# GNM — Garda Națională de Mediu (Hand-off Plan)

**Status:** PARTIAL. Source publishes only aggregate stats. We capture the
publicly-named violators (the headline cases) — full per-CUI fines history is
NOT available without an OUG 109/2007 access-to-info request.

## Sources investigated (2026-05-10)

| Source | URL | Verdict |
|---|---|---|
| gnm.ro homepage | https://www.gnm.ro/ | Only links to PDFs + press releases |
| Annual reports | gnm.ro/rapoarte-si-note-de-activitate/ | Aggregate stats only — `raport_activitate_<an>.pdf` (2012-2024) |
| Monthly synthesis | gnm.ro/wp-content/.../sinteza_<luna>_<an>.pdf | 3-page PDFs: per-judet TOTALS only, no per-firm rows |
| Press releases | gnm.ro/noutati/ | ~358 articles, ~10% enforcement, sporadic firm names |
| RSS feed | gnm.ro/feed/?paged=N | Same articles, structured XML, 36 pages × 10 items |
| data.gov.ro `q=mediu` | 45 datasets | Air quality, IPPC, SEVESO inventories — **no fines dataset** |
| ANPM rapoarte | anpm.ro | IPPC/SEVESO only (already covered by other agents) |

**Why per-CUI is impossible:** GNM is exempt from the open-data registry
obligation (OUG 109/2007). They cite "secret de serviciu" + "operatori
economici personali date" for not publishing the contravention register. The
only legal path is per-firm FOIA requests.

## Schema applied

`services/seap-scraper/sql/037_gnm.sql` — three tables in schema `gnm`:

| Table | Purpose | Rows after first run |
|---|---|---|
| `gnm.comunicate` | Raw archive of every press release (RSS) | **348** |
| `gnm.amenzi_extrase` | Regex-extracted (firm, fine_lei) tuples | **1** (after dedup) |
| `gnm.scrape_log` | Run history (mirrors anre/ancom) | 4 |

`is_enforcement` flag = 36/348 (10.3%) of articles match the
`/amenz|sancțiun|sistare|confiscat|sesizare penal/i` filter.

## Files added

```
services/seap-scraper/sql/037_gnm.sql                 (130 lines)
services/seap-scraper/src/scrape-gnm.ts               (~440 lines)
services/seap-scraper/cron/scrape-gnm.sh              ( 90 lines)
services/seap-scraper/GNM-PLAN.md                     (this file)
```

## Sample ingest stats

First full backfill (2026-05-10):

```
seen=348 inserted=348 updated=0 skipped=0
enforcement=36 violators=2 → 1 after dedup
duration=58s
```

After running the Stage-B fuzzy matcher against `firms.entities`:

```
gnm.amenzi_extrase id=1
  contravenient_name = "Retim Ecologic Service SA"
  contravenient_cui  = 9112229   (RETIM ECOLOGIC SERVICE SA, jud. BIHOR)
  cui_match_score    = 1.0
  suma_lei           = 150000
  context            = "Depozitul de Deșeuri Nepericuloase Ghizela, operat de
                        Retim Ecologic Service SA. Operatorul a fost
                        sancționat cu 150.000 lei amendă..."
```

## Realistic yield estimate

Press-release named violators per year ≈ 50-200 firms (out of ~5,000 actual
fines). Coverage = 1-4%. Acceptable trade-off: the firms that appear in press
releases are the **biggest** offenders (refineries, large landfills, mining
operators) — exactly the firms most likely to also win SEAP contracts. The
tail is invisible but the top of the distribution is captured.

## Cross-source SQL recipes

### 1. Firms with GNM environmental fines that win SEAP construction contracts

```sql
-- Environmental violators winning state contracts.
-- Construction CPV codes start with 45; mining/extraction CPV 14/77.
SELECT
  ge.contravenient_cui,
  ge.contravenient_name,
  ge.suma_lei                                  AS gnm_amenda_lei,
  ge.fapta,
  c.titlu                                      AS gnm_articol,
  c.publicat_la                                AS gnm_data,
  COUNT(DISTINCT a.id)                         AS seap_contracte_castigate,
  SUM(a.contract_value_lei)                    AS seap_valoare_totala_lei,
  STRING_AGG(DISTINCT LEFT(a.cpv_code, 2), ',') AS seap_cpv_prefixes
FROM gnm.amenzi_extrase ge
JOIN gnm.comunicate c           ON ge.comunicat_id = c.id
LEFT JOIN seap.announcements a  ON a.supplier_cui = ge.contravenient_cui
                                AND a.cpv_code LIKE '45%'    -- construction
WHERE ge.contravenient_cui IS NOT NULL
GROUP BY ge.contravenient_cui, ge.contravenient_name, ge.suma_lei, ge.fapta,
         c.titlu, c.publicat_la
HAVING COUNT(DISTINCT a.id) > 0
ORDER BY ge.suma_lei DESC NULLS LAST;
```

### 2. EU funds POIM-Mediu beneficiaries with GNM fines (the double-irony)

```sql
-- POIM = Programul Operațional Infrastructură Mare (Mediu axis).
-- A firm that receives EU money for environmental projects WHILE being fined
-- by GNM for environmental violations is the headline scandal pattern.
SELECT
  ge.contravenient_cui,
  ge.contravenient_name,
  ge.suma_lei                       AS gnm_amenda_lei,
  ge.fapta                          AS gnm_fapta,
  fb.proiect_titlu                  AS eu_proiect,
  fb.valoare_eligibila_eur          AS eu_valoare_eur,
  fb.program_finantator             AS eu_program
FROM gnm.amenzi_extrase ge
JOIN fonduri.beneficiar_proiect fb ON fb.beneficiar_cui = ge.contravenient_cui
WHERE ge.contravenient_cui IS NOT NULL
  AND fb.program_finantator ILIKE '%POIM%'    -- or ILIKE '%mediu%' for broader
ORDER BY fb.valoare_eligibila_eur DESC NULLS LAST;
```

### 3. Top GNM violators sorted by total fines mentioned across press releases

```sql
SELECT
  contravenient_cui,
  MIN(contravenient_name) AS firma,
  COUNT(*)                AS nr_mentions,
  SUM(suma_lei)           AS total_amenzi_lei,
  STRING_AGG(DISTINCT judet, ', ') AS judete_implicate
FROM gnm.amenzi_extrase
WHERE contravenient_cui IS NOT NULL
GROUP BY contravenient_cui
ORDER BY total_amenzi_lei DESC NULLS LAST
LIMIT 50;
```

## Stage-B fuzzy matcher

The scraper stores `contravenient_name_norm` but leaves `contravenient_cui`
NULL. To populate CUIs, run the following after each scrape (idempotent —
only updates rows where CUI is NULL):

```sql
WITH unmatched AS (
  SELECT id, contravenient_name_norm
  FROM gnm.amenzi_extrase
  WHERE contravenient_cui IS NULL AND contravenient_name_norm IS NOT NULL
)
UPDATE gnm.amenzi_extrase a
SET contravenient_cui  = m.cui,
    cui_match_method   = 'fuzzy_name',
    cui_match_score    = m.score,
    matched_at         = now()
FROM (
  SELECT u.id, f.cui,
         similarity(u.contravenient_name_norm, firms.normalize_company_name(f.name)) AS score
  FROM unmatched u
  CROSS JOIN LATERAL (
    SELECT cui, name
    FROM firms.entities
    WHERE firms.normalize_company_name(name) % u.contravenient_name_norm
    ORDER BY similarity(firms.normalize_company_name(name), u.contravenient_name_norm) DESC
    LIMIT 1
  ) f
) m
WHERE a.id = m.id AND m.score >= 0.85;
```

## Operational guidance

* **Cron schedule:** weekly (Sundays 03:00) — RSS rarely changes, ~5-10 new
  articles per week. Use `SINCE_DAYS=14` for incremental runs after the first
  full backfill.
* **Rate limits:** gnm.ro returns `RateLimit-Limit: 100/min`, `1000/hr`. We use
  ~36 requests per full scrape with 800 ms sleep — well within budget.
* **Idempotency:** `gnm.comunicate` UPSERTs on `guid` (WordPress post ID,
  immutable). Skip when `raw_hash` unchanged. Re-extraction wipes only the
  child rows for changed articles.
* **404 on page 36:** harmless — currently 35.8 pages so we 404 on the trailing
  empty fetch. Captured by retry loop, exits cleanly.

## Future enhancements (not in this hand-off)

1. **OCR of monthly synthesis PDFs** — IF in future they add per-judet tabular
   detail (currently 3 pages, totals only, OCR adds nothing).
2. **Annual report PDF** has more granular judet × sector breakdowns
   (waste / air / water / biodiversity) — could add a second extractor for
   `gnm.amenzi_per_judet_sector` aggregates.
3. **Local press archives** (e.g. monitoruldebuzau.ro, focuspress.ro) often
   name specific firms when GNM does press conferences regionally — could
   harvest via a curated whitelist of regional outlets that beat-cover GNM.
   Estimated +50-100 named firms/year. Risk: licensing.
4. **FOIA submissions** via the `gnm@gnm.ro` legea-544 path — could request
   the contravention register annually. Civic-tech precedent: prefectura.ro
   data was successfully unblocked this way in 2024.

## Time spent

~75 minutes:
- 20 min investigation (gnm.ro / data.gov.ro / RSS reconnaissance)
- 5 min schema design + apply
- 35 min scraper write + 3 iterations to tune the regex extractors
- 5 min Stage-B fuzzy match validation
- 10 min documentation