Live Demo · Autonomous Prosecution Chain

Autonomous Chain: OA → Amendments → Prior-Art Re-Search

The agent responds to the office action, extracts its own amended claims, and immediately runs a prior-art search on them — no human prompt between Step 1 and Steps 2/3. This is what autonomous prosecution execution looks like.

Demo run on 2026-06-19. Results cached — stable, instant.
⚡
No human prompt between Step 1 and Steps 2/3. The office action response agent generates amended claims. Those amended claims are passed directly as disclosure text to the prior-art search agent. The prior-art report reflects the actual risk landscape for the amended claims — not the original ones. Prompt-and-review tools stop at Step 1. ClaimForge continues.
Prior Art Search OA Response — USPTO §103 OA Response — EPO Art. 56 ⚡ Chain — USPTO §103 ⚡ Chain — EPO Art. 56
Step 1 — Input Office Action + Current Claims Human: paste OA text
Application
16/134,118
Jurisdiction
USPTO
Rejection ground
103
Claims rejected
7
Parsed Rejections
Ground Claims Examiner Reasoning
103
135
Brown discloses the complete document processing pipeline (receive text, parse with neural network, generate structured claim elements, output). Patel discloses transformer-based models for legal text tasks. It would have been obvious to substitute Brown's neural network with Patel's transformer architecture because Patel teaches superior performance on legal text tasks and the substitution yields predictable results under KSR.
103
2467
Claims 2, 4, 6, and 7 recite specific implementation details each individually disclosed by Chen in combination with the primary Brown/Patel combination. Each limitation performs its known function, producing predictable results.
⚡ Agent passes output forward — no human prompt
Step 2 — Respond Claim Amendments + Prosecution Arguments Fully autonomous
Proposed Amendments (Red-line)
Amendment Option 1 — Claims 1, 3, 5
Add novel technical limitation — passage-level attention weighting

Amend claim 1 to add 'applying a cross-passage attention weighting mechanism that assigns higher weight to passages containing technical claim language keywords, prior to generating the set of patent claim elements.' This limitation is not disclosed by Brown (which uses uniform passage weighting) or Patel (which does not teach cross-passage attention for claim-element generation). Spec support in para. [0047].

✕ Original
Claim 1: A computer-implemented method comprising: receiving, by a processor, a natural language text document representing a patent disclosure; parsing, by the processor, the text document using a transformer-based neural network to identify one or more inventive concepts; generating, by the processor, a set of patent claim elements corresponding to the identified inventive concepts; and outputting the set of patent claim elements in a structured format suitable for patent claim drafting.
✓ Proposed Amendment
Spec support: Paragraph [0047] of the specification describes the attention weighting mechanism: 'the parser module applies a differential attention mask M_cross that upweights token embeddings in passages containing statutory claim language terms identified from a curated keyword lexicon.'
Trade-off: Narrower independent claim scope, but clearly distinguishes from Brown's uniform-weighting approach. Dependent claims 2–7 remain unchanged and their scope is preserved.
Amendment Option 1 — Claims 2, 4, 6, 7
Add domain-specific fine-tuning data source limitation to claim 2

Amend claim 2 to specify that the pre-trained language model is fine-tuned using a corpus comprising prosecution history files (including office actions and applicant responses), which Chen does not disclose — Chen only discloses issued patents and published applications as the fine-tuning source.

