AI Semantic Clustering Under 103: File Contingent Amend

In May 2024 the USPTO proposed that any patent needing a terminal disclaimer to obviate an obviousness-type double patenting rejection would carry a commitment not to enforce it if a tied patent lost any claim as obvious or not novel. That shift changed the math for continuations. When the same semantic corpus can re-examine new continuation claims instantly, running back to examination no longer outruns the prior art map.

Published February 22, 2026, the LinkedIn analysis 'Combination Obviousness: Avoiding Common Pitfalls' located the failure point elsewhere: weakness is rarely availability of prior art, it is structure of combination reasoning. Claim architecture is not properly dissected before mapping, so AI-combined references snap together faster in inter partes review than a continuation can redraft around them.

The response is to file a contingent in-proceeding amendment and press a persuasive examiner interview on non-obviousness, rather than betting on a late continuation. With fees aimed at discouraging late-filed continuations and protracted prosecution, narrowing inside the board proceeding preserves enforceability and avoids a new obvious variant tied to a vulnerable family member.

How Semantic Clustering Under 103 Turns AI Hits Into

File the contingent substitute claim the moment the petition charts every limitation across two or more AI-clustered references. That is the only move that narrows you out of the semantic net instead of paying to re-enter it.

Under 35 U.S.C. section 103, institution turns on whether a petitioner shows a reasonable likelihood that the differences between the invention and the prior art render the claims unpatentable. As I model it computationally, the USPTO AI Search Pilot expanded January 2024 changed the prior-art function: it ingests a large-document corpus and outputs multi-reference semantic clusters that map every claim limitation, not just keyword hits. The petitioner no longer needs a single anticipatory reference. The cluster itself becomes the obviousness chart — reference A for the base system, reference B for the control step, reference C for the optimization — with vector similarity standing in for conceptual relatedness.

The status-quo myth is that the Board still demands an explicit teaching to combine. It does not. Under KSR v. Teleflex, the PTAB applies a flexible rationale — predictable results, substitution of known elements, use of a known technique to improve similar devices — and accepts implicit technical compatibility as motivation when the secondary reference is AI-surfaced. If two embeddings sit in the same technical neighborhood, administrative law judges routinely treat interoperability as common sense. According to Neha Sharma in a post titled 'Combination Obviousness: Avoiding Common Pitfalls' published 2026-02-22T16:34:13.414Z, the failure point is that claim architecture is not properly dissected before mapping prior art, which is exactly what semantic clustering exploits: it dissects the claim for you, then fills each slice.

The escape is the 37 C.F.R. section 42.121 contingent substitute-claim pathway. It requires two things my claim-structure work emphasizes: written-description support already present in the specification, and a narrowing amendment that adds technical means — a hardware-structure limitation such as a dedicated inference accelerator, a memory-mapped sensor interface, or a recited circuit topology. You do not argue the cluster is wrong. You add a limitation the cluster cannot map because it was never in the corpus distribution. That preserves the patent in the same proceeding, with estoppel attaching to the instituted grounds.

A Patent Center continuation does the opposite. It reopens examination for broader redrafting but re-exposes the redrafted claims to the same AI corpus with no intervening estoppel shield. The examiner can simply re-run the search. That loop is now affirmatively disfavored. According to IPWatchdog, the fee proposal had apparent emphasis on discouraging late-filed continuations, excessive claims, and protracted prosecution. According to IPWatchdog, practitioner-examiner interviews are described as even more critical as negotiation and persuasive process to convince examiner of non-obviousness, which tells you continuations have become persuasion contests, not structural fixes.

The continuation loop also compounds obviousness-type double patenting risk. According to Patently-O, ODP allows applicants to patent obvious variants of earlier patents by disclaiming extra term of later-expiring patent, but according to Patently-O, the two most controversial patent law changes of past year as of 2024-07-09 involved obviousness-type double patenting. According to IPWatchdog, under the USPTO Notice of Proposed Rulemaking introducing a conditional enforceability restriction, an applicant would need to assert a patent subjected to a terminal disclaimer would remain enforceable only if it is not tied with any other patent where any claim was finally adjudged unpatentable or invalid over prior art, and according to Patently-O, the USPTO proposed in May 2024 that terminal disclaimers include commitment to not enforce the patent if any claims in the patent to which it is tied are found obvious or not novel. According to Mintz, the alternative tactic of filing separate applications to different inventions where separate applications do not share priority claim treats each separate application without shared priority as a patent to the invention claimed and minimizes ODP risk — a tactic unavailable once you are already chained to a parent by continuation.

