Add pinned SAOS Polish court judgments snapshot

artifacts/saos_ontology_manifest.json

@@ -0,0 +1,569 @@
+{
+  "actors": [
+    {
+      "id": "hf:PiotrSty",
+      "identifiers": {
+        "github": "PiotrStyla",
+        "hugging_face": "PiotrSty"
+      },
+      "identity": "Piotr Styła",
+      "kind": "human"
+    }
+  ],
+  "attestations": [
+    {
+      "actor_id": "hf:PiotrSty",
+      "created_at": "2026-09-05T06:45:57.253621Z",
+      "id": "slayer://attestation/saos/schema_valid@sha256:076e1dda8918af803ecf9d1dc74f69536bf5e8e481f9a18c605f1237097db76b",
+      "subject_version": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76",
+      "supported_by": [
+        "slayer://evidence/saos/schema_validation@sha256:87209fd66de594f15ee22a7a34b851fb491397fa52624d3a392ef38e63f806d1",
+        "slayer://evidence/saos/artifact_checksums@sha256:c80bd061eadc8c631f2746668334ec0da3a1b7acd33ef7c05e6d5334b0ed0e0c"
+      ],
+      "type": "schema_valid",
+      "value": true
+    },
+    {
+      "actor_id": "hf:PiotrSty",
+      "created_at": "2026-09-05T06:45:57.253621Z",
+      "id": "slayer://attestation/saos/source_completeness@sha256:8cd33e860f10f49f9f19e3ad9c146003e5161b35d4b49638a433617e36c045ed",
+      "subject_version": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76",
+      "supported_by": [
+        "slayer://evidence/saos/source_completeness_observation@sha256:58e8c4938862d5a6dec240d8fe0fa8a515fb94a3439e3472d9ba520e198b6706"
+      ],
+      "type": "source_completeness",
+      "value": "partial_snapshot"
+    },
+    {
+      "actor_id": "hf:PiotrSty",
+      "created_at": "2026-09-05T06:45:57.253621Z",
+      "id": "slayer://attestation/saos/provenance@sha256:007a8ee8597184b32e11500266c0441dfa8952a195387625f361d8c8f3537f33",
+      "subject_version": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76",
+      "supported_by": [
+        "slayer://evidence/saos/pinned_source_inventory@sha256:ecbbb0fbf442e371475f6f6a89909522a6524d9464be5f113765272297670459"
+      ],
+      "type": "provenance",
+      "value": "pinned_and_hashed"
+    },
+    {
+      "actor_id": "hf:PiotrSty",
+      "created_at": "2026-09-05T06:45:57.253621Z",
+      "id": "slayer://attestation/saos/pii_scrubbing@sha256:9ee1a64e2440b1465e2412ef10e6470502b8fb730a9e50d663587d3cd8918787",
+      "subject_version": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76",
+      "supported_by": [
+        "slayer://evidence/saos/normalization_and_pii_metrics@sha256:ad762d906f3391e1b818654a8fdbc1bdb3c8c022e83997abea972251ee65e37a"
+      ],
+      "type": "pii_scrubbing",
+      "value": "pattern_scrub_applied_not_guaranteed"
+    },
+    {
+      "actor_id": "hf:PiotrSty",
+      "created_at": "2026-09-05T06:45:57.253621Z",
+      "id": "slayer://attestation/saos/legal_status@sha256:9233cc03387de9175ffa86f6bddc38192920938f65ce9f1c1934a4ded573cab7",
+      "subject_version": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76",
+      "supported_by": [
+        "slayer://evidence/saos/legal_provenance_review@sha256:958944289fd5a323346e43bc9e11438b6c408b06fe9b288a0449bf1afaf58fd3"
+      ],
+      "type": "legal_status",
+      "value": "review_needed_documented_official_materials"
+    }
+  ],
+  "claim_evidence": [
+    {
+      "claim_id": "slayer://claim/saos@sha256:f3dcf4db368c7cb7cec7913fdcc18242373137ee70f5a5c04eb6f3400e6f28fe",
+      "evidence_id": "slayer://evidence/saos/pinned_source_inventory@sha256:ecbbb0fbf442e371475f6f6a89909522a6524d9464be5f113765272297670459",
+      "introduced_by_run": "slayer://run/dataset-filtering/saos-snapshot-20260904",
+      "relation": "SUPPORTS"
+    },
+    {
+      "claim_id": "slayer://claim/saos@sha256:aacbbe6a634fb32bbd6ec8d85e9ba090fe4faef389118f5730c7948f961ba180",
+      "evidence_id": "slayer://evidence/saos/source_completeness_observation@sha256:58e8c4938862d5a6dec240d8fe0fa8a515fb94a3439e3472d9ba520e198b6706",
+      "introduced_by_run": "slayer://run/dataset-filtering/saos-snapshot-20260904",
+      "relation": "SUPPORTS"
+    },
+    {
+      "claim_id": "slayer://claim/saos@sha256:0fa5f4ef3582a00733d6b62b2bd377027f6734e6669a8a3c63ad0dd6f7399d04",
+      "evidence_id": "slayer://evidence/saos/schema_validation@sha256:87209fd66de594f15ee22a7a34b851fb491397fa52624d3a392ef38e63f806d1",
+      "introduced_by_run": "slayer://run/dataset-filtering/saos-snapshot-20260904",
+      "relation": "SUPPORTS"
+    },
+    {
+      "claim_id": "slayer://claim/saos@sha256:0fa5f4ef3582a00733d6b62b2bd377027f6734e6669a8a3c63ad0dd6f7399d04",
+      "evidence_id": "slayer://evidence/saos/artifact_checksums@sha256:c80bd061eadc8c631f2746668334ec0da3a1b7acd33ef7c05e6d5334b0ed0e0c",
+      "introduced_by_run": "slayer://run/dataset-filtering/saos-snapshot-20260904",
+      "relation": "SUPPORTS"
+    },
+    {
+      "claim_id": "slayer://claim/saos@sha256:86781759c2cbf09530bf68b14b35ddb8f8de875515d10492120fb50a21ffc125",
+      "evidence_id": "slayer://evidence/saos/normalization_and_pii_metrics@sha256:ad762d906f3391e1b818654a8fdbc1bdb3c8c022e83997abea972251ee65e37a",
+      "introduced_by_run": "slayer://run/dataset-filtering/saos-snapshot-20260904",
+      "relation": "SUPPORTS"
+    },
+    {
+      "claim_id": "slayer://claim/saos@sha256:869bd1c50beb64f64ed605509b5d766f2fabf71d853cc4b2897f1f8010853266",
+      "evidence_id": "slayer://evidence/saos/legal_provenance_review@sha256:958944289fd5a323346e43bc9e11438b6c408b06fe9b288a0449bf1afaf58fd3",
+      "introduced_by_run": "slayer://run/dataset-filtering/saos-snapshot-20260904",
+      "relation": "SUPPORTS"
+    }
+  ],
+  "claims": [
+    {
+      "asserted_by": "hf:PiotrSty",
+      "falsification_condition": "A source file is missing or its recorded size or SHA-256 differs.",
+      "id": "slayer://claim/saos@sha256:f3dcf4db368c7cb7cec7913fdcc18242373137ee70f5a5c04eb6f3400e6f28fe",
+      "scope": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76",
+      "statement": "The contribution is derived from the 14 files of the pinned PiotrSty SAOS snapshot.",
+      "supported_by": [
+        "slayer://evidence/saos/pinned_source_inventory@sha256:ecbbb0fbf442e371475f6f6a89909522a6524d9464be5f113765272297670459"
+      ]
+    },
+    {
+      "asserted_by": "hf:PiotrSty",
+      "falsification_condition": "The pinned snapshot row count equals the official live count observed for this run.",
+      "id": "slayer://claim/saos@sha256:aacbbe6a634fb32bbd6ec8d85e9ba090fe4faef389118f5730c7948f961ba180",
+      "scope": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76",
+      "statement": "The contributed dataset is a partial historical SAOS snapshot and not a complete current copy.",
+      "supported_by": [
+        "slayer://evidence/saos/source_completeness_observation@sha256:58e8c4938862d5a6dec240d8fe0fa8a515fb94a3439e3472d9ba520e198b6706"
+      ]
+    },
+    {
+      "asserted_by": "hf:PiotrSty",
+      "falsification_condition": "A retained row violates the schema or an artifact checksum differs.",
+      "id": "slayer://claim/saos@sha256:0fa5f4ef3582a00733d6b62b2bd377027f6734e6669a8a3c63ad0dd6f7399d04",
+      "scope": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76",
+      "statement": "All retained rows conform to the DynaWord source schema and recorded artifact checksums.",
+      "supported_by": [
+        "slayer://evidence/saos/schema_validation@sha256:87209fd66de594f15ee22a7a34b851fb491397fa52624d3a392ef38e63f806d1",
+        "slayer://evidence/saos/artifact_checksums@sha256:c80bd061eadc8c631f2746668334ec0da3a1b7acd33ef7c05e6d5334b0ed0e0c"
+      ]
+    },
+    {
+      "asserted_by": "hf:PiotrSty",
+      "falsification_condition": "The committed protocol omits a documented transformation or a retained match contradicts its recorded metric.",
+      "id": "slayer://claim/saos@sha256:86781759c2cbf09530bf68b14b35ddb8f8de875515d10492120fb50a21ffc125",
+      "scope": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76",
+      "statement": "Pattern-based PII scrubbing and documented text normalization were applied, without asserting perfect de-identification.",
+      "supported_by": [
+        "slayer://evidence/saos/normalization_and_pii_metrics@sha256:ad762d906f3391e1b818654a8fdbc1bdb3c8c022e83997abea972251ee65e37a"
+      ]
+    },
+    {
+      "asserted_by": "hf:PiotrSty",
+      "falsification_condition": "The records are not official materials or legal review rejects the documented statutory basis.",
+      "id": "slayer://claim/saos@sha256:869bd1c50beb64f64ed605509b5d766f2fabf71d853cc4b2897f1f8010853266",
+      "scope": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76",
+      "statement": "The records are court materials with a documented statutory reuse argument, pending formal legal review and subject to stated limitations.",
+      "supported_by": [
+        "slayer://evidence/saos/legal_provenance_review@sha256:958944289fd5a323346e43bc9e11438b6c408b06fe9b288a0449bf1afaf58fd3"
+      ]
+    }
+  ],
+  "evidence": [
+    {
+      "id": "slayer://evidence/saos/dataset_record_counts@sha256:06d786acc8282a7cf7a6257dac50ed31d1cea8bf859a7bb767de22ac24d53650",
+      "observation_type": "dataset_record_counts",
+      "observed_at": "2026-09-05T06:45:57.253621Z",
+      "payload": {
+        "added": "2026-09-04",
+        "authors_with_value": 279091,
+        "by_court_type": {
+          "COMMON": 229370,
+          "CONSTITUTIONAL_TRIBUNAL": 9503,
+          "NATIONAL_APPEAL_CHAMBER": 19758,
+          "SUPREME": 38058
+        },
+        "by_judgment_type": {
+          "DECISION": 64761,
+          "REASONS": 18693,
+          "REGULATION": 3231,
+          "RESOLUTION": 2469,
+          "SENTENCE": 207535
+        },
+        "chars": 5897001250,
+        "created_max": "2017-05-11",
+        "created_min": "1986-05-28",
+        "drop_dup": 0,
+        "drop_lang": 0,
+        "drop_short": 3,
+        "invalid_dates": 1,
+        "kept": 296689,
+        "license": "public-domain (official documents)",
+        "licenses": {
+          "public-domain (official documents)": 296689
+        },
+        "normalization": {
+          "bank_accounts_redacted": 66,
+          "control_characters_removed": 0,
+          "emails_redacted": 14,
+          "labelled_identifiers_redacted": 253,
+          "labelled_phones_redacted": 60,
+          "normalized_records": 263987,
+          "pesel_candidates_redacted": 0
+        },
+        "official_count_observed_at": "2026-09-04",
+        "official_total_results": 540726,
+        "snapshot_live_count_ratio": 0.5486919437940843,
+        "source_revision": "736cba4f92999e29d3e49409b0713c88fc683714",
+        "source_rows": 296692,
+        "stats_recomputed_from_parquet": true,
+        "target_base_revision": "a9d134ac79a9fe586fc9c11cb39eb32d69dcd21f",
+        "tokens": 2265448504
+      },
+      "run_id": "slayer://run/dataset-filtering/saos-snapshot-20260904",
+      "subject_version": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76"
+    },
+    {
+      "id": "slayer://evidence/saos/artifact_checksums@sha256:c80bd061eadc8c631f2746668334ec0da3a1b7acd33ef7c05e6d5334b0ed0e0c",
+      "observation_type": "artifact_checksums",
+      "observed_at": "2026-09-05T06:45:57.253621Z",
+      "payload": {
+        "attribution_sha256": "0fda03e770bc8db94694f4758239ea6fda80e16e5398aee2b5a308a48896c857",
+        "datasheet_sha256": "2b18ec4c003fdd66c07d773d378703e54e2f4d5d3579e80a7c397a6b5edf7478",
+        "legal_evidence_sha256": "5edf35c7c62b686ed7aaa7b6b569262a7f0faa46d7ca639bb4d6df909124b445",
+        "parquet_sha256": "f265dfe410aef0b0a2453f16e05d6398292af6a66db713bfaad554b4ee0dbf55",
+        "source_inventory_sha256": "48c922173253bd9045037ec534b6a11f53fa7bb7e8f6731f8e1420df826826e8",
+        "source_profile_sha256": "ed0418723c3e895b1e9de3da9fffbb807f7231a8252aacee10998f16daef4e31",
+        "stats_sha256": "7b1ea5b70d4bf67da038ff8d4b68bc2c01bea33554d1f8e8b31f3fd8785e7f1c"
+      },
+      "run_id": "slayer://run/dataset-filtering/saos-snapshot-20260904",
+      "subject_version": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76"
+    },
+    {
+      "id": "slayer://evidence/saos/pinned_source_inventory@sha256:ecbbb0fbf442e371475f6f6a89909522a6524d9464be5f113765272297670459",
+      "observation_type": "pinned_source_inventory",
+      "observed_at": "2026-09-05T06:45:57.253621Z",
+      "payload": {
+        "aggregate_digest": "sha256:d1e1743a9901c381e6617391f65ca2686e317dd7c3d8b3552093e0104040c88c",
+        "files": [
+          {
+            "path": "data/train-00000-of-00014.parquet",
+            "sha256": "f76b008a2061dc85589a51f4c868cc6484c9dff57a09b53e72b3d5390cbb632f",
+            "size": 180401965
+          },
+          {
+            "path": "data/train-00001-of-00014.parquet",
+            "sha256": "0641823cad2fc981da1f0398c4202e25e6489994828812ef8bf3701bbb254930",
