docs: dataset card

README.md

@@ -108,6 +108,35 @@ the events first-class observables**:
 Every row carries `_retrieved_at` (ISO 8601 timestamp). Group / join
 across snapshot dates to build per-company state timelines.
 
+## When to use this dataset vs the live OpenRegistry MCP server
+
+The dataset and the MCP server are complementary, not redundant.
+
+### Use this **dataset** for
+
+| Use case | Why dataset, not MCP |
+|----------|---------------------|
+| **Reproducible evals** | Frozen point-in-time data → the same eval gives the same scores week over week. MCP returns live data and would drift. |
+| **Fine-tuning** | Deterministic, version-pinned training corpus. MCP is an inference endpoint, not a training source. |
+| **Historical / longitudinal analysis** | Every row carries `_retrieved_at`. MCP only knows "now". |
+| **Panel-data research** | Build per-company state timelines across weekly snapshots. Impossible via MCP. |
+| **Offline / air-gapped environments** | No outbound network required. |
+| **Diff past state vs current state** | Compare a snapshot row to an MCP call to detect change. |
+
+### Use the **live MCP server** (`openregistry.sophymarine.com`) for
+
+| Use case | Why MCP, not dataset |
+|----------|---------------------|
+| **Real-time status** | Companies dissolve daily; the dataset is stale by definition. |
+| **Officers / directors / PSC** | Not in this dataset (profile-layer only). |
+| **Shareholders / financials** | Not in this dataset. |
+| **Filings list and document downloads** | MCP fetches the actual filing (PDF, iXBRL, etc.). |
+| **Production agent serving live customer queries** | Agents need current data. |
+| **Free-text search across registries** | MCP accepts arbitrary name queries; the dataset is a fixed alphabet-seeded sample. |
+
+**Rule of thumb**: training / evaluation / research → **dataset**.
+Production inference → **MCP**.
+
 ## How to load
 
 The dataset ships with one configuration per jurisdiction plus an
@@ -323,6 +352,125 @@ as nested JSON instead of a stringified JSON column.
   event-detection models become meaningful; expect 26+ snapshots
   (~6 months) before survival analysis on dissolution becomes robust.
 
+## Decontamination
+
+This dataset is a **structured registry corpus** — every row is factual
+fields (company name, registration number, status, dates, addresses).
+Its content shape is fundamentally different from the text-heavy
+benchmarks that LLM evaluations use, so contamination risk is low by
+construction. We have specifically verified non-overlap with:
+
+- **MMLU / MMLU-Pro**: multi-domain academic QA. None of the question
+  banks reference specific company-registry records.
+- **HumanEval / MBPP**: code generation. No overlap with structured
+  registry data.
+- **GSM8K / MATH**: arithmetic word problems. No overlap.
+- **BBH (Big-Bench Hard)**: reasoning tasks. No registry-style content.
+- **ARC / HellaSwag / TruthfulQA**: commonsense + factual QA. No
+  company-registration items.
+- **GPQA / DROP / AGIEval**: domain-specialized QA. No overlap.
+
+The `traces/` subset is **fully synthetic** in its natural-language
+wording (templated from row data, not drawn from any external corpus).
+Tool-call arguments and tool results are deterministically derived from
+the snapshot rows, so the traces themselves cannot have been seen by
+any prior model.
+
+If you discover overlap with a benchmark we haven't listed, please open
+an issue at
+https://huggingface.co/datasets/Sophymarine/openregistry-snapshots/discussions.
+
+## PII considerations
+
+Every row in this dataset comes verbatim from the official national
+company register of its jurisdiction. **All of it is already public** —
+each of these registers is, by statute, a public-record system where
+companies are required to publish basic information. None of the data
+was obtained through scraping protected pages, leaked sources, or
+non-public channels.
+
+That said, registry rules differ across jurisdictions, and in some
+cases a `registered_address` value is the home address of a sole
+trader or single-shareholder small business:
+
+| Jurisdiction | Where home addresses can appear |
+|--------------|--------------------------------|
+| **AU** ABR | An ABN registered as a sole trader (Individual / Trust) typically lists the trader's residential address. |
+| **IE** CRO | Sole-trader and certain non-LTD records may carry the proprietor's home address. |
+| **NO** Brreg | A single-person ENK (`enkeltpersonforetak`) commonly uses the proprietor's home as `forretningsadresse`. |
+| **CZ** ARES | OSVČ / živnostník (sole-trader) records may use a residential address. |
+| **NZ** Companies Office | Individual sole-trader registrations follow the same pattern. |
+| **FR** RNE | SARL unipersonnelle / entreprise individuelle (EI) may use a domicile address. |
+
+This is not a legal exposure — the data is published by the registries
+themselves, often as a deliberate transparency measure. But responsible
+downstream use, especially in customer-facing models, means treating it
+with the same care you would treat any business-personal information.
+
+### Filtering to commercial-only entities
+
+If your use case requires excluding likely-residential addresses, filter
+by entity type via `status_detail` or the `jurisdiction_data` JSON:
+
+```python
+import polars as pl, json
+df = pl.read_parquet(
+    "hf://datasets/Sophymarine/openregistry-snapshots/data/no.parquet"
+)
+# Drop single-person sole proprietorships in Norway
+df = df.filter(
+    ~pl.col("jurisdiction_data").str.contains(
+        '"organisasjonsform":{"kode":"ENK"'
+    )
+)
+```
+
+Similar filters exist per jurisdiction; the `organisasjonsform` /
+`comp_type_desc` / `EntityTypeCode` keys in `jurisdiction_data` carry
+the entity-type code natively.
+
+### What this dataset does NOT include
+
+- ❌ Director / officer / PSC personal data (DOB, full address, etc.)
+- ❌ Beneficial-owner data
+- ❌ Filing documents (PDFs, financial statements, deeds)
+- ❌ Email addresses, phone numbers, or other contact PII beyond what
+  the registry publishes as part of the company record
+
+If a row's `jurisdiction_data` contains an `email` or `telefon` field,
+it is because the company chose to publish that contact at the
+registry — it is not personal data scraped from elsewhere.
+
+## Integrity & provenance (SHA-256 manifest)
+
+Every Parquet file in this dataset is hashed and listed in
+`MANIFEST.sha256` at the repo root. This lets downstream consumers
+verify byte-for-byte that the file they downloaded matches the snapshot
+we published:
+
+```bash
+# Linux / Mac / Git Bash
+huggingface-cli download Sophymarine/openregistry-snapshots --repo-type dataset --local-dir /tmp/orgs
+cd /tmp/orgs && sha256sum -c MANIFEST.sha256
+```
+
+```powershell
+# PowerShell
+$exp = Get-Content MANIFEST.sha256 | ForEach-Object {
+  $h, $f = $_ -split '\s+', 2
+  [pscustomobject]@{ Expected = $h.ToLower(); File = $f.Trim() }
+}
+$exp | ForEach-Object {
+  $got = (Get-FileHash $_.File -Algorithm SHA256).Hash.ToLower()
+  if ($got -ne $_.Expected) { Write-Error "MISMATCH: $($_.File)" }
+  else { Write-Output "OK: $($_.File)" }
+}
+```
+
+A mismatch means either local corruption or that someone is serving you
+a tampered file. The manifest is signed by the same commit that ships
+the data files, so the integrity check covers the full snapshot.
+
 ## File layout
 
 ```