HomeCase Studies › AI Prior Art Software Ranking

AI · Software · USPTO §103 · Prior-Art Ranking

AI-Assisted Prior-Art Ranking for Software-Implemented Invention

Four software-implemented patent families — ML feedback-loop cache eviction, distributed-consensus ledger sharding, real-time streaming event deduplication, and federated-learning differential-privacy budget accounting — ranked and re-ranked through ClaimForge's AI retrieval pipeline. 35 U.S.C. §103 obviousness objections surfaced against Espacenet + Google Patents in 90 minutes.

4

Software-Implemented Invention patent families ranked by ClaimForge's AI retrieval pipeline

1

Databricks / Snowflake (ML-based cache eviction feedback loop)

US 11,327,178 B2 · ML-Based Cache Eviction · Feedback-Loop Architecture

ML-Based Cache Eviction
Before / Original Claim (excerpt)
“A cache management system comprising a machine-learned predictive model configured to predict cache utility for items in a cache, wherein items with lower predicted cache utility are evicted from the cache in favor of items with higher predicted cache utility.”
After / Amended Claim (excerpt)
“A cache management system as set forth above, wherein the machine-learned predictive model is retrained at a closed-loop retraining cadence determined as a function of a rolling residual statistic between predicted cache utility values and corresponding measured eviction outcomes, and wherein the cache management system records each eviction outcome and feeds the recorded outcome back to the predictive model at the closed-loop retraining cadence.”
Rejection 35 U.S.C. §103 obviousness rejection over Belady 1966 (MIN optimal-eviction policy formulation) and Megiddo 2003 (randomized online caching with adaptive threshold); ML-based cache-eviction system claimed at the level of "predict cache utility via predictive model" without limitation to a closed-loop retraining signal coupling prediction error to eviction policy
Escape Logic Restrict the ML-cache-eviction genus to a system wherein the predictive model is retrained at a defined cadence using a closed-loop signal derived from eviction-policy outcome residuals (predicted-utility-vs-actual-eviction delta), with the retraining cadence selected as a function of the rolling error statistic — closed-loop retraining cadence-control absent from Belady 1966 and Megiddo 2003 which each disclose static optimal-eviction or threshold-tuning without feedback-driven policy-outcome correction
ClaimForge Action OA-parsing extracts §103 obviousness over Belady 1966 / Megiddo 2003 → embedding-based prior-art retrieval runs cosine-similarity search across ML-cache-eviction patent family → citation-graph pruning removes low-degree prior-art nodes citing only Belady or Megiddo → semantic novelty scoring flags closed-loop retraining cadence as a non-disclosed feature → amendment drafted with retraining-cadence-anchor plus rolling-residual-statistic constraint → re-search confirms clearance via Espacenet + Google Patents
Confidence: 83%
2

NEAR Protocol / Polkadot (distributed-consensus ledger sharding)

US 10,956,915 B2 · Distributed Consensus · Sharding · Throughput Optimization

Distributed Consensus
Before / Original Claim (excerpt)
“A distributed ledger system comprising a plurality of shards, each shard executing transactions independently, wherein transactions crossing shard boundaries are committed after consensus between the affected shards.”
After / Amended Claim (excerpt)
“A distributed ledger system as set forth above, wherein cross-shard transaction commit ordering is determined by reference to a monotonic shard-throughput ledger that is hash-chained across shard-boundary transactions, and wherein the monotonic shard-throughput ledger provides a deterministic ordering index for cross-shard commit sequencing without reliance on a global lock or a single-source transaction-ordering service.”
Rejection 35 U.S.C. §103 obviousness rejection over Lamport 1998 (Paxos distributed-consensus protocol) and Zamani 2018 (blockchain sharding architecture); distributed-consensus sharding system claimed at the level of "shard assignment based on computational load" without limitation to a deterministic cross-shard transaction-commit ordering anchored to a monotonic shard-throughput ledger
Escape Logic Restrict the distributed-consensus sharding system to one wherein cross-shard transaction commit ordering is derived from a monotonic shard-throughput ledger, with the throughput-ledger hash-chained across shard-boundary transactions and used to deterministically resolve cross-shard commit sequencing — monotonic-throughput-ledger-anchored cross-shard commit ordering absent from Lamport 1998 and Zamani 2018 which each disclose independent-shard consensus without coupled throughput-ledger commit ordering
ClaimForge Action OA-parsing flags Lamport 1998 / Zamani 2018 cross-shard-commit anticipation → embedding prior-art retrieval scores distributed-sharding patent candidates on cross-shard-commit novelty → citation-graph pruning strips generic-sharding background art → semantic novelty scoring surfaces monotonic-throughput-ledger + hash-chain anchor as a non-disclosed commit-ordering constraint → amendment drafted with throughput-ledger-anchor and deterministic-ordering no-global-lock features → re-search validates clearance across Espacenet + Google Patents
Confidence: 80%
3

Confluent / Apache Flink (real-time streaming event deduplication)

US 10,956,275 B1 · Streaming Event Processing · Probabilistic Deduplication · Real-Time Pipelines

