oes512q-latch: OES-32 / OES-512 classical residual latch with a self-calibrated threshold (OES latch v2) and a NAB benchmark
Abstract
Standard-library Python research prototype of the OES-32 / OES-512 classical residual latch. A 32-value residual block is scored with S = 0.45·Peak + 0.35·RMS + 0.20·MeanAbs and latches when S ≥ τ. OES-512 applies the same latch to 16 contiguous blocks of 32, which gives a 16-bit coarse syndrome. The fixed default τ = 0.50 is retained. OES latch v2 (oes_calibration) adds an optional, label-free, self-calibrated τ: the 0.99 quantile of warm-up block scores on robust-scaled residuals (1.4826·MAD, with a mean-absolute-deviation fallback). It fails closed on invalid input. An optional amplitude-encoded feature vector (5 or 9 qubits' worth of amplitudes, a classical list) is provided as a software sidecar after the classical latch. Benchmark. The package includes a reproducible block-level benchmark on the 47 labelled real-data series of the Numenta Anomaly Benchmark (NAB, MIT license, pinned commit, SHA-256 manifest). It compares the latch with rolling z-score, CUSUM and IsolationForest using warm-up-calibrated thresholds, paired statistics, block-size sensitivity, IsolationForest seed robustness and NAB's official scorer on a sub-corpus. The results are mixed and reported honestly. CUSUM has the best block F1. The latch has the best mean average precision, but not significantly better than CUSUM. The rolling z-score beats the latch on the NAB score. The fixed τ = 0.50 behaves like flagging every block on this data. Evidence tags. Demo outputs and unit tests: SYNTHETIC. NAB benchmark numbers: REPORTED (measured on public data; results.json records the SHA-256 hashes of the code that produced them). Self-calibrated τ as a contract extension: TARGET. Operation on live telemetry or hardware: UNRUN (not claimed). Relationship to ADR-001. The normative per-block residual remains oes32-residual@b77b612. The weighted latch and its self-calibrated τ are an extension in this repository. Scope and limits. Research prototype. It is not a stabilizer quantum error-correcting code, not a 32- or 512-qubit Hilbert space, and has not been run on quantum hardware, a QPU or a cloud quantum service. It is not a certified safety system, a medical device, field telemetry or a commercial product. No production-readiness claim is made. Evidence tags follow sparkainlp-x/.github (SYNTHETIC / REPORTED / TARGET / UNRUN).
