Configuration Reference
LeanCert has two configuration families because semantic proving and the legacy numerical toolkit expose different contracts.
ProveConfig
import leancert as lc
config = lc.ProveConfig(
taylor_depth=14,
system_root=lc.SystemRootConfig(
max_iterations=12,
max_dimension=4,
precision_bits=28,
),
eventual=lc.EventualConfig(max_checks=2000),
integral=lc.IntegralConfig(
start_partitions=16,
max_partitions=4096,
),
)
| Field | Default | Meaning |
|---|---|---|
taylor_depth |
10 |
non-negative checker effort for supported operations |
system_root.max_iterations |
8 |
Krawczyk candidate-search refinements |
system_root.max_dimension |
4 |
caller ceiling; cannot exceed Bridge capability |
system_root.precision_bits |
20 |
candidate rationalization/search precision |
system_root.candidate |
None |
optional explicit KrawczykCandidate |
eventual.max_checks |
1000 |
cutoff-search check budget |
integral.start_partitions |
32 |
initial untrusted uniform-partition candidate |
integral.max_partitions |
4096 |
maximum partition count considered by discovery |
These settings control effort and candidate search. They do not let a caller invent a checker capability the Bridge did not advertise.
Toolkit Config
Config controls legacy Solver operations: taylor_depth, max_iters,
tolerance, use_monotonicity, timeout_sec, backend, backend-specific
configuration, racing, incremental refinement, target bound, and timeout.
Presets include:
Config.low_precision(),medium_precision(),high_precision();Config.dyadic(),dyadic_fast(),dyadic_high_precision(); andConfig.affine(),affine_compact().
DyadicConfig exposes precision and the compatibility round_after_ops
field, with ieee_double(), high_precision(), and fast() presets.
AffineConfig exposes max_noise_symbols, with default() and compact().
Exact input boundary
ProveConfig never changes the rule that semantic values must be exact.
Toolkit configuration may accept floats for heuristic tolerances or search
inputs; that does not make those floats part of an exact semantic claim.