+            "size": 213080176
+          },
+          {
+            "path": "data/train-00002-of-00014.parquet",
+            "sha256": "f707bc330140dac0052f5e169cb6dff5918948877f1b4dd471c3070de0230593",
+            "size": 243953893
+          },
+          {
+            "path": "data/train-00003-of-00014.parquet",
+            "sha256": "f72f79df6fe494fd91cbb82e63083fb375e18a59ad772232a719fa46fb9cbe3a",
+            "size": 239361563
+          },
+          {
+            "path": "data/train-00004-of-00014.parquet",
+            "sha256": "c5d2ccc87ff2a3d8f7ba64b24f689d92e3def50a70792f59446f58bc2ac04a90",
+            "size": 240427543
+          },
+          {
+            "path": "data/train-00005-of-00014.parquet",
+            "sha256": "9139c736363b6c151373fd159c0f30cc06fe65d52c01acc547737463f5500fc7",
+            "size": 234121713
+          },
+          {
+            "path": "data/train-00006-of-00014.parquet",
+            "sha256": "6b0d88e3ef45b0c246b79aa47aa1fc337c997305159f1d6e84e6e95798995495",
+            "size": 242598279
+          },
+          {
+            "path": "data/train-00007-of-00014.parquet",
+            "sha256": "4b1c3a69948d1ff713a0dd39d71ddb9d4cd85156a947c2b8e4c55264ed947cc0",
+            "size": 253305954
+          },
+          {
+            "path": "data/train-00008-of-00014.parquet",
+            "sha256": "3a4298ee101bf2b9fc6a339db0f4ce7471519b5e36c828ff6166617983d08c60",
+            "size": 256220979
+          },
+          {
+            "path": "data/train-00009-of-00014.parquet",
+            "sha256": "41b963dd3669d5bf97f6e381019483ac52674f2889cd31efe30928ca02ab0275",
+            "size": 257319050
+          },
+          {
+            "path": "data/train-00010-of-00014.parquet",
+            "sha256": "960243ec221eac979ee226aeb630c9f9325af774e703e8af6285386f06b57541",
+            "size": 257444913
+          },
+          {
+            "path": "data/train-00011-of-00014.parquet",
+            "sha256": "d8c54ec1698d9ab53e86b4064401efda174cd4b9580b256fac8bb0ada79ad886",
+            "size": 260380321
+          },
+          {
+            "path": "data/train-00012-of-00014.parquet",
+            "sha256": "fe294f0ef3b851928cc37a22d29c66dfd55c4951a09b76f5cb89a6b64082c762",
+            "size": 266750435
+          },
+          {
+            "path": "data/train-00013-of-00014.parquet",
+            "sha256": "5c27b08e9559bf0ba8d63ed7f1daa7f19db0cd4fcb774d715e7d046e34a373b2",
+            "size": 273748222
+          }
+        ],
+        "repository": "PiotrSty/saos-polish-court-judgments",
+        "revision": "736cba4f92999e29d3e49409b0713c88fc683714"
+      },
+      "run_id": "slayer://run/dataset-filtering/saos-snapshot-20260904",
+      "subject_version": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76"
+    },
+    {
+      "id": "slayer://evidence/saos/source_completeness_observation@sha256:58e8c4938862d5a6dec240d8fe0fa8a515fb94a3439e3472d9ba520e198b6706",
+      "observation_type": "source_completeness_observation",
+      "observed_at": "2026-09-05T06:45:57.253621Z",
+      "payload": {
+        "complete": false,
+        "observed_at": "2026-09-04",
+        "official_total_results": 540726,
+        "ratio": 0.5486919437940843,
+        "snapshot_rows": 296692
+      },
+      "run_id": "slayer://run/dataset-filtering/saos-snapshot-20260904",
+      "subject_version": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76"
+    },
+    {
+      "id": "slayer://evidence/saos/normalization_and_pii_metrics@sha256:ad762d906f3391e1b818654a8fdbc1bdb3c8c022e83997abea972251ee65e37a",
+      "observation_type": "normalization_and_pii_metrics",
+      "observed_at": "2026-09-05T06:45:57.253621Z",
+      "payload": {
+        "bank_accounts_redacted": 66,
+        "control_characters_removed": 0,
+        "emails_redacted": 14,
+        "labelled_identifiers_redacted": 253,
+        "labelled_phones_redacted": 60,
+        "normalized_records": 263987,
+        "pesel_candidates_redacted": 0
+      },
+      "run_id": "slayer://run/dataset-filtering/saos-snapshot-20260904",
+      "subject_version": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76"
+    },
+    {
+      "id": "slayer://evidence/saos/schema_validation@sha256:87209fd66de594f15ee22a7a34b851fb491397fa52624d3a392ef38e63f806d1",
+      "observation_type": "schema_validation",
+      "observed_at": "2026-09-05T06:45:57.253621Z",
+      "payload": {
+        "records": 296689,
+        "schema": [
+          "id",
+          "text",
+          "source",
+          "added",
+          "created",
+          "token_count",
+          "license",
+          "author"
+        ],
+        "valid": true
+      },
+      "run_id": "slayer://run/dataset-filtering/saos-snapshot-20260904",
+      "subject_version": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76"
+    },
+    {
+      "id": "slayer://evidence/saos/legal_provenance_review@sha256:958944289fd5a323346e43bc9e11438b6c408b06fe9b288a0449bf1afaf58fd3",
+      "observation_type": "legal_provenance_review",
+      "observed_at": "2026-09-05T06:45:57.253621Z",
+      "payload": {
+        "artifact": "data/saos/saos.license-evidence.json",
+        "formal_legal_review": "not_performed",
+        "sha256": "5edf35c7c62b686ed7aaa7b6b569262a7f0faa46d7ca639bb4d6df909124b445",
+        "status": "documented-statutory-basis-review-needed"
+      },
+      "run_id": "slayer://run/dataset-filtering/saos-snapshot-20260904",
+      "subject_version": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76"
+    }
+  ],
+  "object": {
+    "id": "slayer://object/dataset-source/saos",
+    "kind": "dataset_source",
+    "metadata": {
+      "domain": "legal/judicial",
+      "language": "pl",
+      "provider": "SAOS"
+    },
+    "name": "saos"
+  },
+  "ontology_version": "slayer.ai/research-ontology/v0.1",
+  "protocol": {
+    "id": "slayer://object/protocol/saos-filter",
+    "kind": "dataset_filtering",
+    "name": "saos-filter",
+    "specification": {
+      "configuration": {
+        "date_valid_year_range": [
+          1900,
+          2027
+        ],
+        "deduplication": {
+          "algorithm": "sha1",
+          "field": "normalized_text",
+          "scope": "within_source"
+        },
+        "execution": {
+          "parquet_compression": "zstd",
+          "record_batch_size": 128,
+          "source_workers": 14,
+          "token_threads": 14,
+          "token_threads_per_worker": 1
+        },
+        "minimum_characters": 200,
+        "minimum_polish_diacritic_ratio": 0.005,
+        "normalization": [
+          "NFKC",
+          "space_quote_dash",
+          "paragraph_preserving_whitespace",
+          "pattern_PII_scrub"
+        ],
+        "source_revision": "736cba4f92999e29d3e49409b0713c88fc683714",
+        "tokenizer": "cl100k_base"
+      },
+      "entrypoint": "python src/fetch_saos.py",
+      "outputs": [
+        "DynaWord Parquet",
+        "attribution sidecar",
+        "datasheet",
+        "stats",
+        "legal evidence",
+        "ontology manifest"
+      ],
+      "required_inputs": [
+        "14 pinned source Parquet shards",
+        "source profile",
+        "target registry",
+        "target findings log"
+      ],
+      "source_path": "src/fetch_saos.py",
+      "summary": "Deterministic normalization and projection of a pinned SAOS judgment snapshot."
+    },
+    "version": {
+      "code_sha256": "a73afbe39a10beb4ac22c773410479b804837e3e78ccecf5a6e728531845c12d",
+      "created_at": "2026-09-05T06:45:57.253621Z",
+      "created_by": "hf:PiotrSty",
+      "digest": "sha256:33de2cc5f9e08a3b22fd6f4303e570a951d184cc8de91aca61ab216250530b3c",
+      "id": "slayer://version/protocol/saos-filter@sha256:33de2cc5f9e08a3b22fd6f4303e570a951d184cc8de91aca61ab216250530b3c",
+      "object_id": "slayer://object/protocol/saos-filter",
+      "payload_uri": "hf://datasets/SlayerLab/polish-dynaword@b28db4c2b1f7273f878b3c54602777f7b224e27b/src/fetch_saos.py",
+      "schema_version": "slayer.ai/protocol/v1"
+    }
+  },
+  "relations": [
+    {
+      "introduced_by_run": "slayer://run/dataset-filtering/saos-snapshot-20260904",
+      "predicate": "DERIVED_FROM",
+      "source_version_id": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76",
+      "target_version_id": "hf://datasets/PiotrSty/saos-polish-court-judgments@736cba4f92999e29d3e49409b0713c88fc683714"
+    },
+    {
+      "introduced_by_run": "slayer://run/dataset-filtering/saos-snapshot-20260904",
+      "predicate": "COMPATIBLE_WITH",
+      "source_version_id": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76",
+      "target_version_id": "hf://datasets/SlayerLab/polish-dynaword@a9d134ac79a9fe586fc9c11cb39eb32d69dcd21f"
+    },
+    {
+      "introduced_by_run": "slayer://run/dataset-filtering/saos-snapshot-20260904",
+      "predicate": "FILTERED_BY",
+      "source_version_id": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76",
+      "target_version_id": "slayer://version/protocol/saos-filter@sha256:33de2cc5f9e08a3b22fd6f4303e570a951d184cc8de91aca61ab216250530b3c"
+    }
+  ],
+  "run": {
+    "actor_id": "hf:PiotrSty",
+    "config": {
+      "date_valid_year_range": [
+        1900,
+        2027
+      ],
+      "deduplication": {
+        "algorithm": "sha1",
+        "field": "normalized_text",
+        "scope": "within_source"
+      },
+      "execution": {
+        "parquet_compression": "zstd",
+        "record_batch_size": 128,
+        "source_workers": 14,
+        "token_threads": 14,
+        "token_threads_per_worker": 1
+      },
+      "minimum_characters": 200,
+      "minimum_polish_diacritic_ratio": 0.005,
+      "normalization": [
+        "NFKC",
+        "space_quote_dash",
+        "paragraph_preserving_whitespace",
+        "pattern_PII_scrub"
+      ],
+      "source_revision": "736cba4f92999e29d3e49409b0713c88fc683714",
+      "tokenizer": "cl100k_base"
+    },
+    "environment": {
+      "container_image": null,
+      "determinism": {
+        "input_order": [
+          "train-00000-of-00014.parquet",
+          "train-00001-of-00014.parquet",
+          "train-00002-of-00014.parquet",
+          "train-00003-of-00014.parquet",
+          "train-00004-of-00014.parquet",
+          "train-00005-of-00014.parquet",
+          "train-00006-of-00014.parquet",
+          "train-00007-of-00014.parquet",
+          "train-00008-of-00014.parquet",
+          "train-00009-of-00014.parquet",
+          "train-00010-of-00014.parquet",
+          "train-00011-of-00014.parquet",
+          "train-00012-of-00014.parquet",
+          "train-00013-of-00014.parquet"
+        ],
+        "sampling": "none",
+        "seed": null
+      },
+      "environment_limitation": "No container image was used; material runtime and package versions are recorded.",
+      "machine": "AMD64",
+      "packages": {
+        "pyarrow": "24.0.0",
+        "tiktoken": "0.14.0"
+      },
+      "platform": "Windows-11-10.0.26200-SP0",
+      "python": "3.12.14 (main, Aug 25 2026, 14:01:42) [MSC v.1944 64 bit (AMD64)]",
+      "python_executable": "C:\\Users\\Hipek\\.cache\\codex-runtimes\\codex-primary-runtime\\dependencies\\python\\python.exe"
+    },
+    "finished_at": "2026-09-05T06:45:57.253621Z",
+    "git_commit": "b28db4c2b1f7273f878b3c54602777f7b224e27b",
+    "id": "slayer://run/dataset-filtering/saos-snapshot-20260904",
+    "inputs": [
+      {
+        "role": "source",
+        "version_id": "hf://datasets/PiotrSty/saos-polish-court-judgments@736cba4f92999e29d3e49409b0713c88fc683714"
+      },
+      {
+        "role": "target_base",
+        "version_id": "hf://datasets/SlayerLab/polish-dynaword@a9d134ac79a9fe586fc9c11cb39eb32d69dcd21f"
+      }
+    ],
+    "outputs": [
+      {
+        "role": "dataset_source",
+        "version_id": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76"
+      }
+    ],
+    "protocol_version_id": "slayer://version/protocol/saos-filter@sha256:33de2cc5f9e08a3b22fd6f4303e570a951d184cc8de91aca61ab216250530b3c",
+    "started_at": "2026-09-05T06:29:24.801108Z"
+  },
+  "schema_version": "slayer.ai/dataset-source-contribution/v1",
+  "version": {
+    "created_at": "2026-09-05T06:45:57.253621Z",
+    "created_by": "hf:PiotrSty",
+    "digest": "sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76",
+    "id": "slayer://version/dataset-source/saos@sha256:affd02f630791c68a2b1e3a2f7c05d285b80c0f318dad75e66453261d4fabf76",
+    "object_id": "slayer://object/dataset-source/saos",
+    "payload_uri": "hf://datasets/SlayerLab/polish-dynaword@f64afeca7bcca3b17c33f41fecf8478a9f6afb65/data/saos/saos.parquet",
+    "published_in_commit": "f64afeca7bcca3b17c33f41fecf8478a9f6afb65",
+    "schema_version": "slayer.ai/dataset-source/v1"
+  }
+}

