---
artifact_kind: model-card
artifact_name: <concise identifier, e.g., "preference-ranker-v2">
version: <semver, e.g., "2.1.0">
owner: <named engineer or team>
last_reviewed: <YYYY-MM-DD>
next_review: <YYYY-MM-DD>
concordia_components: [<list of Concordia subsystems that consume this model>]
related_phase_179_tasks: [<e.g., 179.3.2.1, 179.3.2.2>]
related_cards:
  system: <path to related system card, if any>
  evaluators: [<paths to evaluator cards that score this model>]
---

# Model Card — <artifact_name> v<version>

## 1. Identity

- **Model name:** <name>
- **Provider / developer:** <Anthropic / OpenAI / Concordia / local / etc.>
- **Base model and version:** <e.g., claude-opus-4-7, gpt-5.4, local-qwen-72b>
- **Purpose:** <one sentence describing what this model does inside Concordia>
- **Intended use cases (Concordia use-case classes):** <subset of the 14
  `ConcordiaUseCaseClass` values>
- **Out-of-scope uses:** <what this model is not to be used for — link to
  §179.1.2.2 hard boundaries if relevant>

## 2. Development context

- **Training data summary:** <provider-disclosed data summary, training cutoff,
  language coverage, notable inclusions / exclusions>
- **Fine-tuning:** <none / RLHF / DPO / Concordia-specific SFT — identify
  dataset, consent posture, date>
- **Prompt family:** <link to prompt templates used with this model>
- **Concordia-specific customization:** <any system prompt, tool rigging,
  caching, or routing configuration>

## 3. Inputs and outputs

- **Input schema:** <point to Concordia contracts — e.g.,
  `@concordia/contracts/consent` ConsentRecord, `@concordia/contracts/use-case`
  ConcordiaUseCaseClass>
- **Output schema:** <structured JSON shape; point to TypeScript types>
- **Deterministic settings used in production:** <temperature, seed, top-p, max
  tokens>

## 4. Evaluation

Fill this section with concrete numbers and cite the evaluator cards that
produced them. Placeholder `TBD` values block release.

| Metric                                   | Value                   | Evaluator card           | Date         |
| ---------------------------------------- | ----------------------- | ------------------------ | ------------ |
| preference-stability (paraphrase)        | <e.g., 0.92 Spearman ρ> | <path to evaluator card> | <YYYY-MM-DD> |
| preference-stability (order)             | <...>                   | <...>                    | <...>        |
| preference-stability (prompt template)   | <...>                   | <...>                    | <...>        |
| uncertainty calibration (Brier / ECE)    | <...>                   | <...>                    | <...>        |
| fairness: Nash product shift vs baseline | <...>                   | <...>                    | <...>        |
| fairness: max-min utility                | <...>                   | <...>                    | <...>        |
| benchmark: 2-party suite                 | <...>                   | <...>                    | <...>        |
| benchmark: multi-party suite             | <...>                   | <...>                    | <...>        |
| benchmark: procurement suite             | <...>                   | <...>                    | <...>        |
| benchmark: DAO / governance suite        | <...>                   | <...>                    | <...>        |
| benchmark: restorative suite             | <...>                   | <...>                    | <...>        |
| cost per scoring call (USD)              | <...>                   | <...>                    | <...>        |
| median latency (ms)                      | <...>                   | <...>                    | <...>        |

## 5. Privacy posture

- **Data visibility at the model:** <what fields the model can see>
- **Per-party isolation:** <how §179.5.1.1 physically-isolated prompt contexts
  are enforced for this model>
- **Training-data reuse eligibility:** <true / false; if true, which
  `ConsentPurpose` must be granted>
- **Zero-retention mode:** <yes / no; for which use-case classes>
- **Secure-enclave / confidential-compute posture:** <current / roadmap>

## 6. Safety posture

- **Hard-boundary coverage:** <which §179.1.2.2 hard boundaries this model is
  tested against>
- **Adversarial robustness:** <results from §179.5.2.3 and §179.7.6.4 suites>
- **Coercive-choice refusal:** <does the model refuse to generate countdown
  pressure, hidden-accept anchors, etc.? link to §179.4.3.5 tests>
- **Prompt-injection resistance:** <current red-team result>

## 7. Fairness and bias posture

- **Protected-class disparate treatment:** <test results against §179.3.3.5
  profile>
- **Power-asymmetry handling:** <how the model handles asymmetric BATNA /
  resource / language power>
- **Language / locale coverage:** <which languages pass Concordia's multilingual
  acceptance gate>

## 8. Known limitations

- <plainly stated failure modes — paraphrase sensitivity, order effects,
  adversarial vulnerability, long-context degradation, locale gaps, etc.>

## 9. Human-in-the-loop expectations

- **Required reviewers when this model is used:** <subset of `ReviewerQueue`>
- **Escalation triggers specific to this model:** <if the model emits a
  confidence below X, escalate to Y>
- **Mediator-assist vs autonomous use:** <which operational modes from
  §179.1.2.3 this model is qualified for>

## 10. Incident history

- <links to §179.5.3.4 incident reports that implicated this model; each entry
  includes incident ID, date, root cause, remediation>

## 11. Regulatory posture

- **EU AI Act Article 11 / Annex IV alignment:**
  <notes on technical documentation completeness>
- **JAMS AI Rules expert-inspection readiness:** <artifacts preserved,
  trade-secret flags>
- **NIST AI RMF harm coverage:** <mapped harm families H1–H8 from
  [`../nist-ai-rmf-mapping.md`](../nist-ai-rmf-mapping.md)>

## 12. References

- <links to prompts, datasets, evaluation notebooks, ADRs, and related cards>