Skip to content

ROADMAP.md

Latest commit

 

History

History
76 lines (57 loc) · 2.88 KB

ROADMAP.md

File metadata and controls

76 lines (57 loc) · 2.88 KB
Raw

Roadmap

中文版本: ROADMAP_zh.md

Forward-looking development plan. For per-release narrative see CHANGELOG.md; for MMA-parity status see docs/AUDIT_MMA_PARITY.md.

Active practice — diversity-driven oracle expansion

The MMA-parity audit is closed: every public symbol port has at least one oracle exercising it. Going forward the quality loop is oracle diversification — build new benchmarks along axes the existing suite under-covers, and fix anything that diverges from MMA inline.

The 228 committed oracle triplets currently cover:

Diversity axis Current coverage
Loop number L = 1, 2, 3, 4
Kinematic invariants 1 (s), 2 (s, t), 4 (s, t, mWsq, mZsq)
External legs 2, 3, 4, 5, 6 (pentagon 1L, pentagon 1L massive, pentabox 2L, hexagon 1L)
Mass configurations all-massless, all-equal, mixed at L = 1 / 2 / 3 / 4
Cutkosky cuts 2-, 3-, 4-, 5-particle cuts
ε extremes 1/2, 1/100, 1/1000, 1/10000
Sector size ≤ 8 propagators (pentabox 2L corner)

Target additions: higher-loop multi-invariant cases, larger sectors (≥ 10 propagators), and any topology a downstream user reports.

Each new oracle is roughly one day of work (construct family → MMA reference → C++ run → commit triplet). When a new oracle exposes a divergence, fix it inline; the bug becomes a parity-validated case going forward.

Out of scope — won't be implemented

Deliberate non-goals. The JSON entry-point rejects them with a clear error rather than silently mishandling.

  • SolveIntegralsGaugeLink, HQET / SCET / Wilson-line workflows.
  • IBP backends other than Kira — no FIRE / LiteRed / FiniteFlow / Blade.
  • Complex-valued numeric kinematics. The C++ algebra layer is over Q (FLINT fmpz_mpoly_q_t) and does not carry a complex coefficient ring. The JSON dispatcher rejects the object form {"re":..,"im":..} in amf_options.blackbox.numeric_values; provide only purely-real numeric values. See docs/FAQ.md and docs/AUDIT_MMA_PARITY.md.

Update policy

  • Each new oracle triplet (cpp.json + mma.wl + mma_reference.json) is added to run_perf_audit.sh rotation on landing.
  • When a new oracle exposes a real divergence, fix in the same session if practical and document the relevant assumption inline in code.
  • Do not silently drop items that turn out to be infeasible — document the reason where the user will see it (FAQ or AUDIT).

Performance benchmarking

Wall-clock benchmarking against upstream MMA AMFlow is intentionally deferred until after the project has completed broader correctness validation and the measurements can be run on a dedicated, load-isolated machine. When that pass is performed, the numbers will be added here.