Skip to content

Precision and execution validation records

Historical feature checkpoints from October 2026. These preserve the original configurations, test counts and comparisons; they are not a statement of the current suite size. See current capabilities and testing and CI for maintained instructions.

Validation records

The records below summarize validation at each original feature checkpoint. At those checkpoints, integration additionally ran the routine native, wide-precision, sanitizer and documentation CI suites on the combined trees.

  • Async execution describes scan, preparation and grid lifetime/failure contracts, including completion of all admitted work.
  • Parallel grids records independent scalar comparisons, precision coverage and benchmark conditions.
  • MPFR builds describes provider requirements. Native, wide and sanitizer CI exercise different configurations.
  • Thermodynamic XXX CLI tests compare complete serial, threaded and non-retained rows, including grids larger than the admission window, mixed successful and failed points, threshold counts, native precisions and output preservation.

The MPFR checkpoints #9, #13, #15, #16, #18, #19, #20, #21, #22, #23, #25 and #26 verify precision growth against independent references beyond binary128, typed endpoint/failure values, CLI export/provider contracts and ownership in composed Async calculations. The scheduling checkpoints #11, #12, #14 and #27 compare complete results against serial/baseline calculations and document measured wall/CPU and memory behavior. Finite Hubbard's scalar/LU work is independent of the complex Gamma-ratio functions used by thermodynamic Hubbard.

The composition audit in #17 adds finite Lieb–Liniger owning scan APIs and a compiled Richardson/central-spin sweep. Deferred input destruction, downstream consumers, one/multiple workers, failed solves, cancellation and independent MPFR precisions are covered. Existing XXX scan and structure-factor composition tests were rerun; their APIs need no new wrapper. No new precision support or parallel speedup is claimed by that audit.

Gaudin–Yang's #18 verifies the unreduced complex Bethe equations and two-body accuracy growth at 128/256/333 bits. Free states, zero-update diagnostics and Newton updates work with MPFR scalars, using Uni20 CPU LU when MPLAPACK is unavailable. Mixed interacting populations remain odd, and continuation within each independent solve remains sequential.

Bose–Fermi's #19 validates mixed-species equations, Lieb–Liniger reductions, free states and missing-energy failures. The shared continuum Newton driver uses normal Uni20 dispatch, including MPFR CPU LU. Tests resolve a weak-coupling shift above a nonzero Fermi sea beyond binary128 resolution and compare complete states under owning, mixed-precision grid composition. Mixed interacting shells still require odd fermion number.

Haldane–Shastry and Sutherland's #21 inherit #20's scalar CLI dispatcher. The spin model keeps exact motif ordering; the gas keeps exact MPFR integer gap coefficients until physical scaling. Full MPFR tests run without MPLAPACK, including differential-Hamiltonian references, large-exponent small gaps, precision growth, budget refusals and independent owning Async spectra. Native ED/wavefunction checks and 561 unchanged exported rows validate the existing formats. Individual analytic CLI calculations remain serial.

TB dispersion's #22 uses the same scalar dispatcher from #20 and extends the common sine/massive-band helper rather than duplicating kinematics. Independent dressed-sea and continuum references, scalar-only MPFR physics, positive momenta at the MPFR exponent floor and owning mixed-precision queries pass. Existing TB/XXZ/XYZ checks and 1,710 identical native output rows validate shared-band consumers. That checkpoint did not enable XXZ/XYZ's elliptic models in MPFR; the current inventory records subsequent model-level support.

Sine-Gordon particle dispersion's #23 also uses #20's scalar dispatcher. Independent algebraic breather masses, precision-sensitive stable-species selection, unequal-mass threshold minimization and full owning Async queries pass at 128/256/333 bits. Full MPFR physics runs in scalar-only sanitizers; 1,095 native/binary128 output rows match the baseline. The scattering phase, Bethe–Yang and finite-volume integral equations required separate precision checkpoints; their current coverage is listed in the inventory.

XYZ's #24 prepares the bound coefficient once per branch and reuses it for both parity copies. Independent elliptic/limiting references, owning Async readers and failed preparations/queries pass in native/wide/sanitizer builds. End-to-end binary128 curve time falls from 5.42 to 0.27 seconds with identical output; an independent 2,097,664-energy parameter sweep falls from 4.480 to 0.585 seconds with eight workers. These are distinct reuse and scheduling measurements; the CLI stays serial. Commands and limits are in the PR's docs/parallel-grids.md. That scheduling checkpoint did not extend the theta helper to MPFR.

