> For the complete documentation index, see [llms.txt](https://docs.layerlens.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.layerlens.ai/more-in-this-section-2/eval-patterns-9.md).

# Real estate and PropTech evaluation patterns

Real estate evaluation patterns — primitive ratio, span rules, GEPA labeling, cadence.

Recommended primitive ratio for real estate AI: **60/25/15** — deterministic rules, code assertions, judges. Fair Housing Act, FCRA, and state real-estate forms drive most checks to deterministic rules; bias auditing leans on code scorers.

## Primitive ratio

| Primitive           | Share | Why                                                                                   |
| ------------------- | ----- | ------------------------------------------------------------------------------------- |
| Deterministic rules | 60%   | Protected-class language, agent-review gate, FCRA accuracy, state-form clauses        |
| Code assertions     | 25%   | Demographic fairness, factual match, comp-set checks                                  |
| LLM judges          | 15%   | Steering language detection, neighborhood-claim faithfulness, tone, retaliation guard |

## Pattern table

| Scenario                            | Primary primitive | Specific check                             |
| ----------------------------------- | ----------------- | ------------------------------------------ |
| Listings — protected language       | Rule + Judge      | No protected references direct or indirect |
| Listings — factual accuracy         | Code assertion    | Sq ft, beds, baths match data              |
| Listings — agent gate               | Rule              | No publish without agent sign-off          |
| AVM — fairness                      | Code scorer       | Valuation parity post-attribute control    |
| AVM — comp set transparency         | Rule              | Comps cited                                |
| AVM — confidence interval           | Rule              | CI present                                 |
| AVM — out-of-area refusal           | Rule              | Below threshold = refuse                   |
| Tenant screening — FCRA accuracy    | Rule              | Every item verifiable                      |
| Tenant screening — adverse action   | Rule              | Notice triggered                           |
| Tenant screening — fairness         | Code scorer       | Group outcome parity                       |
| Tenant screening — criminal record  | Rule              | HUD 2016 guidance                          |
| Transactions — state form           | Rule              | Clauses match state standard               |
| Transactions — required disclosures | Rule              | Federal + state surfaced                   |
| Transactions — citation             | Judge             | Output grounded in contract                |
| Property mgmt — tenant rights       | Rule              | Jurisdiction-specific cited                |
| Property mgmt — habitability        | Rule              | Auto-escalate                              |
| Property mgmt — retaliation guard   | Rule + Judge      | Sensitive cases route to human             |

## Span rules

* `listing.lookup` — listing data and source
* `protected.scan` — protected-class language detector
* `comp.set` — comp properties cited
* `fcra.verify` — item verification trail
* `state.form.match` — state-specific clause comparison
* `tenant.rights.lookup` — jurisdiction-specific tenant rights

## GEPA labeling

For each real estate judge, label ≥ 50 paired examples:

* **Steering-language detection:** Fair-housing specialists label; cover subtle steering
* **Neighborhood-claim faithfulness:** Local agents + data analysts label
* **Retaliation-guard sensitivity:** Tenant-rights attorneys label
* **Contract-citation faithfulness:** RE attorneys label per applicable state

Run [judge optimization](/more-in-this-section-9/judge-optimizations-1.md) with `budget="medium"`. For protected-class detectors, run with `budget="heavy"`.

## Cadence

| Cadence        | Activity                                                             |
| -------------- | -------------------------------------------------------------------- |
| **Per-deploy** | CI gate runs scenario suite (rules + assertions)                     |
| **Daily**      | Sample production traces; full evaluation                            |
| **Weekly**     | Demographic-fairness audit on AVM and tenant screening               |
| **Monthly**    | GEPA re-optimization on judges                                       |
| **Quarterly**  | State-form refresh; HUD guidance review; jurisdiction-rights refresh |
| **Annually**   | Full fair-housing audit; insurance E\&O renewal evidence package     |

## See also

* [Real estate scenarios](/more-in-this-section-2/scenarios-9.md)
* [Real estate compliance](/more-in-this-section-2/compliance-9.md)
* [Recipe: GEPA basics](/9.-improve-tune-the-system/gepa-basics.md)