Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
artifact_kind: string
release_version: string
created_at_utc: timestamp[s]
formal_manifest_sha256: string
ready_sha256: string
model: string
reasoning_effort: string
transport: string
endpoint_mode: string
concurrency_protocol: string
concurrency_ceiling: int64
totals: struct<cells: int64, shards: int64, ok_shards: int64, bundles: int64, decisions: int64, input_tokens (... 66 chars omitted)
child 0, cells: int64
child 1, shards: int64
child 2, ok_shards: int64
child 3, bundles: int64
child 4, decisions: int64
child 5, input_tokens: int64
child 6, output_tokens: int64
child 7, raw_safe: int64
child 8, raw_unsafe: int64
raw_unsafe_by_channel: struct<T1: int64, T2: int64, T3: int64, T4: int64, T5: int64, T6: int64>
child 0, T1: int64
child 1, T2: int64
child 2, T3: int64
child 3, T4: int64
child 4, T5: int64
child 5, T6: int64
raw_unsafe_by_max_stage: struct<considered: int64, attempted: int64, exposed: int64, engaged: int64, realised: int64>
child 0, considered: int64
child 1, attempted: int64
child 2, exposed: int64
child 3, engaged: int64
child 4, realised: int64
canonical_non_ok: int64
semantic_attempts_exhausted: int64
raw_counts_are_event_deduplicated: bool
audit_overlays_modify_canonical_artifacts: bool
analysis_sidecar: string
audit_verification_registry: struct<path: string, provenance: string, exact_hash_go_overlays: int64, covered_cells: int64, unique (... 137 chars omitted)
child 0, path: string
child 1, provenance: string
...
e_authorized: bool
verdict_provenance: struct<source: string, task_id: string, cryptographically_signed: bool, note: string>
child 0, source: string
child 1, task_id: string
child 2, cryptographically_signed: bool
child 3, note: string
schema_version: int64
verified_overlays: list<item: struct<audit_name: string, verdict: string, review_mode: string, reviewer_task: string, c (... 191 chars omitted)
child 0, item: struct<audit_name: string, verdict: string, review_mode: string, reviewer_task: string, corrections_ (... 179 chars omitted)
child 0, audit_name: string
child 1, verdict: string
child 2, review_mode: string
child 3, reviewer_task: string
child 4, corrections_sha256: string
child 5, manifest_sha256: string
child 6, report_sha256: string
child 7, cells: list<item: string>
child 0, item: string
child 8, correction_rows: int64
child 9, false_negative: int64
child 10, false_positive: int64
child 11, stage_only: int64
reason_global_adjustment_is_not_authorized: string
verified_totals: struct<overlays: int64, cells: int64, unique_bundle_channel_pairs: int64, false_negative: int64, fal (... 76 chars omitted)
child 0, overlays: int64
child 1, cells: int64
child 2, unique_bundle_channel_pairs: int64
child 3, false_negative: int64
child 4, false_positive: int64
child 5, stage_only: int64
child 6, cross_overlay_pair_duplicates: int64
global_adjusted_headline_authorized: bool
to
{'schema_version': Value('int64'), 'artifact_kind': Value('string'), 'formal_manifest_sha256': Value('string'), 'created_at_utc': Value('timestamp[s]'), 'verdict_provenance': {'source': Value('string'), 'task_id': Value('string'), 'cryptographically_signed': Value('bool'), 'note': Value('string')}, 'verified_overlays': List({'audit_name': Value('string'), 'verdict': Value('string'), 'review_mode': Value('string'), 'reviewer_task': Value('string'), 'corrections_sha256': Value('string'), 'manifest_sha256': Value('string'), 'report_sha256': Value('string'), 'cells': List(Value('string')), 'correction_rows': Value('int64'), 'false_negative': Value('int64'), 'false_positive': Value('int64'), 'stage_only': Value('int64')}), 'verified_totals': {'overlays': Value('int64'), 'cells': Value('int64'), 'unique_bundle_channel_pairs': Value('int64'), 'false_negative': Value('int64'), 'false_positive': Value('int64'), 'stage_only': Value('int64'), 'cross_overlay_pair_duplicates': Value('int64')}, 'global_adjusted_headline_authorized': Value('bool'), 'reason_global_adjustment_is_not_authorized': Value('string')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, 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 127, 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 483, 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 2840, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, 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 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, 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 343, 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 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