Merged Uni20 PR #103 supplies scaled zero-argument rectangular theta constants at native and explicit MPFR precision. Independent product/Gamma references demonstrate accuracy growth through 1,024 bits; native/wide/sanitizer checks and upstream CI passed at that checkpoint. Bethe's XXZ dispersion uses these constants.

XYZ excitation bands now use the shared complex theta values and derivatives in Uni20 #115, replacing Bethe's duplicate evaluator. Explicit MPFR reaches couplings, spinon/bound bands, continuum edges and owning Async readers. The finite XYZ nonlinear solve also carries explicit precision through its controls, complex roots and Newton arithmetic, with independent spin-Hamiltonian and XY references beyond binary128.

Merged Uni20 PR #104 supplies array-valued tanh-sinh quadrature with explicit precision in endpoint distances, accumulation and recoverable error diagnostics. Its validation includes reflected/translated singularities, cutoff sensitivity and accuracy growth through 1,024 bits. The error estimates are not certified bounds. Each Bethe consumer still needs model-specific kernel, tolerance and native-output checks; the current inventory and individual model pages record implemented support.

Update the current inventory with each integrated app checkpoint. A proposed capability or a pending audit is not evidence of implemented support.

Kondo's #29 adds inclusive signed field scans and streams complete rows through the existing bounded grid. Every query owns its series/quadrature scratch; owning deferred/threaded composition and failed responses retain all diagnostics. Native/wide/sanitizer checks pass 15/21/15; 30 single-point cases preserve merged-main outputs. Substantial high-field curves scale 7.6–7.7x at eight workers, including a 65-point binary128 curve taking 74.02 s direct versus 9.64 s parallel. Cheap low-field curves saturate earlier; timings and CPU costs are in that PR's docs/parallel-grids.md. Kondo's Gamma amplitude is positive-real, not complex. MPFR requires precision-aware asymptotic order/shift, series-tail and quadrature error control, as well as explicit constants and reductions; scheduling does not remove that numerical work.

Q-boson's #25 retains the existing continuous scanner and compares complete states, gaps and failed references in serial/threaded execution. Independent occupation-Hamiltonian energies and the unreduced Bethe equations validate MPFR, including a weak-deformation correction beyond binary128 and positive energies at eta=1e-7000. Analytic limits, zero-update diagnostics and actual Newton updates work with MPFR scalars, including Uni20 CPU LU. No new scheduling mechanism is needed.

Quantum-group XXZ's #26 validates both the regular critical real-root family and complete free-endpoint sectors with MPFR scalars alone. Independent spin-Hamiltonian eigenvalues and complex reflection equations verify precision growth at 128/256/333 bits. All 45 states of a 16-site family, including partial and failed scans, agree under serial/threaded execution. Owning regular/free queries also pass deferred and threaded composition tests. Native output comparisons retain 681 identical rows; no massive or new complex-root family is introduced.

Sine-Gordon Bethe–Yang's #27 uses the existing bounded grid for independent interacting levels. One worker, a single level and the free point stay direct and serial. Native/wide/sanitizer CLI checks preserve all fields, partially converged scans and ordered output. Matched baseline comparisons retain 495 selected rows and 2,066 benchmark rows. Eight-worker wall times improve from 6.33 to 0.85 seconds for 2,049 binary64 levels and from 30.82 to 5.71 seconds for 17 binary128 levels. CPU/memory costs, a roughly 4% one-worker binary128 overhead and measurement limits are recorded in the PR's docs/parallel-grids.md. Numerical kernels and precision support are unchanged.

Potts's #30 retains one vacuum solve and reuses each charged-momentum result for both conjugate charge rows. The existing bounded grid streams ordered rows, including missing-reference and iteration failures; default one-worker and selected single-state requests stay serial. Complete owning states match under deferred/one/four-worker execution. Native/wide/sanitizer checks, 33 matched-baseline cases and tutorial exports pass. All 8,100 timed rows match baseline output; eight-worker whole-curve speedups are 7.4x in binary64/binary80 and 6.7x in binary128. CPU/memory costs and shared-workstation limits are documented in the PR's docs/parallel-grids.md. The underlying real-root XXZ equations and precision support are unchanged.

