The full application

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.

DISCLOSURESEARCH PLAN01 · PLAN · YOU APPROVE02 · RETRIEVE · PATENT OFFICESUSEPWOJPCNIN30M+ PATENTS · WORLDWIDESEMANTIC LANE03 · THE GATENOVEL04 · MAP05 · REPORT · ~15 MIN
Fig. E2Novelty Assessment

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

per-zone soil-moisture sensingcore
forecast-driven reschedulingnovelty
independent valve actuationimpl
A01G 25/16G05B 15/02+ 14 synonyms
Approve & runEdit terms

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

DISCLOSUREUSEPWOJPCNIN+KEYWORD LANE · PATENT OFFICESSEMANTIC LANE · EMBEDDINGSRERANKONE LIST

Coverage · 30M+ patents · publications from

USPTOEPOWIPO · PCTJapanChinaKoreaIndiaAustraliaUKGermanyCanada+ more
Google ScholarSemantic ScholarCrossrefarXiv

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

120 CANDIDATESRELEVANCE GATEEVIDENCE QUALITY · HIGH / MEDIUM / LOWACCEPTBORDERLINE · QUOTA 5REJECT · NO BUDGET SPENTDEEPANALYSIS≤ 60 REFS

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

D1D2D3F1F2F3 · NOVELF4NOT DISCLOSED● DISCLOSED · ○ PARTIAL · — ABSENT

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

1.7Key feature analysis matrix
2.1Relevant patent citations
3Applicant landscape
5Claim-positioning analysis
7Limitations & next steps
Novelty risk · MediumCombination risk · flagged

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.

US 10,842 B2TITLE · ABSTRACT · CLAIMS ONLYPRESENT“…VERBATIM, ≤ 18 WORDS”0.92?UNKNOWNTHIN EVIDENCE — NEVER GUESSEDEVIDENCE, OR IT DID NOT HAPPEN
Fig. E2.1 — Verbatim evidence extraction

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.

CapabilityKeyword databasesGeneric AI searchPatentNest
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

DimensionA typical search toolPatentNest
The search termsA black box — you discover what was searched after the results arrive.
An editable plan you approve before anything runs.
RetrievalKeyword 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.
RelevanceA single opaque score sorts everything; every result gets equal, shallow treatment.
An evidence-quality gate routes deep analysis only to references that earn it.
EvidenceAI-written summaries you cannot verify without reading each patent yourself.
Verbatim quotes, character-for-character, for every disclosure claim — checkable in seconds.
UncertaintyForced verdicts — thin evidence quietly becomes a confident answer.
Unknown is a first-class verdict. The pipeline shows its uncertainty instead of hiding it.
The deliverableA 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.

Start a novelty search

Free to start · No credit card required