Streaming Event Processing
Before / Original Claim (excerpt)
“A stream-processing system comprising a probabilistic data structure configured to deduplicate events received on a real-time event stream, wherein events determined to have been previously seen by the probabilistic data structure are suppressed from downstream processing.”
After / Amended Claim (excerpt)
“A stream-processing system as set forth above, wherein a false-positive rate parameter of the probabilistic data structure is dynamically adjusted based on a backpressure signal received from a downstream event sink, and wherein the false-positive rate parameter is updated at a sliding-window granularity derived from downstream sink throughput telemetry, such that the false-positive rate targets a worst-case acceptable deduplication error budget under measured backpressure.”
Rejection 35 U.S.C. §103 obviousness rejection over Bloom 1970 (Bloom filter probabilistic-membership structure) and Flajolet-Martin 1985 (Flajolet-Martin sketch cardinality estimation); streaming event deduplication claimed at the level of "deduplicate events using a probabilistic data structure" without limitation to a tunable false-positive-rate-anchored decision threshold coupled to downstream-sink backpressure
Escape Logic Restrict the streaming-event-deduplication system to one wherein the probabilistic-data-structure false-positive-rate parameter is dynamically adjusted as a function of the downstream consumer sink's measured backpressure, and the false-positive-rate decision threshold is updated at a sliding-window granularity derived from the sink-throughput telemetry — dynamic-FPR-backpressure coupling absent from Bloom 1970 and Flajolet-Martin 1985 which each disclose static-FPR probabilistic structures without downstream-coupled threshold adjustment
ClaimForge Action OA-parsing extracts Bloom 1970 / Flajolet-Martin 1985 §103 obviousness over static-FPR probabilistic structures → embedding retrieval surfaces candidates across streaming-deduplication patent space → citation-graph pruning filters commonly-cited background art → semantic novelty scoring flags downstream-backpressure-coupled dynamic-FPR threshold adjustment → amendment grounds the genus in backpressure-signal-input + sliding-window-update features → re-search confirms clearance against Espacenet + Google Patents
Confidence: 78%
4

Google Research / Apple (federated-learning DP-budget accountant)

US 11,238,193 B2 · Federated Learning · Differential Privacy · Budget Accounting

Federated Learning
Before / Original Claim (excerpt)
“A federated-learning system comprising a differential-privacy budget manager configured to allocate a per-round differential-privacy budget to client update contributions across training rounds.”
After / Amended Claim (excerpt)
“A federated-learning system as set forth above, wherein the differential-privacy budget manager maintains a client-noise-consumption ledger that decrements a per-client epsilon budget as a function of an empirical noise-magnitude distribution observed across prior rounds, and wherein the per-round budget allocation is rate-adaptively bounded by a global epsilon-cap tolerance that is updated as a function of the client-noise-consumption ledger.”
Rejection 35 U.S.C. §103 obviousness rejection over Dwork 2006 (epsilon-differential privacy framework) and McMahan 2017 (federated-learning client-update aggregation); federated-learning DP-budget accountant claimed at the level of "track differential-privacy budget across training rounds" without limitation to a per-round budget allocation strategy coupled to a client-noise-consumption ledger with rate-adaptive allocation
Escape Logic Restrict the federated-learning DP-budget accountant to one wherein the per-round DP-budget allocation is computed by a client-noise-consumption ledger that decrements the per-client epsilon budget as a function of the empirical noise-magnitude distribution observed in prior rounds, with the rate-adaptive allocation bounded by a global epsilon-cap tolerance — client-noise-consumption-ledger rate-adaptive allocation absent from Dwork 2006 and McMahan 2017 which each disclose DP budget framing without ledger-driven per-round epsilon-allocation dynamics
ClaimForge Action OA-parsing identifies Dwork 2006 / McMahan 2017 §103 obviousness over generic DP-budget framing → embedding retrieval combines FL+DP patent-candidate scoring → citation-graph pruning removes foundational-ML background art → semantic novelty scoring flags client-noise-consumption-ledger + rate-adaptive allocation as coupled novelty constraints → amendment drafted with ledger-anchor + global-cap features → re-search confirms unique combination across Espacenet + Google Patents
Confidence: 81%
Pipeline

How ClaimForge Automates Each Software-Implemented Invention Prior-Art Ranking Step

The autonomous ClaimForge loop covers every step from embedding-based prior-art retrieval through citation-graph pruning and amendment drafting across all four software patent families.

Step 1
Embedding Prior-Art Retrieval
Indexes software-implemented invention claims + abstracts into text embeddings; top-K candidate retrieval over Espacenet + Google Patents with cosine-similarity threshold.
Step 2
Semantic Ranking Score
Ranks retrieved candidates using semantic novelty scoring against the §103 obviousness framework; flags feedback-loop, throughput-ledger, backpressure-coupled, and rate-adaptive features as novelty anchors.
Step 3
Citation Graph Pruning
Prunes low-degree prior-art nodes that cite only foundational background art (Belady, Lamport, Bloom, Dwork); surfaces genuinely-coupled novelty constraints from cited prior art.
Step 4
Amendment Drafting
Generates amended claims grounded in embedding novelty gaps, citation-graph-pruned prior art, and software-implementation-specific feature anchors (closed-loop cadence, monotonic throughput ledger, dynamic-FPR backpressure, client-noise-consumption ledger).
Step 5
Re-Search Validation
Runs Espacenet + Google Patents on amended claims; final confidence band 75–85% based on candidate-prior-art clearance counts after pruning.

Run Your Own Software-Implemented Invention OA Loop

Upload a 35 U.S.C. §103 office action — ClaimForge ranks the most relevant prior art, prunes citation-graph noise, and drafts an amendment grounded in your software-implemented invention's specific feature novelty. 90 minutes vs. 20+ hours manually.