Datasets:

BoscoLabs
/

counterparty-check-benchmark

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
date: timestamp[s]
model: string
api_base: string
cap_usd: double
truth_checked: timestamp[s]
totals: struct<web: struct<n: int64, accuracy: struct<status: struct<correct: int64, n: int64>, sanctioned:  (... 702 chars omitted)
  child 0, web: struct<n: int64, accuracy: struct<status: struct<correct: int64, n: int64>, sanctioned: struct<corre (... 273 chars omitted)
      child 0, n: int64
      child 1, accuracy: struct<status: struct<correct: int64, n: int64>, sanctioned: struct<correct: int64, n: int64>, insol (... 39 chars omitted)
          child 0, status: struct<correct: int64, n: int64>
              child 0, correct: int64
              child 1, n: int64
          child 1, sanctioned: struct<correct: int64, n: int64>
              child 0, correct: int64
              child 1, n: int64
          child 2, insolvent: struct<correct: int64, n: int64>
              child 0, correct: int64
              child 1, n: int64
      child 2, all_correct: int64
      child 3, missed_sanctions: int64
      child 4, false_sanction_flags: int64
      child 5, no_answer: int64
      child 6, errors: int64
      child 7, usd: double
      child 8, api_usd: int64
      child 9, seconds: double
      child 10, searches: int64
      child 11, fetches: int64
      child 12, no_tool_use: int64
  child 1, api: struct<n: int64, accuracy: struct<status: struct<correct: int64, n: int64>, sanctioned: struct<corre (... 274 chars omitted)
      child 0, n: int64
      child 1, accuracy: struct<st
...
              child 4, sanctions_entity: string
              child 5, sanctions_people: string
              child 6, vat_number: string
          child 4, answer_text: string
          child 5, model_usd: double
          child 6, usd: double
          child 7, seconds: double
          child 8, usage: struct<input_tokens: int64, output_tokens: int64, cache_creation_input_tokens: int64, cache_read_inp (... 58 chars omitted)
              child 0, input_tokens: int64
              child 1, output_tokens: int64
              child 2, cache_creation_input_tokens: int64
              child 3, cache_read_input_tokens: int64
              child 4, web_searches: int64
              child 5, web_fetches: int64
          child 9, answer: struct<status: string, sanctioned: bool, sanctions_lists: list<item: string>, insolvent: bool>
              child 0, status: string
              child 1, sanctioned: bool
              child 2, sanctions_lists: list<item: string>
                  child 0, item: string
              child 3, insolvent: bool
          child 10, score: struct<status: bool, sanctioned: bool, insolvent: bool>
              child 0, status: bool
              child 1, sanctioned: bool
              child 2, insolvent: bool
      child 6, usd: double
null_means: string
fields: struct<status: string, sanctioned: string, insolvent: string>
  child 0, status: string
  child 1, sanctioned: string
  child 2, insolvent: string
version: int64
checked: timestamp[s]
name: string
to
{'name': Value('string'), 'version': Value('int64'), 'checked': Value('timestamp[s]'), 'fields': {'status': Value('string'), 'sanctioned': Value('string'), 'insolvent': Value('string')}, 'null_means': Value('string'), 'companies': List({'id': Value('string'), 'name': Value('string'), 'country': Value('string'), 'identifier': List(Value('string')), 'api_input': {'query': Value('string'), 'country': Value('string')}, 'truth': {'status': Value('string'), 'sanctioned': Value('bool'), 'sanctions_lists': List(Value('string')), 'insolvent': Value('bool')}, 'sources': List({'url': Value('string'), 'checked': Value('timestamp[s]'), 'what': Value('string')}), 'note': Value('string')})}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 544, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3052, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 558, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 361, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2381, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              date: timestamp[s]
              model: string
              api_base: string
              cap_usd: double
              truth_checked: timestamp[s]
              totals: struct<web: struct<n: int64, accuracy: struct<status: struct<correct: int64, n: int64>, sanctioned:  (... 702 chars omitted)
