edukacja_medialna_pl: drop 4 lessons with embedded third-party quotations (252 -> 248)

artifacts/source_findings.md

@@ -1758,3 +1758,43 @@ Licence read-only 2026-10-06: <https://nonsa.pl/wiki/Nonsensopedia:Prawa_autorsk
 
 **Lesson** — a tag name is not a content type. `<poem>` is used for jokes and logs as
 well as verse, so measure what a rule drops before quoting its size.
+
+## `edukacja_medialna_pl`: four lessons with embedded third-party quotations dropped
+
+**Investigated:** 2026-10-06
+**Verdict:** 252 → 248 lessons stay CC BY-SA 3.0 (728,987 → 713,052 tokens; the Michera pair is 7,282
+tokens, the profile pair 8,653).
+`src/clean_edukacja_medialna_pl.py` filters the SHA-pinned PR #26 Parquet and sidecars. Tracks the
+2026-10 rights audit line "optionally remove the 3 quote passages".
+
+### Verified
+
+- Two lessons quote the film scholar Wojciech Michera by name, credited to a University of
+  Warsaw course ("cytat z materiałów do kursu ..."): `montaz-materialu-filmowego` and
+  `scenografia-charakteryzacja-kostiumy-i-aktorzy`, one passage each. Dropped whole.
+- Two lessons (`kim-jestem-w-internecie`, `wizerunek-w-sieci`) share an invented-profile
+  exercise with six "Ulubione cytaty" lines, each credited to a named person (Marilyn Monroe,
+  Winston Churchill and others). Dropped whole as a fail-closed guess (see below).
+- Kept on purpose: the attributed CC BY-SA Wikipedia excerpt (`ekonomiczne-aspekty-...`),
+  the Constitution and ECHR article quotes (statutory text, Art. 4), lessons that only name
+  films, books or point to a handout by URL (`uzaleznienia` links the Jakubik article, it does
+  not embed it).
+
+### Not established
+
+- The audit says "2-3 short quotes" and "3 quote passages". Only the two Michera passages are
+  reproducible. No third third-party quotation was found by reading every lesson that contains
+  "cytat", "źródło" or a quote followed by an attribution. The profile exercise is a guess at
+  what the audit counted. To keep those two lessons, delete the two `PROFILES` entries from
+  `DROP` and run `git checkout 5ec3e18 -- data/edukacja_medialna_pl` first (+8,653 tokens).
+- The raw XML is not in the checkout, so the filter works on the Parquet. The datasheet's
+  "Reproduction" section still describes the original build.
+
+### Side effects
+
+- `src/test_edukacja_medialna_contribution.py` pinned the contributed data files by SHA-256
+  through `artifacts/edukacja_medialna_pl_ontology_manifest.json`. It now compares the manifest
+  pins with the input pins in the clean script, so the manifest stays untouched.
+
+**Lesson:** an audit count ("3 quotes") is a lead, not a list. Reproduce it from the data before
+writing the filter, and say which drops are confirmed.

data/edukacja_medialna_pl/NOTICE.md

@@ -7,4 +7,4 @@ License statement at the origin: every lesson page ends with "Licencja: Creative
 
 Per-record text/scenario/expert creators, canonical URLs, source dates and rights evidence are preserved in `artifacts/attribution.jsonl`. Raw source XML and acquisition hashes are preserved in `artifacts/raw_xml.jsonl.gz`.
 
-Preparation: Piotr Styla with OpenAI Codex. Changes: extracted the authored `powiesc` lesson body from source XML; omitted external reading lists; normalized Unicode and whitespace; applied language, length, creator, license, limited PII-pattern and deduplication gates. No endorsement by Fundacja Nowoczesna Polska is implied.
+Preparation: Piotr Styla with OpenAI Codex. Changes: extracted the authored `powiesc` lesson body from source XML; omitted external reading lists; normalized Unicode and whitespace; applied language, length, creator, license, limited PII-pattern and deduplication gates. Four lessons that embed quotations credited to named third parties are omitted. No endorsement by Fundacja Nowoczesna Polska is implied.