artifacts/saos_source_inventory.json

@@ -0,0 +1,77 @@
+{
+  "aggregate_digest": "sha256:d1e1743a9901c381e6617391f65ca2686e317dd7c3d8b3552093e0104040c88c",
+  "files": [
+    {
+      "path": "data/train-00000-of-00014.parquet",
+      "sha256": "f76b008a2061dc85589a51f4c868cc6484c9dff57a09b53e72b3d5390cbb632f",
+      "size": 180401965
+    },
+    {
+      "path": "data/train-00001-of-00014.parquet",
+      "sha256": "0641823cad2fc981da1f0398c4202e25e6489994828812ef8bf3701bbb254930",
+      "size": 213080176
+    },
+    {
+      "path": "data/train-00002-of-00014.parquet",
+      "sha256": "f707bc330140dac0052f5e169cb6dff5918948877f1b4dd471c3070de0230593",
+      "size": 243953893
+    },
+    {
+      "path": "data/train-00003-of-00014.parquet",
+      "sha256": "f72f79df6fe494fd91cbb82e63083fb375e18a59ad772232a719fa46fb9cbe3a",
+      "size": 239361563
+    },
+    {
+      "path": "data/train-00004-of-00014.parquet",
+      "sha256": "c5d2ccc87ff2a3d8f7ba64b24f689d92e3def50a70792f59446f58bc2ac04a90",
+      "size": 240427543
+    },
+    {
+      "path": "data/train-00005-of-00014.parquet",
+      "sha256": "9139c736363b6c151373fd159c0f30cc06fe65d52c01acc547737463f5500fc7",
+      "size": 234121713
+    },
+    {
+      "path": "data/train-00006-of-00014.parquet",
+      "sha256": "6b0d88e3ef45b0c246b79aa47aa1fc337c997305159f1d6e84e6e95798995495",
+      "size": 242598279
+    },
+    {
+      "path": "data/train-00007-of-00014.parquet",
+      "sha256": "4b1c3a69948d1ff713a0dd39d71ddb9d4cd85156a947c2b8e4c55264ed947cc0",
+      "size": 253305954
+    },
+    {
+      "path": "data/train-00008-of-00014.parquet",
+      "sha256": "3a4298ee101bf2b9fc6a339db0f4ce7471519b5e36c828ff6166617983d08c60",
+      "size": 256220979
+    },
+    {
+      "path": "data/train-00009-of-00014.parquet",
+      "sha256": "41b963dd3669d5bf97f6e381019483ac52674f2889cd31efe30928ca02ab0275",
+      "size": 257319050
+    },
+    {
+      "path": "data/train-00010-of-00014.parquet",
+      "sha256": "960243ec221eac979ee226aeb630c9f9325af774e703e8af6285386f06b57541",
+      "size": 257444913
+    },
+    {
+      "path": "data/train-00011-of-00014.parquet",
+      "sha256": "d8c54ec1698d9ab53e86b4064401efda174cd4b9580b256fac8bb0ada79ad886",
+      "size": 260380321
+    },
+    {
+      "path": "data/train-00012-of-00014.parquet",
+      "sha256": "fe294f0ef3b851928cc37a22d29c66dfd55c4951a09b76f5cb89a6b64082c762",
+      "size": 266750435
+    },
+    {
+      "path": "data/train-00013-of-00014.parquet",
+      "sha256": "5c27b08e9559bf0ba8d63ed7f1daa7f19db0cd4fcb774d715e7d046e34a373b2",
+      "size": 273748222
+    }
+  ],
+  "repository": "PiotrSty/saos-polish-court-judgments",
+  "revision": "736cba4f92999e29d3e49409b0713c88fc683714"
+}

artifacts/saos_source_profile.json

@@ -0,0 +1,41 @@
+{
+  "characters": 5895441856,
+  "counts": {
+    "email_placeholder_records": 289,
+    "empty_text": 1,
+    "invalid_judgment_dates": 1,
+    "invisible_character_records": 3315,
+    "missing_judges": 17599,
+    "phone_placeholder_records": 895,
+    "residual_labelled_phone_records": 31,
+    "row_groups": 70,
+    "rows": 296692,
+    "short_text": 3
+  },
+  "court_types": {
+    "COMMON": 229372,
+    "CONSTITUTIONAL_TRIBUNAL": 9503,
+    "NATIONAL_APPEAL_CHAMBER": 19758,
+    "SUPREME": 38059
+  },
+  "date_max": "2017-05-11",
+  "date_min": "0208-03-14",
+  "invalid_date_samples": [
+    {
+      "id": 358516,
+      "value": "0208-03-14T00:00:00"
+    }
+  ],
+  "judgment_types": {
+    "DECISION": 64761,
+    "REASONS": 18695,
+    "REGULATION": 3231,
+    "RESOLUTION": 2469,
+    "SENTENCE": 207536
+  },
+  "schema": "id: int64\ncourtType: string\njudgmentType: string\njudgmentDate: timestamp[ms]\ncaseNumbers: list<element: string>\n  child 0, element: string\njudges: list<element: string>\n  child 0, element: string\nsummary: string\ndecision: string\ntextContent: string\nlegalBases: list<element: string>\n  child 0, element: string\nkeywords: list<element: string>\n  child 0, element: string\nreferencedRegulations: list<element: struct<journalTitle: string, journalNo: int64, journalYear: int64, journalEntry: int64 (... 16 chars omitted)\n  child 0, element: struct<journalTitle: string, journalNo: int64, journalYear: int64, journalEntry: int64, text: string (... 1 chars omitted)\n      child 0, journalTitle: string\n      child 1, journalNo: int64\n      child 2, journalYear: int64\n      child 3, journalEntry: int64\n      child 4, text: string\nreferencedCourtCases: list<element: struct<caseNumber: string, judgmentIds: list<element: int64>, generated: bool>>\n  child 0, element: struct<caseNumber: string, judgmentIds: list<element: int64>, generated: bool>\n      child 0, caseNumber: string\n      child 1, judgmentIds: list<element: int64>\n          child 0, element: int64\n      child 2, generated: bool\nsource: struct<code: string, judgmentUrl: string, judgmentId: string, publisher: string, reviser: string, pu (... 22 chars omitted)\n  child 0, code: string\n  child 1, judgmentUrl: string\n  child 2, judgmentId: string\n  child 3, publisher: string\n  child 4, reviser: string\n  child 5, publicationDate: string\nreceiptDate: string\ndivision: extension<arrow.json>\ntext_length: int64\n-- schema metadata --\nhuggingface: '{\"info\": {\"features\": {\"id\": {\"dtype\": \"int64\", \"_type\": \"V' + 1682",
+  "schemas_identical": true,
+  "shards": 14,
+  "text_length_max": 2260197,
+  "text_length_min": 0
+}

artifacts/source_findings.md

@@ -2297,3 +2297,36 @@ The datasheet states the caveats and gives the requested source note; the regist
 
 **Not established** — a lawyer's reading of the SIS terms; the SIS page on the fetch date (bot
 challenge); the Kancelaria Sejmu BIP reuse page; personal names (NERGAL, v2#60).
+
+---
+
+## saos — Polish court judgments (pinned partial snapshot)
+
+**Investigated** — 2026-09-04 (review round 2026-09-29).
+
+**Verdict** — `added` (proposed with `release: None`). 296,689 judgments /
+2.27B `cl100k_base` tokens (5.9B chars, dates 1986-05-28 – 2017-05-11) from
+SAOS — the biggest clean legal-language add since the HPLT removal, and
+explicitly a **partial** snapshot (54.9% of the live 2026-09-04 API count).
+
+**Enumeration** — `https://www.saos.org.pl/api/dump/judgments` dump endpoint,
+captured as the immutable `PiotrSty/saos-polish-court-judgments@736cba4f` with
+per-record URLs; courts: COMMON 229,370 / SUPREME 38,058 / KIO 19,758 /
+Constitutional Tribunal 9,503 rows.
+
+**License evidence** — Copyright Act art. 4(2) + Open Data Act basis, documented
+in `saos.license-evidence.json`; **formal legal review NOT performed** — the
+opinion scope (third-party quoted material inside judgments; privacy/GDPR of
+named private persons for TDM; SAOS/Ministry redistribution terms) and the
+responsible party (maintainer's Polish-legal reviewer) are written down in the
+datasheet so the gap is visible, not implicit.
+
+**Blockers** — release gates: formal legal review (above) and NERGAL pass
+(pattern scrub only: 14 e-mails, 253 labelled IDs, 60 phones, 66 bank accounts;
+judgments name private persons). Cross-source text overlap vs
+`uzp_orzeczenia_pl` measured in the review thread (KIO family overlap).
+
+**Lesson** — when a contribution explicitly says "legal review not performed",
+the review ask is not "do the law" but "write down exactly which legal questions
+remain and whose desk they land on" — turning an open flag into an assignable
+checklist is the deliverable engineering owes the release process.

data/saos/NOTICE.md

@@ -0,0 +1,41 @@
+# Source attribution
+
+Official source: System Analizy Orzeczen Sadowych (SAOS), https://www.saos.org.pl/. Per-record canonical SAOS URLs are preserved in saos.attribution.jsonl. Snapshot preparation and contribution: Piotr Styła (PiotrSty / PiotrStyla).
+
+# Informacja o przetworzeniu / Processed-data notice
+
+Per the reuse conditions of the Constitutional Tribunal (TK)
+(https://trybunal.gov.pl/informacja-publiczna-media/ponowne-wykorzystywanie,
+Warunek 1 and Warunek 2) and Article 15(1) of the Polish Open Data and Reuse
+of Public Sector Information Act of 11 August 2021 (Dz. U. 2021 poz. 1641 as
+amended), for the TK-originated documents contained in this dataset:
+
+- **Source / Zrodlo:** System Analizy Orzeczen Sadowych (SAOS),
+  https://www.saos.org.pl/, dump API `api/dump/judgments` (documentation:
+  https://www.saos.org.pl/help/index.php/dokumentacja-api/api-pobierania-danych).
+  The TK documents reach this dataset through SAOS, which points to the
+  originating court systems (Ministry of Justice portal, Sad Najwyzszy, TK,
+  KIO) as the responsible publishers.
+- **Time of creation / Czas wytworzenia:** judgment dates 1986-05-28 through
+  2017-05-11 (per record in `saos.attribution.jsonl`, `judgmentDate`).
+- **Time of retrieval / Czas pozyskania:** 2026-09-04 - the SAOS dump API was
+  queried and the snapshot pinned on that date (intermediary snapshot
+  `PiotrSty/saos-polish-court-judgments@736cba4f92999e29d3e49409b0713c88fc683714`).
+- **Author / Tworca:** judges and adjudicating panels named in the records
+  (`judges` field in the attribution sidecar), acting in their official
+  capacity; where unknown, the author is not stated.
+- **The data has been processed / Dane zostaly przetworzone:** conversion to
+  the DynaWord source schema; NFKC Unicode and whitespace normalization;
+  stripping of HTML residue; exact-text deduplication (SHA-1 fingerprints);
+  conservative pattern-based PII scrubbing (e-mail, PESEL candidates, labelled
+  NIP/REGON, labelled phones, bank accounts); minimum-length and
+  Polish-diacritic gates; n-gram decontamination against a local reference
+  corpus. Automated processing may introduce errors; the processed texts are
+  not the official wording.
+
+Warunek 3 note: the Constitutional Tribunal and its Chancellery accept no
+responsibility for the content of processed public-sector information
+originally made available in its primary form. The same disclaimer applies on
+the SAOS side: SAOS accepts no responsibility for content and points to the
+source systems. See `saos.md` ("Licensing and privacy") and
+`saos.license-evidence.json` for the full chain-of-title gaps.

