Model card what it knows, how precisely, and where it stops

1. Intended purpose (AI Act Art. 13(3)(b)(ii))

The assistant answers questions about the hydrological plans of Spain’s river basin districts (3rd cycle 2022-2027) with citations to official documents. It is designed for informational and exploratory use.

Not designed for:

• Binding legal advice
• Expert assessments
• Formal administrative decisions
• Replacing direct consultation of the official documentation

2. Data and sources

The indexed corpus covers the hydrological plans of all 25 river basin districts of Spain (mainland confederations, intra-community basins, the Balearic Islands, the Canary Islands, Ceuta and Melilla). All 25 are fully covered: each district’s hydrological plan, technical annexes and regulation, plus its Drought Special Plan where applicable (the Canary islands publish none of their own, because they fold drought management into their island plan). Since the August 2026 indexing run, two instruments that only the Tagus used to have are now in: the Flood Risk Management Plan (PGRI) for all 25 and the Key Issues Scheme (ETI) for 24. Sources are cited in every answer with an identifier, label, and relevance metric.

On earlier cycles. Year ranges are not universal: the Ebro’s 1st cycle ran 2010-2015, not 2009-2015, and the Guadiana’s and Catalonia’s 2nd ran 2016-2021. Every citation states the cycle it comes from with that district’s real range, and answers resting on a repealed plan say so. Retrieval defaults to the plan in force and only widens to other cycles when the question asks for it.

Total indexed: ~535 district documents plus the shared legal framework (~104,000 fragments) for the cycle in force. Documents are re-indexed periodically; the current snapshot date is the one heading the last column of the table above. The full district-by-district inventory is published at corpus coverage.

El Hierro’s missing ETI is declared a gap, not “not applicable”: the document was published at the time, and the download service that distributed it no longer responds. The distinction matters, because “not applicable” would claim that authority does not publish one.

3. Capabilities and metrics (AI Act Art. 13(3)(b)(iii) + (vi))

Metrics from the latest baseline are indicative and updated manually with each release.

Automated live publication of these values is on the roadmap; until then they are updated manually with each release.

4. Known limitations (AI Act Art. 13(3)(b)(iv))

• Does not generate binding legal advice. Responses are model interpretations based on the indexed corpus.
• The corpus may not reflect administrative revisions made after the snapshot date, which is given in §2.
• Supported language: Spanish. Other languages receive best-effort responses without quality guarantees.
• The underlying language and reranking models introduce the inherent biases of their respective training data. The reranker acts as a relevance filter but does not eliminate corpus biases.
• There is no guarantee of a response for questions outside the domain of the indexed hydrological plans; the system may decline to answer.

5. Disclosure and traceability

Each assistant response carries:

• system_notice: AI Act Art. 50(1) notice (AI-generation disclosure).
• confidence_level: high / medium / low / null, ALTAI #1 Human Agency.
• sources[]: array of citations with rerank_score (Tier 3.3 Explainability).
• metadata.trace_id: trace identifier for support and audit.
• HTTP header X-AI-Generated: true: AI Act Art. 50(2) disclosure.

The system validates the response envelope before rendering any content; malformed responses are automatically rejected.

5.bis Per-query environmental footprint

Every response carries a footprint object (metadata.footprint) with the estimated energy and water consumed to produce it. Values are computed by a deterministic module that captures the real token count and applies an auditable formula:

kWh_facility = model_energy × tokens/1 000 × PUE
liters       = kWh_facility × (WUE_onsite + WUE_source)

Coefficients are calibrated (version 2026-05-27) against:

• Google 2024 Environmental Report: PUE and on-site WUE per region.
• Patterson et al. 2021 (arXiv:2104.10350): WUE_source by power mix and per-token energy for transformer inference.
• Li et al. 2023 (arXiv:2304.03271): separation between on-site cooling water and electricity-embedded water.

The uncertainty_pct field explicitly documents the ±30 % band that’s reasonable to expect given the coefficients are published aggregates, not in-situ measurements.

For the full methodology, step-by-step formulas and per-region coefficient tables, see the environmental transparency page (ALTAI #6).

6. Contact and reporting issues

For serious incidents (AI Act Art. 73): see /en/incident or, if you have a session, click the flag icon at the foot of any response — that is the recommended route, because it attaches the trace and an excerpt of that response to the report. You can also write to fluxia.soporte@gmail.com.

For general quality feedback (unhelpful response), use 👍 / 👎 on the response.

7. Deployed version

The version and environment of the deployment you are looking at appear in the footer of every page, in the form Production · v1.4.0. Each release to production increments that number according to the nature of the changes (a fix, a new feature, or a breaking change).

The language model version and the corpus snapshot date are independent of that number and are updated by hand in this card with each release: the current values are the ones in §2 and §3.