A Hit Chart is not a model's opinion about a product. It is a limitation-by-limitation mapping of your claim onto the public record — each element tied to something you can open and read, or honestly marked as not found. This is what happens between choosing a target and getting that chart.
A claim is not one idea — it is a list of limitations, each of which must be met. So the first thing we do is decompose it: independent claims separated from dependent, preamble from body, and each limitation isolated as its own testable element. Terms are bound across the claim, so "the said receiver" is tracked back to the receiver it refers to.
We then classify each element: is this the inventive core, or is it scaffolding? A claim that recites "a processor" and "a memory" alongside a genuinely novel step should not score those three the same way. Getting this distinction right is what stops a chart from flattering itself later.
Rather than firing one web search per element, we assemble a corpus for the target: its product pages, specification sheets, manuals and technical documentation. That corpus is chunked and indexed so each claim element can retrieve the passages that actually bear on it.
For standards-essential work the corpus is the specification itself — the real 3GPP, ETSI or ITU document, read in full, so the mapping lands on the mandatory provisions rather than a summary of them.
Each element is rated FULL, PARTIAL or NONE against the retrieved evidence — and a rating only survives if a real excerpt supports it. Citations are validated against the passage they claim to come from, so a plausible-sounding reference that does not actually contain the support is rejected rather than shipped.
If nothing documents an element, it is NONE. Not "likely". Not "probably present". The engine is built to refuse rather than to infer, because a chart is only useful if a reader can check every line.
When an element is not fully covered, re-running the same query just returns more of the same marketing page. Instead we cast document-type-targeted nets at the places the missing detail actually lives: the manufacturer's own datasheet, the user manual, the application note, the FCC filing (test reports, internal photos, bundled manuals), the distributor catalogue, the FDA record. PDFs are read deeply, so a spec table buried on page 40 is not truncated away.
There is a language problem here too: patents say "resolving carrier-phase integer ambiguity", datasheets say "RTK fixed, cm-level accuracy". So the element is rewritten into the vocabulary a datasheet would actually use before retrieval — otherwise the very document we just fetched never surfaces.
Manufacturers publish what a product does. They rarely publish how it does it — the signal path, the internal architecture, the stage ordering. That detail is often visible in exactly one place: a block diagram, a schematic, or a teardown photograph.
So for a stubborn element we retrieve those images and a vision model reads them against that specific limitation. It reports only what is visibly shown; if the picture does not show it, the result is discarded rather than guessed. The citation is the image itself. In our testing this recovered implementation-internal elements that no text source anywhere disclosed.
Some limitations are real but undocumented because nobody would bother writing them down: a housing, a power source, a down-conversion stage in a receiver that plainly receives. Treating these as failures makes a chart pessimistic in a way that is just as misleading as inventing evidence.
So a narrow, high-bar judge may mark a non-inventive element as inherently present, with its reasoning stated — and it is labelled that way in the chart, never dressed up as a documented citation. Inventive elements are never eligible. A claim cannot read on scaffolding alone.
Elements are checked against each other for consistency and re-verified before anything is published. The verdict then weights coverage by significance: an independent claim that meets the boilerplate but misses the inventive step has not been shown to read on the target, and the chart says so.
ClaimHit does not declare infringement. It shows you, limitation by limitation, what the public record supports — and where it runs out. The legal conclusion is your counsel's to draw.
The public sweep runs first and builds the best chart the open record allows. Then, if you hold something the public doesn't — a teardown report, an internal specification, a datasheet you were given — you can add it, and we fold its evidence in the same way: read, retrieved, cited.
A claim chart is only worth the parts a sceptical reader can verify. These constraints cost us coverage on purpose — they are what makes the rest of it trustworthy.
The fastest way to judge a claim chart is to read one about a patent you already know inside out.