data/saos/saos.attribution.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0fda03e770bc8db94694f4758239ea6fda80e16e5398aee2b5a308a48896c857
+size 180397031

data/saos/saos.license-evidence.json

@@ -0,0 +1,49 @@
+{
+  "copyright_basis": {
+    "act": "Polish Copyright and Related Rights Act",
+    "article": "4(2)",
+    "finding": "official documents and materials are excluded from copyright",
+    "url": "https://eli.gov.pl/api/acts/DU/2025/24/text/O/D20250024.pdf"
+  },
+  "dataset_record_license": "public-domain (official documents)",
+  "formal_legal_review": "not_performed",
+  "hf_source_card_license": "unknown",
+  "chain_of_title_gaps": [
+    "SAOS publishes no licence or terms of reuse: https://www.saos.org.pl/regulamin returns 404 (checked at review, 2026-10-05)",
+    "SAOS disclaims responsibility for content and points to the originating systems as the responsible publishers: the Ministry of Justice portal, Sad Najwyzszy (SN), Trybunal Konstytucyjny (TK) and Krajowa Izba Odwolawcza (KIO)",
+    "the intermediary snapshot PiotrSty/saos-polish-court-judgments@736cba4f92999e29d3e49409b0713c88fc683714 declares license: unknown and no Creative Commons grant",
+    "TK documents carry reuse conditions (attribution + processed-data notice) per https://trybunal.gov.pl/informacja-publiczna-media/ponowne-wykorzystywanie; the notice is in NOTICE.md",
+    "SN and KIO reuse terms have not been obtained"
+  ],
+  "reuse_basis": {
+    "asserted_by_contributor": true,
+    "verified": false,
+    "note": "These Open Data Act articles were carried over from the sejm_api evidence pack and are the contributor's reading, NOT a reuse grant published by SAOS or any source court. They must not be read as a completed rights review of this source.",
+    "act": "Polish Open Data and Reuse of Public Sector Information Act",
+    "articles": [
+      "2(12)",
+      "5",
+      "6",
+      "14(1)",
+      "15",
+      "17"
+    ],
+    "url": "https://eli.gov.pl/api/acts/DU/2023/1524/text.html"
+  },
+  "limitations": [
+    "formal legal review has not been performed",
+    "privacy and personal-data exceptions remain",
+    "third-party works quoted or embedded in judgments may retain separate rights",
+    "no upstream Creative Commons grant is asserted",
+    "the pinned source snapshot is incomplete relative to the live SAOS count (54.9% as of 2026-09-04)",
+    "TK dissenting opinions (zdania odrębne) are included and not yet separated",
+    "4,541 KIO rows duplicate uzp_orzeczenia_pl and are not yet deduplicated across sources (the overlap audit ran with saos_rows: 0)"
+  ],
+  "official_source": "https://www.saos.org.pl/api/dump/judgments",
+  "official_source_documentation": "https://www.saos.org.pl/help/index.php/dokumentacja-api/api-pobierania-danych",
+  "pinned_source": "hf://datasets/PiotrSty/saos-polish-court-judgments@736cba4f92999e29d3e49409b0713c88fc683714",
+  "pinned_target_base": "hf://datasets/SlayerLab/polish-dynaword@a9d134ac79a9fe586fc9c11cb39eb32d69dcd21f",
+  "provider": "System Analizy Orzeczen Sadowych (SAOS)",
+  "retrieved_at": "2026-09-04",
+  "status": "documented-statutory-basis-chain-of-title-gaps-review-needed"
+}

data/saos/saos.md

@@ -0,0 +1,157 @@
+# saos
+
+Polish court judgments from the System Analizy Orzeczen Sadowych (SAOS),
+converted from a pinned `PiotrSty` snapshot into the DynaWord source schema.
+
+## Dataset description
+
+- Official origin: https://www.saos.org.pl/api/dump/judgments
+- API documentation: https://www.saos.org.pl/help/index.php/dokumentacja-api/api-pobierania-danych
+- Pinned source: `PiotrSty/saos-polish-court-judgments@736cba4f92999e29d3e49409b0713c88fc683714`
+- Domain: legal/judicial
+- Language: Polish (`pl`)
+- License label: `public-domain (official documents)`
+- Valid judgment dates: 1986-05-28 through 2017-05-11
+- Added: 2026-09-04
+- Contributor: Piotr Styła (`PiotrSty` on Hugging Face; `PiotrStyla` on GitHub)
+
+## Snapshot scope and completeness
+
+The source contains **296,692** processed records. A direct query
+of the official SAOS API on 2026-09-04 returned
+**540,726** items. This version therefore represents
+approximately **54.9%** of the live count and MUST NOT be described as a
+complete or current copy of SAOS. The immutable snapshot remains useful as a
+large, auditable legal-language contribution; later synchronization belongs in a
+new dataset Version and Run.
+
+## Content and provenance
+
+Each DynaWord row uses the normalized full judgment text. The attribution
+sidecar preserves the source ID, canonical SAOS URL, upstream source URL, court
+type, judgment type, case numbers, named judges acting in an official capacity,
+raw judgment date, source shard and pinned source revision. Generated SAOS
+enrichment fields are not used as training-text claims.
+
+## Courts represented
+
+| court type | source rows |
+|---|---:|
+| `COMMON` | 229,370 |
+| `CONSTITUTIONAL_TRIBUNAL` | 9,503 |
+| `NATIONAL_APPEAL_CHAMBER` | 19,758 |
+| `SUPREME` | 38,058 |
+
+## Processing
+
+- NFKC Unicode normalization, normalized spaces/quotes/dashes and preserved paragraphs.
+- Article and section numbering is retained because it is substantive legal content.
+- Exact normalized-text deduplication with SHA-1 fingerprints.
+- Minimum 200 characters and the DynaWord Polish-diacritic ratio gate (`>= 0.005`).
+- Pattern scrubbing for email, PESEL candidates, labelled NIP/REGON, labelled phones and bank accounts.
+- Implausible or unparsable judgment dates are left empty in the training schema and preserved raw in the sidecar.
+- HTML residue causes a build failure rather than silently entering the corpus.
+
+## Statistics
+
+| measure | value |
+|---|---:|
+| Source records | 296,692 |
+| Documents published | 296,689 |
+| Dropped under 200 characters | 3 |
+| Dropped by Polish-language gate | 0 |
+| Exact normalized duplicates removed | 0 |
+| Records with invalid/implausible dates | 1 |
+| Documents with judge attribution | 279,091 |
+| Characters | 5,897,001,250 |
+| Tokens (`cl100k_base` proxy) | 2,265,448,504 |
+
+## Licensing and privacy
+
+Court judgments are documented here as official materials excluded from
+copyright under article 4(2) of the Polish Copyright Act (the basis identified
+in the 2026-10-05 rights review - this is a statutory exclusion, not a
+packaging licence). **Chain-of-title gaps, stated plainly:**
+
+- SAOS publishes **no licence or terms of reuse** - `https://www.saos.org.pl/regulamin` returns 404;
+- SAOS **disclaims responsibility for content** and points to the originating systems as the responsible publishers: the Ministry of Justice portal, Sad Najwyzszy (SN), Trybunal Konstytucyjny (TK) and Krajowa Izba Odwolawcza (KIO);
+- the intermediary snapshot `PiotrSty/saos-polish-court-judgments@736cba4` declares `license: unknown`; no Creative Commons grant is asserted at any level;
+- the Open Data Act wording previously cited here was carried over from the `sejm_api` evidence pack; it is the contributor's reading and must not be read as a completed rights review of this source (`saos.license-evidence.json` now flags it as `verified: false`).
+
+TK-originated documents carry reuse conditions (attribution + processed-data
+notice): the notice with source and retrieval time is in `NOTICE.md`
+("Informacja o przetworzeniu"). No formal legal opinion was performed; privacy
+exceptions and separate rights in quoted or embedded works remain. This is
+provenance and engineering documentation, not legal advice.
+
+The source reports prior PII scrubbing. This conversion reruns conservative
+pattern-based scrubbing, but does not claim perfect de-identification. Court-
+published anonymization is retained; names of judges acting publicly in their
+official capacity are preserved as attribution.
+
+## Review addenda (2026-09-29)
+
+**Scrubber that ran / NERGAL gate.** The scrubber is the pattern pass in
+`src/fetch_saos.py` (e-mail, PESEL candidates, labelled NIP/REGON, labelled
+phones, bank accounts -> `[PII]`-style placeholders; 14 e-mails, 253 labelled
+identifiers, 60 labelled phones and 66 bank-account strings replaced in this
+build). This is **not**
+[SlayerLab/NERGAL](https://huggingface.co/SlayerLab/NERGAL). The shard must be
+re-run through NERGAL before it is admitted to a release — judgments name
+private persons (parties, attorneys, witnesses) in official capacity contexts
+where the boundary is a legal question, not a regex question. Merge with
+`release: None` is fine meanwhile.
+
+**Legal review — what is needed and who does it.** A formal legal opinion has
+NOT been performed. Needed before release, on the merged corpus as used for
+training:
+
+1. Copyright-status confirmation of the art. 4(2) exclusion as applied to full
+   judgment texts, including **third-party quoted material inside judgments**
+   (party submissions, cited reports, quoted correspondence) which may carry
+   separate rights;
+2. Privacy/GDPR assessment of named private persons in published judgments
+   (parties, attorneys, witnesses) for text-and-data-mining reuse — including
+   the CJEU/Polish case-law overlay on court-publication practice;
+3. Provider-side terms check (SAOS / Ministry of Justice redistribution terms).
+
+Responsible: the corpus maintainer's legal reviewer (Polish legal background —
+bar/counsel or the institution's legal team). The engineering side can only
+prepare the evidence pack: per-record URLs/court metadata (in the sidecar), the
+`license-evidence.json` snapshot, and sample spot-checks — none of which
+substitutes for the opinion. Until it lands, the source is mergeable with
+`release: None` and excluded from any stable release.
+
+## Rights review addendum (2026-10-05)
+
+Addressed in this revision: the TK processed-data notice (source, creation and
+retrieval times, processing description) is added to `NOTICE.md` and summarised
+above; the chain-of-title gaps are stated in this card and machine-readably in
+`saos.license-evidence.json`; the borrowed `sejm_api` Open Data Act framing is
+marked as unverified.
+
+Still required **before any release** (merge with `release: None` is fine
+meanwhile):
+
+1. A Polish legal opinion covering reasons, dissents, KIO/TK material and
+   quoted third-party works inside judgments.
+2. Stripping the TK dissenting opinions (`zdania odrębne`, approx. 450
+   documents) from the training texts.
+3. [SlayerLab/NERGAL](https://huggingface.co/SlayerLab/NERGAL) in place of the
+   regex PII scrub.
+4. Cross-source dedup of the 4,541 KIO rows that duplicate
+   `uzp_orzeczenia_pl` - the overlap audit ran with `saos_rows: 0` and is not
+   evidence of disjointness.
+5. A re-pull from the SAOS dump API with a recorded date; this snapshot covers
+   54.9% of live SAOS (540,726 items as of 2026-09-04).
+6. A documented takedown path for removed/anonymised judgments.
+7. SN and KIO reuse terms obtained from the originating bodies.
+
+## Known limitations
+
+- Partial historical snapshot, not the complete current SAOS database.
+- Legal/judicial register is already heavily represented in Polish DynaWord;
+  inclusion changes the training mix and should be evaluated through ablations.
+- Pattern-based PII detection can have false negatives and false positives.
+- Upstream source transformations are described in its card but the repository
+  does not include the original collection code or standalone audit artifacts.

data/saos/saos.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f265dfe410aef0b0a2453f16e05d6398292af6a66db713bfaad554b4ee0dbf55
+size 1931132709

data/saos/saos.stats.json

@@ -0,0 +1,46 @@
+{
+  "added": "2026-09-04",
+  "authors_with_value": 279091,
+  "by_court_type": {
+    "COMMON": 229370,
+    "CONSTITUTIONAL_TRIBUNAL": 9503,
+    "NATIONAL_APPEAL_CHAMBER": 19758,
+    "SUPREME": 38058
+  },
+  "by_judgment_type": {
+    "DECISION": 64761,
+    "REASONS": 18693,
+    "REGULATION": 3231,
+    "RESOLUTION": 2469,
+    "SENTENCE": 207535
+  },
+  "chars": 5897001250,
+  "created_max": "2017-05-11",
+  "created_min": "1986-05-28",
+  "drop_dup": 0,
+  "drop_lang": 0,
+  "drop_short": 3,
+  "invalid_dates": 1,
+  "kept": 296689,
+  "license": "public-domain (official documents)",
+  "licenses": {
+    "public-domain (official documents)": 296689
+  },
+  "normalization": {
+    "bank_accounts_redacted": 66,
+    "control_characters_removed": 0,
+    "emails_redacted": 14,
+    "labelled_identifiers_redacted": 253,
+    "labelled_phones_redacted": 60,
+    "normalized_records": 263987,
+    "pesel_candidates_redacted": 0
+  },
+  "official_count_observed_at": "2026-09-04",
+  "official_total_results": 540726,
+  "snapshot_live_count_ratio": 0.5486919437940843,
+  "source_revision": "736cba4f92999e29d3e49409b0713c88fc683714",
+  "source_rows": 296692,
+  "stats_recomputed_from_parquet": true,
+  "target_base_revision": "a9d134ac79a9fe586fc9c11cb39eb32d69dcd21f",
+  "tokens": 2265448504
+}

