Process
From technical disclosure to a reviewable first draft
Start with a structured description of the invention. Patendra builds a matter record that connects search results, claim versions, examination issues, score inputs, and draft sections. You can inspect each stage and keep unresolved decisions with the work.
The workflow
Every run leaves a trail you can follow
From a disclosure to a defended draft, each stage writes its evidence to disk, so nothing in the report is a black box.
Overview
The matter workflow at a glance
Ideas flow left to right; evidence flows with them. Nothing reaches the expensive stages without surviving the cheap ones, and nothing reaches the report without a citation trail.
domain description
|
v
[ IDEATION ] gap scan · anomalies · cross-class transplants
| hypothesis generation
v
[ PRIOR-ART USPTO ODP + Espacenet + Google Patents (BigQuery)
RETRIEVAL ] + local ChromaDB --> RRF fusion (+ reranking)
|
v
[ CLAIM-SPACE best-first search over candidate claim sets
SEARCH ] Tier 0 structural validation (free)
| Tier 1 quick anticipation check (cheap)
| Tier 2 adversarial search + exam (full)
v
[ EXAMINATION ] §102 anticipation · §103 obviousness
| §112(a) enablement · §112(b) definiteness
v
[ SCORING ] novelty + non-obviousness + enablement + breadth
| (deterministic -- the LLM never makes the number)
v
[ DRAFTING ] background · detailed description · claims
| Graphviz figures · PDF export
v
[ LEARNING ] human assessment --> weight + confidence
recalibration across runsFIG. 1: workflow overview
Stage by stage
Eight connected stages
Every run executes the same sequence. You watch it happen live in the dashboard over Server-Sent Events: real phase transitions and real logs, not a fake progress bar.
-
Configuration
The run starts from your settings: which model provider to use, how much search budget to spend, and where outputs go. Everything is local-first. Configuration, keys, and results live in your per-user application directory, not on our servers. See the security model.
-
LLM client
Patendra connects to the model you brought: Anthropic Claude or Google Gemini, with your own API key. The workflow is provider-agnostic. The model does the reasoning, but every conclusion it proposes still has to survive the deterministic checks downstream.
-
Component initialization
The retrieval backends come online: USPTO Open Data Portal, Espacenet (EPO OPS), Google Patents Public Data on BigQuery, and the local ChromaDB vector store. A backend that fails to initialize is recorded as unavailable. Search failure is distinguished from empty results, so a dead API can never masquerade as "no prior art found".
-
Ideation
From your domain description, the ideation stage scans for gaps in the existing art, flags anomalies, tries cross-class transplants (techniques from one patent class applied in another), and generates concrete invention hypotheses. Each hypothesis is a candidate worth testing, not a claim worth trusting.
-
Claim refinement
Each hypothesis becomes a claim set, and the claim sets enter a best-first search over claim space: broaden what survives, narrow what gets rejected. Three tiers of validation, detailed below, keep the search honest and the token bill sane.
-
Simulated examination
The winning claim set faces a full simulated Office action: element-by-element §102 anticipation against single, ID-cited references; §103 obviousness combinations with an articulated motivation to combine; §112(a) enablement and written description; §112(b) definiteness. Then the deterministic scorer computes patentability from four signals: novelty, non-obviousness, enablement, breadth. How that works: patentability scoring.
§ 102 § 103 § 112(a) § 112(b)
-
Drafting
The write-up stage turns the examined claims into an application: a background section written as a prosecution argument, a detailed description drafted against a §112(a) enablement checklist, the refined claims, Graphviz patent figures, and a formatted PDF export. Details: how Patendra drafts.
-
Saving outputs & learning
Every artifact is written to your local data directory: hypotheses, search logs, examination memos, score tables, the draft itself. When you later rate the run Strong / Moderate / Weak / Rejected, that assessment recalibrates scoring weights and confidence, feeds the search-strategy optimizer, and lands in experiential memory. No fine-tuning; the improvement lives in your data, on your machine.
Cost design
Verifier-first: cheap checks gate expensive ones
Exploring claim space with an LLM is expensive if every candidate gets the full treatment. Patendra's answer is a three-tier funnel: deterministic checks run first and for free, so tokens are only spent on claims that have earned them.
Structural validation
Pure deterministic checks, no model calls: antecedent basis, claim dependencies, and element consistency. A claim that would draw a formalities objection in the first Office action never costs a token.
Quick anticipation check
A fast novelty screen against the already-retrieved art. Candidates that read directly on a single reference are pruned here, before the expensive tier ever sees them.
Adversarial search + examination
Only the strongest survivors get a fresh adversarial prior-art search aimed at killing the claim, followed by the full simulated examination. The result is a defensible verdict, purchased only where it matters.
The consequence: the best-first search can afford to explore many claim formulations, because most of them are eliminated by grammar and a cheap screen, not by a full examination each. More on the adversarial stage: prior-art search.
Deliverables
What you hold at the end of a run
You get a file set, not a chat transcript. Everything below is saved locally and stays available after the run finishes.
Complete draft, exported as PDF
Background written as a prosecution argument, detailed description checked against a §112(a) enablement checklist, the refined claim set, and Graphviz figures, assembled into a formatted PDF package.
Per-statute findings
Results reported separately for §102, §103, §112(a), and §112(b), never collapsed into a single "ALLOWABLE" badge, with every asserted rejection citing a retrieved, ID-verified reference.
Deterministic patentability score
Computed from novelty, non-obviousness, enablement, and breadth, with Strong / Moderate / Weak confidence labels and the Bayesian evidence chain that explains why the system believes what it believes.
Hypotheses, search logs, memos
The ideation output, every prior-art query and its results, and the refinement history: the paper trail that lets an attorney audit the run instead of taking its word for it.
Every report carries a professional-review notice. Patendra surfaces research and simulated examination issues, but it does not provide legal advice or predict an examiner outcome. Continue with the guidance for attorneys, for inventors and R&D teams, or the documentation.
Start with a real disclosure
Install the desktop app, add a Claude or Gemini API key, and watch the eight stages stream live. Plans and limits are on the pricing page.