ASEP's #31 extracts the existing rate solver around an owning PreparedGap. One TASEP seed feeds independent rate queries through the shared Async preparation/grid APIs. Full fields, roots, failure budgets and independent analytic endpoints agree in direct, deferred and threaded checks. Native/wide/ sanitizer checks pass 29/39/29; 60 matched-main cases preserve frontend outputs and physical/control metadata. All 38,070 complete benchmark state rows match. Seed reuse removes 24–32% of serial work; prepared eight-worker curves achieve 8.9–11.0x combined speedup over independent direct solves. The example and docs/parallel-grids.md in that PR record separate reuse/concurrency costs, output retention and shared seed-counter semantics. The physical branch, single-request analytic paths and precision support are unchanged.

Checkpoint #32 validates owning TASEP and finite XYZ requests, including successful and failed states, deferred/one/multiple workers, and independent error handling when both models share a scheduler. Substantial filling/eta curves scale 7.3–7.6x with eight workers across native fixed precisions; all complete output states match direct calls. Scalar kernels and single-request CLIs are unchanged. Their MPFR paths remain unimplemented; the complex-function and Newton audit is separate from this measured concurrency checkpoint.

Checkpoint #33 validates owning TB, SU(N) fermion and t–J requests with a compiled example and complete-state comparisons under deferred and one/four-worker scheduling. Native/wide/sanitizer checks pass 57/79/57. All 90,888 timed state rows match direct calls; independent-request sweeps scale 6.4–7.5x at eight workers, with documented CPU/memory costs. The binary64 t–J benchmark includes two numerical failures in every variant. A larger 512-site curve has 12 failures; all states and diagnostics match direct/eight-worker execution, and a bounded higher-precision probe does not resolve the representative dilute failure. This convergence limitation is separate from Async correctness. Model kernels, frontends and precision support are unchanged; no new state family is claimed.

Checkpoint #34 adds ladder::PreparedSectors<Real>: one full-sector scan owns its ordered solved/failing branch prefix and later field/rung queries replay the same selection rules without new Newton solves. Whole-scan budgets, partial candidates, analytic product bypasses and exact tie/large-shift handling are preserved. Copies share immutable storage; dependent Async grids outlive the preparation handle. Native/wide/sanitizer checks pass 51/67/52, including the independent ladder oracles and shared-grid held-query lifetime checks. Matched-main comparisons retain 60 requests / 1,491 identical rows; all 86,088 benchmark state rows converge and match direct calls. Reuse is the primary gain (roughly 90–270x on these many-parameter curves); subsequent query parallelism saves only hundredths of a second. Fixed-sector scans and native precision support remain unchanged. Catalogue/output retention costs are documented.

Checkpoint #35 extends the biquadratic frontend's --threads to six ferromagnetic bound-cluster families and AF module minima. A shared CLI adapter owns its sequential label generator and per-candidate requests, uses existing continuous runners, and waits on the final Async result before ranking or output. Numerical families, budgets, ordering and first-failure rows are preserved; all candidates are still retained before ranking. Native/wide/sanitizer checks pass 18/18/18, including held-peer failure completion and complete CLI comparisons. Fresh-main comparison covers 135 requests / 3,192 rows; standalone long-double bound-pair fields have documented rounding-scale differences, while worker-count outputs match exactly. All 45 substantial benchmark runs converge and their 6,309,585 exported rows match main and serial output. Whole-frontend eight-worker scaling is 5.9–7.4x, with unchanged serial cost. Native precision support is unchanged; no MPFR claim is made. Single-state and dependent Q-system/singlet modes remain serial; cheap scans do not justify changing the one-worker default.

Bethe–Yang precision

The same-charge sine-Gordon pair solver and its real phase/complex Fourier kernel carry explicit MPFR precision through constants, quadrature, tail bounds and failed-state diagnostics. These paths use scalar arithmetic without MPLAPACK. The existing bounded level grid also accepts owning requests at different precisions. The vacuum NLIE also supports explicit MPFR/MPC, including independently checked Fourier tables, free-Dirac energies and interacting D3 reference comparisons. Both vacuum and selected excited NLIE frontends retain explicit precision in all controls and success/failure diagnostics. Source-table and free-Dirac level checks reach 333 bits. Arithmetic precision and Fourier, hole, nonlinear, mesh and cutoff budgets remain independent; more bits can require more work.