                child 0, web: struct<n: int64, accuracy: struct<status: struct<correct: int64, n: int64>, sanctioned: struct<corre (... 273 chars omitted)
                    child 0, n: int64
                    child 1, accuracy: struct<status: struct<correct: int64, n: int64>, sanctioned: struct<correct: int64, n: int64>, insol (... 39 chars omitted)
                        child 0, status: struct<correct: int64, n: int64>
                            child 0, correct: int64
                            child 1, n: int64
                        child 1, sanctioned: struct<correct: int64, n: int64>
                            child 0, correct: int64
                            child 1, n: int64
                        child 2, insolvent: struct<correct: int64, n: int64>
                            child 0, correct: int64
                            child 1, n: int64
                    child 2, all_correct: int64
                    child 3, missed_sanctions: int64
                    child 4, false_sanction_flags: int64
                    child 5, no_answer: int64
                    child 6, errors: int64
                    child 7, usd: double
                    child 8, api_usd: int64
                    child 9, seconds: double
                    child 10, searches: int64
                    child 11, fetches: int64
                    child 12, no_tool_use: int64
                child 1, api: struct<n: int64, accuracy: struct<status: struct<correct: int64, n: int64>, sanctioned: struct<corre (... 274 chars omitted)
                    child 0, n: int64
                    child 1, accuracy: struct<st
              ...
                            child 4, sanctions_entity: string
                            child 5, sanctions_people: string
                            child 6, vat_number: string
                        child 4, answer_text: string
                        child 5, model_usd: double
                        child 6, usd: double
                        child 7, seconds: double
                        child 8, usage: struct<input_tokens: int64, output_tokens: int64, cache_creation_input_tokens: int64, cache_read_inp (... 58 chars omitted)
                            child 0, input_tokens: int64
                            child 1, output_tokens: int64
                            child 2, cache_creation_input_tokens: int64
                            child 3, cache_read_input_tokens: int64
                            child 4, web_searches: int64
                            child 5, web_fetches: int64
                        child 9, answer: struct<status: string, sanctioned: bool, sanctions_lists: list<item: string>, insolvent: bool>
                            child 0, status: string
                            child 1, sanctioned: bool
                            child 2, sanctions_lists: list<item: string>
                                child 0, item: string
                            child 3, insolvent: bool
                        child 10, score: struct<status: bool, sanctioned: bool, insolvent: bool>
                            child 0, status: bool
                            child 1, sanctioned: bool
                            child 2, insolvent: bool
                    child 6, usd: double
              null_means: string
              fields: struct<status: string, sanctioned: string, insolvent: string>
                child 0, status: string
                child 1, sanctioned: string
                child 2, insolvent: string
              version: int64
              checked: timestamp[s]
              name: string
              to
              {'name': Value('string'), 'version': Value('int64'), 'checked': Value('timestamp[s]'), 'fields': {'status': Value('string'), 'sanctioned': Value('string'), 'insolvent': Value('string')}, 'null_means': Value('string'), 'companies': List({'id': Value('string'), 'name': Value('string'), 'country': Value('string'), 'identifier': List(Value('string')), 'api_input': {'query': Value('string'), 'country': Value('string')}, 'truth': {'status': Value('string'), 'sanctioned': Value('bool'), 'sanctions_lists': List(Value('string')), 'insolvent': Value('bool')}, 'sources': List({'url': Value('string'), 'checked': Value('timestamp[s]'), 'what': Value('string')}), 'note': Value('string')})}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Counterparty check benchmark