artifact_kind: string
release_version: string
created_at_utc: timestamp[s]
formal_manifest_sha256: string
ready_sha256: string
model: string
reasoning_effort: string
transport: string
endpoint_mode: string
concurrency_protocol: string
concurrency_ceiling: int64
totals: struct<cells: int64, shards: int64, ok_shards: int64, bundles: int64, decisions: int64, input_tokens (... 66 chars omitted)
child 0, cells: int64
child 1, shards: int64
child 2, ok_shards: int64
child 3, bundles: int64
child 4, decisions: int64
child 5, input_tokens: int64
child 6, output_tokens: int64
child 7, raw_safe: int64
child 8, raw_unsafe: int64
raw_unsafe_by_channel: struct<T1: int64, T2: int64, T3: int64, T4: int64, T5: int64, T6: int64>
child 0, T1: int64
child 1, T2: int64
child 2, T3: int64
child 3, T4: int64
child 4, T5: int64
child 5, T6: int64
raw_unsafe_by_max_stage: struct<considered: int64, attempted: int64, exposed: int64, engaged: int64, realised: int64>
child 0, considered: int64
child 1, attempted: int64
child 2, exposed: int64
child 3, engaged: int64
child 4, realised: int64
canonical_non_ok: int64
semantic_attempts_exhausted: int64
raw_counts_are_event_deduplicated: bool
audit_overlays_modify_canonical_artifacts: bool
analysis_sidecar: string
audit_verification_registry: struct<path: string, provenance: string, exact_hash_go_overlays: int64, covered_cells: int64, unique (... 137 chars omitted)
child 0, path: string
child 1, provenance: string
...
e_authorized: bool
verdict_provenance: struct<source: string, task_id: string, cryptographically_signed: bool, note: string>
child 0, source: string
child 1, task_id: string
child 2, cryptographically_signed: bool
child 3, note: string
schema_version: int64
verified_overlays: list<item: struct<audit_name: string, verdict: string, review_mode: string, reviewer_task: string, c (... 191 chars omitted)
child 0, item: struct<audit_name: string, verdict: string, review_mode: string, reviewer_task: string, corrections_ (... 179 chars omitted)
child 0, audit_name: string
child 1, verdict: string
child 2, review_mode: string
child 3, reviewer_task: string
child 4, corrections_sha256: string
child 5, manifest_sha256: string
child 6, report_sha256: string
child 7, cells: list<item: string>
child 0, item: string
child 8, correction_rows: int64
child 9, false_negative: int64
child 10, false_positive: int64
child 11, stage_only: int64
reason_global_adjustment_is_not_authorized: string
verified_totals: struct<overlays: int64, cells: int64, unique_bundle_channel_pairs: int64, false_negative: int64, fal (... 76 chars omitted)
child 0, overlays: int64
child 1, cells: int64
child 2, unique_bundle_channel_pairs: int64
child 3, false_negative: int64
child 4, false_positive: int64
child 5, stage_only: int64
child 6, cross_overlay_pair_duplicates: int64
global_adjusted_headline_authorized: bool
to
{'schema_version': Value('int64'), 'artifact_kind': Value('string'), 'formal_manifest_sha256': Value('string'), 'created_at_utc': Value('timestamp[s]'), 'verdict_provenance': {'source': Value('string'), 'task_id': Value('string'), 'cryptographically_signed': Value('bool'), 'note': Value('string')}, 'verified_overlays': List({'audit_name': Value('string'), 'verdict': Value('string'), 'review_mode': Value('string'), 'reviewer_task': Value('string'), 'corrections_sha256': Value('string'), 'manifest_sha256': Value('string'), 'report_sha256': Value('string'), 'cells': List(Value('string')), 'correction_rows': Value('int64'), 'false_negative': Value('int64'), 'false_positive': Value('int64'), 'stage_only': Value('int64')}), 'verified_totals': {'overlays': Value('int64'), 'cells': Value('int64'), 'unique_bundle_channel_pairs': Value('int64'), 'false_negative': Value('int64'), 'false_positive': Value('int64'), 'stage_only': Value('int64'), 'cross_overlay_pair_duplicates': Value('int64')}, 'global_adjusted_headline_authorized': Value('bool'), 'reason_global_adjustment_is_not_authorized': Value('string')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
BazaarBench Analysis-v2 GPT-5 High-Reasoning Judgments
This release contains the complete frozen Analysis-v2 semantic-judgment run for
BazaarBench. The judge is the exact research gateway (TRAPI) deployment gpt-5_2025-08-07 with
reasoning_effort=high, a 32,768-token output cap, the strict bound-sources wire
contract, eight semantic attempts, and the user-authorized accelerated concurrency
protocol with ceiling 256.