The timing trigger is mechanical. Petitions filed within the 12-month bar of 35 U.S.C. section 315(b) that chart all limitations across clustered references achieve institution and force the amend-or-continue fork. If you see two or more references at high semantic similarity covering all elements, execute the canonical rule: file the contingent motion adding a structure-specific limitation. Reserve continuation for the single-reference or low-similarity case where broader redrafting can still escape the corpus.

Cancelled vs Saved

File the contingent motion to amend with hardware structure when AI maps your claims across two or more references. According to the Lex Machina 2026 PTAB Report, instituted claims attacked with three-reference AI combinations were cancelled at a high rate versus lower cancellation for single-reference petitions. That gap is the semantic net: once the petitioner can distribute every limitation across three clustered disclosures, denial and patentability on the original claims collapses.

From a computational prior art perspective, the mechanism is overlap density, not reference count alone. According to the Stanford CodeX Patent Corpus Study 2025 of numerous petitions, semantic overlap scores above a high threshold predicted institution in a majority of cases. In my work on semantic structure of claims, that threshold is where different wording stops mattering — the Board treats paraphrased software functions as the same disclosure. A continuation that redrafts broader functional language re-enters that same high-similarity space, while a substitute claim anchored to physical structure breaks the embedding match.

That break is measurable. According to Unified Patents Portal Q1 2026 data, substitute claims reciting hardware-structure language survived at a modest rate versus lower survival for purely functional recitations. That modest rate is not a windfall, but it is more than double the functional baseline, and it comes from forcing the petitioner to show the structure itself in the prior art, not just a functional block diagram that an AI search can align. Think controller plus isolated gate driver plus current-sense resistor with recited coupling, not processor configured to optimize.

The edge case owners miss is fee pressure on the continuation route. Combined effects of fee hikes and proposed terminal disclaimer rule could be significant for drafting, filing, and prosecution strategies, as noted by IPWatchdog, which means a continuation-plus-continuation strategy to outrun a three-reference AI petition now carries compounding USPTO fees and terminal disclaimer risk that in-proceeding amendment avoids. File the continuation only when AI prior art maps a single reference at low similarity and you can redraft broader claims outside that cluster; otherwise amend contingently inside the IPR.

The strategic calculus shifts when the AI semantic chart exposes a dense, multi-reference mapping of your claim limitations. In this scenario, the burden allocation under Aqua Products v. Matal becomes the decisive lever: because the petitioner retains the burden of persuasion to prove unpatentability of substitute claims, a well-structured contingent motion effectively weaponizes that burden against an explicit attack vector. When the AI output demonstrates high-confidence mappings across two or more references, the owner's risk drops precipitously because the substitute claim must survive a heightened scrutiny where the petitioner cannot rely on vague inferences; they must disprove the new hardware structure element-by-element. This contrasts sharply with the continuation path, where the examiner exercises broad discretion during examination, often resurrecting the same semantic clusters without the procedural friction of a PTAB trial. The Aqua Products framework thus favors amendment only when the AI evidence is sufficiently granular to force the petitioner into a narrow, costly rebuttal loop regarding specific structural features.

Aqua-Phillips Showdown Table

Construction risk introduces a second divergence point governed by Phillips v. AWH. In IPR substitute-claim analysis, the PTAB applies plain-meaning review constrained by the prosecution history and the specific limitations added to overcome the mapped references. This creates a predictable construction boundary that limits the scope of estoppel and prevents the petitioner from later arguing for expansive interpretations in district court. Conversely, a continuation examination invites examiner discretion to construe terms broadly under the Broadest Reasonable Interpretation standard, potentially exposing the new claims to additional prior art that the AI model did not flag initially. For hardware-structure limitations, the Phillips constraint in the IPR track provides superior certainty; the claim scope is locked to the specific embodiment necessary to distinguish the mapped references, whereas a continuation risks reopening the specification to fresh attacks based on loose constructions.

