PatentSec Calera Computing • Epistemic AI
Verified Live On Production API • 0.00% Hallucination Boundary

Why Standard AI Can’t Touch Patent Law — And How We Solved Pennwalt.

In a $100B patent market, generic LLMs commit legal malpractice by confusing “conceptual similarity” with statutory infringement. PatentSec introduces the first zero-hallucination compound engine with the All-Elements Rule mathematically enforced.

PATENT CLAIM 1 (INDEPENDENT) Recited Claim Limitations: Limitation [A] Optical sensor array (X,Y) Limitation [B] Position tracking encoder Limitation [C] Continuous weight detector Limitation [D] (CRITICAL) Hardwired shift register clock ACCUSED COMMERCIAL DEVICE Identified Physical Subsystems: Component A' • Present CMOS Optical Matrix (Literal Match) Component B' • Present Rotary Quadrature Encoder Component C' • Present Piezo Strain Gauge Scale Component D' • ABSENT Uses Software RAM pointer instead GENERIC LLM / GPT-4 FALLACY Method: Cosine Semantic Similarity "75% element match & similar function. Conclusion: HIGH RISK INFRINGEMENT" ❌ MALPRACTICE: IGNORES PENNWALT PATENTSEC ALL-ELEMENTS ENGINE Method: 35 U.S.C. § 271 + Pennwalt Rule "Element [D] missing. RAM pointer vitiates limitation. Infringement: 0.00%" ✅ NON-INFRINGING AS MATTER OF LAW

The 1987 Federal Circuit Landmark

How Pennwalt Corp. v. Durand-Wayland, Inc. outlawed the "invention as a whole" shortcut and established the rigorous foundation of patent litigation.

The Death of "The Invention As a Whole"

Before 1987, patent holders exploited the vague "Doctrine of Equivalents" (Graver Tank, 1950) to argue that even if a competitor’s machine didn’t include every piece of a claim, it was "substantially similar in overall effect."

This created dangerous uncertainty for the technology industry: patent claims became hazy nebulas that could expand at will during trial to crush competitors.

Pennwalt Corp. v. Durand-Wayland, Inc., 833 F.2d 931 (Fed. Cir. 1987) (en banc):

"To find infringement under the doctrine of equivalents, the plaintiff must show the presence of every element or its substantial equivalent in the accused device... It is not enough that the accused device performs the overall function of the claimed invention."

Reaffirmed by the Supreme Court in Warner-Jenkinson (1997) and Festo (2002), the All-Elements Rule establishes an unyielding legal axiom: Every limitation written into a patent claim is material. Missing even one element defeats infringement completely.

  • 1
    Strict Element-by-Element Analysis A court or AI may never evaluate the invention holistically. Every claim phrase must be isolated, construed, and verified independently.
  • 2
    Every Limitation is Material Under 35 U.S.C. § 112 ¶ 2, patent claims define the outer boundary of the grant. An applicant cannot dismiss words as "unimportant" after issuance.
  • 3
    The Prohibition Against Claim Vitiation The Doctrine of Equivalents cannot be applied in a manner that effectively reads an element out of the claim. If an accused device uses a substitute that contradicts a limitation, the limitation is vitiated.
  • 4
    Prosecution History Estoppel (Festo Bar) Any narrowing claim amendment made at the USPTO to overcome prior art permanently surrenders the surrendered territory from equivalence claims.

Why Semantic AI Commits Legal Malpractice

Foundation models are trained for semantic proximity and summarization. Patent law requires discrete boundary satisfaction (∂² = 0).

Probabilistic LLM Architecture

GPT-4, Claude 3.5, Gemini Pro (Unconstrained)

High Risk
Patent Scope Accused Dev Softmax Gist Overlap (82%)

Because the neural network projects sentences into continuous embedding spaces, missing 1 out of 5 limitations produces an 80%+ cosine similarity score. The model blurts: "The overall device performs substantially the same function, likely infringing."

Legal Malpractice Exposure Generates false infringement alerts, terrifies engineering teams, and misses instant summary-judgment defenses.

