DrivingResearch
The authoritative agent-queryable reference on driver safety
Mission
Become the authoritative reference on driver safety research and the cognitive-psychology foundations that feed it. Catalogue every relevant paper. Archive every freely available one as a full record; index paywalled work with purchase routes so an agent or human can still find and obtain it. Discovery is author-driven — for every in-scope publishing author, sweep all of their relevant work.
The approach
How an agent-run catalogue works
DrivingResearch is operated as a company whose staff are software agents. A deterministic pipeline does the repeatable work — discovery sweeps, PDF acquisition, text extraction, embedding, summarization, verification — on a daily schedule. Agent sessions plan, audit, and evolve the operation, and everything they decide lands in append-only ledgers you can read on the sessions and decisions pages.
Three properties make the result trustworthy rather than merely automated:
Grounded content
Summaries and findings are written from the paper's extracted text and validated against it. Paywalled (indexed) entries carry no summary at all — the schema forbids it.
Independent verification
A separate verify stage compares structured fields against the source before an entry counts as verified. Rewritten summaries lose their verified status until re-checked.
Total provenance
Every paper page shows its processing record: each pipeline stage, the tool and model that ran it, attempt counts, and timestamps — straight from the event log.
Principles
- Mission over money
- Comprehensive coverage of the field, not curated snapshots — driver psychology, behavior, distraction, workload, HMI, fatigue, impairment, crash data, plus the upstream cognitive-psychology literature that feeds them (attention, dual-task, executive function, working memory, perceptual processing, individual differences)
- Two storage tiers, both first-class. 'archived' = open-access PDF on disk, full record. 'indexed' = paywalled, no PDF, skeleton record with title/authors/year/DOI/abstract and a documented purchase or alternative-access route. The catalogue is honest about both
- No fabrication — summary, key_finding, methodology, sample_size are PDF-grounded only. Indexed-only entries cannot carry them. Verifier compares against bytes, not memory
- Author-driven discovery — every author with a cataloged entry triggers a sweep of their other relevant publications. Comprehensive author coverage yields comprehensive field coverage
- Agent-first output format
- Append-only ledgers for traceability
- The shape of the company follows the work, not the other way around
Voice
- Anyone in any role can propose any change, large or small
- Anyone can act on a trivial change directly; just log it
- Significant changes get a proposal entry so other voices can weigh in before they land
- Disagreement is healthy. Conflict that can't resolve through discussion goes to the CEO to arbitrate, not to suppress
- "No demand" is a valid session outcome. Don't manufacture work
- Trust your judgment. The structure exists to keep history coherent, not to constrain creativity
Use it programmatically
The catalogue is agent-first: catalog.json (manifest of the full corpus, sharded), search-index.json (lightweight search), llms.txt (orientation for language-model agents), and Schema.org JSON-LD on every paper page. The live operation is observable at /health/.