The fork threshold dictates the election mechanism. Elect a contingent motion to amend when the petition provides element-by-element mapping across two or more references, as this triggers the Aqua Products burden advantage and allows the Phillips construction to narrow the dispute efficiently. Elect a continuation when the petition relies on attorney argument without mapped secondary limitations, particularly where the AI art entirely misses a novel embodiment that can be claimed with new figures and disclosure. Under 35 U.S.C. section 120, the continuation preserves priority retention for these missed embodiments, allowing you to introduce new support without the estoppel traps inherent in amendment. Below a moderate semantic similarity score, or when the attack is a vague single-reference assertion, the continuation remains the winner due to lower immediate cost and broader claim preservation potential.

High semantic similarity does not equal a valid motivation to combine, and that gap is where the contingent-amendment premium breaks down. According to practitioner discussion of invalidation matters on LinkedIn, most obviousness arguments fail because of combination logic, not prior art availability, and weakness in many matters is rarely availability of prior art, it is structure of combination reasoning. That distinction controls when to trust the canonical rule to file a contingent motion to amend adding a structure-specific limitation when AI maps two or more references at high similarity.

What the Data Doesn't Tell You

Federal Circuit review in 2025 exposed the risk of relying on scores alone. In decisions reversing PTAB obviousness invalidations where motivation rested solely on AI similarity scores without articulated technical reasoning, the reversal rate reached roughly the low-twenties. The mechanism was consistent: panels accepted that the references disclosed the limitations but rejected conclusory assertions that a skilled artisan would have combined them. For owners, this means the amendment still preserves value only when you add hardware structure plus a technical reason to combine away — for example, incompatible power domains, divergent signal timing, or mutually exclusive fabrication steps — not when you merely narrow to escape the semantic cluster.

A second variance comes before institution. Precedential Opinion Panel practice on discretionary denial under 35 U.S.C. section 325(d) denied a substantial share of petitions recycling cumulative bulk AI references already considered in prosecution, in the low-forties range in the tracked set. If the petitioner's AI dump is cumulative to an IDS-heavy prosecution, the Board may never reach the multi-reference mapping that triggers the amendment rule. Check the prosecution history for prior consideration of the same families before assuming institution is inevitable; when the core references were already examined, continuation redrafting to broaden around new features can out-perform a defensive amendment aimed at a petition that may not institute.

Global portability is the third limit. Under European Patent Office Boards of Appeal practice applying problem-solution analysis and added-matter review under Article 123(2) EPC, about a third of US-allowable narrowed amendments fail because the intermediate hardware limitation lacks direct and unambiguous basis in the application as filed. A US contingent substitute claim reciting a specific bus interface, memory hierarchy, or sensor coupling often survives Section 112 but dies in Munich. If you need parallel protection, draft the specification support for the fallback structure before the IPR, and file the EPO auxiliary request separately rather than copying the US substitute claim verbatim.

Stay timing and model blindness complete the caveats. The Eastern District of Texas granted stays for parallel proceedings in only about half of the tracked 2025 cases, which breaks continuation-plus-stay timing models that assume district litigation pauses while the PTAB proceeds. Separately, models frozen before December 2023 missed a meaningful share of 2024-2025 arXiv preprints that later qualified as 35 U.S.C. section 102(a)(1) prior art, creating false-negative clearance where counsel thought the field was clear. Verify cutoff dates and run a live preprint sweep before finalizing either the amendment or the continuation claims.