Can an AI agent tell whether a company is still registered, sanctioned or insolvent? This dataset has the 25 companies we asked about, a hand-checked ground truth with a source URL for every field, and the full answers of two agents in two runs:

web: Claude Sonnet 5.5 with Anthropic's server-side web search and web fetch (at most 4 of each, boscolabs.dev blocked).
api: the same model, answering from the result of one call to the Bosco Labs counterparty-check API.

Bosco Labs makes that API. We built the API arm and wrote the ground truth, so treat this as a vendor benchmark and check our labels against the sources.

Results

	Run 2026-11-08: web	Run 2026-11-08: API	Run 2026-11-09: web	Run 2026-11-09: API
All three fields correct	23/25	17/25	19/25	20/25
Register status correct	23/24	21/24	20/24	24/24
Sanctioned correct	24/25	24/25	21/25	25/25
Insolvent correct	25/25	18/25	20/25	20/25
Sanctioned companies not flagged (of 6)	1	1	2	0
Clean companies flagged as sanctioned	0	0	0	0
Cost, total	$1.13	$1.66	$1.20	$1.72

The first run showed four bugs in the API (UK insolvency history, a deleted Estonian company returning HTTP 400, French and German liquidations reported as active, sanctions matched on name only). They were fixed before the second run; nothing changed on the web arm. One run per company per date and 25 companies: a difference of one or two companies is noise, and the web arm itself moved by four between the runs.

Files

ground_truth.json: the 25 companies. Per company: id, name, country, identifier, api_input (what the API arm sent), truth (status, sanctioned, sanctions_lists, insolvent), sources (URL, date checked, what it showed) and sometimes a note. Field definitions are in fields; a null truth field was not verified and is not scored.
runs/<date>-counterparty.json: one run. totals per arm, and per company the truth and for each arm the full answer_text, the parsed answer, the score per field, cost in USD, seconds, token usage, and for the web arm the number of searches and fetches; for the API arm the HTTP status, the charge and the check verdicts (api_checks).
runs/<date>-counterparty.md: the same run as a readable report.

The question

Both arms got the same prompt per company, built from ground_truth.json:

Counterparty check on {name} (country {country}, {identifier type} {identifier}). As of today:
1. Register status. {fields.status}.
2. Sanctioned: is the company itself on the UK Sanctions List, the OFAC SDN List or the EU consolidated financial sanctions list? Which lists?
3. Insolvent: {fields.insolvent}.
End your answer with one JSON object on its own line: {"status": "active|winding_up|dissolved|unknown", "sanctioned": true|false|null, "sanctions_lists": ["UK"|"OFAC SDN"|"EU", ...], "insolvent": true|false|null}

The API arm's system prompt tells the model to answer only from the tool result and to answer unknown or null where the result says nothing.

Scoring

The last JSON object in each answer is compared field by field with the ground truth. status must be equal; sanctioned and insolvent must be the same true or false. null or unknown counts as not correct; the run files count those apart from wrong answers. No model judges the answers.

Limits

25 hand-picked companies from 10 countries, 12 of them sanctioned, dissolved, in liquidation or bankrupt: far more hard cases than a normal supplier list.
Ground truth checked by hand on 2026-11-08. Where a register had no free access (KvK, Handelsregister, CVR) we used GLEIF, VIES or press, or left the field unscored. Camaïeu's status is arguable: Sirene says active and the liquidation has no closing notice, so we labelled it winding_up.
One model at low effort, capped searches. Some web-arm misses came from the tools (search cap reached, a fetch refused, a run that ended mid-search).
The API arm returns unverified for insolvency in Denmark, the US and for deleted Estonian companies; the model then answers unknown. That accounts for all 5 API unknowns in the second run.
Sanctions lists, registers and companies change. These labels are as of 2026-11-08 and will go stale.

Licence and sources

CC BY 4.0, © Bosco Labs. The labels are ours; the facts come from public sources, linked per field: UK Sanctions List and Companies House (Open Government Licence v3.0), OFAC SDN List (US government, public domain), EU consolidated financial sanctions list (European Commission, reuse under Decision 2011/833/EU), GLEIF (CC0), Brønnøysundregistrene (NLOD 2.0), PRH (CC BY 4.0), BODACC and Sirene (Licence Ouverte 2.0), VIES, the Estonian e-Business Register and KRS. Model outputs were generated with Claude (Anthropic).

Nothing here is compliance or legal advice. A sanctions answer from either arm is a screening aid, not a determination.

Contact

Corrections, especially a label you think is wrong: support@boscolabs.dev. More about the API: counterparty-check.

@misc{boscolabs_counterparty_benchmark_2026,
  title        = {Counterparty check benchmark: web search vs. counterparty-check API},
  author       = {{Bosco Labs}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/BoscoLabs/counterparty-check-benchmark}},
}

Downloads last month: 16

Total file size:

269 kB