Beta-Disclaimer: Dieses Case Study zeigt einen synthetischen Demonstrations-Loop. Alle Inhalte dienen der Illustration und ersetzen keine Rechtsberatung. “Synthetisches Beispiel — dient der Illustration, kein reales Patent.”
USPTO §103 • Software Loop

ML-Based Cache Eviction
Prosecution Loop

US20220318887A1 (synthetisch) • 35 U.S.C. §103 Obviousness

US20220318887A1 →

47/100 — Novelty Gap

Multi-Reference-§103-Rejection über US8626791B1 + US20190251040A1

Pilot anfragen → Live-Demo ansehen
1

Ausgangslage

Technisches Umfeld und Art der Ablehnung.

Patent US20220318887A1 — Machine Learning Cache Eviction (synthetisch)
Tech Field Cache management in distributed computing, ML-driven eviction policy
OA Type Non-Final, 35 U.S.C. §103 Obviousness Rejection
Rejection Claims 1-12 rejected over US8626791B1 (Predictive Model Caching, Salesforce) combined with US20190251040A1 (Google, ML Cache Eviction)
Key Issue Multi-reference combination; no single reference teaches all claim elements
Agent: OA empfangen → §103-Multi-Reference-Rejection identifiziert → Gap-Analyse gestartet.
2

Original-Claims

Die beanspruchten Ansprüche 1-12, wie ursprünglich eingereicht.

Typ Nr. Inhalt
Unabhängig 1 A computer-implemented method for cache eviction comprising: receiving a cache miss for a data item D; computing, via a machine learning model, a predicted eviction accuracy score for D; comparing the predicted eviction accuracy score to a threshold; evicting at least one cached item based on the comparison.
Abhängig 2 The method of claim 1, further comprising storing the predicted eviction accuracy score in a metadata table associated with the cache.
Unabhängig 3 A system comprising: a processor; a memory storing instructions that, when executed, cause the processor to perform the method of claim 1.
Abhängig 4–12 Claims 4–12: threshold mechanisms, incremental learning updates, distributed cache clusters, etc.
Synthetisches Beispiel — dient der Illustration, kein reales Patent.
3

OA-Parsing-Output

ClaimForge’s OA-Agent hat die Office Action strukturiert extrahiert.

35 U.S.C. §103(a) — Obviousness

Rejection Summary

§103(a), Claims 1-12 obvious over US8626791B1 + US20190251040A1

Primary Ref [1] — US8626791B1 (Salesforce)

  • Ref 1 Passage: col.4 ll.20-35 — request frequency as eviction proxy (NOT predicted eviction accuracy score)
  • Note: Frequency ist rückwärtsgewandt; predicted accuracy ist vorwärts-gerichtete ML-Prognose.

Secondary Ref [2] — US20190251040A1 (Google)

  • Ref 2 Passage: [0032] — ML prediction for cache eviction

Motivation to Combine

KSR Rationale A+B; both in same field; combination yields predictable results

Gap Elements

  • (a) per-miss ML prediction
  • (b) metadata table with per-item scores
  • (c) threshold-based eviction with comparison
Novelty Gap Score
47
/ 100
Agent: OA analysiert → Gap-Elemente identifiziert → Amendment-Vorschlag generiert.
4

Claim-Amendments (rot markiert)

Rot markierte Änderungsvorschläge für Anspruch 1 mit Spec-Ankern.

Amended Claim 1 (Neu) 
A computer-implemented method for cache eviction comprising:
   receiving a cache miss for a data item D in a distributed cache system;
   retrieving, from a metadata table, a predicted eviction accuracy score associated
   with D, wherein the predicted eviction accuracy score is computed by a machine
   learning model trained on historical cache access patterns;
   comparing the predicted eviction accuracy score to a dynamically adjustable threshold,
   wherein the dynamically adjustable threshold is determined based on a current cache
   hit rate and a target cache hit rate;
   evicting at least one cached item from the cache system based on the comparison;
   updating the metadata table to reflect the eviction outcome;
   and retraining the machine learning model incrementally based on the eviction outcome.
Spec-Anker: [0045] (threshold adjustment) [0062] (metadata table) [0078] (incremental retraining)
Key Additions vs. Original
Element Original Amended
1 Closed-loop feedback architecture eviction outcome → metadata update → incremental retraining
2 Dynamic threshold based on current/target cache hit rate
3 Per-miss retrieval from metadata table (not compute-from-scratch)
Agent: Anspruch 1 geändert → 3 Gap-Elemente adressiert → Counter-Arguments werden generiert.
5