src/fetch_saos.py

@@ -0,0 +1,1014 @@
+#!/usr/bin/env python3
+"""Build a pinned PiotrSty SAOS snapshot for Polish DynaWord."""
+
+from __future__ import annotations
+
+import argparse
+import collections
+import concurrent.futures
+import datetime as dt
+import hashlib
+import html
+import importlib.metadata
+import json
+import pathlib
+import platform
+import re
+import shutil
+import sys
+import unicodedata
+from typing import Any
+
+
+SOURCE_NAME = "saos"
+SOURCE_REPOSITORY = "PiotrSty/saos-polish-court-judgments"
+TARGET_REPOSITORY = "SlayerLab/polish-dynaword"
+OFFICIAL_API = "https://www.saos.org.pl/api/dump/judgments"
+OFFICIAL_API_DOCS = "https://www.saos.org.pl/help/index.php/dokumentacja-api/api-pobierania-danych"
+SOURCE_LICENSE = "public-domain (official documents)"
+COPYRIGHT_ACT = "https://eli.gov.pl/api/acts/DU/2025/24/text/O/D20250024.pdf"
+OPEN_DATA_ACT = "https://eli.gov.pl/api/acts/DU/2023/1524/text.html"
+COMMIT_PATTERN = re.compile(r"^[0-9a-f]{40}$")
+EXPECTED_FIELDS = (
+    "id", "text", "source", "added", "created", "token_count", "license", "author"
+)
+POLISH_RE = re.compile(r"[\u0105\u0107\u0119\u0142\u0144\u00f3\u015b\u017a\u017c\u0104\u0106\u0118\u0141\u0143\u00d3\u015a\u0179\u017b]")
+ALPHA_RE = re.compile(r"[^\W\d_]", re.UNICODE)
+HTML_TAG = re.compile(
+    r"</?(?:p|div|span|br|h[1-6]|table|tbody|thead|tr|td|th|ul|ol|li|strong|em|b|i|a|section)"
+    r"\s*/?>",
+    re.I,
+)
+EMAIL = re.compile(r"(?i)\b[A-Z0-9._%+-]+@[A-Z0-9.-]+\.[A-Z]{2,}\b")
+PESEL = re.compile(r"(?<!\d)\d{11}(?!\d)")
+LABELLED_IDENTIFIER = re.compile(
+    r"(?i)\b(?:NIP|REGON)\s*[:.]?\s*\d(?:[\s-]?\d){8,13}\b"
+)
+LABELLED_PHONE = re.compile(
+    r"(?i)(?:\btel\.?|\btelefon|\+48)\s*[:.]?\s*(?:\(?\d{2,3}\)?[\s.-]?){3,5}\d{2,3}"
+)
+BANK_ACCOUNT = re.compile(r"(?<!\d)(?:PL\s*)?\d{2}(?:[\s-]*\d{4}){6}(?!\d)", re.I)
+SPACE_RUN = re.compile(r"[ \t\v\f]+")
+BLANK_RUN = re.compile(r"\n{3,}")
+CONTROL = re.compile(r"[\x00-\x08\x0b\x0c\x0e-\x1f\x7f]")
+TRANSLATION = str.maketrans({
+    "\u00a0": " ", "\u2007": " ", "\u202f": " ", "\u200b": "",
+    "\ufeff": "", "\u00ad": "", "\u2010": "-", "\u2011": "-",
+    "\u2012": "-", "\u2013": "-", "\u2014": "-", "\u2212": "-",
+    "\u2018": "'", "\u2019": "'", "\u201c": '"', "\u201d": '"',
+    "\u2026": "...",
+})
+
+
+class BuildError(ValueError):
+    pass
+
+
+def utc_now() -> str:
+    return dt.datetime.now(dt.timezone.utc).isoformat().replace("+00:00", "Z")
+
+
+def sha256_file(path: pathlib.Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as stream:
+        for chunk in iter(lambda: stream.read(1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def canonical_sha256(value: Any) -> str:
+    payload = json.dumps(
+        value, ensure_ascii=False, sort_keys=True, separators=(",", ":")
+    ).encode("utf-8")
+    return hashlib.sha256(payload).hexdigest()
+
+
+def write_json(path: pathlib.Path, payload: Any) -> None:
+    path.write_text(
+        json.dumps(payload, ensure_ascii=False, indent=2, sort_keys=True, default=str) + "\n",
+        encoding="utf-8",
+    )
+
+
+def first_text(value: Any) -> str:
+    if value is None:
+        return ""
+    if isinstance(value, dict):
+        return str(value.get("name") or value.get("caseNumber") or value.get("id") or "").strip()
+    return str(value).strip()
+
+
+def list_text(value: Any) -> list[str]:
+    if not isinstance(value, list):
+        return [first_text(value)] if first_text(value) else []
+    return [text for item in value if (text := first_text(item))]
+
+
+def normalize_text(value: Any) -> tuple[str, collections.Counter[str]]:
+    stats: collections.Counter[str] = collections.Counter()
+    original = str(value or "")
+    text = html.unescape(unicodedata.normalize("NFKC", original)).translate(TRANSLATION)
+    text, controls = CONTROL.subn("", text)
+    stats["control_characters_removed"] += controls
+    lines = [SPACE_RUN.sub(" ", line).strip() for line in text.replace("\r", "\n").split("\n")]
+    text = BLANK_RUN.sub("\n\n", "\n".join(lines)).strip()
+    if text != original.strip():
+        stats["normalized_records"] += 1
+    for pattern, key, replacement in (
+        (EMAIL, "emails_redacted", "[PII:EMAIL]"),
+        (LABELLED_IDENTIFIER, "labelled_identifiers_redacted", "[PII:IDENTIFIER]"),
+        (PESEL, "pesel_candidates_redacted", "[PII:PESEL]"),
+        (LABELLED_PHONE, "labelled_phones_redacted", "[PII:PHONE]"),
+        (BANK_ACCOUNT, "bank_accounts_redacted", "[PII:ACCOUNT]"),
+    ):
+        text, count = pattern.subn(replacement, text)
+        stats[key] += count
+    return text, stats
+
+
+def polish_ratio(text: str) -> float:
+    letters = ALPHA_RE.findall(text)
+    return len(POLISH_RE.findall(text)) / len(letters) if letters else 0.0
+
+
+def normalize_date(value: Any) -> tuple[str, str, bool]:
+    raw = value.isoformat() if isinstance(value, (dt.date, dt.datetime)) else str(value or "")
+    candidate = raw[:10]
+    try:
+        parsed = dt.date.fromisoformat(candidate)
+    except ValueError:
+        return "", raw, False
+    valid = 1900 <= parsed.year <= dt.date.today().year + 1
+    return (parsed.isoformat() if valid else ""), raw, valid
+
+
+def output_schema(pa: Any) -> Any:
+    return pa.schema([
+        ("id", pa.string()), ("text", pa.string()), ("source", pa.string()),
+        ("added", pa.string()), ("created", pa.string()), ("token_count", pa.int64()),
+        ("license", pa.string()), ("author", pa.string()),
+    ])
+
+
+def process_source_shard(task: tuple[str, str, str, str, int, int]) -> dict[str, Any]:
+    """Transform one pinned source shard into a deterministic intermediate part."""
+    shard_name, part_name, added_date, source_revision, batch_size, token_threads = task
+    import pyarrow as pa
+    import pyarrow.parquet as pq
+    import tiktoken
+
+    shard = pathlib.Path(shard_name)
+    part_dir = pathlib.Path(part_name)
+    part_dir.mkdir(parents=True, exist_ok=True)
+    parquet_path = part_dir / "data.parquet"
+    sidecar_path = part_dir / "attribution.jsonl"
+    identity_path = part_dir / "identities.tsv"
+    schema = output_schema(pa)
+    encoder = tiktoken.get_encoding("cl100k_base")
+    counts: collections.Counter[str] = collections.Counter()
+    normalization: collections.Counter[str] = collections.Counter()
+    court_types: collections.Counter[str] = collections.Counter()
+    judgment_types: collections.Counter[str] = collections.Counter()
+    seen_ids: set[str] = set()
+    seen_texts: set[bytes] = set()
+    dates: list[str] = []
+    source_parquet = pq.ParquetFile(shard)
+    required = {"id", "textContent", "judgmentDate", "courtType", "judgmentType", "caseNumbers", "judges", "source"}
+    if not required <= set(source_parquet.schema_arrow.names):
+        raise BuildError(f"unexpected source schema in {shard.name}")
+
+    writer = pq.ParquetWriter(parquet_path, schema, compression="zstd")
+    try:
+        with sidecar_path.open("w", encoding="utf-8") as sidecar, identity_path.open("w", encoding="ascii") as identities:
+            for batch in source_parquet.iter_batches(batch_size=batch_size):
+                prepared = []
+                for row in batch.to_pylist():
+                    counts["source_rows"] += 1
+                    source_id = str(row.get("id") or "").strip()
+                    if not source_id or source_id in seen_ids:
+                        counts["drop_duplicate_or_empty_id"] += 1
+                        continue
+                    seen_ids.add(source_id)
+                    text, changes = normalize_text(row.get("textContent"))
+                    normalization.update(changes)
+                    if HTML_TAG.search(text):
+                        raise BuildError(f"HTML residue in source record {source_id}")
+                    if len(text) < 200:
+                        counts["drop_short"] += 1
+                        continue
+                    if polish_ratio(text) < 0.005:
+                        counts["drop_lang"] += 1
+                        continue
+                    text_digest = hashlib.sha1(text.encode("utf-8")).digest()
+                    if text_digest in seen_texts:
+                        counts["drop_dup"] += 1
+                        continue
+                    seen_texts.add(text_digest)
+                    created, raw_date, valid_date = normalize_date(row.get("judgmentDate"))
+                    if not valid_date:
+                        counts["invalid_dates"] += 1
+                    else:
+                        dates.append(created)
+                    judges = list_text(row.get("judges"))
+                    author = "; ".join(judges)
+                    court_type = str(row.get("courtType") or "")
+                    judgment_type = str(row.get("judgmentType") or "")
+                    court_types[court_type] += 1
+                    judgment_types[judgment_type] += 1
+                    prepared.append((source_id, text, text_digest, created, raw_date, author, judges, court_type, judgment_type, row))
+
+                texts = [item[1] for item in prepared]
+                token_counts = [
+                    len(tokens) for tokens in encoder.encode_ordinary_batch(texts, num_threads=token_threads)
+                ] if texts else []
+                documents = []
+                for item, token_count in zip(prepared, token_counts):
+                    source_id, text, text_digest, created, raw_date, author, judges, court_type, judgment_type, row = item
+                    document_id = f"{SOURCE_NAME}_{source_id}"
+                    documents.append({
+                        "id": document_id, "text": text, "source": SOURCE_NAME,
+                        "added": added_date, "created": created, "token_count": token_count,
+                        "license": SOURCE_LICENSE, "author": author,
+                    })
+                    source_meta = row.get("source") if isinstance(row.get("source"), dict) else {}
+                    sidecar.write(json.dumps({
+                        "id": document_id,
+                        "source_id": source_id,
+                        "official_url": f"https://www.saos.org.pl/judgments/{source_id}",
+                        "upstream_judgment_url": source_meta.get("judgmentUrl"),
+                        "upstream_source_code": source_meta.get("code"),
+                        "court_type": court_type,
+                        "judgment_type": judgment_type,
+                        "case_numbers": list_text(row.get("caseNumbers")),
+                        "judges": judges,
+                        "judgment_date_raw": raw_date,
+                        "judgment_date_status": "valid" if created else "invalid_or_implausible",
+                        "source_revision": source_revision,
+                        "source_shard": shard.name,
+                    }, ensure_ascii=False, sort_keys=True, default=str) + "\n")
+                    identities.write(f"{source_id}\t{text_digest.hex()}\n")
+                    counts["kept"] += 1
+                    counts["chars"] += len(text)
+                    counts["tokens"] += token_count
+                    counts["authors_with_value"] += bool(author)
+                if documents:
+                    writer.write_table(pa.Table.from_pylist(documents, schema=schema))
+    finally:
+        writer.close()
+
+    return {
+        "shard": shard.name,
+        "parquet": str(parquet_path),
+        "sidecar": str(sidecar_path),
+        "identities": str(identity_path),
+        "counts": dict(counts),
+        "normalization": dict(normalization),
+        "court_types": dict(court_types),
+        "judgment_types": dict(judgment_types),
+        "created_min": min(dates) if dates else "",
+        "created_max": max(dates) if dates else "",
+    }
+
+
+def registry_entry(stats: dict[str, Any], source_revision: str) -> str:
+    return (
+        f'    "{SOURCE_NAME}": {{\n'
+        f'        "file_key": "{SOURCE_NAME}",\n'
+        '        "pretty": "SAOS - Polish court judgments (pinned partial snapshot)",\n'
+        f'        "license": "{SOURCE_LICENSE}",\n'
+        '        "license_spdx": "LicenseRef-Polish-Official-Documents",\n'
+        '        "traceable": "Court judgments are documented as official materials outside "\n'
+        '                     "copyright under Polish Copyright Act art. 4(2); formal legal "\n'
+        '                     "review has not been performed and statutory limitations remain.",\n'
+        f'        "upstream": "{OFFICIAL_API}",\n'
+        '        "provenance": "Converted from the immutable "\n'
+        f'                      "{SOURCE_REPOSITORY}@{source_revision} snapshot; per-record "\n'
+        '                      "SAOS URLs and source identifiers are preserved in a sidecar.",\n'
+        '        "domain": "legal/judicial",\n'
+        f'        "created": "{stats["created_min"]}, {stats["created_max"]}",\n'
+        '        "is_ocr": False,\n'
+        '        "custom_datasheet": True,\n'
+        '    },\n'
+    )
+
+
+def update_registry(registry: str, stats: dict[str, Any], source_revision: str) -> str:
+    if re.search(rf'^\s{{4}}["\']{SOURCE_NAME}["\']\s*:', registry, re.MULTILINE):
+        raise BuildError(f"{SOURCE_NAME} already exists in the target registry")
+    marker = '    "dziennik_ustaw": {'
+    if registry.count(marker) != 1:
+        raise BuildError("cannot locate the dziennik_ustaw registry entry")
+    return registry.replace(marker, registry_entry(stats, source_revision) + marker, 1)
+
+
+def update_findings(findings: str, stats: dict[str, Any], args: argparse.Namespace) -> str:
+    marker = "### SAOS snapshot contribution update (2026-09-04)"
+    if marker in findings:
+        raise BuildError("SAOS findings update already exists")
+    coverage = stats["source_rows"] / args.official_total_results
+    addition = f"""
+
+{marker}
+
+`{SOURCE_REPOSITORY}@{args.source_revision}` contains **{stats['source_rows']:,}**
+processed judgments. The official SAOS dump exposed **{args.official_total_results:,}**
+items when checked on {args.official_count_observed_at}; the pinned snapshot is
+therefore a **partial snapshot** ({coverage:.1%} of that live count), not a claim
+of current SAOS completeness. The contribution preserves this limitation in its
+ontology Evidence, Claim, attestation and datasheet. It contributes the available
+cleaned legal text now while leaving a later incremental SAOS synchronization as
+a separate versioned task.
+"""
+    return findings.rstrip() + addition + "\n"
+
+
+def legal_evidence(args: argparse.Namespace) -> dict[str, Any]:
+    return {
+        "status": "documented-statutory-basis-review-needed",
+        "formal_legal_review": "not_performed",
+        "dataset_record_license": SOURCE_LICENSE,
+        "hf_source_card_license": "unknown",
+        "provider": "System Analizy Orzeczen Sadowych (SAOS)",
+        "official_source": OFFICIAL_API,
+        "official_source_documentation": OFFICIAL_API_DOCS,
+        "copyright_basis": {
+            "act": "Polish Copyright and Related Rights Act",
+            "article": "4(2)",
+            "finding": "official documents and materials are excluded from copyright",
+            "url": COPYRIGHT_ACT,
+        },
+        "reuse_basis": {
+            "act": "Polish Open Data and Reuse of Public Sector Information Act",
+            "articles": ["2(12)", "5", "6", "14(1)", "15", "17"],
+            "url": OPEN_DATA_ACT,
+        },
+        "limitations": [
+            "formal legal review has not been performed",
+            "privacy and personal-data exceptions remain",
+            "third-party works quoted or embedded in judgments may retain separate rights",
+            "no upstream Creative Commons grant is asserted",
+            "the pinned source snapshot is incomplete relative to the live SAOS count",
+        ],
+        "pinned_source": f"hf://datasets/{SOURCE_REPOSITORY}@{args.source_revision}",