data/edukacja_medialna_pl/edukacja_medialna_pl.attribution.jsonl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8522652f776bd59c115453331b4a5c771f01e6467fab2938798c3aad0065e55d
-size 347519
+oid sha256:f4263644a00b2cb0b65ecbc2bedb5a6433ba8297b9b67acd838fedf6c013e7d9
+size 341996

data/edukacja_medialna_pl/edukacja_medialna_pl.decisions.jsonl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:960bffd3e684a52a119633efbedc7d1ef0cd6d8c70bfbb16bb2c03c551da9962
-size 67322
+oid sha256:ab37785368d6b417cea223c5728c122e7bb4fe312819927561559ad7a2935b19
+size 67366

data/edukacja_medialna_pl/edukacja_medialna_pl.md

@@ -16,8 +16,8 @@ configs:
 A text-only snapshot of Polish media-literacy lesson explanations and scenarios acquired directly from official lesson XML.
 
 - Discovered: 255 lesson URLs
-- Retained: 252 records
-- Tokens: 728,987 (`cl100k_base` proxy)
+- Retained: 248 records
+- Tokens: 713,052 (`cl100k_base` proxy)
 - License: CC BY-SA 3.0 per retained record
 - Author coverage: 100.0%
 
@@ -49,7 +49,7 @@ See `NOTICE.md`, `artifacts/attribution.jsonl`, `artifacts/decisions.jsonl`, `ar
 - **Register**: prepared didactic Polish — formal written prose with classroom
   instructions, exercises and check questions; gender-inclusive paired forms
   ("uczestnicy i uczestniczki"); standard spelling. Human-authored, not OCR'd,
-  not machine-translated; 62 credited authors across 252 records (roles
+  not machine-translated; 62 credited authors across 248 records (roles
   `textbook`, `scenario`, `expert`, `methodologist`).
 - **Time span**: upstream publication dates 2012-11-09 – 2015-01-12 (per-record
   `created`). The material includes historically dated technology and legal
@@ -78,7 +78,13 @@ pinned raw XML, so the shard is reproducible:
   creator (3 dropped: `missing_creator`), body ≥ 300 chars (0 dropped), langid
   Polish (7-language set, confidence ≥ 0.99; 0 dropped), exact dedup on
   normalized text (0 dropped), near-dedup seeded MinHash 128 / LSH 0.8 with
-  exact 5-word Jaccard ≥ 0.9 (0 dropped). 252 of 255 retained.
+  exact 5-word Jaccard ≥ 0.9 (0 dropped). 252 of 255 passed these gates.
+- **Third-party quotations** (rights audit 2026-10): 4 further lessons are dropped whole by
+  `src/clean_edukacja_medialna_pl.py` (reason `third_party_quote` in `artifacts/decisions.jsonl`).
+  Two quote Wojciech Michera, credited to a University of Warsaw course
+  (`montaz-materialu-filmowego`, `scenografia-charakteryzacja-kostiumy-i-aktorzy`); two reuse an
+  exercise with six attributed quotations of named people (`kim-jestem-w-internecie`,
+  `wizerunek-w-sieci`). 248 of 255 retained.
 
 Before/after excerpts of the extraction and normalization are in the PR
 description.
@@ -105,7 +111,7 @@ The records are CC BY-SA 3.0 works; the corpus is redistributed as CC BY-SA 4.0.
 What this adds: authored didactic prose — objectives, exercises, classroom
 procedure, glossary definitions — a school-teaching register the corpus does not
 otherwise draw from (its registers are encyclopedic, legal, parliamentary,
-journalistic and fiction). 728,987 tokens is a diversity addition, not a scale
+journalistic and fiction). 713,052 tokens is a diversity addition, not a scale
 one.
 
 Bias and quality caveats: a single publisher (Fundacja Nowoczesna Polska) and 62
@@ -121,3 +127,18 @@ The source includes historically dated technology and legal guidance. Dates are
 ## Reproduction
 
