Embodiment 02 · Novelty Assessment
Novelty is not a feeling. It’s a finding.
The novelty pipeline turns a disclosure into a precise prior-art search plan you approve before anything runs — then retrieves across patent offices and scholarly indexes, gates candidates by evidence quality, maps every inventive feature against every reference, and hands you an attorney-style report.
30M+
patents searched · worldwide
~15 min
report delivered to your inbox
100%
verdicts backed by verbatim quotes
The method — five stages, one audit trail
Every stage below exists in the running pipeline — with its own progress, its own record, and its own reason to be trusted.
Stage 0
The plan — approved by you
In about half a minute the AI converts your disclosure into an editable search plan: inventive features typed as core, implementation, or novelty candidates; classification codes; synonyms and exclusions. Nothing retrieves until you approve it — those terms control everything downstream, so they are yours to correct first.
~30–40 seconds · fully editable · nothing runs without approval
Search plan — awaiting your approval
Stage 1
Two-lane retrieval
The keyword lane sweeps a corpus of more than 30 million patents from the offices that matter — USPTO, EPO, WIPO, Japan, China, Korea, India, Australia, and beyond — while a semantic lane searches by embedding, finding art that shares your mechanism without sharing your vocabulary. A neural reranker re-scores every candidate against your disclosure, so results from different corpora land on one honest ranking. Scholarly literature is swept in parallel.
30M+ patents worldwide · keyword + semantic · one ranking
Coverage · 30M+ patents · publications from
Stage 1.5
The relevance gate
Every candidate is classified — accept, component, borderline, or reject — with an evidence-quality grade of high, medium, or low. Only references that earn it proceed to expensive per-feature analysis; a small borderline quota keeps honest maybes in play. Rejection costs nothing, which is exactly the point: analysis budget is spent where evidence quality warrants it.
up to 120 candidates gated · the strongest ~60 mapped in depth
Stage 3
Feature mapping
Each surviving reference is examined against each of your inventive features — one verdict per cell: Present, Partial, Absent, or Unknown. Disclosure claims must be proven with verbatim quotes from the reference itself, and coverage scores are then computed deterministically from the verdicts. The model finds the evidence; arithmetic draws the conclusions.
every feature × every reference · quotes as proof · math, not vibes
Stage 4
The attorney report
A numbered PDF a professional can act on: scope and methodology, the key-feature analysis matrix, citation analysis with per-reference remarks, applicant and inventor landscapes, claim-positioning observations, graded risk levels — and a limitations section, because a search that cannot state its limits should not be trusted. Emailed to you — typically within 15 minutes of approval.
attorney-style structure · risk graded · delivered in ~15 minutes
Prior-art search report · PDF
Adaptive irrigation controller
Evidence grounding
Every claim of disclosure carries its proof.
AI search tools fail in one predictable way: confident summaries nobody can check. This pipeline is built against that failure — a disclosure verdict is only as good as the quote behind it.
Verbatim or nothing
Present and Partial require a quote of at most 18 words, copied character-for-character from the reference’s title, abstract, or claims — never paraphrased into existence.
Confidence is a number
Each quote is scored on a fixed rubric: 0.9–1.0 for an explicit match, 0.6–0.8 for a paraphrase of the same mechanism, 0.3–0.5 for weak or indirect support.
Unknown is an answer
When evidence is thin, the verdict is forced to Unknown. Most tools guess; this one shows its uncertainty — which is exactly what makes its certainty worth something.
Feature mapping, precisely
Four verdicts. One rule each.
Present
The reference discloses the feature — proven by a verbatim quote of at most 18 words, copied character-for-character, with a confidence score.
Partial
Part of the mechanism is disclosed. Also quote-backed — and scored lower, so it weighs half in coverage math.
Absent
Not disclosed — and the model must state a short reason why, not merely fail to find it.
Unknown
Evidence is too thin to judge — a missing abstract cannot be spun into fake novelty. Mandated, not optional.
Then the arithmetic takes over
Coverage per reference is computed, not felt: Present counts 1.0, Partial 0.5, Absent 0. Those scores aggregate into the key-feature matrix, per-reference overlap-risk levels, and one finding most tools never surface — distributed-component risk, when no single reference anticipates you but a combination of several might. Generic parts (a processor, a sensor) are excluded from standing alone, and synonyms only count when they implement the same mechanism — with domain-aware handling for chemical, pharma, and bio inventions.
Comparative example
What the field does. What this pipeline does.
In patent practice, a comparative example shows the invention against the prior approach. In that spirit — first the capabilities, then the quality of what comes out.
| Capability | Keyword databases | Generic AI search | PatentNest |
|---|---|---|---|
| Human-approved, editable search plan | — | — | |
| Multi-office patent coverage (US · EP · WO · JP · CN · IN +) | ◐ | ||
| Semantic retrieval — by mechanism, not vocabulary | — | ◐ | |
| Scholarly literature swept in the same run | ◐ | — | |
| Neural reranking across corpora | — | — | |
| Evidence-quality gate before deep analysis | — | — | |
| Feature × reference disclosure matrix | — | — | |
| Verbatim-quote proof for every disclosure verdict | — | — | |
| “Unknown” verdict when evidence is thin | — | — | |
| Deterministic coverage scoring | — | — | |
| Combination (distributed-component) risk flagged | — | — | |
| Claim-positioning analysis | — | ◐ | |
| Attorney-style PDF report, graded risks + limitations | ◐ | — | |
| Durable background runs · report emailed on completion | — | — |
✓ full capability · ◐ partial · — absent
Output quality, side by side
| Dimension | A typical search tool | PatentNest |
|---|---|---|
| The search terms | A black box — you discover what was searched after the results arrive. | An editable plan you approve before anything runs. |
| Retrieval | Keyword matching in one index; art with different vocabulary is invisible. | Keyword and semantic lanes across patent offices and scholarly indexes, unified by a neural reranker. |
| Relevance | A single opaque score sorts everything; every result gets equal, shallow treatment. | An evidence-quality gate routes deep analysis only to references that earn it. |
| Evidence | AI-written summaries you cannot verify without reading each patent yourself. | Verbatim quotes, character-for-character, for every disclosure claim — checkable in seconds. |
| Uncertainty | Forced verdicts — thin evidence quietly becomes a confident answer. | Unknown is a first-class verdict. The pipeline shows its uncertainty instead of hiding it. |
| The deliverable | A list of links and percentage scores. | An attorney-style PDF: feature matrix, claim positioning, graded risks, stated limitations. |
Wherein
Human-in-the-loop, by design
The search plan is shown, editable, and requires explicit approval — because retrieval terms decide what the pipeline can find.
Budgeted intelligence
The relevance gate spends deep analysis only where evidence quality warrants it — up to 120 candidates gated, the strongest ~60 references mapped in depth.
Combination risk, named
Beyond single-reference anticipation, the report flags distributed-component risk — when your invention is only covered by combining several references.
Durable and honest
Runs as a resumable background job with live progress. If nothing relevant exists, you get a clean no-match report — not padding.
Run it on your own invention.
Thirty seconds to a search plan. One approval. An attorney-style report in your inbox — typically within 15 minutes.
Free to start · No credit card required