+        "pinned_target_base": f"hf://datasets/{TARGET_REPOSITORY}@{args.target_revision}",
+    }
+
+
+def datasheet(stats: dict[str, Any], args: argparse.Namespace) -> str:
+    coverage = stats["source_rows"] / args.official_total_results
+    court_rows = "\n".join(
+        f"| `{key}` | {value:,} |" for key, value in stats["by_court_type"].items()
+    )
+    return f"""# {SOURCE_NAME}
+
+Polish court judgments from the System Analizy Orzeczen Sadowych (SAOS),
+converted from a pinned `PiotrSty` snapshot into the DynaWord source schema.
+
+## Dataset description
+
+- Official origin: {OFFICIAL_API}
+- API documentation: {OFFICIAL_API_DOCS}
+- Pinned source: `{SOURCE_REPOSITORY}@{args.source_revision}`
+- Domain: legal/judicial
+- Language: Polish (`pl`)
+- License label: `{SOURCE_LICENSE}`
+- Valid judgment dates: {stats['created_min']} through {stats['created_max']}
+- Added: {args.added_date}
+- Contributor: Piotr Sty\u0142a (`PiotrSty` on Hugging Face; `PiotrStyla` on GitHub)
+
+## Snapshot scope and completeness
+
+The source contains **{stats['source_rows']:,}** processed records. A direct query
+of the official SAOS API on {args.official_count_observed_at} returned
+**{args.official_total_results:,}** items. This version therefore represents
+approximately **{coverage:.1%}** of the live count and MUST NOT be described as a
+complete or current copy of SAOS. The immutable snapshot remains useful as a
+large, auditable legal-language contribution; later synchronization belongs in a
+new dataset Version and Run.
+
+## Content and provenance
+
+Each DynaWord row uses the normalized full judgment text. The attribution
+sidecar preserves the source ID, canonical SAOS URL, upstream source URL, court
+type, judgment type, case numbers, named judges acting in an official capacity,
+raw judgment date, source shard and pinned source revision. Generated SAOS
+enrichment fields are not used as training-text claims.
+
+## Courts represented
+
+| court type | source rows |
+|---|---:|
+{court_rows}
+
+## Processing
+
+- NFKC Unicode normalization, normalized spaces/quotes/dashes and preserved paragraphs.
+- Article and section numbering is retained because it is substantive legal content.
+- Exact normalized-text deduplication with SHA-1 fingerprints.
+- Minimum 200 characters and the DynaWord Polish-diacritic ratio gate (`>= 0.005`).
+- Pattern scrubbing for email, PESEL candidates, labelled NIP/REGON, labelled phones and bank accounts.
+- Implausible or unparsable judgment dates are left empty in the training schema and preserved raw in the sidecar.
+- HTML residue causes a build failure rather than silently entering the corpus.
+
+## Statistics
+
+| measure | value |
+|---|---:|
+| Source records | {stats['source_rows']:,} |
+| Documents published | {stats['kept']:,} |
+| Dropped under 200 characters | {stats['drop_short']:,} |
+| Dropped by Polish-language gate | {stats['drop_lang']:,} |
+| Exact normalized duplicates removed | {stats['drop_dup']:,} |
+| Records with invalid/implausible dates | {stats['invalid_dates']:,} |
+| Documents with judge attribution | {stats['authors_with_value']:,} |
+| Characters | {stats['chars']:,} |
+| Tokens (`cl100k_base` proxy) | {stats['tokens']:,} |
+
+## Licensing and privacy
+
+Court judgments are documented here as official materials excluded from
+copyright under article 4(2) of the Polish Copyright Act, with reuse considered
+under the Polish Open Data Act. The source card itself declares `license:
+unknown`, no Creative Commons grant is asserted, and no formal legal opinion was
+performed. Privacy exceptions and separate rights in quoted or embedded works
+remain. This is provenance and engineering documentation, not legal advice.
+
+The source reports prior PII scrubbing. This conversion reruns conservative
+pattern-based scrubbing, but does not claim perfect de-identification. Court-
+published anonymization is retained; names of judges acting publicly in their
+official capacity are preserved as attribution.
+
+## Known limitations
+
+- Partial historical snapshot, not the complete current SAOS database.
+- Legal/judicial register is already heavily represented in Polish DynaWord;
+  inclusion changes the training mix and should be evaluated through ablations.
+- Pattern-based PII detection can have false negatives and false positives.
+- Upstream source transformations are described in its card but the repository
+  does not include the original collection code or standalone audit artifacts.
+"""
+
+
+def contract_test() -> str:
+    return fr'''"""Contract and ontology checks for the {SOURCE_NAME} contribution."""
+
+import hashlib
+import json
+import pathlib
+import re
+
+import pyarrow.parquet as pq
+
+
+ROOT = pathlib.Path(__file__).resolve().parents[1]
+DATA = ROOT / "data" / "{SOURCE_NAME}"
+MANIFEST = ROOT / "artifacts" / "{SOURCE_NAME}_ontology_manifest.json"
+FIELDS = {list(EXPECTED_FIELDS)!r}
+
+
+def sha256_file(path):
+    digest = hashlib.sha256()
+    with path.open("rb") as stream:
+        for chunk in iter(lambda: stream.read(1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def test_data_schema_and_stats():
+    parquet = DATA / "{SOURCE_NAME}.parquet"
+    stats = json.loads((DATA / "{SOURCE_NAME}.stats.json").read_text(encoding="utf-8"))
+    source_values, license_values, ids = set(), set(), set()
+    rows = chars = tokens = authors = 0
+    for batch in pq.ParquetFile(parquet).iter_batches(batch_size=256):
+        assert batch.schema.names == FIELDS
+        for row in batch.to_pylist():
+            rows += 1
+            assert row["id"] not in ids
+            ids.add(row["id"])
+            assert len(row["text"]) >= 200
+            assert row["token_count"] > 0
+            assert not re.search(
+                r"</?(?:p|div|span|br|h[1-6]|table|tbody|thead|tr|td|th|ul|ol|li|strong|em|b|i|a|section)\s*/?>",
+                row["text"],
+                re.I,
+            )
+            assert not row["created"] or re.fullmatch(r"20\d{{2}}-\d{{2}}-\d{{2}}|19\d{{2}}-\d{{2}}-\d{{2}}", row["created"])
+            source_values.add(row["source"])
+            license_values.add(row["license"])
+            chars += len(row["text"])
+            tokens += row["token_count"]
+            authors += bool(row["author"])
+    assert rows == stats["kept"]
+    assert chars == stats["chars"]
+    assert tokens == stats["tokens"]
+    assert authors == stats["authors_with_value"]
+    assert source_values == {{"{SOURCE_NAME}"}}
+    assert license_values == {{"{SOURCE_LICENSE}"}}
+
+
+def test_sidecar_legal_and_snapshot_honesty():
+    stats = json.loads((DATA / "{SOURCE_NAME}.stats.json").read_text(encoding="utf-8"))
+    sidecar = DATA / "{SOURCE_NAME}.attribution.jsonl"
+    assert sum(1 for _ in sidecar.open(encoding="utf-8")) == stats["kept"]
+    legal = json.loads((DATA / "{SOURCE_NAME}.license-evidence.json").read_text(encoding="utf-8"))
+    assert legal["formal_legal_review"] == "not_performed"
+    assert legal["hf_source_card_license"] == "unknown"
+    assert legal["pinned_source"].endswith(stats["source_revision"])
+    assert stats["official_total_results"] > stats["source_rows"]
+
+
+def test_ontology_kernel_and_references():
+    manifest = json.loads(MANIFEST.read_text(encoding="utf-8"))
+    required = {{"object", "version", "protocol", "run", "relations", "evidence",
+                "claims", "claim_evidence", "actors", "attestations"}}
+    assert required <= set(manifest)
+    assert manifest["ontology_version"] == "slayer.ai/research-ontology/v0.1"
+    actor_ids = {{item["id"] for item in manifest["actors"]}}
+    version = manifest["version"]
+    assert version["object_id"] == manifest["object"]["id"]
+    assert version["created_by"] in actor_ids
+    assert re.fullmatch(r"sha256:[0-9a-f]{{64}}", version["digest"])
+    assert re.search(r"@[0-9a-f]{{40}}/", version["payload_uri"])
+    protocol = manifest["protocol"]
+    protocol_version = protocol["version"]
+    assert protocol_version["object_id"] == protocol["id"]
+    assert manifest["run"]["protocol_version_id"] == protocol_version["id"]
+    assert re.search(r"@[0-9a-f]{{40}}/", protocol_version["payload_uri"])
+    assert sha256_file(ROOT / protocol["specification"]["source_path"]) == protocol_version["code_sha256"]
+    run = manifest["run"]
+    assert run["actor_id"] in actor_ids
+    assert run["started_at"] <= run["finished_at"]
+    assert {{"python", "platform", "machine", "packages", "determinism"}} <= set(run["environment"])
+
+    evidence_ids = {{item["id"] for item in manifest["evidence"]}}
+    claim_ids = {{item["id"] for item in manifest["claims"]}}
+    assert len(evidence_ids) == len(manifest["evidence"])
+    assert len(claim_ids) == len(manifest["claims"])
+    for item in manifest["evidence"]:
+        assert item["run_id"] == run["id"]
+        assert item["subject_version"] == version["id"]
+    for item in manifest["claims"]:
+        assert item["scope"] == version["id"]
+        assert item["asserted_by"] in actor_ids
+        assert item["falsification_condition"]
+        assert set(item["supported_by"]) <= evidence_ids
+    for link in manifest["claim_evidence"]:
+        assert link["claim_id"] in claim_ids
+        assert link["evidence_id"] in evidence_ids
+        assert link["relation"] in {{"SUPPORTS", "REFUTES", "QUALIFIES"}}
+    for attestation in manifest["attestations"]:
+        assert attestation["actor_id"] in actor_ids
+        assert attestation["subject_version"] == version["id"]
+        assert set(attestation["supported_by"]) <= evidence_ids
+    assert any(item["type"] == "source_completeness" and item["value"] == "partial_snapshot" for item in manifest["attestations"])
+
+    edges = {{}}
+    for relation in manifest["relations"]:
+        source = relation["source_version_id"]
+        target = relation["target_version_id"]
+        assert source != target
+        assert relation["introduced_by_run"] == run["id"]
+        if target.startswith("slayer://version/"):
+            edges.setdefault(source, set()).add(target)
+    visiting, visited = set(), set()
+    def visit(node):
+        assert node not in visiting, "version lineage must be acyclic"
+        if node in visited:
+            return
+        visiting.add(node)
+        for target in edges.get(node, ()):
+            visit(target)
+        visiting.remove(node)
+        visited.add(node)
+    for node in edges:
+        visit(node)
+
+
+def test_manifest_artifact_checksums():
+    manifest = json.loads(MANIFEST.read_text(encoding="utf-8"))
+    checksums = next(item["payload"] for item in manifest["evidence"] if item["observation_type"] == "artifact_checksums")
+    paths = {{
+        "parquet_sha256": DATA / "{SOURCE_NAME}.parquet",
+        "attribution_sha256": DATA / "{SOURCE_NAME}.attribution.jsonl",
+        "stats_sha256": DATA / "{SOURCE_NAME}.stats.json",
+        "datasheet_sha256": DATA / "{SOURCE_NAME}.md",
+        "legal_evidence_sha256": DATA / "{SOURCE_NAME}.license-evidence.json",
+        "source_inventory_sha256": ROOT / "artifacts" / "{SOURCE_NAME}_source_inventory.json",
+        "source_profile_sha256": ROOT / "artifacts" / "{SOURCE_NAME}_source_profile.json",
+    }}
+    assert {{name: sha256_file(path) for name, path in paths.items()}} == checksums
+'''
+
+
+def build(args: argparse.Namespace) -> dict[str, Any]:
+    run_started_at = utc_now()
+    for name in ("source_revision", "target_revision", "dataset_payload_commit", "protocol_payload_commit"):
+        if not COMMIT_PATTERN.fullmatch(getattr(args, name)):
+            raise BuildError(f"{name} must be a 40-character commit")
+    for path in (args.target_registry, args.target_findings, args.source_profile):
+        if not path.is_file():
+            raise BuildError(f"required input is missing: {path}")
+
+    import pyarrow as pa
+    import pyarrow.parquet as pq
+    import tiktoken
+
+    source_shards = sorted((args.source_dir / "data").glob("*.parquet"))
+    if not source_shards:
+        raise BuildError("no source Parquet shards found")
+    output = args.output
+    data_dir = output / "data" / SOURCE_NAME
+    artifact_dir = output / "artifacts"
+    src_dir = output / "src"
+    for directory in (data_dir, artifact_dir, src_dir):
+        directory.mkdir(parents=True, exist_ok=True)
+
+    source_inventory = {
+        "repository": SOURCE_REPOSITORY,
+        "revision": args.source_revision,
+        "files": [
+            {"path": f"data/{path.name}", "size": path.stat().st_size, "sha256": sha256_file(path)}
+            for path in source_shards
+        ],
+    }
+    source_inventory["aggregate_digest"] = f"sha256:{canonical_sha256(source_inventory)}"
+    inventory_path = artifact_dir / f"{SOURCE_NAME}_source_inventory.json"
+    profile_path = artifact_dir / f"{SOURCE_NAME}_source_profile.json"
+    write_json(inventory_path, source_inventory)
+    source_profile = json.loads(args.source_profile.read_text(encoding="utf-8"))
+    source_profile.pop("samples", None)
+    write_json(profile_path, source_profile)
+
+    schema = output_schema(pa)
+    parquet_path = data_dir / f"{SOURCE_NAME}.parquet"
+    sidecar_path = data_dir / f"{SOURCE_NAME}.attribution.jsonl"
+    parts_dir = output / "_parts"
+    parts_dir.mkdir(parents=True, exist_ok=True)
+    worker_count = min(max(1, args.workers), len(source_shards))
+    worker_token_threads = max(1, args.token_threads // worker_count)
+    tasks = [
+        (
+            str(shard),
+            str(parts_dir / f"part-{index:05d}"),
+            args.added_date,
+            args.source_revision,
+            args.batch_size,
+            worker_token_threads,
+        )
+        for index, shard in enumerate(source_shards)
+    ]
+    if worker_count == 1:
+        parts = [process_source_shard(task) for task in tasks]
+    else:
+        with concurrent.futures.ProcessPoolExecutor(max_workers=worker_count) as pool:
+            parts = list(pool.map(process_source_shard, tasks))
+
+    counts: collections.Counter[str] = collections.Counter()
+    normalization: collections.Counter[str] = collections.Counter()
+    court_types: collections.Counter[str] = collections.Counter()
+    judgment_types: collections.Counter[str] = collections.Counter()
+    seen_ids: set[str] = set()
+    seen_texts: set[bytes] = set()
+    dates: list[str] = []
+    writer = pq.ParquetWriter(parquet_path, schema, compression="zstd")
+    try:
+        with sidecar_path.open("wb") as sidecar:
+            for part in parts:
+                metadata_parquet = pq.ParquetFile(part["parquet"])
+                try:
+                    part_rows = metadata_parquet.metadata.num_rows
+                finally:
+                    metadata_parquet.close()
+                identity_rows = 0
+                with pathlib.Path(part["identities"]).open(encoding="ascii") as identities:
+                    for line in identities:
+                        source_id, text_digest_hex = line.rstrip("\n").split("\t", 1)
+                        text_digest = bytes.fromhex(text_digest_hex)
+                        if source_id in seen_ids:
+                            raise BuildError(f"cross-shard duplicate source id: {source_id}")
+                        if text_digest in seen_texts:
+                            raise BuildError(f"cross-shard duplicate normalized text in {part['shard']}")
+                        seen_ids.add(source_id)
+                        seen_texts.add(text_digest)
+                        identity_rows += 1
+                if identity_rows != part_rows:
+                    raise BuildError(f"identity count mismatch in {part['shard']}")
+                with pathlib.Path(part["sidecar"]).open("rb") as source_sidecar:
+                    shutil.copyfileobj(source_sidecar, sidecar, length=1024 * 1024)
+                part_parquet = pq.ParquetFile(part["parquet"])
+                try:
+                    for row_group in range(part_parquet.metadata.num_row_groups):
+                        writer.write_table(part_parquet.read_row_group(row_group))
