# Onsite FIDAS PM10 versus Sofia-Mladost

Study period: 15 February 2024 through 13 October 2026. Agency clock: Europe/Sofia. Agency portal data are preliminary.

## Matched daily results

```csv
year,paired_days,fidas_mean,eea_mean,bias_fidas_minus_eea,pearson_r

2024,94,21.673395965276114,22.576276595744687,-0.9028806304685646,0.9498621923951007

2025,357,23.140346387648016,26.6966106442577,-3.5562642566096887,0.9635573696853295

2026,279,18.568537883165337,19.940465949820787,-1.3719280666554525,0.905601883991306
```

PM inputs are regular nominal 15-minute observations (14-16 minutes), including averages derived from minute measurements. Export intervals enter concurrent comparisons by UTC midpoint. Missing observations are not interpolated; measured zeros are retained.

Daily eligibility: >=72 accepted slots. For 2024/2026 require >=10 measured valid minutes per quarter and >=1080 per day; unknown minute counts do not qualify. For 2025 use record-slot coverage; valid-minute fields remain blank. Hourly eligibility: >=3 accepted slots, plus >=45 measured minutes in 2024/2026. These distinct quality bases limit comparisons.

Pair on identical dates/hours; equal weight per accepted quarter and per paired date. Use published EEA daily means, not rounded hourly recomputations. Winter combines Jan/Feb/Dec of the displayed year. Mean interval ratios precede aggregation. Zero exclusion is only a sensitivity; minute weighting is unavailable for 2025.

EEA labels 1-24 are hour ends. All 120 source hourly rows on five DST dates are excluded from hourly matching; agency daily values remain. No DST repair. Diurnal SD is between matched hourly observations, not uncertainty of a mean. Daily confidence intervals use 600 seven-calendar-day circular block-bootstrap draws; gaps retained, no multiplicity correction.

Current matched samples: 730 daily pairs, 17508 hourly pairs. Independent standard-library verification is recorded in validation.json, not implied by this build. All dates, quality denominators, source hashes, event thresholds, matched values and sensitivity tables are inspectable in downloads.

These are different sites and methods, not calibration or causal-source attribution. Reproduce with work/compare_eea_mladost_pm10.py and work/verify_eea_mladost_pm10.py.