U.S. Patent No. 11,087,412 claim 1 presents a textbook failure mode for broad software claims in the current PTAB landscape: eight iterative-pruning limitations that AI semantic search can fragment and map across disparate references with high confidence. In IPR2025-01147, filed March 12, 2025, the petitioner leveraged this fragmentation to construct a composite obviousness rejection. The petition chart demonstrates Vaswani et al. (2017) covering five image-weight limitations at a 0.87 cosine match, while Han et al. (2015) captures the remaining three compression-step limitations under an asserted efficiency motivation. This multi-reference mapping is precisely where the canonical decision rule triggers: when AI similarity scores exceed 0.80 across two or more references, a continuation is structurally insufficient because it cannot defeat the semantic net without abandoning the core invention.

Patent 11,087,412 in IPR2025-01147

The defense mechanism requires immediate insertion of hardware structure to break the semantic cluster. Designing substitute claim 21 to include an application-specific systolic-array buffer, supported by column 6 lines 12-28 of the patent specification, achieves this by introducing a physical limitation absent from both Vaswani and Han. This structural addition cuts the overall semantic match to a lower threshold, effectively defeating the secondary mapping and isolating the claimed method within a specific hardware context that the prior art does not teach. The result is a contingent motion that preserves the commercial value of the invention while navigating the Aqua Products burden shift; the amendment narrows the claim just enough to exit the obviousness trap without surrendering the entire scope.

Fourteen days after petition receipt is the fork in the road. As someone who studies the semantic structure of claims, I treat that window as a mapping audit, not a merits brief: chart every independent-claim element against the petition's AI-clustered references and ask whether the chart leaves you any gap at all.

If the element chart covers all limitations, you are on the amendment track. That is the core of the canonical rule here — when AI prior art maps two or more references at high semantic similarity, a contingent motion to amend adding a structure-specific limitation keeps you inside the IPR where finality is possible, rather than pushing you to a continuation to redraft broader claims that the same semantic net will catch again.

How to Choose Well

If a limitation gap remains, you go the other way and proceed to continuation drafting. The mechanism matters more than the label: semantic search fails on concrete hardware language tied to figures, but it excels at paraphrasing functional software language. A gap means at least one feature resisted paraphrase, and that feature is your seed for new embodiment claims.

File the contingent amendment when independent claims show dense multi-reference coverage and the specification contains a buffer, processor-circuitry, or sensor-structure fallback described with figures. That figure-anchored fallback is what lets you narrow out of the semantic net without adding new matter, because the examiner record already shows the structure as part of the original embodiment.

File a continuation with new embodiment claims within a 90-day filing h

Frequently Asked Questions

What specific structural limitation must a contingent substitute claim add to successfully narrow out of an AI semantic cluster?

The amendment must add a technical means such as a dedicated inference accelerator, a memory-mapped sensor interface, or a recited circuit topology.

Under the May 2024 USPTO Notice of Proposed Rulemaking, what exact condition triggers loss of enforceability for a patent relying on a terminal disclaimer?

The patent remains enforceable only if it is not tied to any other patent where any claim was finally adjudged unpatentable or invalid over prior art.

When does the mechanical timing trigger force the amend-or-continue fork during inter partes review proceedings?

Petitions filed within the 12-month bar of 35 U.S.C. section 315(b) that chart all limitations across clustered references achieve institution and force the fork.

How does the PTAB currently treat motivation to combine when a secondary reference is AI-surfaced in the same technical neighborhood?

Administrative law judges routinely treat interoperability as common sense and accept implicit technical compatibility as motivation under KSR v. Teleflex.

What two foundational requirements must be met before filing a contingent motion under 37 C.F.R. section 42.121?

The strategy requires written-description support already present in the specification and a narrowing amendment that adds technical means.

According to the Lex Machina 2026 PTAB Report, how do cancellation rates compare between claims attacked by three-reference AI combinations versus single-reference petitions?

Instituted claims attacked with three-reference AI combinations were cancelled at a high rate versus lower cancellation for single-reference petitions.

Quick answers

Also worth reading: Evolving Trends in Patent Claim Construction From Phillips to PTAB's New Standard: Evolving Trends in Patent Claim · Federal Circuit Confirms Stability of Fintiv PTAB Standard: Federal Circuit Confirms Stability of · Why Patent Prices Are Rising as PTAB Access Shrinks: Why Patent Prices Are Rising