How KPsphere research is made
KPsphere research is found, gathered, written and checked by an automated monitoring system built on AI models. This page sets out what that system watches, how often it looks, how the AI writes from the evidence, how the writing is checked, and what to do if you find something wrong.
Written with AI. Built to be checked.
We do not ask you to trust an author. We show you where the evidence came from, how it was checked, and how to tell us when something is wrong.
AI does the research and the writing
Topics are chosen, evidence is gathered, and reports are drafted and checked by AI models (currently Anthropic's Claude) running inside an automated pipeline. No one writes these reports by hand, and they are not edited by a person before they go out.
The evidence is public record
Every report starts from records retrieved from public research, patent and funding databases, and is written to link the specific papers, patents and awards it relies on back to those records.
Accountability is the method
The sources are named below, the checks are described plainly, including what they do not catch, and errors you report are corrected.
Twenty-two public databases and registries
Each search goes to these sources in parallel. Some are asked only in the fields where they hold relevant records, because asking them elsewhere returns unrelated results.
Research papers and preprints
OpenAlex, arXiv, Semantic Scholar, CrossRef, DOAJ and DOE OSTI. For biotechnology and neurotechnology, also PubMed, Europe PMC and bioRxiv.
Patents
Google Patents and USPTO PatentsView. Duplicates between the two are merged.
Research funding
NSF awards, USASpending federal awards, UKRI, OpenAIRE, and Grants.gov open funding calls. For biotechnology and neurotechnology, also NIH.
Regulatory and company records
SEC EDGAR company filings. For biotechnology and neurotechnology, also ClinicalTrials.gov and openFDA.
Data and code
The NIST data repository, for materials, quantum, semiconductors, energy, robotics and cybersecurity; and public GitHub repositories.
Commercial activity
The public deep-research reports also run a web search for company announcements, investment rounds and partnerships. A result is kept only if it comes back with the web address it was found at.
Searches run every day, on a fixed schedule
Scheduled jobs query the sources at set times each day, each for a specific piece of research. Every search records which databases answered, which returned nothing and which failed, so a quiet result can be told apart from a broken one.
- Twice a day
- The daily briefing, on a new topic in a different sector each time. One is emailed each morning; the other is published on the site.
- Every day
- A public deep-research report, rotating through the twelve sectors, with a second report on Tuesdays, Thursdays and Saturdays that is also emailed.
- Every day
- For each sector a paid subscriber follows, a sweep of the last three days of records. Anything already alerted in the previous 60 days is skipped, and an alert goes out only when the AI judges something significant. On many days, nothing is sent.
- Every week
- A subscriber deep-research report for each followed sector, and a check on each company or lab a subscriber tracks by name.
- Every month
- A landscape of each followed sector: what moved, what stalled, what to watch.
- Every 30 minutes
- The queue of custom reports that subscribers have ordered is checked for new questions.
Six steps behind every report
The same steps produce the public reports and the subscriber reports. Each one leaves a record of what was searched, what was kept and what the checks found.
-
Choose
Pick a topic
The day's sector comes from a fixed rotation. An AI model proposes a specific, narrow topic within it and writes the search query, steering away from topics covered recently.
-
Gather
Search the sources
The query goes to the databases above. A report runs two differently worded searches, then merges them and removes duplicates.
-
Screen
Check links and relevance
Records with malformed links, or links that fail a live check, are dropped. A relevance filter sets aside records that do not match the topic's technical terms; if too few match, the full set is kept rather than writing from almost nothing.
-
Draft
Write from the evidence
The AI writes from the retrieved records. It is instructed to link each specific paper, patent or award to its record, use only links present in the evidence, separate what the sources show from its own inference, label background knowledge that did not come from the sources, and say where the evidence is thin.
-
Check
Fact-check against the evidence
A separate panel tests the report's key statements against the same records, and rewrites the ones it cannot support. How it works.
-
Publish
Ship with its sources
The report goes out with its citations. Subscriber reports normally arrive as a PDF with an evidence workbook of the records behind them.
Three reviewers, each told to disagree
A single reviewer asked "is this right?" tends to agree with fluent text. So before a report is published or delivered, a separate AI panel is asked to refute it, judging only against the evidence gathered for that report.
First, the panel pulls out up to ten statements a reader would most likely check: figures, attributions such as who raised, filed or demonstrated something, and claims of status such as "first" or "approved". Three reviewers then test every one, each through a single lens:
Arithmetic
Does the figure appear in the evidence, with the same magnitude, unit, currency and period?
Attribution
Does the evidence say this organisation did this, and not a university, partner or competitor?
Overstatement
Is the claim stronger than the evidence allows: a lab result called deployment, a filing called an approval?
Each reviewer is told to treat a claim as wrong until the gathered records show otherwise, and not to fall back on its own knowledge. A statement that two of the three reject is flagged. Flagged sentences are rewritten to what the evidence supports (softened, re-attributed or cut back), and only those sentences change.
The panel runs on the daily briefings, the public deep-research reports, and every subscriber report, including the first report after signing up and custom reports.
What it does not do. It checks the statements most likely to matter, not every sentence, and on a long report it draws them from the opening part of the text. It does not hold a report back: if a flagged sentence cannot be rewritten, or the panel cannot run, the report still goes out and KPsphere is alerted. Videos, podcasts, weekly digests, monthly landscapes and signal alerts are made from the same sources but do not pass through this panel.
Follow a key finding back to its record
The point of citing is that you can check. The writer is given the links of the records the system retrieved and is instructed to use no others.
Inline citations
The writer links specific papers, patents and awards where it mentions them. The source list on a public report is built from the records the report actually cites, and subscriber PDFs end with a numbered reference list built from their inline citations.
The evidence workbook
Subscriber reports and custom reports are normally sent with an Excel workbook of the records gathered for them, including ones the report did not cite, so you can see what the writer had in front of it. Papers, patents, and funding and commercial records sit on separate sortable, filterable sheets, each row with its source database, date and a link to the original.
Want to see one before subscribing? Read a sample report→
Know the limits before you rely on it
The system is built to be checked because it can be wrong. Read it as a well-sourced starting point, not a final word.
- Not advice. Nothing here is investment, legal, financial or procurement advice, or a recommendation to buy, fund or acquire anything.
- AI can make mistakes. The fact-check reduces errors; it does not eliminate them. For any decision, open the linked source and read it yourself.
- Coverage is not complete. A report reflects what the databases returned on the day. The funding sources are US, UK and EU records, so coverage leans toward them, and patent applications are generally not visible until they publish, typically about eighteen months after filing. The absence of a record is not evidence that none exists.
- Background is labelled. Where a report adds context that did not come from the retrieved sources, the writer is instructed to mark it as background. Treat those passages as unverified.
- Figures are as reported. Funding amounts are reproduced as the source states them and are not independently audited.
Tell us, and we will check it
Email info@kpsphere.com with a link to the report or briefing, the statement you think is wrong, and the source that shows it. You can also use the contact form. If the report is wrong, we correct it.