 Install `src/requirements.txt`, decompress `artifacts/raw_xml.jsonl.gz` to a work directory as `raw_xml.jsonl`, and run the builder against the preserved inventory. Live acquisition creates a new source Version and must use a fresh work directory.
+
+The builder yields the 252 records contributed in PR #26. The published shard is 248: after the build, run
+`python3 src/clean_edukacja_medialna_pl.py --input <dir> --out data/edukacja_medialna_pl`, where `<dir>` holds the
+six builder outputs under their Dynaword names (`edukacja_medialna_pl.parquet`, `.attribution.jsonl`,
+`.decisions.jsonl`, `.sample.jsonl`, `.stats.json`, `.qa.json`). The script drops the four lessons listed under
+"Third-party quotations" above and rewrites those six files, statistics included. It exits before writing
+anything unless the SHA-256 of every input file equals its pin in `INPUT_SHA256`, so a rebuild that is not
+byte-identical to PR #26 stops there instead of being filtered silently; it also exits unless the quotation
+markers match exactly the four ids in `DROP`. To rerun it from the PR #26 files without rebuilding, restore them
+and run the script with its defaults:
+
+```
+git checkout 5ec3e18 -- 'data/edukacja_medialna_pl/*.parquet' 'data/edukacja_medialna_pl/*.jsonl' 'data/edukacja_medialna_pl/*.json'
+python3 src/clean_edukacja_medialna_pl.py
+```

data/edukacja_medialna_pl/edukacja_medialna_pl.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:dbce26650981e8e552c7fdac48bd35c02e9bd0a56fced4ea4871489b5a2365c4
-size 778646
+oid sha256:1de35ab7c6c4db5fad67704c3f79afa65db6bfd411a1a55291bea05a4c940222
+size 761117

data/edukacja_medialna_pl/edukacja_medialna_pl.qa.json

@@ -16,5 +16,14 @@
   "pii_pattern_matches": {
     "email": 1,
     "labelled_phone": 0
+  },
+  "third_party_quotes": {
+    "dropped": {
+      "edukacja_medialna_pl_kim-jestem-w-internecie": "candidate: six attributed quotations of named people in the profile exercise",
+      "edukacja_medialna_pl_montaz-materialu-filmowego": "confirmed: quotes Wojciech Michera, credited to a University of Warsaw course",
+      "edukacja_medialna_pl_scenografia-charakteryzacja-kostiumy-i-aktorzy": "confirmed: quotes Wojciech Michera, credited to a University of Warsaw course",
+      "edukacja_medialna_pl_wizerunek-w-sieci": "candidate: six attributed quotations of named people in the profile exercise"
+    },
+    "rule": "whole lesson dropped when it embeds a quotation credited to a named third party"
   }
 }

data/edukacja_medialna_pl/edukacja_medialna_pl.stats.json

@@ -2,10 +2,20 @@
   "acquired": 255,
   "added": "2026-09-06",
   "author_coverage": 1.0,
-  "characters": 2033576,
+  "characters": 1988751,
   "discovered": 255,
-  "kept": 252,
-  "rejected": 3,
+  "drop_third_party_quote": 4,
+  "input_sha256": {
+    "edukacja_medialna_pl.attribution.jsonl": "8522652f776bd59c115453331b4a5c771f01e6467fab2938798c3aad0065e55d",
+    "edukacja_medialna_pl.decisions.jsonl": "960bffd3e684a52a119633efbedc7d1ef0cd6d8c70bfbb16bb2c03c551da9962",
+    "edukacja_medialna_pl.parquet": "dbce26650981e8e552c7fdac48bd35c02e9bd0a56fced4ea4871489b5a2365c4",
+    "edukacja_medialna_pl.qa.json": "95e42a29bd2f1d328efac5f20378311a1806bd934b5bd88d6002a3bbf2ec78dc",
+    "edukacja_medialna_pl.sample.jsonl": "1d3277d93a5a1b9281ef1c71e745c9669eb2fac4cc78c7e60946557aedc7c1fa",
+    "edukacja_medialna_pl.stats.json": "38c7a8418b798021ce217a8bd9c9c120a7b92a7ac8e0cdbf56cb76757eb7b8cf"
+  },
+  "kept": 248,
+  "rejected": 7,
   "sample_count": 12,
-  "tokens": 728987
+  "stats_recomputed_from_parquet": true,
+  "tokens": 713052
 }

