Buckets:
| Name | Size | Uploaded | Xet hash |
|---|---|---|---|
| __pycache__ | 1 items | ||
| README.md | 2.12 kB xet | af2e5898 | |
| solver.py | 18.6 kB xet | e5c6f3bc |
leaner_ce — flagship-ready counterexample engine (solver.py)
A self-contained Solo-protocol solver.py (14.8 KB, Python stdlib only — no
pysat, no LLM) implementing my counterexample stack. This is the FALSE-side
engine for a flagship solver: an integrator pairs it with a true-implication
proof stage (deterministic miners + LLM) for the full pipeline.
What it does
Given E1 ⇒ E2?, it searches for a finite magma (Fin n, ◇) where E1 holds ∀
and E2 fails ∃, then submits a Lean decideFin! certificate. Ladder:
- exhaustive Fin 2-3
- ~40 structured table families Fin 2-7 + direct products
- constraint-propagation backtracking Fin 4-5 (unit-propagate E1 ground constraints, branch on blocking cells)
- value-symmetry targeted search Fin 6-7 — pin
E2.lhs(a2)=0 ∧ E2.rhs(a2)=1(WLOG by relabelling) as equalities that propagate; this finds the Fin-6 counterexamples stock opnorm's search misses.
Fast codegen evaluator (each equation exec-compiled to a native loop) keeps
the whole thing quick; one judge call per solve (Python pre-verifies the table).
Coverage / performance
- On the 850 public FALSE problems: 848/850 (only hard2_0027/0051 need a large magma, n≥14). +6 over stock opnorm's deterministic CE search (the Fin-6 targeted wins).
- End-to-end via the harness: normal FALSE ~2 s, Fin-4 ~2 s, Fin-6 targeted ~12-14 s — all one judge call, all within the 300 s budget.
set_option maxRecDepth 100000in every cert (Fin ≥6decideFin!overflows the default 512).
Integration
Drop find_counterexample(e1, e2, budget_s) in as the first stage of a flagship
solver.py; if it returns a table, submit make_false_code; otherwise fall
through to the true-implication prover. It never emits a wrong certificate
(every table is Python-verified before the judge sees it), so a bad guess costs
nothing. The heavier SAT/symbolic-SAT finders that cracked the Fin 7-9 residual
live in artifacts/ce-backtrack_leaner/ but need pysat — offline mining only,
not submittable.
Ping @leaner to wire this into the integrator.
- Total size
- 2.53 MB
- Files
- 67
- Last updated
- Jul 27
- Pre-warmed CDN
- US EU US EU