Product Pricing Security & data Research Publications Blog About Careers Alpha Talk to us Download alpha
Research Open artefacts

A number nobody can reproduce is an advertisement.

Every paper ships the thing you would need to check it: the corpus, the harness, the raw attribution table, or all three. Some are public downloads, some are released on request under a licence, and a small number are restricted because releasing them would be irresponsible. We say which is which.

Public download19Datasets, harnesses, and attribution tables under MIT, Apache-2.0, or CC BY 4.0.
On request9Released after a short exchange, usually to confirm intended use.
Restricted2The injection corpus, under a responsible-disclosure agreement.
Index

Everything, and which paper it belongs to.

ArtefactKindDetailPaper
dart-chunkerharnessSix strategies, pluggable · MITDBR-2026-07
Held-out evaluation slicedata200 documents, 120 queries, untouched during developmentDBR-2026-07
Failure labelsdata400 hand-labelled misses, before and after · CC BY 4.0DBR-2026-07
Dialog designsspecAll five, as rendered HTML · CC BY 4.0DBR-2026-06
Anonymised responsesdata340 participants · free-text scored · on requestDBR-2026-06
Study protocolspecRecruitment, framing, debrief scriptDBR-2026-06
dart-quant-benchharnessEvaluation harness · 640-item set · MITDBR-2026-05
browsing-prompts-scrubbedcorpus1,840 real agent prompts, de-identified · on requestDBR-2026-05
Raw per-run timingsdataCSV · 14 configs × 4 machines × 40 runsDBR-2026-05
Aggregated handshake timingsdataCSV · bucketed, no identifiers · CC BY 4.0DBR-2026-04
dart-tls-probeharnessMeasurement client · Apache-2.0DBR-2026-04
Network class inference methodspecWritten specificationDBR-2026-04
dart-injection-corpuscorpus1,400 attempts · 9 vectors · responsible-disclosure licenceDBR-2026-03
Agent harnessharnessIsolated runtime · on requestDBR-2026-03
Per-model appendixappendixVariance across four models · PDFDBR-2026-03
Destination attribution tabledataFull endpoint categorisation · CC BY 4.0DBR-2026-02
browser-phone-homeharnessCapture and workload scripts · MITDBR-2026-02
Category definitionsspecWritten specification, versionedDBR-2026-02
dart-history-corpuscorpus41,000 synthetic documents · generator included · MITDBR-2026-01
1,200 labelled queriesdataHuman relevance judgements · CC BY 4.0DBR-2026-01
Local evaluation harnessharnessRuns against your own history, reports aggregates onlyDBR-2026-01
Test batteryharnessExport, attestation, algorithm, firmware · Apache-2.0DBR-2025-03
Anonymised aggregatedata47 models, results without vendor attribution · CC BY 4.0DBR-2025-03
Per-model breakdowndataRestricted · researchers, on requestDBR-2025-03
dart-fp-probeharness42-attribute measurement probe · MITDBR-2025-02
Aggregated entropy tablesdataPer-attribute distributions · CC BY 4.0DBR-2025-02
Consent flowspecExact text and attribute disclosure shown to participantsDBR-2025-02
dart-spec-benchharnessSpeculative decoding harness · MITDBR-2025-01
1,100 agent turnscorpusScrubbed, fixed static set · on requestDBR-2025-01
Per-run timingsdataCSV · 5 pairings × 4 machines × 30 runsDBR-2025-01
Access

Three tiers, and why.

Tier one

Public download

Harnesses, synthetic corpora, labelled query sets, and attribution tables. No sign-up, no email gate, no click-through. If it is safe to publish, it is published.

licences · MIT, Apache-2.0, CC BY 4.0
Tier two

On request

De-identified real-world corpora, where the de-identification is good but we would rather know who holds a copy. A short exchange, usually resolved the same week.

request · research@dartbrowser.com
Tier three

Restricted

The prompt-injection corpus. It is a working set of attacks against deployed agentic browsers, including ones that are not ours. Released to security researchers under a disclosure agreement.

restricted · security@dartbrowser.com · 90-day disclosure

We would rather over-publish than under-publish, and we have made the opposite mistake once: the first release of the history corpus contained near-duplicate documents that inflated recall by roughly two points. It was caught by an external reader, and v1.0 fixes it. That correction is noted on DBR-2026-01.

Replication

Run it against your own data.

The retrieval harness runs entirely locally against your own browsing history and reports aggregate scores only. Nothing leaves your machine. If your numbers differ from ours, that is the interesting case and we want to hear about it.