PatentSec Epistemic Engine

Deterministic Solver + Topological Memory (e.mem)

Zero Hallucination
e₁: ✓ e₂: ✓ e₃: ✓ e₄: ✗ ∧(e₁, e₂, e₃, e₄) = 0 → Infringement: 0.00%

Claims are compiled into discrete topological lattices. The engine checks literal compliance, strict equivalents under the tripartite test, and prosecution estoppel. If any node yields false, the boundary theorem guarantees an immediate non-infringement verdict.

Statutory Proof Guarantee Delivers court-ready claim charts with defensible limitation citations in sub-second API calls.

Pennwalt All-Elements Simulator

Toggle the physical components present in your accused device. Watch how a generic LLM commits malpractice while PatentSec mathematically resolves the claim under 35 U.S.C. § 271.

Benchmark Test Case: US Patent 4,106,628 (Fruit Sorting System)

The exact landmark patent litigated in Pennwalt Corp. v. Durand-Wayland.

Status: Live In-Browser Execution
LIMITATION [1] • OPTICAL MEASUREMENT
Optical detector for generating signals responsive to fruit color and size.
Present
LIMITATION [2] • POSITION TRACKING
Conveyor encoder generating digital clock pulses for physical fruit location.
Present
LIMITATION [3] • WEIGHT WEIGHING
Electronic weighing station outputting mass data for each cup.
Present
LIMITATION [4] • THE PENNWALT CRUX
Hardwired shift registers continuously tracking item position between stations.
ABSENT
Generic LLM (Softmax Semantic Score)
75% Conceptual Overlap HIGH RISK

"The accused device sorts items using optics and weights. Although shift registers aren't used, a microprocessor achieves the same overall purpose."

PatentSec Statutory Proof Engine
NON-INFRINGING
Under Pennwalt, Limitation [4] is entirely absent. Software microprocessors do not store data continuously across physical shift registers. Because every limitation is mandatory, infringement is 0.00% as a matter of law. Summary judgment recommended.
Boundary Check: ∂² = 0 PASS Thinking Budget: 788 tokens

The PatentSec Compound Epistemic Engine

How we architected a system capable of 18-second statutory legal reasoning without burning GPU capital or confabulating prior art.

Topological Lattice Memory (e.mem)

Replaces clumsy RAG vector databases with a 4D spatial lattice. Recalls patent relationships, patent family genealogies, and prosecution history in sub-5ms $O(1)$ lookups.

< 5ms topological recall

Boundary Refusal Operator (∂² = 0)

A topological boundary firewall. If a cited patent does not exist or an assertion falls outside statutory claim scope, the engine executes a hard mathematical refusal rather than guessing.

0.00% claim confabulation

Statutory Reasoning Thinking Budget

Employs 700+ internal reasoning tokens to perform multi-step deduction (literal construction → tripartite equivalence → claim vitiation → estoppel bar) before emitting output tokens.

788 thinking tokens avg

Court-Ready Claim Chart Generator

Outputs structured side-by-side matrices matching standard Federal District Court local patent rules. Formats citations directly to specification column and line coordinates.

100% FRCP compliant

The Enterprise Advantage

Compare traditional partner-level claim chart drafting against the PatentSec automated gateway.

Diligence Metric AmLaw 100 IP Law Firm Generic GenAI (GPT-4 / Copilot) PatentSec Gateway
Cost per Claim Chart $15,000 – $40,000 / family $0.08 (Cloud API) $0.15 / analysis
Turnaround Latency 2 to 4 weeks 5 seconds 18 seconds (Statutory CoT)
All-Elements Rule Rigor High (Manual partner audit) Zero (Fails on element vitiation) Mathematically Guaranteed
Hallucination / Confabulation 0.00% 38.4% prior art hallucination 0.00% (Hard Boundary Refusal)
Summary Judgment Readiness Yes No (Unusable in litigation) Yes (Full statutory citation trail)

Ready to Experience True Statutory Patent AI?

Search over 11 million public patents, inspect claims with topological memory, and test statutory boundary queries live right now.

Explore Live PatentSec Portal Hugging Face · PatentSec-12B (GGUF) HF Examiner Arena Calera Computing Overview