20

Appendix D. The committed-set sampling floor

  • The smallest uncapped readout effect, 4.53 points on CC24_300, exceeds that battery’s 2.9-point committed-set sampling floor.A floor-corrected difference would shrink the CC24_300 effects to roughly 1.6 to 4.4 points.Source: Appendix D · p. 20

Appendix E. Psychological account

  • The proposed psychological account is not causally identified because the studies manipulate synthetic response contracts, not human ones.The account is consistent with elicitation literature, but the passage also notes possible costs of repeated probability judgments and their omission of co-selection and cardinality.Source: Appendix E · p. 20

Appendix F. Earlier platform-based studies

  • The reported results were re-collected on named models through public APIs; earlier commercial-platform runs are not reported and underpin no claim.Source: Appendix F · p. 20

Appendix G. Coverage statistics

The appendix defines threshold-free coverage statistics and reports support collapse under committed-set readouts. It also documents coding, safeguard, and cohort-construction boundaries that qualify interpretation of the comparisons.

Appendix G. Coverage statistics

  • Coverage counts exact zero means, degeneracy counts exact zero standard deviations, and only summed human prevalence of zero-covered options requires human targets.The descriptive surprise screen is not a valid p-value because prompts are dependent, calibration is not assured, and human-target uncertainty is omitted.Source: Appendix G · pp. 20–21
  • 85 of 206 option slots have zero coverage under the direct readout, versus 0 of 206 under probability; 100 direct slots have no between-respondent variance.The latter includes 15 slots selected by every respondent; the probability comparison reflects marginal expressivity, not realized answer-set behavior.Source: Appendix G · p. 21
  • Zero-coverage options include salient responses such as emergency-room visits at 22.1% human prevalence, and the worst run leaves 108.0 points of human prevalence unexpressed.Government debt at 12.0% is zero-covered on all three German models.Source: Appendix G · p. 21

H.1 Scoring formalism and base restriction

  • CC24_300d is sensitivity-only because eligibility depends on personas’ own committed sets and empty-array scoring changes the direct MAE from 13.16 to 24.98 under alternative treatments.The corresponding literal-empty score is 16.30, so no CC24_300d figure contributes to a headline summary.Source: §H.1 · p. 22

H.2 Safeguards and what they do not cover

  • Study notes describe design locks and partial target blinding, but no trusted timestamp or external registry verifies their chronology; no public instrument is model-held-out.SHA-256 digests verify current bytes, not when they existed or who had seen them.Source: §H.2 · p. 22

H.3 Cohort construction check

  • The first-500-row cohort has one-way deviations up to 5.4 points, while reallocating rows over generated-file profiles moves the paired difference by 0.17 points without changing coverage.This bounds sensitivity for one battery and model, not population-representative joint composition.Source: §H.3 · p. 22

Appendix I. Held-out unpublished-cell targets

The held-out analysis tests panel conditioning on unpublished demographic-cell targets and paraphrased item wording, then examines subgroup patterns across thirty selected cells. The reported findings remain descriptive and bounded by target selection and cell-sampling choices.

Appendix I. Held-out unpublished-cell targets

  • The three unpublished CES demographic-cell targets were selected from microdata by divergence from national marginals, and the national table scored 6.97 against the panel’s best readout at 7.52 on selection-free targets.Because the same microdata selected cells and estimated targets, the original 7.00-point divergence is subject to winner’s curse.Source: Appendix I · p. 23
  • Five of six held-out cell comparisons favored probabilities, but gpt-4o-mini reversed on the 30–44 cell by 2.77 points; probabilities also returned zero for 8 of 13 options on the 65+ cell.The reversal is attributed to that cell’s rare-option regime, where the probability floor is costly under absolute error.Source: Appendix I · p. 24
  • After question and option labels were paraphrased, five of six comparisons still favored probabilities, with effect sizes correlated at r = 0.94 against the original wording.This tests dependence on published phrasing, not whether the paraphrased instrument measures the same construct.Source: Appendix I · p. 24

Appendix J. Subgroup conditioning at thirty cells

  • Across thirty divergence-spread cells, neither contract had lower MAE than the national-table baseline, although centered deviations correlated with human deviations for probability models.Rank-spread selection inflates available variation, and the post hoc decomposition is hypothesis-generating rather than evidence of reliable subgroup recovery.Source: Appendix J · p. 25

J.1 Does anchoring the levels repair it?

The appendix evaluates post hoc per-option offsets, a second battery, and profile-based grounding within selected-cell pools. Results show some within-pool error reductions, but do not establish transfer to new cells or batteries.

J.1 Does anchoring the levels repair it?

  • 10 anchors reduce claude-haiku-4-5 probability MAE from 6.08 to 4.42, with 95.0% of held-cell draws below the unadjusted national table; one anchor reduces gemini-3.7-flash MAE from 5.09 to 3.52.These are conditional resampling frequencies, not probabilities of success on a new battery; committed sets remain above the national table at displayed budgets.Source: §J.1 · pp. 25–26

J.2 Replication on a second battery

  • On the second battery, probability MAEs were 11.18, 8.21, 6.67, and 6.33 against the national table’s 7.57, while one of five committed-set options was identical across cells for every model.Two models already outperformed the table before offsets.Source: §J.2 · p. 27

J.3 Richer grounding on the thirty cells

  • Generated profiles lowered probability MAE from 6.08 to 4.52, below the table’s 4.85, while committed-set MAE moved from 11.36 to 9.63 and remained worse.The post hoc arm adds no individual evidence; a 3.39 level bias remains and grounding on real individual evidence is untested.Source: §J.3 · p. 28

Appendix K. Reproducibility

  • The release maps tables to scripts and provides checks that rebuild manuscript values from retained responses and targets where possible, but model-call records do not preserve complete provider transport envelopes.Digests verify archived bytes, not when targets were fixed or who inspected them.Source: Appendix K · p. 28
Loading 2609.07305v1…