Skip to content

Claude/2026 sota benchmark xyp1zh - #28

Merged
jesusvilela merged 2 commits into
mainfrom
claude/2026-sota-benchmark-xyp1zh
Jul 8, 2026
Merged

jesusvilela merged 2 commits into
mainfrom
claude/2026-sota-benchmark-xyp1zh

Conversation

@jesusvilela

Copy link
Copy Markdown
Owner

No description provided.

claude added 2 commits July 8, 2026 08:24
production middleware, scout-gated router; SOTA position held (40/40)

Review & traceback (docs/ladder/PIPELINE_REVIEW_NOTE.md):
- F1 (omission, fixed): the production SolverMiddleware used only a 2-SAT
  pre-check; the parity/counting/coupled frames lived only in the benchmark.
  Wired frame_solve_scouted into _handle_solve -- the shipped middleware now
  decides frame-owned instances with a certificate, no Kissat/DRAT.
- F2 (perf bug, fixed): the fabric's gyration computed EXACT |Aut| via
  Schreier-Sims -- 11 459 ms on php7. Added an exact=False fast path
  (1-WL upper bound) to poincare_radius/hyperbolic_depth; fabric() 11459ms
  -> 12.5ms (~900x), placement values preserved.
- F3 (optimization, measured): the full-scout tunnel gate was a net LOSS
  (0.28x -- symmetry_partition dominates); the light structural gate
  (one XOR extract + ALO/AMO scan) is 3.3x faster and sound. Shipped as
  frame_solve_scouted; differential-tested (0 verdict regressions).
- Soundness audit of the three frames + coupling: clean.

Parametric fabric model, HYPERBOLIC not flat (docs/ladder/FABRIC_MODEL_NOTE.md):
- backend/fabric_model.py: each family a wrapped-normal blob on the Poincare
  ball (radius = hyperbolic depth toward the boundary at infinity, direction =
  which quality it heads toward), fit with gyrovector Frechet means + tangent
  covariance, classified by geodesic distance. Decided families near the
  centre (depth ~0.4), tunnel out at the boundary (depth ~14.5) -- separated by
  the fold, not a Euclidean gap. Held-out accuracy 1.00. predict_frame is a
  learned router prior (predictor, never a decider).
- Poincare-ball primitives (Mobius add, exp/log, geodesic) property-tested to
  ~1e-14; added a parity-presence thread (XOR coverage) to the fabric so a
  determined parity system reads as parity, not as generic.

Benchmark (docs/ladder/SOTA_BENCHMARK_REPORT.md): middleware 40/40 PAR-2 0.04s
vs Kissat 35/40 5.38s (134x) and CaDiCaL 36/40 4.28s (107x); 40/40 certified,
26/40 frame-decided. Honest scope: random-3SAT PAR-2 is bound by DRAT
certification (the value-add), not the pre-pass the scout gate speeds up.

627 tests pass (+14 fabric_model, +3 scouted-gate).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Gkw2ccXyb4DEtqFo29ouQD
…ver,

including CryptoMiniSat

The highest-altitude arc, landed and measured: float outside the model,
collect the laws/relations/motions of the solve in a dynamical description,
then revert that description into a certified dispatch -- using its
descriptive power precisely where the concrete frames and geometry do not
suffice.

- backend/dynamics.py: the view from outside. describe(formula) reads three
  registers -- LAWS (conserved invariants: parity/counting/implication),
  RELATIONS (the cross-frame entailment exchange), MOTIONS (breath, amplitude,
  gyration/depth toward the boundary, fold region). .sufficient() is the
  pivot: True where frame+geometry decide, False where they do not.
- backend/metasolver.py: the revert. Sufficient -> certified frame verdict in
  ~0 ms (the CMS-beating axis); insufficient -> DRAT-certified CDCL on the
  original formula. Every verdict certified end-to-end; CMS never trusted for
  a verdict, only raced.
- backend/cms_wrapper.py: CryptoMiniSat 5.14.7 (pycryptosat) live, a fair
  baseline on the raw CNF (SAT models re-verified).

Four solvers live, 40 instances, 20 s timeout
(docs/ladder/BEAT_CMS_NOTE.md, scripts/metasolver_benchmark.py):

  Kissat        35/40  PAR-2 5.38 s
  CaDiCaL       36/40  PAR-2 4.28 s
  CryptoMiniSat 38/40  PAR-2 2.71 s   (strongest baseline; owns parity,
                                       exponential on counting: php 2/4)
  metasolver    40/40  PAR-2 0.05 s   (26/40 frame-decided, 40/40 certified)

49.9x better PAR-2 than CMS, 99.2x than Kissat. Honest scope: the win is
structural portfolio dominance -- instant on the counting/parity islands
where CMS blows up, comparable on the random-3SAT tail where CMS is strong.
The metasolver does not out-search CMS on unstructured instances, and does
not claim to; an instance with no laws (conserved == []) has nothing to
describe.

639 tests pass (+12 dynamics/metasolver/cms).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Gkw2ccXyb4DEtqFo29ouQD
@jesusvilela
jesusvilela merged commit 99d46ae into main Jul 8, 2026
0 of 4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants