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.
Everything, and which paper it belongs to.
| Artefact | Kind | Detail | Paper |
|---|---|---|---|
| dart-chunker | harness | Six strategies, pluggable · MIT | DBR-2026-07 |
| Held-out evaluation slice | data | 200 documents, 120 queries, untouched during development | DBR-2026-07 |
| Failure labels | data | 400 hand-labelled misses, before and after · CC BY 4.0 | DBR-2026-07 |
| Dialog designs | spec | All five, as rendered HTML · CC BY 4.0 | DBR-2026-06 |
| Anonymised responses | data | 340 participants · free-text scored · on request | DBR-2026-06 |
| Study protocol | spec | Recruitment, framing, debrief script | DBR-2026-06 |
| dart-quant-bench | harness | Evaluation harness · 640-item set · MIT | DBR-2026-05 |
| browsing-prompts-scrubbed | corpus | 1,840 real agent prompts, de-identified · on request | DBR-2026-05 |
| Raw per-run timings | data | CSV · 14 configs × 4 machines × 40 runs | DBR-2026-05 |
| Aggregated handshake timings | data | CSV · bucketed, no identifiers · CC BY 4.0 | DBR-2026-04 |
| dart-tls-probe | harness | Measurement client · Apache-2.0 | DBR-2026-04 |
| Network class inference method | spec | Written specification | DBR-2026-04 |
| dart-injection-corpus | corpus | 1,400 attempts · 9 vectors · responsible-disclosure licence | DBR-2026-03 |
| Agent harness | harness | Isolated runtime · on request | DBR-2026-03 |
| Per-model appendix | appendix | Variance across four models · PDF | DBR-2026-03 |
| Destination attribution table | data | Full endpoint categorisation · CC BY 4.0 | DBR-2026-02 |
| browser-phone-home | harness | Capture and workload scripts · MIT | DBR-2026-02 |
| Category definitions | spec | Written specification, versioned | DBR-2026-02 |
| dart-history-corpus | corpus | 41,000 synthetic documents · generator included · MIT | DBR-2026-01 |
| 1,200 labelled queries | data | Human relevance judgements · CC BY 4.0 | DBR-2026-01 |
| Local evaluation harness | harness | Runs against your own history, reports aggregates only | DBR-2026-01 |
| Test battery | harness | Export, attestation, algorithm, firmware · Apache-2.0 | DBR-2025-03 |
| Anonymised aggregate | data | 47 models, results without vendor attribution · CC BY 4.0 | DBR-2025-03 |
| Per-model breakdown | data | Restricted · researchers, on request | DBR-2025-03 |
| dart-fp-probe | harness | 42-attribute measurement probe · MIT | DBR-2025-02 |
| Aggregated entropy tables | data | Per-attribute distributions · CC BY 4.0 | DBR-2025-02 |
| Consent flow | spec | Exact text and attribute disclosure shown to participants | DBR-2025-02 |
| dart-spec-bench | harness | Speculative decoding harness · MIT | DBR-2025-01 |
| 1,100 agent turns | corpus | Scrubbed, fixed static set · on request | DBR-2025-01 |
| Per-run timings | data | CSV · 5 pairings × 4 machines × 30 runs | DBR-2025-01 |
Three tiers, and why.
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.
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.
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.
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.
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.