Patti Insight is a mechanism-gap technology report. It reads prior art against the mechanism behind a technical problem: what each document discloses, which part it leaves unsolved, and what remains open.
A Preview, not a freedom-to-operate opinion. Patti Insight is a Preview study. It does not guarantee novelty, non-infringement or freedom to operate, and it is not legal advice.
Helps engineers analyze large amounts of technology, patent and market information related to their own work, cutting the time and effort of competitive technology analysis.
Designed against AI hallucination
Built to reduce false numbers and facts produced by AI.
Possible routes
Technology-expansion and industry-expansion research that gives readers initial possible routes for market and technology development.
Reports come in three editions: Public, Advanced and Professional. Some sections below appear only in the Professional edition, as noted.
1. Designed against hallucination
Each rule below is checked either by an automated gate on the finished report or by a separate checker.
Rule
What it means in a report
Reading depth is labeled
Every source is labeled full text, abstract, search snippet or secondary. Snippets never carry an argument.
Numbers need a source
Every factual sentence has a source code that resolves to the report's own source list: sentence by sentence in the Professional edition, collected at the end of each paragraph in the Public and Advanced editions. A sentence without a source is removed.
Counts are recounted
Numbers reported by research agents are recounted from the underlying list before they are used.
Inferences are labeled
A sentence that combines sources into a conclusion must be marked as an inference or an estimate, and every source sentence behind it must be found word for word in the stored source files. An unmarked inference fails the build check.
Verbatim spot checks
Key claims are traced back to the original passage and marked verified, unverified or corrected. Qualifiers such as “only”, “at least” or “all” must match the source sentence word for word.
Independent physics re-check
Any passage written or rewritten after fact-checking gets a separate physics re-check with a before-and-after comparison: did the edit change the object, temperature, range, condition or quantity?
Patent numbers are verified
Every patent number in the final report is checked against bibliographic records. Numbers that cannot be verified are left out.
“Not searched” is not “not found”
A source that was blocked or unreachable is recorded as not searched. “No public data found” may be written only when a query log exists: which sources, which query strings, the date and the result.
White space needs full text
A white-space conclusion requires the closest prior art to have been read in full. If it was not, the report says “full text not read; no white-space judgement” and lists the numbers. The report never says “no prior art”; it says “not seen within the documents read” and states that scope.
Each report discloses how many documents were read, how many were read in full, how many readers took part, how many checks were run and how many corrections were made.
2. How the patent line is counted
Professional editions include an integrated reading of patents, markets and players. The patent counts behind it are measured, and their error is stated.
One family, one count. Each patent family (DOCDB simple family) is counted once, in its earliest priority year. The earliest priority may be a U.S. provisional application, which can be up to about a year earlier than the formal filing; the methods note says so.
Precision. For every line, a stratified sample of at least 20 families is judged against a written inclusion rule. A weighted precision below 80% is labeled low confidence.
Queries tuned on known families, recall measured on others (V1 absolute recall). Each line needs at least 10 known relevant families from independent sources, such as verified patents in the report appendix, their citations, a third-party report or the client. They may not be taken from the line's own search hits, and relevance is judged without seeing whether the family was hit. Families used to tune the query form a tuning set; final recall is measured on a separate validation set.
Query words come from the patents, not from us. Candidate query terms are taken only from words that actually appear in the titles and abstracts of the known relevant families and the families one citation step away. A model may judge whether a candidate is in scope, but may not add words of its own.
What recall allows. Recall of 80% or more passes. From 60% to 80%, the chart and text say “insufficient recall”. Below 60%, or with fewer than 10 validation families, the line gets no chart; its direction is described in words only.
Era-by-era precision calibration. A line whose V1 recall passes but whose precision is low can still be charted only if precision was calibrated era by era: at least 30 families per era, Wilson confidence intervals, a consistency test between eras, and a calibrated series shown with its interval. The chart is then labeled an approximate analysis. This rule is currently checked by hand, not yet by an automated gate.
“Where the counts are inaccurate.” An approximate chart is preceded by a short section that states the precision of each era, the recall, mixing-in of neighboring technologies, misses, side effects of the calibration, and which direction the error leans.
Second judge and spot checks. Ten families per line are re-judged by a second judge (a different model or prompt); agreement below 80% is flagged. Ten families per report are checked against their public patent pages for earliest priority date and members; all ten must match.
Relative and sensitivity views. Each line is also measured as a share of its main patent classes, and recounted with only families filed at two or more offices and without utility models. If these views disagree with the main count, the text says so.
Our thresholds are labeled as ours. Thresholds such as ±5% per year, 80% precision, 80%/60% recall and 80% agreement are marked “custom, no literature source”. Reference values published by WIPO are shown alongside, not substituted.
Patent activity is not market activity. A patent curve describes patent filing activity only, never market growth or decline. Market figures appear under their own heading and on their own chart axis, each from a named source. Patent counts are never used to estimate a market.
3. Possible routes, with signposts
Readings of a trend differ, and only future facts settle them. So the Professional edition does not give one answer. From where each line stands today, it lays out up to three possible routes. It gives no probabilities and does not pick a side.
Where routes come from. Routes are derived from shifts between four combinations of family counts and applicant counts: staying in the current combination, or moving to another. Routes do not overlap.
Every route has signposts. At least two per route. Each signpost is an indicator, a direction, a threshold and an observation window. Thresholds are marked as set by the report.
Only complete years count as evidence. Signposts observed so far use years with complete data. The most recent 18 months may be shown but are labeled “not used as evidence”.
What it means for you. Each route has options for two kinds of reader: those who can shape the industry (large companies, standards participants) and those who cannot (smaller companies, new entrants). Every option points back to a chapter of the report. Wording is limited to “could consider”, “could prepare”, “could first check”; no legal conclusions.
Numbers come from the data file. Figures in the routes are filled in by program from the same data as the report tables, not typed by hand, and the finished text is compared with the rendering.
4. China, Japan, Korea and Taiwan coverage
A white space written from U.S. patents alone can be contradicted by filings elsewhere. Before publication, every report checks patents in China, Japan, Korea and Taiwan. The count for each, and the query that produced it, are recorded. Zero is a valid result (searched, none found), kept distinct from not searched. Each edition states this coverage in its text. A report that misses either requirement fails its check and is not delivered.
5. Verification IDs
Every edition carries a Verification ID in the form PI-XXXX-XXXX, with a check character, printed on the cover and on every page. One ID belongs to exactly one content: when a report is revised, it gets a new ID. You can look up an ID at patti.empryon.ai/verify.
6. Limits
Publication delay. Patent data for the most recent 18 months is not yet complete. A dip at the end of a curve does not mean a decline; the affected years are shaded and stated.
Market numbers are shown alongside, not derived. Revenue, shipments, installed base and price come only from named sources. When none is found, the report says “No public market data found” and logs the search.
Counts across lines. Comparing filing counts between different lines has no market meaning.
Preview positioning. Patti Insight reports are technology and market analysis. They are not legal opinions on infringement, validity or freedom to operate, not investment advice, and not a substitute for a licensed patent attorney.