How we work
Our numbers are only as good as how we gather and check them. This is exactly how a figure becomes a published data point — and what we refuse to do.
The data verification flow
- 1
Collect
arXiv · FDA · clinical registries · official announcements (varies by field)
- 2
Normalize
into one standard shape: value · date · actor · source
- 3
Candidate
the end of automation — the figure plus the exact source sentence
- 4
Human verification
Never automateda person checks the primary source, then approves or rejects
- 5
Record as a time series
append-only — never overwritten, corrections kept in history
- 6
Publish with its source
linked to the primary source, with date and actor
What sources we use
Structured, citable sources first: arXiv, OpenAlex, official company announcements and roadmaps, public registries. API-provided data wherever possible.
Which metrics, and why
Only objective, industry-agreed indicators — qubit counts, gate fidelity, clinical phases, benchmark scores. We never invent a composite “% there” score.
How a number gets published
Automation only ever produces candidates. A person checks the primary source and approves before anything is published — and we record who verified what, when.
What we don’t do
No invented scores. No unverified figures. No cross-field comparisons (quantum vs drugs is meaningless). No copied article text — only the figure and its source link.
Fact vs interpretation
Tracker numbers are neutral and achromatic. Analysis is labelled and coloured — clearly “our read”, never disguised as fact.
The tech desk
Tech stories are drafted by an automated desk. It reads technology feeds, picks stories with a concrete technical development, and has Claude (an AI model by Anthropic) find and read the primary source — the company announcement, paper or filing — and independent coverage. Every fact is recorded with the exact sentence it came from.
Software, not the model, then decides whether a piece can go out on its own: every number must appear in a source we fetched ourselves, every quoted sentence must be found in its page, the piece needs a primary document or two independent outlets, it must be written in its own words rather than rewritten, and it may not contain investment advice. Anything that misses a single check, or that the model itself is unsure about, waits for an editor.
Each article says which path it took: "Verified by Claude against the cited sources" or "Reviewed by an editor". Tech stories never change the tracker's figures, which stay human-verified. If we get something wrong, we correct it on the corrections page.