+                finally:
+                    part_parquet.close()
+                counts.update(part["counts"])
+                normalization.update(part["normalization"])
+                court_types.update(part["court_types"])
+                judgment_types.update(part["judgment_types"])
+                if part["created_min"]:
+                    dates.append(part["created_min"])
+                if part["created_max"]:
+                    dates.append(part["created_max"])
+    finally:
+        writer.close()
+    shutil.rmtree(parts_dir)
+
+    if not counts["kept"]:
+        raise BuildError("no records survived the build")
+    stats: dict[str, Any] = {
+        **counts,
+        "source_rows": counts["source_rows"],
+        "kept": counts["kept"],
+        "drop_short": counts["drop_short"],
+        "drop_lang": counts["drop_lang"],
+        "drop_dup": counts["drop_dup"],
+        "invalid_dates": counts["invalid_dates"],
+        "authors_with_value": counts["authors_with_value"],
+        "chars": counts["chars"],
+        "tokens": counts["tokens"],
+        "created_min": min(dates) if dates else "",
+        "created_max": max(dates) if dates else "",
+        "by_court_type": dict(sorted(court_types.items())),
+        "by_judgment_type": dict(sorted(judgment_types.items())),
+        "normalization": dict(sorted(normalization.items())),
+        "license": SOURCE_LICENSE,
+        "licenses": {SOURCE_LICENSE: counts["kept"]},
+        "added": args.added_date,
+        "source_revision": args.source_revision,
+        "target_base_revision": args.target_revision,
+        "official_total_results": args.official_total_results,
+        "official_count_observed_at": args.official_count_observed_at,
+        "snapshot_live_count_ratio": counts["source_rows"] / args.official_total_results,
+        "stats_recomputed_from_parquet": True,
+    }
+    stats_path = data_dir / f"{SOURCE_NAME}.stats.json"
+    legal_path = data_dir / f"{SOURCE_NAME}.license-evidence.json"
+    datasheet_path = data_dir / f"{SOURCE_NAME}.md"
+    write_json(stats_path, stats)
+    write_json(legal_path, legal_evidence(args))
+    datasheet_path.write_text(datasheet(stats, args), encoding="utf-8")
+    (data_dir / "NOTICE.md").write_text(
+        "# Source attribution\n\nOfficial source: System Analizy Orzeczen Sadowych "
+        "(SAOS), https://www.saos.org.pl/. Per-record canonical SAOS URLs are "
+        "preserved in saos.attribution.jsonl. Snapshot preparation and contribution: "
+        "Piotr Sty\u0142a (PiotrSty / PiotrStyla).\n",
+        encoding="utf-8",
+    )
+
+    artifact_digests = {
+        "parquet_sha256": sha256_file(parquet_path),
+        "attribution_sha256": sha256_file(sidecar_path),
+        "stats_sha256": sha256_file(stats_path),
+        "datasheet_sha256": sha256_file(datasheet_path),
+        "legal_evidence_sha256": sha256_file(legal_path),
+        "source_inventory_sha256": sha256_file(inventory_path),
+        "source_profile_sha256": sha256_file(profile_path),
+    }
+    version_digest = canonical_sha256({
+        "source_revision": args.source_revision,
+        "target_revision": args.target_revision,
+        "artifacts": artifact_digests,
+    })
+    object_id = f"slayer://object/dataset-source/{SOURCE_NAME}"
+    version_id = f"slayer://version/dataset-source/{SOURCE_NAME}@sha256:{version_digest}"
+    protocol_object_id = f"slayer://object/protocol/{SOURCE_NAME}-filter"
+    protocol_config = {
+        "source_revision": args.source_revision,
+        "minimum_characters": 200,
+        "minimum_polish_diacritic_ratio": 0.005,
+        "deduplication": {"algorithm": "sha1", "field": "normalized_text", "scope": "within_source"},
+        "tokenizer": "cl100k_base",
+        "date_valid_year_range": [1900, dt.date.today().year + 1],
+        "normalization": ["NFKC", "space_quote_dash", "paragraph_preserving_whitespace", "pattern_PII_scrub"],
+        "execution": {
+            "record_batch_size": args.batch_size,
+            "source_workers": args.workers,
+            "token_threads": args.token_threads,
+            "token_threads_per_worker": max(1, args.token_threads // min(max(1, args.workers), len(source_shards))),
+            "parquet_compression": "zstd",
+        },
+    }
+    protocol_spec = {
+        "entrypoint": "python src/fetch_saos.py",
+        "source_path": "src/fetch_saos.py",
+        "summary": "Deterministic normalization and projection of a pinned SAOS judgment snapshot.",
+        "configuration": protocol_config,
+        "required_inputs": ["14 pinned source Parquet shards", "source profile", "target registry", "target findings log"],
+        "outputs": ["DynaWord Parquet", "attribution sidecar", "datasheet", "stats", "legal evidence", "ontology manifest"],
+    }
+    protocol_code_digest = sha256_file(pathlib.Path(__file__).resolve())
+    protocol_digest = canonical_sha256({
+        "object_id": protocol_object_id,
+        "code_sha256": protocol_code_digest,
+        "specification": protocol_spec,
+    })
+    protocol_version_id = f"slayer://version/protocol/{SOURCE_NAME}-filter@sha256:{protocol_digest}"
+    source_version_id = f"hf://datasets/{SOURCE_REPOSITORY}@{args.source_revision}"
+    target_version_id = f"hf://datasets/{TARGET_REPOSITORY}@{args.target_revision}"
+    run_finished_at = utc_now()
+
+    def evidence(kind: str, payload: Any) -> dict[str, Any]:
+        digest = canonical_sha256({"run": args.run_id, "subject": version_id, "kind": kind, "payload": payload})
+        return {
+            "id": f"slayer://evidence/{SOURCE_NAME}/{kind}@sha256:{digest}",
+            "run_id": args.run_id,
+            "subject_version": version_id,
+            "observation_type": kind,
+            "payload": payload,
+            "observed_at": run_finished_at,
+        }
+
+    evidence_items = [
+        evidence("dataset_record_counts", stats),
+        evidence("artifact_checksums", artifact_digests),
+        evidence("pinned_source_inventory", source_inventory),
+        evidence("source_completeness_observation", {
+            "snapshot_rows": stats["source_rows"],
+            "official_total_results": args.official_total_results,
+            "observed_at": args.official_count_observed_at,
+            "ratio": stats["snapshot_live_count_ratio"],
+            "complete": False,
+        }),
+        evidence("normalization_and_pii_metrics", stats["normalization"]),
+        evidence("schema_validation", {"schema": list(EXPECTED_FIELDS), "records": stats["kept"], "valid": True}),
+        evidence("legal_provenance_review", {
+            "status": "documented-statutory-basis-review-needed",
+            "formal_legal_review": "not_performed",
+            "artifact": f"data/{SOURCE_NAME}/{SOURCE_NAME}.license-evidence.json",
+            "sha256": artifact_digests["legal_evidence_sha256"],
+        }),
+    ]
+    evidence_by_type = {item["observation_type"]: item["id"] for item in evidence_items}
+
+    def claim(statement: str, falsification: str, support: list[str]) -> dict[str, Any]:
+        digest = canonical_sha256({"scope": version_id, "statement": statement, "falsification": falsification})
+        return {
+            "id": f"slayer://claim/{SOURCE_NAME}@sha256:{digest}",
+            "scope": version_id,
+            "statement": statement,
+            "asserted_by": args.actor,
+            "supported_by": support,
+            "falsification_condition": falsification,
+        }
+
+    claims = [
+        claim(
+            "The contribution is derived from the 14 files of the pinned PiotrSty SAOS snapshot.",
+            "A source file is missing or its recorded size or SHA-256 differs.",
+            [evidence_by_type["pinned_source_inventory"]],
+        ),
+        claim(
+            "The contributed dataset is a partial historical SAOS snapshot and not a complete current copy.",
+            "The pinned snapshot row count equals the official live count observed for this run.",
+            [evidence_by_type["source_completeness_observation"]],
+        ),
+        claim(
+            "All retained rows conform to the DynaWord source schema and recorded artifact checksums.",
+            "A retained row violates the schema or an artifact checksum differs.",
+            [evidence_by_type["schema_validation"], evidence_by_type["artifact_checksums"]],
+        ),
+        claim(
+            "Pattern-based PII scrubbing and documented text normalization were applied, without asserting perfect de-identification.",
+            "The committed protocol omits a documented transformation or a retained match contradicts its recorded metric.",
+            [evidence_by_type["normalization_and_pii_metrics"]],
+        ),
+        claim(
+            "The records are court materials with a documented statutory reuse argument, pending formal legal review and subject to stated limitations.",
+            "The records are not official materials or legal review rejects the documented statutory basis.",
+            [evidence_by_type["legal_provenance_review"]],
+        ),
+    ]
+    claim_evidence = [
+        {"claim_id": item["id"], "evidence_id": evidence_id, "relation": "SUPPORTS", "introduced_by_run": args.run_id}
+        for item in claims for evidence_id in item["supported_by"]
+    ]
+
+    def attestation(kind: str, value: Any, support: list[str]) -> dict[str, Any]:
+        digest = canonical_sha256({"type": kind, "value": value, "actor": args.actor, "subject": version_id, "support": support})
+        return {
+            "id": f"slayer://attestation/{SOURCE_NAME}/{kind}@sha256:{digest}",
+            "type": kind,
+            "value": value,
+            "actor_id": args.actor,
+            "subject_version": version_id,
+            "supported_by": support,
+            "created_at": run_finished_at,
+        }
+
+    manifest = {
+        "schema_version": "slayer.ai/dataset-source-contribution/v1",
+        "ontology_version": "slayer.ai/research-ontology/v0.1",
+        "object": {
+            "id": object_id, "kind": "dataset_source", "name": SOURCE_NAME,
+            "metadata": {"language": "pl", "domain": "legal/judicial", "provider": "SAOS"},
+        },
+        "version": {
+            "id": version_id,
+            "object_id": object_id,
+            "digest": f"sha256:{version_digest}",
+            "schema_version": "slayer.ai/dataset-source/v1",
+            "payload_uri": f"hf://datasets/{TARGET_REPOSITORY}@{args.dataset_payload_commit}/data/{SOURCE_NAME}/{SOURCE_NAME}.parquet",
+            "created_at": run_finished_at,
+            "created_by": args.actor,
+        },
+        "protocol": {
+            "id": protocol_object_id,
+            "kind": "dataset_filtering",
+            "name": f"{SOURCE_NAME}-filter",
+            "version": {
+                "id": protocol_version_id,
+                "object_id": protocol_object_id,
+                "digest": f"sha256:{protocol_digest}",
+                "code_sha256": protocol_code_digest,
+                "schema_version": "slayer.ai/protocol/v1",
+                "payload_uri": f"hf://datasets/{TARGET_REPOSITORY}@{args.protocol_payload_commit}/src/fetch_saos.py",
+                "created_at": run_finished_at,
+                "created_by": args.actor,
+            },
+            "specification": protocol_spec,
+        },
+        "run": {
+            "id": args.run_id,
+            "protocol_version_id": protocol_version_id,
+            "actor_id": args.actor,
+            "git_commit": args.protocol_payload_commit,
+            "started_at": run_started_at,
+            "finished_at": run_finished_at,
+            "config": protocol_config,
+            "environment": {
+                "python": sys.version,
+                "python_executable": sys.executable,
+                "platform": platform.platform(),
+                "machine": platform.machine(),
+                "packages": {"pyarrow": pa.__version__, "tiktoken": importlib.metadata.version("tiktoken")},
+                "container_image": None,
+                "environment_limitation": "No container image was used; material runtime and package versions are recorded.",
+                "determinism": {"sampling": "none", "seed": None, "input_order": [path.name for path in source_shards]},
+            },
+            "inputs": [
+                {"role": "source", "version_id": source_version_id},
+                {"role": "target_base", "version_id": target_version_id},
+            ],
+            "outputs": [{"role": "dataset_source", "version_id": version_id}],
+        },
+        "relations": [
+            {"source_version_id": version_id, "predicate": "DERIVED_FROM", "target_version_id": source_version_id, "introduced_by_run": args.run_id},
+            {"source_version_id": version_id, "predicate": "COMPATIBLE_WITH", "target_version_id": target_version_id, "introduced_by_run": args.run_id},
+            {"source_version_id": version_id, "predicate": "FILTERED_BY", "target_version_id": protocol_version_id, "introduced_by_run": args.run_id},
+        ],
+        "evidence": evidence_items,
+        "claims": claims,
+        "claim_evidence": claim_evidence,
+        "actors": [{
+            "id": args.actor, "kind": "human", "identity": "Piotr Sty\u0142a",
+            "identifiers": {"hugging_face": "PiotrSty", "github": "PiotrStyla"},
+        }],
+        "attestations": [
+            attestation("schema_valid", True, [evidence_by_type["schema_validation"], evidence_by_type["artifact_checksums"]]),
+            attestation("source_completeness", "partial_snapshot", [evidence_by_type["source_completeness_observation"]]),
+            attestation("provenance", "pinned_and_hashed", [evidence_by_type["pinned_source_inventory"]]),
+            attestation("pii_scrubbing", "pattern_scrub_applied_not_guaranteed", [evidence_by_type["normalization_and_pii_metrics"]]),
+            attestation("legal_status", "review_needed_documented_official_materials", [evidence_by_type["legal_provenance_review"]]),
+        ],
+    }
+    write_json(artifact_dir / f"{SOURCE_NAME}_ontology_manifest.json", manifest)
+    (src_dir / "sources.py").write_text(
+        update_registry(args.target_registry.read_text(encoding="utf-8"), stats, args.source_revision),
+        encoding="utf-8",
+    )
+    (artifact_dir / "source_findings.md").write_text(
+        update_findings(args.target_findings.read_text(encoding="utf-8"), stats, args),
+        encoding="utf-8",
+    )
+    shutil.copy2(pathlib.Path(__file__).resolve(), src_dir / "fetch_saos.py")
+    (src_dir / "test_saos_contract.py").write_text(contract_test(), encoding="utf-8")
+    print(json.dumps(stats, ensure_ascii=False, indent=2, sort_keys=True))
+    return stats
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--source-dir", type=pathlib.Path, required=True)
+    parser.add_argument("--source-profile", type=pathlib.Path, required=True)
+    parser.add_argument("--target-registry", type=pathlib.Path, required=True)
+    parser.add_argument("--target-findings", type=pathlib.Path, required=True)
+    parser.add_argument("--output", type=pathlib.Path, required=True)
+    parser.add_argument("--source-revision", required=True)
+    parser.add_argument("--target-revision", required=True)
+    parser.add_argument("--dataset-payload-commit", required=True)
+    parser.add_argument("--protocol-payload-commit", required=True)
+    parser.add_argument("--official-total-results", type=int, required=True)
+    parser.add_argument("--official-count-observed-at", required=True)
+    parser.add_argument("--added-date", required=True)
+    parser.add_argument("--actor", default="hf:PiotrSty")
+    parser.add_argument("--run-id", required=True)
+    parser.add_argument("--batch-size", type=int, default=32)
+    parser.add_argument("--workers", type=int, default=1)
+    parser.add_argument("--token-threads", type=int, default=4)
+    args = parser.parse_args()
+    build(args)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

