Read the filing.
A published research proposal
AI systems generate answers. ASCQ proposes a memory for their corrections.
The application proposes a persistent correction memory: a record of prior AI-output corrections, retained beyond the immediate exchange. Similarity retrieval connects a current input with relevant earlier fixes, giving later processing a history to draw on rather than starting from the same mistake.
An audit-trail module is proposed to record the correction and its application. Tamper-evident records connect a retrieved fix with the decision it informed. The filing describes verifiable evidence of that relationship, rather than a claim that every AI output is automatically correct.
Effectiveness information feeds back into retrieval and scheduling parameters. The proposed loop joins correction memory, audit records and workload planning. Its intended result is fewer recurring errors; this page describes the application, not a measured production outcome.
A corrected answer should not be the end of a conversation. It can become context for the next decision.
That is the proposal behind ASCQ.
The published application describes adaptive quality management for AI processing. A persistent memory stores prior corrections. Similarity retrieval recalls relevant fixes. An audit-trail module records their application, while a feedback controller updates how corrections are retrieved and workloads are scheduled. These parts are proposed as a closed loop, not a collection of disconnected checks.
Mahmoud Alshraideh and Ihab Odeh are named as founders and inventors on ASCQ's public site. The site states patent-pending status. The published application is the source for the architecture described here; no customer deployments, independent benchmarks or granted-patent claims are implied.
Four proposed modules connect correction, evidence and feedback.
The application describes a persistent store of prior corrections to AI outputs. A current input is compared with stored records to retrieve a relevant correction. Claim 3 describes embedding vectors and an approximate nearest-neighbor index, with similarity weighted by recency and prior effectiveness. The intended role is to carry earlier fixes into later AI processing, rather than losing them at the end of an exchange.
The audit-trail module is proposed to generate a tamper-evident record of a correction and its application. Claim 4 describes Merkle-tree records and external verification, including a timestamp authority or public blockchain. The purpose is traceability: connecting a correction with where it was applied, with selective disclosure of audit proofs to external reviewers.
A predictive resource manager is proposed to schedule AI workloads using service-level constraints, carbon-intensity information and quality metrics. Claim 5 describes deferring non-latency-critical work while maintaining service-level requirements. This is an architectural proposal for balancing workload demands, not a published claim of achieved cost or emissions savings.
A feedback controller is proposed to update retrieval and scheduling parameters using effectiveness information from the audit trail. Claim 6 describes changing retrieval weights in proportion to the effectiveness of applied corrections. The proposed loop links what was recalled, what was applied and what happened next. Its effectiveness requires implementation and independent evaluation.
The inventors
The proposal joins persistent correction memory, an audit-trail module, a predictive resource manager and a feedback controller in a closed-loop adaptive system. Each part is described in the published application. This is a research summary, not a customer testimonial or proof of production performance.
Prior AI-output corrections are stored in persistent memory and retrieved by similarity to a current input.
Claim 3 describes correction embeddings in an approximate nearest-neighbor index, weighted by similarity, recency and earlier effectiveness.
Claim 4 describes Merkle-tree correction records and anchoring their root to an external verification service, with selective disclosure of proofs to reviewers. The design is proposed in the filing; this summary does not claim a deployed verification service.