Current execution status
No external generative AI output was used in the current scientific assessment.
The current AQ26 run uses deterministic acquisition, normalisation, quality-control, freshness, linkage and communication rules. Configured AI credentials or catalogue entries do not constitute model execution or scientific evidence.
Automated publication authority: no. Human review required: yes.
Permitted roles
- Suggest source-discovery queries and candidate datasets for human approval.
- Assist metadata normalisation and document/entity extraction with deterministic validation against originals.
- Rank evidence-review workload using transparent, reproducible features.
- Draft bounded summaries that are checked against the frozen public release contract.
- Propose anomaly candidates only when deterministic statistical gates independently reproduce them.
Prohibited roles
- Causal source attribution or source-apportionment determination.
- Permit, legal, regulatory or compliance determination.
- Exposure, safety, medical or health-impact determination.
- Inventing, repairing or replacing missing measurements.
- Granting website, email, social or scientific-publication authority.
- Treating model confidence, fluent text or credential availability as validation.
Controlled analytics roadmap
Each capability remains disabled for scientific inference until its minimum gate and evaluation requirements are met.
| Priority | Capability | Method | Minimum gate |
|---|---|---|---|
| 1 | data-latency and missingness diagnostics | deterministic lateness baselines and change-point alerts by provider, station and pollutant | timestamp provenance, stable expected cadence and alert evaluation on historical outages |
| 2 | robust anomaly candidate detection | seasonal robust statistics with multiplicity control, never a single uncalibrated black-box score | fresh quality-eligible series, station-role compatibility and out-of-sample false-alert evaluation |
| 3 | spatiotemporal episode clustering | distance/time constrained clustering followed by independent meteorology and competing-source checks | multiple physical sites, synchronized timestamps and documented spatial uncertainty |
| 4 | document and register extraction | schema-constrained extraction with page-level citations, hashes and human verification | original authoritative document retained and every extracted field traceable to source location |
| 5 | regional-model residual screening | compare compatible ground observations with CAMS or equivalent model context without treating residuals as attribution | unit, averaging period, grid, time and uncertainty compatibility established |
VVIP provenance
Assurance receipt: 83cc64313eea4d66704de3d6e7ab4b1147b7621db3ea5cbc5dc701e67e795845
All current inputs are listed with presence, size and SHA-256 identity in assets/data/ai-ml-assurance.json.
AI or ML may assist retrieval, triage and drafting only under frozen evidence, deterministic checks and human review. It cannot establish source contribution, causation, compliance, exposure, safety or health effects.