Post-v1.0 audit-driven correctness pass + Phase 3 oracle diversity expansion. Together they close all silent-wrong-result paths surfaced by the post-release line-level MMA parity audit and open the project's five "we-didn't-think-of-this" oracle diversity axes.
Highlights
Audit table after v1.1.0: 86 🟢 verified / 0 🟡 unverified / 7 🔴 (6 fully fixed, 1 D5 deferred indefinitely) / 17 ⚪ not ported. Started at v1.0.0 with 65 🟢 / 21 🟡 / 6 🔴; the post-v1.0 pass closed every 🟡, fixed 6 of the 7 🔴, and surfaced D7 as the seventh — now fixed too.
Oracle suite: expanded from 12 (v1.0.0) to 20 cases. All 5 ROADMAP diversity axes have at least one parity-validated oracle:
| Axis | New oracle (v1.1.0) | C++ vs MMA |
|---|---|---|
| Loop number L=4 | banana_4loop_eps001 |
rel 7.2 × 10⁻³¹ |
| ≥3 kinematic invariants (electroweak) | ewbox_1loop_eps001 |
rel 3.8 × 10⁻³⁰ |
| Multi-cut Cutkosky (5-particle) | cutbanana_4L_eps001 |
rel 2.2 × 10⁻³⁰ |
| Mixed mass (some massive, some massless) | bn3mix_eps001 (3-loop) |
rel 2.9 × 10⁻³⁰ |
| ε extremes (D = 3 / eps = 10⁻⁴) | cutbubble_1L_eps2 / cutbubble_1L_eps10000 |
rel 1.9 × 10⁻⁶⁴ / 1.0 × 10⁻³² |
Notable architectural fix
D7 — dual-Kira-call master-count divergence (surfaced by the L=4 banana oracle). ibp::reduce and ibp::diffeq were restructured to mirror upstream's BlackBoxReduce / BlackBoxDiffeq two-call pattern: a Masters-mode preheat call (sector-wide enumeration via select_mandatory_recursively) followed by a Reduce-mode call (select_mandatory_list for the specific targets), both at the same (rank, dot). Each call runs in its own subdirectory because the Kira 2.x release refuses to share a $ReductionDirectory between the two calls.
See CHANGELOG.md, docs/AUDIT_MMA_PARITY.md, and docs/ROADMAP.md for the full picture.
Notes
- Upstream MMA AMFlow limitation surfaced at eps = 1 (D = 2): an initial ε-extremes oracle attempt at eps = 1 hit an upstream DESolver edge case where the boundary returns a partially-symbolic expression instead of a numeric. This is an upstream-package limitation, not a C++ divergence; the v1.1.0 ε-extremes oracle uses eps = 1/2 (D = 3) as the next-most-extreme rational that yields a clean numeric.
- D5 (complex-numeric kinematics) remains deferred indefinitely — see audit report for the architectural trade-off.
Attribution
This C++17 implementation re-implements the algorithms presented in the Mathematica package AMFlow by Xiao Liu and Yan-Qing Ma (arXiv:2201.11669, Comput. Phys. Commun. 283 (2023) 108565). All algorithmic credit belongs to the upstream authors.