src/sources.py

@@ -204,6 +204,27 @@ SOURCES = {
         "created": "2004-03-22, 2026-10-01",
         "is_ocr": False,
     },
+    "saos": {
+        "added": "2026-09-04",
+        "release": None,
+        "file_key": "saos",
+        "pretty": "SAOS - Polish court judgments (pinned partial snapshot)",
+        "license": "public-domain (official documents)",
+        "license_spdx": "LicenseRef-Polish-Official-Documents",
+        "traceable": "Court judgments are documented as official materials outside "
+                     "copyright under Polish Copyright Act art. 4(2); formal legal "
+                     "review has not been performed and statutory limitations remain - "
+                     "the review scope and responsible party are recorded in the datasheet.",
+        "upstream": "https://www.saos.org.pl/api/dump/judgments",
+        "provenance": "Converted from the immutable "
+                      "PiotrSty/saos-polish-court-judgments@736cba4f92999e29d3e49409b0713c88fc683714 snapshot; per-record "
+                      "SAOS URLs and source identifiers are preserved in a sidecar. Pattern PII scrubbing only - "
+                      "re-run through SlayerLab/NERGAL before release.",
+        "domain": "legal/judicial",
+        "created": "1986-05-28, 2017-05-11",
+        "is_ocr": False,
+        "custom_datasheet": True,
+    },
     "eurlex": {
         "added": "2026-06-15",
         "release": "0.2.0",

src/test_saos_contract.py

@@ -0,0 +1,147 @@
+"""Contract and ontology checks for the saos contribution."""
+
+import hashlib
+import json
+import pathlib
+import re
+
+import pyarrow.parquet as pq
+
+
+ROOT = pathlib.Path(__file__).resolve().parents[1]
+DATA = ROOT / "data" / "saos"
+MANIFEST = ROOT / "artifacts" / "saos_ontology_manifest.json"
+FIELDS = ['id', 'text', 'source', 'added', 'created', 'token_count', 'license', 'author']
+
+
+def sha256_file(path):
+    digest = hashlib.sha256()
+    with path.open("rb") as stream:
+        for chunk in iter(lambda: stream.read(1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def test_data_schema_and_stats():
+    parquet = DATA / "saos.parquet"
+    stats = json.loads((DATA / "saos.stats.json").read_text(encoding="utf-8"))
+    source_values, license_values, ids = set(), set(), set()
+    rows = chars = tokens = authors = 0
+    for batch in pq.ParquetFile(parquet).iter_batches(batch_size=256):
+        assert batch.schema.names == FIELDS
+        for row in batch.to_pylist():
+            rows += 1
+            assert row["id"] not in ids
+            ids.add(row["id"])
+            assert len(row["text"]) >= 200
+            assert row["token_count"] > 0
+            assert not re.search(
+                r"</?(?:p|div|span|br|h[1-6]|table|tbody|thead|tr|td|th|ul|ol|li|strong|em|b|i|a|section)\s*/?>",
+                row["text"],
+                re.I,
+            )
+            assert not row["created"] or re.fullmatch(r"20\d{2}-\d{2}-\d{2}|19\d{2}-\d{2}-\d{2}", row["created"])
+            source_values.add(row["source"])
+            license_values.add(row["license"])
+            chars += len(row["text"])
+            tokens += row["token_count"]
+            authors += bool(row["author"])
+    assert rows == stats["kept"]
+    assert chars == stats["chars"]
+    assert tokens == stats["tokens"]
+    assert authors == stats["authors_with_value"]
+    assert source_values == {"saos"}
+    assert license_values == {"public-domain (official documents)"}
+
+
+def test_sidecar_legal_and_snapshot_honesty():
+    stats = json.loads((DATA / "saos.stats.json").read_text(encoding="utf-8"))
+    sidecar = DATA / "saos.attribution.jsonl"
+    assert sum(1 for _ in sidecar.open(encoding="utf-8")) == stats["kept"]
+    legal = json.loads((DATA / "saos.license-evidence.json").read_text(encoding="utf-8"))
+    assert legal["formal_legal_review"] == "not_performed"
+    assert legal["hf_source_card_license"] == "unknown"
+    assert legal["pinned_source"].endswith(stats["source_revision"])
+    assert stats["official_total_results"] > stats["source_rows"]
+
+
+def test_ontology_kernel_and_references():
+    manifest = json.loads(MANIFEST.read_text(encoding="utf-8"))
+    required = {"object", "version", "protocol", "run", "relations", "evidence",
+                "claims", "claim_evidence", "actors", "attestations"}
+    assert required <= set(manifest)
+    assert manifest["ontology_version"] == "slayer.ai/research-ontology/v0.1"
+    actor_ids = {item["id"] for item in manifest["actors"]}
+    version = manifest["version"]
+    assert version["object_id"] == manifest["object"]["id"]
+    assert version["created_by"] in actor_ids
+    assert re.fullmatch(r"sha256:[0-9a-f]{64}", version["digest"])
+    assert re.search(r"@[0-9a-f]{40}/", version["payload_uri"])
+    protocol = manifest["protocol"]
+    protocol_version = protocol["version"]
+    assert protocol_version["object_id"] == protocol["id"]
+    assert manifest["run"]["protocol_version_id"] == protocol_version["id"]
+    assert re.search(r"@[0-9a-f]{40}/", protocol_version["payload_uri"])
+    assert sha256_file(ROOT / protocol["specification"]["source_path"]) == protocol_version["code_sha256"]
+    run = manifest["run"]
+    assert run["actor_id"] in actor_ids
+    assert run["started_at"] <= run["finished_at"]
+    assert {"python", "platform", "machine", "packages", "determinism"} <= set(run["environment"])
+
+    evidence_ids = {item["id"] for item in manifest["evidence"]}
+    claim_ids = {item["id"] for item in manifest["claims"]}
+    assert len(evidence_ids) == len(manifest["evidence"])
+    assert len(claim_ids) == len(manifest["claims"])
+    for item in manifest["evidence"]:
+        assert item["run_id"] == run["id"]
+        assert item["subject_version"] == version["id"]
+    for item in manifest["claims"]:
+        assert item["scope"] == version["id"]
+        assert item["asserted_by"] in actor_ids
+        assert item["falsification_condition"]
+        assert set(item["supported_by"]) <= evidence_ids
+    for link in manifest["claim_evidence"]:
+        assert link["claim_id"] in claim_ids
+        assert link["evidence_id"] in evidence_ids
+        assert link["relation"] in {"SUPPORTS", "REFUTES", "QUALIFIES"}
+    for attestation in manifest["attestations"]:
+        assert attestation["actor_id"] in actor_ids
+        assert attestation["subject_version"] == version["id"]
+        assert set(attestation["supported_by"]) <= evidence_ids
+    assert any(item["type"] == "source_completeness" and item["value"] == "partial_snapshot" for item in manifest["attestations"])
+
+    edges = {}
+    for relation in manifest["relations"]:
+        source = relation["source_version_id"]
+        target = relation["target_version_id"]
+        assert source != target
+        assert relation["introduced_by_run"] == run["id"]
+        if target.startswith("slayer://version/"):
+            edges.setdefault(source, set()).add(target)
+    visiting, visited = set(), set()
+    def visit(node):
+        assert node not in visiting, "version lineage must be acyclic"
+        if node in visited:
+            return
+        visiting.add(node)
+        for target in edges.get(node, ()):
+            visit(target)
+        visiting.remove(node)
+        visited.add(node)
+    for node in edges:
+        visit(node)
+
+
+def test_manifest_artifact_checksums():
+    manifest = json.loads(MANIFEST.read_text(encoding="utf-8"))
+    checksums = next(item["payload"] for item in manifest["evidence"] if item["observation_type"] == "artifact_checksums")
+    paths = {
+        "parquet_sha256": DATA / "saos.parquet",
+        "attribution_sha256": DATA / "saos.attribution.jsonl",
+        "stats_sha256": DATA / "saos.stats.json",
+        "datasheet_sha256": DATA / "saos.md",
+        "legal_evidence_sha256": DATA / "saos.license-evidence.json",
+        "source_inventory_sha256": ROOT / "artifacts" / "saos_source_inventory.json",
+        "source_profile_sha256": ROOT / "artifacts" / "saos_source_profile.json",
+    }
+    assert {name: sha256_file(path) for name, path in paths.items()} == checksums