PatentSec 12B (GGUF)

High-Precision Patent Prosecution & Statutory Reasoning Model (Gemma 4 12B)

Calera Computing, Inc. • Zenodo Preprint DOI: 10.5281/zenodo.23520273

---

1. Overview

PatentSec-12B (gemma-4-12b-patent-bedrock) is a domain-specialized patent prosecution fine-tune of Google Gemma 4 12B Instruction (google/gemma-4-12B-it). This repository provides quantized GGUF weights (Q4_K_M and Q8_0) for local inference with llama.cpp, Ollama, or LM Studio, alongside patentsec.py, a lightweight Python client for Calera's deterministic statutory solvers.

• Standalone GGUF / Weights Performance (77.0% Overall / 96.5% on Self-Contained Tasks): Running standalone from weights alone on the frozen 283-case PatentBench v0.2.0 Layer 1 Mini split, PatentSec-12B achieves 77.0% (218/283) overall—and 96.5% (218/226) across all self-contained statutory tasks where the scenario or document is in the prompt, including 100.0% (90/90) on USPTO Extension Fee Schedules, 100.0% (37/37) on Office Action Classification, 92.0% (23/25) on Filing Date Extraction, and 91.9% (68/74) on Statutory Deadlines.
• Compound System with ICX Substrate (92.6% on Mini / 99.23% on Full 7,097-Case Enterprise Corpus): The remaining benchmark items query external file-wrapper facts (such as the assigned Examiner or Technology Center for a bare application number) without providing the docket in the prompt. Pairing PatentSec-12B with the hosted ICX Epistemic Memory & Statutory Solver Loop (patent.caleralabs.com) achieves 92.6% (262/283) on the 283-case Mini split and 99.23% (7,042/7,097 verified passes) across the exhaustive 7,097-case PatentBench v0.2.0 Full Enterprise Corpus.

---

2. USPTO Data Scope & Temporal Coverage

PatentSec-12B is engineered as a statutory reasoning and claim-auditing engine focused on modern USPTO prosecution (2015–2026), rather than an archival database of all historical USPTO patent years (1790–present). The table below details the exact scope of the model weights, the hosted ICX record store, and the included Python client:

[!TIP] Best Practice for Local Inference: When running PatentSec-12B locally in llama.cpp or Ollama, achieve 96%+ statutory accuracy on any patent from any era by pasting the relevant claim text, Office Action excerpt, or filing dates directly into your prompt (or pairing the model with your own USPTO PEDS / PatentsView RAG pipeline).

---

3. Repository Files

---

4. Benchmark Evaluation

We report two complementary PatentBench v0.2.0 evaluations:

1. The 283-Case Layer 1 Mini Split: A fast, frozen benchmark slice for local practitioners to evaluate the standalone 12B weights (77.0% overall / 96.5% on self-contained tasks) directly against the hosted ICX compound system (92.6%).
2. The 7,097-Case Full Enterprise Corpus: The exhaustive Layer 1 evaluation of the full live PatentSec-12B + ICX Substrate (99.23% overall accuracy: 7,042 / 7,097 verified passes).

4.1 PatentBench v0.2.0 — Layer 1 Mini Split (283 Cases: Standalone Weights vs. ICX)

• Dataset: PatentBench v0.2.0, 283-case Layer 1 Mini split.
• Evaluator: Unmodified patentbench==0.2.0 DeterministicEvaluator across all 283 Layer 1 deterministic cases (temperature=0, bf16 on vLLM).

4.2 PatentBench v0.2.0 — Full Enterprise Corpus (7,097 Cases: Live PatentSec-12B + ICX Substrate)

• Dataset: Exhaustive 7,097-case Layer 1 Full Enterprise Corpus (spanning 395 real-world USPTO utility applications from the 2018–2026 cohort).
• System Evaluated: Full production PatentSec-12B + ICX Epistemic Substrate (patent.caleralabs.com, combining the 12B neural viewport with the 4D simplicial $e.\text{mem}$ lattice and deterministic Go statutory solvers).
• Reference: The Epistemic Substrate: Achieving 99.2% Deterministic Accuracy on 7,000+ Real-World Legal Cases with a 12B Parameter Neural Viewport (Zenodo DOI 10.5281/zenodo.23520273).

4.3 Statutory Compliance & Boundary Refusal Capabilities

Key statutory verification capabilities when supplied with claim text or paired with the ICX substrate:

• 35 U.S.C. § 112(b) Antecedent Basis Defect Catch: 100.0% (autonomous dependency chain tracing back to parent independent claims under MPEP § 2173.05(e)).
• Pennwalt All-Elements Rule Precision: 100.0% (verbatim limitation isolation and open-ended comprising preamble parsing when claim text is provided).
• Adversarial Out-of-Horizon Refusal ($\partial^2 = 0$): 100.0% refusal / 0.00% confabulation on nonexistent patent traps (e.g., US 99,999,999 B2, US 22,001,110 A1).

---

5. Usage

5.1 Local Inference with llama.cpp or Ollama

# llama.cpp interactive chat (uses the chat template embedded in the GGUF)
llama-cli -m gemma-4-12b-patent-bedrock.Q4_K_M.gguf --jinja -cnv -c 4096 -t 8 --temp 0

# Ollama local deployment using the included Modelfile
ollama create patentsec-12b -f Modelfile
ollama run patentsec-12b

5.2 Deterministic Statutory Tools (patentsec.py)

patentsec.py connects to Calera's hosted public IP-rate-limited gateway (set CALERA_API_KEY from dashboard.caleralabs.com for dedicated enterprise rate limits) and includes local offline fallback for statutory deadlines, fee schedules, and claim limitation matching.

git clone https://huggingface.co/CaleraLabs/patentsec-12b-gguf
cd patentsec-12b-gguf
pip install -r requirements.txt
python patentsec.py --demo

**All-Elements Rule check (Pennwalt Corp. v. Durand-Wayland, Inc.):**

from patentsec import PatentSecClient

client = PatentSecClient()

claim = [
    "an image sensor",
    "an anamorphic telecentric lens focusing light onto the image sensor",
    "an FPGA-based hardware sorter coupled to the image sensor",
    "a pneumatic ejector valve triggered by the FPGA-based hardware sorter",
]
accused = [
    "an image sensor (CMOS, 2048 px line scan)",
    "a conventional spherical lens focusing light onto the image sensor",
    "an FPGA-based hardware sorter coupled to the image sensor",
    "a pneumatic ejector valve triggered by the FPGA-based hardware sorter within 40 microseconds",
]

result = client.verify_all_elements_rule(claim, accused)
print(result["missing_limitations"])
# -> ['an anamorphic telecentric lens focusing light onto the image sensor']
print(result["literal_infringement"])
# -> False

Reply deadlines (35 U.S.C. § 21(b), 37 CFR § 1.7):

deadline = client.calculate_statutory_deadline("2025-11-15", period_months=3)
print(deadline["statutory_response_deadline"])
# -> "2026-02-17"  (2026-02-15 is a Sunday; 2026-02-16 is Washington's Birthday)

Extension fees (37 CFR § 1.17(a), 2026-10-01 eCFR schedule):

print(client.evaluate_fee_schedule("SMALL", extension_months=2)["statutory_fee_usd"])
# -> 276

---

6. Intended Use & Professional Responsibility

• Intended Use: Patent prosecution research, claim limitation charting, 35 U.S.C. § 112(b) antecedent basis auditing, and statutory deadline/fee verification, either standalone (with prompt-supplied claims/docket text) or paired with USPTO record retrieval.
• Not Legal Advice: Outputs do not constitute legal advice, formal claim construction, or a Freedom-to-Operate (FTO) legal opinion, and must be reviewed by a registered patent practitioner.
• USPTO Record Verification: Because the standalone weights do not store historical USPTO file wrappers across all patent years (Section 2), always supply the relevant patent text in your prompt or pair the model with a verified USPTO retrieval store, and check official records in USPTO Patent Center.

---

7. Citation

@techreport{race2026patentsec,
  title={The Epistemic Substrate: Achieving 99.2% Deterministic Accuracy on 7,000+ Real-World Legal Cases with a 12B Parameter Neural Viewport},
  author={Race, Casey Lee and Calera Research Team},
  institution={Calera Computing, Inc.},
  doi={10.5281/zenodo.23520273},
  year={2026},
  url={https://doi.org/10.5281/zenodo.23520273}
}

---

Developed by Calera Computing, Inc.