Skip to content

Claude/2026 sota benchmark xyp1zh - #29

Merged
jesusvilela merged 3 commits into
mainfrom
claude/2026-sota-benchmark-xyp1zh
Jul 8, 2026
Merged

jesusvilela merged 3 commits into
mainfrom
claude/2026-sota-benchmark-xyp1zh

Conversation

@jesusvilela

Copy link
Copy Markdown
Owner

No description provided.

claude added 3 commits July 8, 2026 09:30
minimax mixed strategy (game-theoretic self-reflection)

Self-reflection (docs/ladder/GAME_THEORY_NOTE.md): the moving-frame regime is
a zero-sum game, Solver vs Nature. The decidability FOLD is the boundary
between two solution concepts -- on the island a dominant pure strategy
exists (the owning frame, ~0 ms); on the tunnel none does (heavy-tailed
runtime), so the equilibrium is MIXED. Von Neumann minimax + Gomes-Selman
heavy tails explain the win, and bound it honestly (diminishing returns,
No-Free-Lunch = minimax regret, k x CPU cost).

backend/metametasolver.py: the fractal, n-caged portfolio. Frames-if-
sufficient (instant certified island), else a PARALLEL seed-diversified CDCL
portfolio (Kissat x breadth seeds + CaDiCaL, first certified arm wins, losers
killed). No CMS in the pool -- a win is a genuine out-search. Certified
winner (DRAT/model); k x CPU cost reported.

Measured (docs/ladder/scripts/metametasolver_benchmark.py), hard random-3SAT
n=220-280, all at 30 s:
  CryptoMiniSat (1 core)   14/16  PAR-2 10.01 s
  fractal portfolio (~9c)  15/16  PAR-2  6.22 s   beats CMS 13/16, 1.6x
Large wins on the heavy-tailed instances (n260s3 5.83 s vs CMS 16.69 s).
Honest caveats, all measured: loses trivial instances to launch overhead;
does not beat a single strong Kissat on core-limited hardware; some tails
(n280s1) are unescapable. The operator's prediction -- that a fractal
portfolio out-searches CMS on random-3SAT -- confirmed, by playing the mixed
strategy a single CMS thread cannot.

("ACAF" was not identifiable as a named SAT method; the multi-armed,
multi-distribution mixed strategy IS the multi-omni realization -- noted, not
fabricated.)

644 tests pass (+5 metametasolver).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Gkw2ccXyb4DEtqFo29ouQD
universal game solver

Answers the two standing gaps (trivial-instance overhead; only tying the
strongest single engine) with an adaptive policy, and the conceptual
question -- what is our lambda logic as a universal game solver.

Lambda logic (docs/ladder/ACAF_NOTE.md): the actor-critic loop is a fixpoint
(Bellman = the Y combinator), already in the repo as lambda_sat.fixpoint_sat.
Lambda is the universal policy language and the zero-overhead in-process
kernel; the native engines are the fast moves it dispatches. Universality
without efficiency: it expresses every strategy, executes the cheap ones.

backend/acaf.py -- four organs:
- CRITIC: dynamics.describe + a cheap hardness proxy (does frame+geometry
  suffice; how heavy the tail).
- ACTOR: staged policy -- frame (in-process, instant) -> single arm (no swarm
  overhead on trivial) -> cores-sized diversified portfolio (heavy tail).
- AMBIGATOR: polysemy/hardness sets breadth = round(1 + hardness*(cores-1)).
- FUZZER: emits decorrelated engine+seed configs.

Winning trivial, measured: the actor's single-arm stage removes the swarm
overhead; the residual gap to CMS is the CERTIFICATE, not the search
(trivial UNSAT: solve 54ms, DRAT-certify 117ms -- proof is 2-3x the solve).
Lambda/frames win structured trivial outright with a constructive certificate
(php12 0.03s certified vs CMS 30s). We pay proof only where CDCL is forced.

Besting the strongest, measured: sizing the mixed strategy to nproc (not a
fixed 9) beats single Kissat 5/5 on the hard band (k=4: 34.0s vs k=1: 37.0s);
the naive portfolio had merely oversubscribed the cores.

650 tests pass (+6 acaf).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Gkw2ccXyb4DEtqFo29ouQD
Rewrite the front door in a spare, modern voice (Japanese section markers,
negative space) with three mermaid diagrams exposing the real architecture:
the ACAF staged solve path, the island/tunnel decidability fold, and the
fractal stack (ACAF -> metasolver/meta-metasolver -> frame router + certified
CDCL -> observer). Bind every claim to the shipped modules and their measured
results; foreground the discipline (verify, don't trust; Kissat not in the
TCB). Update the module map and quick-start to the current surface.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Gkw2ccXyb4DEtqFo29ouQD
@jesusvilela
jesusvilela merged commit 0f77851 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