src/clean_edukacja_medialna_pl.py

@@ -0,0 +1,148 @@
+#!/usr/bin/env python3
+"""Drop lessons that embed third-party quotations from edukacja_medialna_pl (rights audit 2026-10).
+
+The audit found embedded third-party text "negligible" but listed "2-3 short quotes" and advised to
+"optionally remove the 3 quote passages". The raw XML is no longer in the checkout, so this script
+filters the pinned Parquet instead of rebuilding it. A lesson is dropped whole (fail closed), and each
+drop names the exact string that identifies the quotation:
+
+  confirmed: two lessons quote the film scholar Wojciech Michera by name ("cytat z materiałów do kursu ...
+             Uniwersytet Warszawski"); the quotation is credited to him and not to the lesson authors.
+  candidates: the audit names no third quote, so none was confirmed. The two lessons that reuse the
+             "Kim jestem w internecie" exercise carry six attributed quotations of named people
+             ("Ulubione cytaty: ... (Winston Churchill)") inside invented profiles. They are dropped
+             on the same fail-closed reasoning; delete both from DROP to keep them.
+
+`build()` exits unless the markers match exactly the DROP ids, so a changed or extended input cannot slip
+through. Kept on purpose: attributed CC BY-SA excerpts (a Wikipedia article and CC BY-SA photo credits),
+a one-sentence GMO definition cited to a ministry page, and lesson plans that only name films or books.
+
+The input is the Parquet and sidecars as contributed in PR #26 (unchanged since), pinned by SHA-256.
+Restore them before a rerun (the three globs match exactly the six pinned files, so the datasheet and
+NOTICE.md are left alone):
+  git checkout 5ec3e18 -- 'data/edukacja_medialna_pl/*.parquet' 'data/edukacja_medialna_pl/*.jsonl' 'data/edukacja_medialna_pl/*.json'
+  python3 src/clean_edukacja_medialna_pl.py
+"""
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import os
+from pathlib import Path
+
+import pyarrow as pa
+import pyarrow.parquet as pq
+
+SOURCE = "edukacja_medialna_pl"
+ROOT = Path(__file__).resolve().parents[1]
+DATA = ROOT / "data" / SOURCE
+INPUT_SHA256 = {
+    f"{SOURCE}.parquet": "dbce26650981e8e552c7fdac48bd35c02e9bd0a56fced4ea4871489b5a2365c4",
+    f"{SOURCE}.attribution.jsonl": "8522652f776bd59c115453331b4a5c771f01e6467fab2938798c3aad0065e55d",
+    f"{SOURCE}.decisions.jsonl": "960bffd3e684a52a119633efbedc7d1ef0cd6d8c70bfbb16bb2c03c551da9962",
+    f"{SOURCE}.sample.jsonl": "1d3277d93a5a1b9281ef1c71e745c9669eb2fac4cc78c7e60946557aedc7c1fa",
+    f"{SOURCE}.stats.json": "38c7a8418b798021ce217a8bd9c9c120a7b92a7ac8e0cdbf56cb76757eb7b8cf",
+    f"{SOURCE}.qa.json": "95e42a29bd2f1d328efac5f20378311a1806bd934b5bd88d6002a3bbf2ec78dc",
+}
+REASON = "third_party_quote"
+SAMPLE_SIZE = 12
+
+MICHERA = "Wojciech Michera"
+PROFILES = "Ulubione cytaty"
+# id -> (string that must occur in the lesson, why it is dropped)
+DROP = {
+    f"{SOURCE}_montaz-materialu-filmowego": (MICHERA, "confirmed: quotes Wojciech Michera, credited to a University of Warsaw course"),
+    f"{SOURCE}_scenografia-charakteryzacja-kostiumy-i-aktorzy": (MICHERA, "confirmed: quotes Wojciech Michera, credited to a University of Warsaw course"),
+    f"{SOURCE}_kim-jestem-w-internecie": (PROFILES, "candidate: six attributed quotations of named people in the profile exercise"),
+    f"{SOURCE}_wizerunek-w-sieci": (PROFILES, "candidate: six attributed quotations of named people in the profile exercise"),
+}
+
+
+def sha(value: bytes | object) -> str:
+    if not isinstance(value, bytes):
+        value = json.dumps(value, ensure_ascii=False, sort_keys=True, separators=(",", ":")).encode()
+    return hashlib.sha256(value).hexdigest()
+
+
+def read_lines(path: Path) -> list[dict]:
+    return [json.loads(line) for line in path.read_text(encoding="utf-8").splitlines() if line]
+
+
+def lines(rows: list[dict]) -> str:
+    return "".join(json.dumps(row, ensure_ascii=False, sort_keys=True) + "\n" for row in rows)
+
+
+def dump(value: dict) -> str:
+    return json.dumps(value, ensure_ascii=False, sort_keys=True, indent=2) + "\n"
+
+
+def write_text(path: Path, text: str, eol: str) -> None:
+    path.write_bytes(text.replace("\n", eol).encode("utf-8"))
+
+
+def build(input_dir: Path, out_dir: Path) -> dict:
+    paths = {name: input_dir / name for name in INPUT_SHA256}
+    for name, expected in INPUT_SHA256.items():
+        if sha(paths[name].read_bytes()) != expected:
+            raise SystemExit(f"{paths[name]} is not the PR #26 input; restore it with the command in this script's docstring")
+    # The contributed text files use CRLF; keep it so the diff shows only the dropped records.
+    eol = "\r\n"
+    if any(b"\r\n" not in paths[name].read_bytes() for name in INPUT_SHA256 if not name.endswith(".parquet")):
+        raise SystemExit("expected CRLF line endings in the PR #26 text files")
+    table = pq.read_table(paths[f"{SOURCE}.parquet"])
+    rows = table.to_pylist()
+    ids = [row["id"] for row in rows]
+    attribution = read_lines(paths[f"{SOURCE}.attribution.jsonl"])
+    decisions = read_lines(paths[f"{SOURCE}.decisions.jsonl"])
+    if len(set(ids)) != len(ids) or [r["id"] for r in attribution] != ids:
+        raise SystemExit("Parquet and attribution sidecar do not list the same ids in the same order")
+    if stale := DROP.keys() - set(ids):
+        raise SystemExit(f"DROP names ids that are not in the input: {sorted(stale)}")
+    markers = {marker for marker, _ in DROP.values()}
+    hits = {row["id"] for row in rows if any(marker in row["text"] for marker in markers)}
+    if hits != DROP.keys():
+        raise SystemExit(f"markers match {sorted(hits ^ DROP.keys())} differently from DROP")
+    if missing := [i for i, row in zip(ids, rows) if i in DROP and DROP[i][0] not in row["text"]]:
+        raise SystemExit(f"DROP markers missing from {missing}")
+
+    kept = [row for row in rows if row["id"] not in DROP]
+    kept_ids = {row["id"] for row in kept}
+    samples = sorted(kept, key=lambda row: sha(("sample:" + row["id"]).encode()))[:SAMPLE_SIZE]
+    decisions = [{**d, "selected": False, "reason": REASON} if d["id"] in DROP else d for d in decisions]
+    stats = {
+        **json.loads(paths[f"{SOURCE}.stats.json"].read_text(encoding="utf-8")),
+        "kept": len(kept), "rejected": len(decisions) - len(kept),
+        "tokens": sum(row["token_count"] for row in kept), "characters": sum(len(row["text"]) for row in kept),
+        "sample_count": len(samples), "author_coverage": sum(bool(row["author"]) for row in kept) / len(kept),
+        "drop_third_party_quote": len(DROP), "stats_recomputed_from_parquet": True, "input_sha256": INPUT_SHA256,
+    }
+    qa = {**json.loads(paths[f"{SOURCE}.qa.json"].read_text(encoding="utf-8")),
+          "third_party_quotes": {"rule": "whole lesson dropped when it embeds a quotation credited to a named third party",
+                                 "dropped": {i: why for i, (_, why) in sorted(DROP.items())}}}
+    staged = {
+        f"{SOURCE}.parquet": lambda p: pq.write_table(pa.Table.from_pylist(kept, schema=table.schema), p, compression="zstd"),
+        f"{SOURCE}.attribution.jsonl": lambda p: write_text(p, lines([r for r in attribution if r["id"] in kept_ids]), eol),
+        f"{SOURCE}.decisions.jsonl": lambda p: write_text(p, lines(decisions), eol),
+        f"{SOURCE}.sample.jsonl": lambda p: write_text(p, lines(samples), eol),
+        f"{SOURCE}.stats.json": lambda p: write_text(p, dump(stats), eol),
+        f"{SOURCE}.qa.json": lambda p: write_text(p, dump(qa), eol),
+    }
+    out_dir.mkdir(parents=True, exist_ok=True)
+    for name, write in staged.items():
+        write(out_dir / (name + ".tmp"))
+    for name in staged:
+        os.replace(out_dir / (name + ".tmp"), out_dir / name)
+    return {**stats, "dropped": {i: why for i, (_, why) in sorted(DROP.items())}}
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__.split("\n")[0])
+    parser.add_argument("--input", type=Path, default=DATA, help="directory holding the PR #26 Parquet and sidecars")
+    parser.add_argument("--out", type=Path, default=DATA)
+    args = parser.parse_args()
+    print(json.dumps(build(args.input, args.out), indent=2, ensure_ascii=False))
+
+
+if __name__ == "__main__":
+    main()

