EVALUATION_REPORT.md

# Evaluation Report: Spain Reference Personas Frontier v0.1

## 1. Scope

This report documents the measured release-level properties of spain-reference-personas-2025-v0.1. It evaluates the shipped bundle itself: package integrity, composition fidelity, household utility, token-budget behavior, benchmark structure, and disclosure metadata. It does not claim that cross-model benchmark lift has already been measured inside this release.

## 2. Evaluation protocol

| Dimension | What was checked | Primary metric |
| --- | --- | --- |
| Package integrity | Required configs, docs, and row counts | Artifact inventory |
| Composition fidelity | Match between intended and released population shares | MAE and max absolute error in percentage points |
| Weight stability | Whether weights remain mild rather than extreme | min, p05, p50, p95, max |
| View efficiency | Whether public prompt views remain compact and predictable | average tokens, max tokens, pass rate |
| Benchmark completeness | Whether families and held-out splits are fully populated | counts by family and split |
| Household usefulness | Whether economic and housing context is rich enough for analysis | tenure, burden, constraint distributions |
| Governance signals | Whether uncertainty and disclosure remain explicit | tagged-row shares |

## 3. Headline results

| Metric | Result | Reading |
| --- | ---: | --- |
| Region share MAE | 0.022 pp | Strong macro fidelity by region |
| Age share MAE | 2.95 pp | Main remaining calibration gap |
| View budget compliance | 100% | All public persona views pass their limits |
| Benchmark matrix | 9 families / 4 splits | Full matrix populated |
| Weight stability | 0.9889 - 1.0551 | No extreme design effects visible |
| High disclosure-risk rows | 0.418% | Small flagged tail remains exposed |

## 4. Package integrity

| Artifact | Rows |
| --- | ---: |
| persona_core.parquet | 1,000,000 |
| household_core.parquet | 536,741 |
| persona_views.parquet | 6,350,524 |
| actor_state_init.parquet | 1,000,000 |
| benchmark_tasks.parquet | 1,800 |
| source_registry.parquet | 11 |
| field_provenance.parquet | 13 |

## 5. Composition fidelity

### 5.1 Metric definition

- MAE_pp = mean(abs(target_share - observed_share)) across the tested categories.
- Max absolute error is the largest category-level deviation in percentage points.
- Shares are interpreted as release-level composition checks, not downstream model outputs.

### 5.2 Regional fidelity

| Region | Target | Observed | Error |
| --- | ---: | ---: | ---: |
| Andalucía | 17.876% | 17.895% | +0.019 pp |
| Cataluña | 16.360% | 16.345% | -0.015 pp |
| Madrid | 14.244% | 14.194% | -0.051 pp |
| Comunidad Valenciana | 10.656% | 10.617% | -0.039 pp |
| Galicia | 5.695% | 5.726% | +0.032 pp |
| Castilla y León | 5.028% | 5.068% | +0.040 pp |
| País Vasco | 4.672% | 4.700% | +0.028 pp |

~~~text
Andalucia              17.90%  ##################
Cataluna               16.35%  #################-
Madrid                 14.19%  ##############----
Comunidad Valenciana   10.62%  ###########-------
Galicia                 5.73%  ######------------
Castilla y Leon         5.07%  #####-------------
Pais Vasco              4.70%  #####-------------
~~~

Interpretation: regional alignment is strong enough for subgroup slicing and macro simulation by territory.

### 5.3 Age fidelity

| Age group | Target | Observed | Error |
| --- | ---: | ---: | ---: |
| 18-24 | 8.000% | 10.484% | +2.48 pp |
| 25-34 | 13.000% | 16.996% | +4.00 pp |
| 35-44 | 17.000% | 15.843% | -1.16 pp |
| 45-54 | 19.000% | 14.840% | -4.16 pp |
| 55-64 | 17.000% | 13.460% | -3.54 pp |
| 65+ | 26.000% | 28.377% | +2.38 pp |

~~~text
18-24   +2.48 pp
25-34   +4.00 pp
35-44   -1.16 pp
45-54   -4.16 pp
55-64   -3.54 pp
65+     +2.38 pp
~~~

Interpretation: 25-34 is overrepresented and 45-64 is underrepresented. This is the main statistical weakness of v0.1 and the first target for a future recalibration pass.

## 6. Weight stability

| Statistic | Value |
| --- | ---: |
| Mean | 1.0000 |
| Min | 0.9889 |
| p05 | 0.9939 |
| p50 | 1.0009 |
| p95 | 1.0066 |
| Max | 1.0551 |

Interpretation: weights stay close to 1.0 and do not imply extreme survey-style reweighting behavior.