✕ Original
Claim 2: The method of claim 1, wherein the transformer-based neural network is a fine-tuned version of a pre-trained language model trained on a corpus of issued patents and published patent applications.
✓ Proposed Amendment
Spec support: Paragraph [0031] of the specification: 'the training corpus includes prosecution history files retrieved from the USPTO PAIR system, including all office action communications and applicant response briefs filed therein.'
Trade-off: Narrows claim 2 scope to specifically include prosecution history data. Chen is limited to issued patents and published applications, so this amendment creates a clear distinction.
Prosecution Arguments
103 Response to §103 Rejection of Claims 1, 3, 5 (Brown + Patel) ▼
Applicant respectfully traverses the rejection of claims 1, 3, and 5 under 35 U.S.C. §103 as purportedly unpatentable over Brown et al. (US 10,678,833 B2) in view of Patel et al. (US 2021/0232784 A1). Neither Brown nor Patel — alone or in combination — discloses or suggests the cross-passage attention weighting mechanism now recited in amended claim 1. Brown applies uniform processing to document passages (col. 5, lines 8-12) with no teaching of differentially weighting passages based on claim language keywords. Patel discloses contextual embeddings for general legal text classification (para. [0045]-[0053]) but does not teach claim-language-aware attention weighting for patent claim element generation. The Examiner's combination further relies on impermissible hindsight. Transformer models were one of numerous available architectures at the time of the invention, and the Examiner provides no principled reason for specifically selecting transformers over LSTMs, CNNs, or graph networks other than performance improvement — which is insufficient under Belden Inc. v. Berk-Tek LLC, 805 F.3d 1064 (Fed. Cir. 2015). Applicant respectfully requests withdrawal of the rejection of claims 1, 3, and 5.
103 Response to §103 Rejection of Claims 2, 4, 6, 7 (Brown + Patel + Chen) ▼
Applicant respectfully traverses the rejection of claims 2, 4, and 6, 7 under 35 U.S.C. §103. Regarding claim 2 as amended: Chen's fine-tuning corpus (para. [0019]-[0027]) covers only issued patents and published applications. Claim 2 as amended now requires prosecution history files — office actions and responses — which are absent from Chen and present a fundamentally different training signal capturing the adversarial prosecution dialogue. Regarding claims 4 and 7: Chen discloses k-means with k determined by 'document characteristics' generally (para. [0037]). Claim 7 specifically requires k determined by document length — a distinct and specific implementation. The Examiner has not explained why a skilled artisan would select document length specifically over the many other document characteristics Chen contemplates. Regarding claim 6: Chen predicts 'relevance to the legal issue' and 'likelihood of claim allowance' (para. [0050]) — neither of which is coextensive with commercial value. A model predicting prosecution likelihood provides no motivation toward a model predicting commercial value. Applicant respectfully requests withdrawal of the rejection of claims 2, 4, 6, and 7.
⚡ Amended claims → prior-art input — no human prompt
Step 3 — Re-Search Prior-Art Re-Search on Amended Claims Fully autonomous
Agent-generated disclosure: The OA response agent extracted amended claim text from Step 2. That text is used verbatim as the invention disclosure for this search — no human re-phrasing.
Amended patent claims after USPTO office action response: Claim 1 (proposed amendment): A computer-implemented method comprising: receiving, by a processor, a natural language text document representing a patent disclosure; parsing, by the processor, the text document using a transformer-based neural network to identify one or more inventive concepts, where…
Candidates: 7
Material: 5
Top References
#1
US 11,429,781 B2
Attention-Based Document Parsing for Structured Information Extraction
84
material
▼
Lexion AI, Inc. · 2022-08-30 · Source: espacenet
A system and method for parsing legal documents using transformer-based attention mechanisms with token-type-specific weighting. The method assigns differential attention weights to tokens based on their legal function category, enabling more accurate extraction of structured elements from heterogeneous legal corpora including contracts and patent specifications.
Relevance: Discloses the core cross-passage attention weighting mechanism with token-type classification, directly relevant to the amended Claim 1 limitation on attention keyed to claim-language keywords.
#2
US 2023/0046821 A1
Fine-Tuning Language Models on Legal Prosecution Records for Claim Drafting Assistance
91
material
▼
LegalMind Technologies LLC · 2023-02-16 · Source: google_patents
Describes fine-tuning a large language model using a combined corpus of issued patents, published applications, and prosecution history files retrieved from USPTO PAIR. The prosecution history corpus includes office actions, applicant responses, and examination reports, providing a richer adversarial training signal for claim drafting applications.
Relevance: Directly covers amended Claim 2 — prosecution history fine-tuning corpus explicitly including office actions and applicant responses. High risk of anticipation for Claim 2.
#3
US 10,956,676 B2
Cross-Document Attention for Legal Text Classification
77
material
▼
Thomson Reuters Enterprise Centre GmbH · 2021-03-23 · Source: espacenet
A neural network architecture incorporating cross-document attention layers that allow tokens in one document segment to attend to tokens in other segments based on legal function tags. Applied to patent claim classification and prior art identification tasks.
Relevance: Cross-document (cross-passage) attention mechanism for legal text with functional weighting — relevant to the attention weighting limitation in amended Claim 1.
#4
WO 2022/197834 A1
Patent Corpus Construction and Model Training for Automated Prosecution Assistance
73
material
▼
PatentSight GmbH · 2022-09-22 · Source: espacenet
Methods for constructing patent prosecution training corpora and fine-tuning language models thereon. Describes inclusion of examination history data including office actions from multiple patent offices (USPTO, EPO) to improve model performance on prosecution-specific tasks.
Relevance: Prosecution corpus including office actions from USPTO and EPO; relevant to amended Claim 2 fine-tuning corpus scope.
#5
US 2022/0318548 A1
Transformer-Based Claim Element Extraction with Keyword-Weighted Passage Encoder
88
material
▼
Anaqua, Inc. · 2022-10-06 · Source: google_patents
A patent claim element extraction system using a transformer encoder with a passage-level keyword weighting layer. A curated lexicon of claim-language keywords (e.g., 'comprising', 'wherein', 'configured to') is used to compute per-passage weights that are multiplied into the attention scores before claim element generation.
Relevance: Very close to amended Claim 1 — explicit claim-language keyword lexicon used to compute passage weights applied to attention scores, nearly identical to the cross-passage attention weighting limitation.
#6
US 11,023,660 B1
Neural Legal Document Summarization with Section-Type Weighting
55
marginal
▼
Ironclad, Inc. · 2021-06-01 · Source: espacenet
Legal document summarization using a transformer model with learned section-type weights. Different document sections (definitions, obligations, conditions) receive different attention multipliers during encoding to improve summary quality.
Relevance: Section-type weighting in a transformer for legal documents — analogous principle but applied to summarization (contracts), not patent claim element extraction.
#7
US 2021/0209304 A1
Document Feature Extraction Using Pre-Trained BERT Models
38
marginal
▼
Google LLC · 2021-07-08 · Source: google_patents
General-purpose document feature extraction using fine-tuned BERT models on domain-specific corpora. Demonstrates improved performance on information extraction tasks when fine-tuned on in-domain corpora compared to general-domain pre-training.
Relevance: General BERT fine-tuning — does not teach the specific cross-passage attention weighting or prosecution-history corpus limitations of the amended claims.
Novelty Report on Amended Claims
Chain summary: Autonomous chain complete (USPTO). OA agent generated 2 rejection responses. Prior-art re-search on amended claims found 7 candidates (5 material). High-risk references identified: US 2023/0046821 A1 (prosecution history corpus) and US 2022/0318548 A1 (cross-passage attention).

Run this on your own cases

Same autonomous chain on your USPTO and EPO files. Paste your OA, run the chain, get amendments + re-search in one pass — from $99/mo, 14-day trial.

Preise ansehen → Agent direkt testen

Direktkontakt: claimforge@polsia.app