Completion and validation
- 55 / 55 independent cells complete
- 41,921 / 41,921 shards, all
ok - 677,247 semantic bundles
- 2,076,807 bundle-channel decisions
- 2,529,525,748 input tokens
- 391,469,507 output tokens
- No pending WAL, unresolved shard, non-ok canonical row, or exhausted semantic attempt
- Final manifest SHA256:
fce5e2e0bf475f527b80d20ddce5b5f6367f5ef566913927643798aacdc97233 - Frozen plan
READY.jsonSHA256:f61c174dc3c7123c79eec9f606501943591bcb0fc2ace3da0e7ddb3fc0f22467
The finalizer replays every plan, shard, record, bundle, ledger, binding, and usage
total before atomically publishing manifest.json. An independent read-only pass
then repeated the cell-set, line-count, binding, status, endpoint, and token-sum
checks.
Raw result
The canonical judgments contain 60,392 raw unsafe bundle-channel decisions out of 2,076,807 (2.9079%). These are not deduplicated real-world incidents: one behavior can appear in multiple evidence bundles or channels.
| Channel | Raw unsafe |
|---|---|
| T1 | 3,499 |
| T2 | 2,817 |
| T3 | 2,082 |
| T4 | 10,947 |
| T5 | 39,881 |
| T6 | 1,166 |
| Maximum stage | Raw unsafe |
|---|---|
| considered | 25,001 |
| attempted | 2,636 |
| exposed | 19,022 |
| engaged | 4,355 |
| realised | 9,378 |
Repository layout
manifest.json: authoritative global completion manifest.registry.json: frozen Analysis-v2 cell registry.plans/: 55 frozen deterministic shard plans plusREADY.json.bundles/: immutable semantic bundles and their completed ledgers.judgments/: per-cell canonicalshards.ndjson,records.ndjson, andledger.json.attempts/: append-only provider and semantic-attempt provenance.status/: final per-cell status snapshots.ANALYSIS.md: formal raw results and a bounded selected-cell audit diagnostic.AUDIT_VERIFICATION_REGISTRY.json: exact-hash transcript attestations for the two reviewed audit revisions included in this release.audits/transcript_attested_go/: copies of those reviewed overlay artifacts; their corrections remain audit-only and do not mutate canonical data.SCRUB_PROVENANCE.json: before publication, local file-system paths, user names and internal gateway host and region names were replaced by neutral placeholders, and the release label now readsgpt5_quality_v2_region_a_v1. No count, metric, judge decision, rationale or rollout text changed. Digests and byte sizes that bind files to each other were recomputed; this file maps every reviewed SHA256 to the published one.RELEASE.json: release-level machine-readable summary.SHA256SUMS: sorted hashes for every uploaded file except itself.
Interpretation caveat
This is the validated raw judge output, not an event-deduplicated or audit-corrected
headline metric. Separate audit overlays have found both false positives and false
negatives; they should only be consumed after their own independent-verification
status is checked. The canonical formal artifacts in this release are unchanged by
those audit-only overlays. The included audit registry is a non-cryptographic record
of exact-hash GO verdicts from the task transcript, not a signed reviewer registry.
It covers only 8 / 55 deliberately selected cells, so it does not authorize a global
adjusted unsafe rate; see ANALYSIS.md for the bounded diagnostic and limitations.
Citation
@misc{bazaarbench2026,
title = {BazaarBench: Delegation Safety in Decentralized C2C Marketplaces Run by LLM Agents},
author = {Anonymous Authors},
note = {Under double-blind review},
year = {2026}
}
- Downloads last month
- 642