## 7. View-layer efficiency

| View | Count | Avg tokens | Max tokens | Utilization | Pass rate |
| --- | ---: | ---: | ---: | ---: | ---: |
| micro_card | 1,000,000 | 99.8 | 120 | 83.1% | 100.0% |
| standard_card | 1,000,000 | 175.7 | 212 | 70.3% | 100.0% |
| policy_view | 1,000,000 | 89.2 | 97 | 49.5% | 100.0% |
| consumer_view | 1,000,000 | 95.6 | 113 | 53.1% | 100.0% |
| culture_view | 1,000,000 | 105.8 | 153 | 58.8% | 100.0% |
| dialogue_view | 1,000,000 | 83.4 | 93 | 46.3% | 100.0% |
| extended_profile | 350,524 | 364.5 | 407 | 60.8% | 100.0% |

~~~text
micro_card         99.8 / 120   ############--
standard_card     175.7 / 250   ##########----
policy_view        89.2 / 180   #######-------
consumer_view      95.6 / 180   #######-------
culture_view      105.8 / 180   ########------
dialogue_view      83.4 / 180   ######--------
extended_profile  364.5 / 600   ########------
~~~

Interpretation: compact cards remain genuinely compact, and long-form context stays optional rather than becoming the default burden for every prompt.

## 8. Household and economic usefulness

| Household metric | Result |
| --- | ---: |
| Average adults per household | 1.863 |
| Average minors per household | 0.560 |
| Households with minors | 38.013% |

| Tenure band | Share |
| --- | ---: |
| private_rent | 39.471% |
| mortgage | 21.837% |
| owner_outright | 21.602% |
| family_transfer | 11.990% |
| protected_rent | 5.100% |

| Housing-cost burden | Share |
| --- | ---: |
| moderate | 36.710% |
| low | 33.614% |
| high | 29.676% |

| Consumption constraint | Share |
| --- | ---: |
| managed | 36.882% |
| comfortable | 29.454% |
| tight | 22.080% |
| affluent | 11.584% |

| Tenure band | High burden | Moderate burden | Low burden |
| --- | ---: | ---: | ---: |
| private_rent | 53.2% | 37.8% | 8.9% |
| mortgage | 22.0% | 50.0% | 27.9% |
| owner_outright | 10.0% | 28.0% | 62.1% |
| family_transfer | 10.0% | 28.3% | 61.7% |
| protected_rent | 9.8% | 28.0% | 62.2% |

Interpretation: housing context is materially useful for policy, inflation, and consumer-choice work because rent, mortgage, and owner-outright households are not treated as interchangeable.

## 9. Benchmark completeness

| Family | Tasks |
| --- | ---: |
| policy_opinion | 200 |
| election_turnout | 200 |
| poll_response | 200 |
| event_reaction | 200 |
| media_trust | 200 |
| consumer_choice | 200 |
| culture_identity | 200 |
| multi_turn_social | 200 |
| future_expectations | 200 |

| Split regime | Tasks |
| --- | ---: |
| in_distribution | 450 |
| heldout_persona_seen_task | 450 |
| seen_persona_heldout_task | 450 |
| heldout_persona_heldout_task | 450 |

## 10. Governance signals

| Disclosure risk | Share |
| --- | ---: |
| low | 82.948% |
| moderate | 16.634% |
| high | 0.418% |

| Uncertainty level | Share |
| --- | ---: |
| low | 66.283% |
| medium | 20.049% |
| high | 13.668% |

## 11. Expert reading

- Sociologists: strongest for subgroup design, scenario prototyping, and synthetic survey rehearsal.
- Poll analysts: useful for persona-conditioned open-ended response simulation and split-based benchmark design, but not a substitute for field polling.
- Policy analysts: strongest on household and actor-state layers for event-reaction and tradeoff prompts.
- Economists and consumer researchers: useful for inflation shocks, trade-down behavior, housing-policy response, and price sensitivity scenarios.

## 12. Main remaining gaps

- Live benchmark-lift comparisons against actual models are not yet included.
- Cross-model transfer studies are not yet included.
- Prompt-sensitivity reruns are not yet included.
- Time-instability reruns are not yet included.
- Human-rater studies of narrative plausibility are not yet included.
- Formal disclosure attack studies beyond release-level tagging are not yet included.

## 13. Bottom line

- Strong package design for LLM evaluation and simulation.
- Strong regional fidelity and household/economic usefulness.
- Full token-budget compliance across the public view layer.
- Full benchmark matrix population.
- Age calibration remains the main open statistical improvement area.
- Cross-model benchmark lift still needs to be measured by downstream experiments rather than inferred from packaging alone.