src/test_edukacja_medialna_contribution.py

@@ -5,6 +5,8 @@ import re
 
 import pyarrow.parquet as pq
 
+from clean_edukacja_medialna_pl import INPUT_SHA256
+
 SOURCE = "edukacja_medialna_pl"
 LICENSE = "CC-BY-SA-3.0"
 FIELDS = ["id", "text", "source", "added", "created", "token_count", "license", "author"]
@@ -48,19 +50,22 @@ def test_ontology_is_addressed_and_falsifiable():
     manifest = json.loads(MANIFEST.read_text(encoding="utf-8"))
     evidence = {item["id"]: item for item in manifest["evidence"]}
     checks = next(item["payload"] for item in manifest["evidence"] if item["observation_type"] == "checksums")
+    # The manifest pins the contributed build. src/clean_edukacja_medialna_pl.py later rewrote the data files,
+    # so their pins must equal the input hashes that script checks (the cleaned files are checked in
+    # test_edukacja_medialna_pl_contract.py). This test file changed with the cleaning and cannot pin itself.
+    contributed = {
+        "data/train-00000-of-00001.parquet": f"{SOURCE}.parquet",
+        "artifacts/attribution.jsonl": f"{SOURCE}.attribution.jsonl",
+        "artifacts/decisions.jsonl": f"{SOURCE}.decisions.jsonl",
+        "artifacts/sample.jsonl": f"{SOURCE}.sample.jsonl",
+        "artifacts/stats.json": f"{SOURCE}.stats.json",
+        "artifacts/qa.json": f"{SOURCE}.qa.json",
+    }
     for name, expected in checks.items():
-        mapping = {
-            "data/train-00000-of-00001.parquet": DATA / f"{SOURCE}.parquet",
-            "artifacts/attribution.jsonl": DATA / f"{SOURCE}.attribution.jsonl",
-            "artifacts/decisions.jsonl": DATA / f"{SOURCE}.decisions.jsonl",
-            "artifacts/sample.jsonl": DATA / f"{SOURCE}.sample.jsonl",
-            "artifacts/stats.json": DATA / f"{SOURCE}.stats.json",
-            "artifacts/qa.json": DATA / f"{SOURCE}.qa.json",
-            "src/build_edukacja_medialna_pl.py": ROOT / "src/build_edukacja_medialna_pl.py",
-            "src/test_edukacja_medialna_contribution.py": ROOT / "src/test_edukacja_medialna_contribution.py",
-        }
-        if name in mapping:
-            assert digest(mapping[name].read_bytes()) == expected
+        if name in contributed:
+            assert INPUT_SHA256[contributed[name]] == expected
+        elif name == "src/build_edukacja_medialna_pl.py":
+            assert digest((ROOT / name).read_bytes()) == expected
     assert all(item["falsification_condition"] for item in manifest["claims"])
     assert all(set(item["supported_by"]) <= set(evidence) for item in manifest["claims"])
     assert any(item["predicate"] == "VALIDATED_AGAINST" for item in manifest["relations"])

src/test_edukacja_medialna_pl_contract.py

@@ -0,0 +1,138 @@
+"""Contract checks for the third-party-quote drop in edukacja_medialna_pl (src/clean_edukacja_medialna_pl.py)."""
+
+import json
+from pathlib import Path
+
+import pyarrow as pa
+import pyarrow.parquet as pq
+import pytest
+
+import clean_edukacja_medialna_pl as clean
+
+SOURCE = "edukacja_medialna_pl"
+DATA = Path(__file__).resolve().parents[1] / "data" / SOURCE
+# 255 lessons discovered, 3 without a creator, 4 dropped for quotations of named third parties.
+KEPT, TOKENS, CHARACTERS = 248, 713_052, 1_988_751
+# Confirmed: lessons that quote Wojciech Michera by name. Candidates: the "Kim jestem w internecie" exercise.
+CONFIRMED = {f"{SOURCE}_montaz-materialu-filmowego", f"{SOURCE}_scenografia-charakteryzacja-kostiumy-i-aktorzy"}
+CANDIDATES = {f"{SOURCE}_kim-jestem-w-internecie", f"{SOURCE}_wizerunek-w-sieci"}
+
+
+def jsonl(name):
+    return [json.loads(line) for line in (DATA / f"{SOURCE}.{name}.jsonl").read_text(encoding="utf-8").splitlines() if line]
+
+
+def test_the_four_quoting_lessons_are_gone_and_recorded_as_rejected():
+    assert clean.DROP.keys() == CONFIRMED | CANDIDATES
+    rows = pq.read_table(DATA / f"{SOURCE}.parquet").to_pylist()
+    ids = {row["id"] for row in rows}
+    assert not (CONFIRMED | CANDIDATES) & ids
+    assert not (CONFIRMED | CANDIDATES) & {r["id"] for r in jsonl("attribution")}
+    assert not (CONFIRMED | CANDIDATES) & {r["id"] for r in jsonl("sample")}
+    decisions = {d["id"]: d for d in jsonl("decisions")}
+    assert {i for i, d in decisions.items() if d["reason"] == clean.REASON} == CONFIRMED | CANDIDATES
+    assert not any(decisions[i]["selected"] for i in CONFIRMED | CANDIDATES)
+    # No lesson that is kept still carries one of the quotation markers.
+    assert [r["id"] for r in rows if any(marker in r["text"] for marker, _ in clean.DROP.values())] == []
+
+
+def test_counts_stats_and_sidecars_agree():
+    table = pq.read_table(DATA / f"{SOURCE}.parquet")
+    rows = table.to_pylist()
+    stats = json.loads((DATA / f"{SOURCE}.stats.json").read_text(encoding="utf-8"))
+    qa = json.loads((DATA / f"{SOURCE}.qa.json").read_text(encoding="utf-8"))
+    attribution, decisions, sample = jsonl("attribution"), jsonl("decisions"), jsonl("sample")
+    ids = [row["id"] for row in rows]
+    assert table.column_names == ["id", "text", "source", "added", "created", "token_count", "license", "author"]
+    assert len(rows) == stats["kept"] == len(attribution) == sum(d["selected"] for d in decisions) == KEPT
+    assert stats["tokens"] == sum(r["token_count"] for r in rows) == TOKENS
+    assert stats["characters"] == sum(len(r["text"]) for r in rows) == CHARACTERS
+    assert stats["discovered"] == stats["acquired"] == len(decisions) == 255
+    assert stats["rejected"] == len(decisions) - KEPT == 3 + stats["drop_third_party_quote"] == 7
+    assert [a["id"] for a in attribution] == ids and len(set(ids)) == len(ids)
+    assert {r["license"] for r in rows} == {a["license"] for a in attribution} == {"CC-BY-SA-3.0"}
+    assert len(sample) == stats["sample_count"] == 12 and all(s in rows for s in sample)
+    assert sample == sorted(rows, key=lambda r: clean.sha(("sample:" + r["id"]).encode()))[:12]
+    assert stats["input_sha256"] == clean.INPUT_SHA256
+    assert qa["third_party_quotes"]["dropped"].keys() == clean.DROP.keys()
+
+
+def write_input(directory, rows):
+    """A miniature input with CRLF text files, plus the pins for it."""
+    directory.mkdir(parents=True)
+    names = {kind: directory / f"{SOURCE}.{kind}" for kind in ("parquet", "attribution.jsonl", "decisions.jsonl", "sample.jsonl", "stats.json", "qa.json")}
+    schema = pa.schema([(f, pa.int64() if f == "token_count" else pa.string()) for f in clean.pq.read_table(DATA / f"{SOURCE}.parquet").column_names])
+    pq.write_table(pa.Table.from_pylist(rows, schema=schema), names["parquet"])
+    text = {
