Skip to benchmark content
← Leaderboard

OpenDisclosure Redaction

SystemMaintainer

OpenDisclosure's model and rule engine, run together by the benchmark harness.

Vendor
OpenDisclosure
Provider
Mistral La Plateforme
Region
EU
Version
pilot 2026-10-14
Composition
Mistral Small 2603 at Mistral EU, with OpenDisclosure's rule engine and detector stages

F2 score

73.6%95% CI 67.3%–79.1%

Leak recall

71.8%95% CI 64.8%–78.0%

Over-redaction

18.2%95% CI 14.6%–23.9%

Output precision

64.7%95% CI 55.3%–71.1%

Cost per 1,000 pages

EUR 0.5195% CI not applicable · measured list-price cost

Recall by bucket

Universal identifiers
75.6%95% CI 66.7%–82.9%
Names
92.6%95% CI 86.9%–96.1%
Contextual personal data
48.1%95% CI 35.2%–55.9%

Breakdowns

Groups below the reporting floor are not shown.

LanguageDocumentsF2RecallOver-redaction
de2473.4%71.7%19.0%
en2474.8%74.2%23.0%
es2276.5%74.7%15.7%
fr2080.3%79.6%16.6%
it2359.6%57.4%29.9%
nl2274.6%72.6%16.2%

Provenance

Dataset version and split
1.0.0-r3+labels.20261006.1 · tuning
Scoring
1.3
Harness version
pilot-2026-10-15
Run date
2026-10-14
Documents
135
Pages
892
Repeats
1
Price date
2026-09-29
Spend
EUR 0.46
Credits
None · paid by OpenDisclosure

Cite this

@misc{eu-redaction-opendisclosure-redaction-no-roster,
  author = {OpenDisclosure},
  title = {EU Redaction Benchmark: OpenDisclosure Redaction},
  year = {2026},
  url = {https://opendisclosure.eu/benchmark/systems/opendisclosure-redaction/no-roster/},
  note = {r3-tuning-pilot-2026-10-15; tuning split; scoring 1.3}
}
Download JSON