whitepaperMar 2026
Cited answers, end-to-end: a reference architecture
VeriCite research
How source intake, answer rules, citations, handoffs, and review events fit into one inspectable product flow.
research at vericite
The work behind the citation chip.
VeriCite treats source quality, safe handoffs, and review history as product surfaces. The system is built so an institution can inspect why an answer appeared.
recent research notes
Open by default.
request
benchmarkFeb 2026
Evaluating answers backed by policy
VeriCite evaluation notes
A practical pattern for checking whether an answer cites the right source, handles uncertainty well, and avoids unsupported claims.
taxonomyJan 2026
When the product should say I don't know
VeriCite product and engineering
A product taxonomy for missing sources, access limits, low confidence, conflicting sources, and sensitive records.
postmortemDec 2025
What breaks when source ownership is unclear
VeriCite field notes
A deployment note on why source ownership, escalation routing, and answer review should be designed before broad rollout.
benchmark posture
The metrics we inspect before a release ships.
Citation quality, fallback behavior, and unsupported-answer risk need to be reviewed together. A model that answers everything is not the goal.
| System | Citation-correct | Fallback behavior | Unsupported-answer risk |
|---|---|---|---|
| VeriCite trusted-answer flow | source reviewed | named handoff | controlled by answer rules |
| Generic AI assistant | often a postscript | prompt-dependent | requires manual review |
| Search-only support surface | source visible | no answer synthesis | not applicable |
| Unscoped chatbot | unsupported | rare | high operational risk |
answer-quality review
Every deployment deserves a named review loop.
The highest-value work happens after the first source set is live: replaying misses, naming fallback reasons, and deciding which source owner should close each policy gap.
meet the foundersthe bulletin
One useful answer-quality note, every few weeks.
No generic AI commentary. Just source-quality findings, fallback notes, benchmark refreshes, and postmortems when something does not work.
Email for the bulletin
a 30-minute walkthrough
Bring a source set and inspect the evidence.
A walkthrough can focus on the questions your stakeholders already ask: which sources were used, why they were trusted, and when the product should route to a person.