+        "attribution.jsonl": clean.lines([{"id": r["id"], "license": r["license"]} for r in rows]),
+        "decisions.jsonl": clean.lines([{"id": r["id"], "selected": True, "reason": "include", "page_sha256": "p", "xml_sha256": "x"} for r in rows]),
+        "sample.jsonl": clean.lines(rows),
+        "stats.json": clean.dump({"acquired": len(rows), "discovered": len(rows), "kept": len(rows), "rejected": 0, "added": "2026-09-06"}),
+        "qa.json": clean.dump({"license_gate": "x"}),
+    }
+    for kind, content in text.items():
+        clean.write_text(names[kind], content, "\r\n")
+    return {path.name: clean.sha(path.read_bytes()) for path in names.values()}
+
+
+def row(slug, text):
+    return {"id": f"{SOURCE}_{slug}", "text": text, "source": SOURCE, "added": "2026-09-06", "created": "2013-01-01",
+            "token_count": len(text.split()), "license": "CC-BY-SA-3.0", "author": "Anna Autor"}
+
+
+@pytest.fixture
+def mini(tmp_path, monkeypatch):
+    rows = [row("a", "plain lesson"), row("b", "as Wojciech Michera said"), row("c", "another plain lesson here")]
+    pins = write_input(tmp_path / "in", rows)
+    monkeypatch.setattr(clean, "INPUT_SHA256", pins)
+    monkeypatch.setattr(clean, "DROP", {f"{SOURCE}_b": ("Wojciech Michera", "confirmed: test")})
+    return tmp_path
+
+
+def test_build_drops_the_listed_lesson_and_rewrites_every_file(mini):
+    out = mini / "out"
+    stats = clean.build(mini / "in", out)
+    assert pq.read_table(out / f"{SOURCE}.parquet").column("id").to_pylist() == [f"{SOURCE}_a", f"{SOURCE}_c"]
+    assert (stats["kept"], stats["rejected"], stats["tokens"], stats["drop_third_party_quote"]) == (2, 1, 6, 1)
+    decisions = {d["id"]: d for d in map(json.loads, (out / f"{SOURCE}.decisions.jsonl").read_text(encoding="utf-8").splitlines())}
+    assert decisions[f"{SOURCE}_b"]["reason"] == clean.REASON and not decisions[f"{SOURCE}_b"]["selected"]
+    assert decisions[f"{SOURCE}_a"]["selected"]
+    assert b"\r\n" in (out / f"{SOURCE}.stats.json").read_bytes()
+    assert not list(out.glob("*.tmp"))
+
+
+def test_build_refuses_an_input_that_is_not_the_pinned_one(mini):
+    (mini / "in" / f"{SOURCE}.qa.json").write_bytes(b"{}\r\n")
+    with pytest.raises(SystemExit, match="not the PR #26 input"):
+        clean.build(mini / "in", mini / "out")
+    assert not (mini / "out").exists()
+
+
+def test_build_fails_closed_on_a_stale_id_or_an_unlisted_quotation(mini, monkeypatch):
+    monkeypatch.setattr(clean, "DROP", {f"{SOURCE}_zzz": ("Wojciech Michera", "stale")})
+    with pytest.raises(SystemExit, match="not in the input"):
+        clean.build(mini / "in", mini / "out")
+    # The marker also matches lesson b, which DROP does not list.
+    monkeypatch.setattr(clean, "DROP", {f"{SOURCE}_a": ("Wojciech Michera", "wrong lesson")})
+    with pytest.raises(SystemExit, match="differently from DROP"):
+        clean.build(mini / "in", mini / "out")
+
+
+def test_build_refuses_the_already_cleaned_data(tmp_path):
+    with pytest.raises(SystemExit, match="not the PR #26 input"):
+        clean.build(DATA, tmp_path / "out")
+
+
+def test_datasheet_reproduction_steps_include_the_clean_script_and_its_pin_check():
+    text = (DATA / f"{SOURCE}.md").read_text(encoding="utf-8")
+    assert text.count("## Reproduction") == 1
+    reproduction = text.split("## Reproduction")[1]
+    for needle in ("src/clean_edukacja_medialna_pl.py", "INPUT_SHA256", "DROP", "5ec3e18"):
+        assert needle in reproduction
+
+
+def test_restore_command_is_the_same_in_docstring_and_datasheet_and_touches_only_the_pinned_files():
+    command = next(line.strip() for line in clean.__doc__.splitlines() if line.strip().startswith("git checkout"))
+    assert command in (DATA / f"{SOURCE}.md").read_text(encoding="utf-8")
+    globs = [arg.strip("'").rsplit("/", 1)[1] for arg in command.split(" -- ")[1].split()]
+    assert sorted(p.name for pattern in globs for p in DATA.glob(pattern)) == sorted(clean.INPUT_SHA256)