Counter-Arguments-Draft

Vollständiger Entwurf der Erwiderung auf die Office Action.

RESPONSE TO OFFICE ACTION — US Patent Application No. [App#] Claims 1-12 rejected under 35 U.S.C. § 103(a) as obvious over US8626791B1 in view of US20190251040A1. Applicant respectfully traverses. I. NO PRIMA FACIE CASE ESTABLISHED A. US8626791B1 Does Not Teach "Predicted Eviction Accuracy Score" Request frequency ≠ predicted eviction accuracy score. Frequency is backward-looking; predicted accuracy is forward-looking ML prediction. B. No Motivation to Combine Under KSR Rationale A US8626791B1 explicitly teaches away from computational overhead (minimizes ML invocation to preserve latency). US20190251040A1's per-miss ML computation contradicts this teaching. C. New Element: Closed-Loop Feedback Architecture The amended claims add: eviction outcome → metadata update → incremental retraining. Neither reference discloses this feedback loop — this is a novel system architecture, not a mere combination of known elements. II. SECONDARY CONSIDERATIONS 23% improvement in cache hit rate over frequency-based methods in production testing. Unexpected given prior art ML-based eviction showed only marginal gains. III. CONCLUSION Withdrawal of §103 rejection and allowance of claims 1-12 requested.
Agent: Counter-Arguments generiert (6 Min.) → Autonomous Re-Search auf amended Claims gestartet.
6

Autonomous Re-Search (Amended Claims)

Re-Search nach automatischer Änderung der Ansprüche — Espacenet + Google Patents.

Rank Patent Score Unterscheidung
1 US10671435B1
NVIDIA, 2020		 31/100 GPU data transform caching — no ML prediction, no feedback
2 US10311372B1
Akamai, 2016		 28/100 Cluster routing — no per-item accuracy score
3 US11983806B1
Adobe, 2024		 24/100 ML image generation — unrelated
4 EP3965362A1
checkpoint, 2023		 19/100 Domain reputation ML — unrelated
5 US11887367B1
OpenAI, 2023		 15/100 Video ML — unrelated
Novelty Verdict
No single reference or combination teaches closed-loop ML cache eviction with dynamic threshold + incremental retraining.
Confidence: 78%
Agent: Re-Search abgeschlossen. Keine Einzelreferenz und keine Kombination deckt Closed-Loop-Architektur ab → Chain abgeschlossen.
7

Verdict

Ergebnis der automatischen Chain-Analyse.

Action: Refine
OA-Agent's response would likely overcome rejection on amended claims. Feedback loop is genuinely novel; dynamic threshold has technical character. Human check: §101 subject matter eligibility + spec support for amended elements.
8

Was der Loop Gespart Hat

Zeitvergleich: manuell vs. ClaimForge Loop.

Szenario Stunden
Ohne ClaimForge 18–24 h
ClaimForge Loop 2.5 h
Netto-Ersparnis 15–20 h
Caveat: Real-world review, spec-checking, attorney review add 4–8 h. Total mit ClaimForge: 6–10 h.
9

Limits

Was ClaimForge in diesem Loop nicht abdeckt.

10

Probieren Sie den Loop

ML-Based Cache Eviction als Demo-Szenario

ClaimForge’s autonomer Loop kann über den gesamten Prosecution-Workflow eingesetzt werden — von der Prior-Art-Recherche über Claim-Tree-Generierung bis zur Office-Action-Erwiderung.

Pilot anfragen → Live-Demo: Amended Claim 1 Features