Input
Expert-authored partially funded pension proposals A, B, and C with published fiscal assumptions.
Pension · Pre-verification
Synthetic Citizen Pre-verification
A virtual simulation that entered expert pension-reform proposals and checked execution and recordkeeping paths at a scale of 100 synthetic citizens.
No real people participated in this survey or event.
The responses came from model-generated synthetic personas. They do not represent or predict Korean public opinion, a representative sample, actual behaviour, or policy effects. This artifact checks whether the policy-development engine completed and recorded a 100-person synthetic run.
Run
Two rounds of responses from 100 synthetic citizens and additional-option generation by Korean-language AI models.
Verified scope
Completion, records, model-role separation, and limitation disclosure—not real acceptance or deliberative effects.
Data lineage
How the 100 synthetic citizens were made
This engine-test input combines an official age allocation with NVIDIA synthetic data.
Set the age allocation
MOIS population data dated May 31, 2026 set 28 aged 20–39, 37 aged 40–59, and 35 aged 60+.
Select NVIDIA personas
We selected 100 records from the nvidia/Nemotron-Personas-Korea dataset v1.0 (2026-04-20) and preserved each upstream source_uuid.
Freeze the exact input
The JSON preserves age, sex, occupation, region, education, family, housing, and a synthetic narrative for each record.
Assign response models
Run seed 20260722 assigned the 100 inputs to 20 LGAI-EXAONE/K-EXAONE-236B-A23B, 50 solar-pro3, and 30 HCX-005 response roles.
Both the 100 inputs and the upstream NVIDIA dataset are synthetic, not real people.
MOIS data affected only the age allocation. No income-distribution weighting was applied, and no numeric personal-income field was used. The selection file records seed=7, but its extraction script was not preserved: exact UUID reuse is possible, while replaying the full selection procedure is not. These records are not a representative sample or a reconstruction of real joint distributions.
Only recorded request model IDs are stated.
solar-pro3 was a provider alias, and the resolved fixed version was not retained in the run log. Neither solar-open2 nor an NVIDIA inference model was used in this pension run.
Run contract
The run was recorded in six stages
No 2045 future-scenario generation stage was used.
1. Policy input
Structured expert alternatives A, B, and C with publishable fiscal assumptions.
2. First responses
Generated accept, conditional, or reject judgments and reasons for every profile and alternative.
3. Aggregate
Grouped model outputs by age and alternative without treating them as public opinion.
4. Additional AI option
Three models produced independent drafts, cross-strengthening, consolidation, revision, voting, and incorporation.
5. Second responses
Presented the three originals and the additional AI option to the same profiles.
6. Audit record
Stored inputs, full responses, aggregates, model roles, and unresolved issues.
The additional AI option is not a human-approved policy proposal.
The runner recorded approval automatically to continue the next round. The option changed the fixed 43% replacement-rate premise to 40% and triggered a precise-number guard warning. This publication retains that failure as part of the pre-verification record.
Inspect · reuse
Inspect the data and code directly
Only publishable derivatives are provided; the experts' source HWP and XLSX remain excluded.