Measured operating envelope

On 2026-10-17, a bounded study accepted repeated sparse local/short-cascade reuse for the first release. It does not establish that preprocessing benefits every one-shot solve. The measured decomposition candidate keeps root preprocessing default off, and full certificate verification/export and detailed rollback reports remain explicit operations.

The shareable aggregate contains all 440 paired cost summaries, all five differences per regime, workflow outcomes, source revisions and machine/dependency metadata. It supports checking the table arithmetic. The producing QUBOBenchmarks harness is private; this aggregate is not a claim that public users can independently rerun that private source. The retained private campaign includes raw archives, manifests, source hashes, failed preflights and archive-member digests.

First calls and algorithmic budgets

Package precompilation runs a small PrecompileTools workload for the budgeted entrypoints in Strong and Weak modes. The deterministic fixtures include a frustrated graph that reaches every scheduler family, a productive residual-SCC contraction, productive flow/Max-Cut/block/probing reductions, and both Float64 and exact-rational witness preparation. The largest fixture has 20 variables; no benchmark corpus, solver or optional consumer dependency is loaded.

This reduces compilation on common Float64-model call paths; it does not promise that every input storage type, label type or productive deduction compiles during installation. Measure package load, workspace construction and scheduling separately when first-call latency matters. For steady-state comparisons, warm representative productive paths before timing and report that policy. A warmed algorithmic budget and fresh-process wall time answer different questions.

The scheduler's time_limit still starts after workspace construction and always allows the cheap floor to finish. Precompilation does not interrupt compilation or garbage collection, make the soft cutoff a hard wall-time guarantee, or remove exact-arithmetic allocation costs on large models.

Repeated boundary updates

The current workspace uses persistent integer numerators on a shared dyadic grid. Construction, integer update preparation, promotion/exponent extension, public rational conversion and rational proof/flow preparation have separate costs. This backend has not yet been measured under the historical study below; the pinned results remain evidence for their recorded implementation. Fresh matched construction and corpus measurements are needed to quantify its effect.

Each comparison performs ten identical branch/propagate/rollback cycles. Reuse pays for one cold construction and first-growth allocations; the baseline pays for a fresh workspace and root propagation at each cycle. Both include returned fixing data and trace construction/hashing; the recorded traces are checked for equality outside timing. Five paired deterministic trials alternate execution order; labels 41–45 are not backend RNG seeds. Hot paths have discarded warmups. Detailed reports are disabled in this table.

Negative differences favor reuse. Allocation totals are cumulative bytes allocated, not peak resident memory. The range is the observed paired range, not a population confidence interval.

Sparse workloadnMedian complete time difference (ms)Paired range (ms)Median allocation difference (MiB)
local2,000-105.301-134.781 to -49.322-277.496
local10,000-3689.948-4735.476 to -2405.385-6391.492
short cascade2,000-112.136-125.885 to -99.962-279.002
short cascade10,000-3741.479-3789.532 to -2759.845-6397.117

All five paired time and allocation differences in each required regime favored reuse. Cold initialization is charged over exactly ten cycles. Recorded break-even estimates use average measured cycle costs; they are not observed first-crossover positions or guarantees for a different workload.

Median per-trial average warmed no-change allocation was 3,873.6 bytes per cycle at n=10, 2,000 and 10,000, including harness trace/output overhead. Median per-trial average latency was 7.1, 11.3 and 14.0 microseconds respectively. Separate CI guards check that no-change/local structural and allocation work does not scan unrelated graph data. Cold construction still scales with the graph. Dense degree, unavoidable cascades and non-LIFO rebuild/replay retain their full costs in the aggregate; they are not promised constant-time updates. Retained object size is measured outside timing and is not peak RSS.

Complete direct and decomposition workflows

The nine tiny fixtures use the same deterministic Rational exhaustive child, zero start, capacity and shared work allowances across each off/on comparison. Exact enumeration refuses more than eight compiled variables before exponentiation. Each run has 256 available assignment units, including a final original-state audit, and bounded structural work. Complete totals include construction/compilation, loading, solve, lifting and final original audit; independent reference enumeration and campaign/archive/export are recorded separately. Direct off uses identity preprocessing through the same adapter and therefore includes adapter overhead.

All 180 outcomes were retained: 135 completed jobs and 45 failed/refused jobs. Only successful incumbents enter statistical exports. A capacity refusal is not a speedup. On the split fixture, default-off decomposition stops at its capacity limit with an incumbent, while certified root reduction permits two small children and an original optimum. On no-reduction, near-fallback, constant and other fixtures, preprocessing can add cost; the aggregate retains those losses. Unsafe Float64 export is refused by the direct path, and decomposition can fall back to the original objective when its budget permits. Approximate/local results retain their qualified status.

Evidence limits and choices

The study used Julia 1.10.11 and one Julia/OpenBLAS thread, an 8 GiB address-space limit and a 600-second command limit. The full measurement took 320.474 seconds and retained about 8.72 MB of output. The first preflight exposed an archive restriction and projected excessive cost for 100 cycles; the protocol was amended and committed to ten cycles before measurement. A second status-recording failure and the successful third preflight remain separate historical attempts.

Producer revisions were QUBOPreprocessing 3aa2923a4b5fe6cda21fbb4e331e9198b603029f and QUBODecomposition 1a08bfbd14bca1eff6d6f80326ea5cc270282cec, with QUBODrivers 0.6.6, QUBOTools 0.16.2 and ToQUBO 0.7.1. Source-pinned candidate evidence does not prove General installation. The 0.1.0 runtime algorithm code matches the source used for this envelope; later release/documentation changes require their own publication checks.

Boundary observations include ordinary interval bounds, forced exact fallback, certificate production/verification, materialization and reconstruction. These observations have no separate per-boundary discarded warmup and may include first-call compilation. They are inclusive measurements, not steady-state isolated phase costs. Full verification/export remains an explicit consumer boundary; expensive automatic use on every hot update is deferred. Structural units omit arithmetic bit complexity and hashing/sorting overhead; elapsed times include them.

Choose reuse for the measured repeated sparse regime, preserve cheap off/bounded paths for one-shot callers, and measure a new consumer before changing its default. No universal speed ratio or noisy timing threshold is enforced by CI. For corpus-wide size reductions, $n = 2{,}000$ to $32{,}000$ scaling timings, and solution audits across 6,292 benchmark instances, see the QUBOLib, QOBLIB & MINLPLib corpus benchmarks.