eml-rewrite 0.1.0a0

F-family fusion pattern rewriter for symbolic expressions and Python source.

eml-rewrite

Pre-release. Patent pending. Not for redistribution.

F-family fusion pattern rewriter for symbolic expressions. Detects equivalent rewrites that strictly reduce predicted EML cost (per eml-cost). Provable non-regression: every proposed rewrite either improves the cost or is filtered out.

Status

Pre-release. Covered by Monogate Research patent applications #11 and #12 (filed 2026-04-25). Public release follows post-prosecution licensing.

Library API

from eml_rewrite import suggest, best, score, Suggestion
import sympy as sp

x = sp.Symbol("x", real=True)

# Find improving rewrites (provably non-regressing)
sugg = suggest(sp.exp(x) / (1 + sp.exp(x)))
# [Suggestion(pattern_name="sigmoid",
#             rewritten=1/(1 + exp(-x)),
#             score_before=3, score_after=2, reduction=1)]

# Pick the lowest-score rewrite, or original if no improvement
best(sp.sinh(x) / sp.cosh(x))
# tanh(x)

best(sp.exp(sp.exp(x)))
# exp(exp(x))   ← original; nothing in library matches

Pattern library (9)

1. exp(x)/(1+exp(x)) → 1/(1+exp(-x)) — sigmoid canonicalization
2. sinh(x)/cosh(x) → tanh(x) — F-family fusion
3. (exp(x)+exp(-x))/2 → cosh(x) — hyperbolic fusion
4. (exp(x)-exp(-x))/2 → sinh(x) — hyperbolic fusion
5. sin(x)² + cos(x)² → 1 — Pythagorean identity
6. cosh(x)² - sinh(x)² → 1 — hyperbolic identity
7. exp(log(x)), log(exp(x)) → x — inverse pair
8. log(a/b) → log(a) - log(b) — when score improves
9. log(x^n) → n·log(x) — when score improves

Command line

eml-rewrite scan FILE [FILE ...]    # report rewrites without applying
eml-rewrite fix  FILE [FILE ...]    # show rewrites that would be applied
eml-rewrite analyze "exp(sin(x))"   # full Pfaffian profile + suggestions

Example:

$ eml-rewrite analyze "exp(x)/(1 + exp(x))"
Expression:           exp(x)/(1 + exp(x))
  pfaffian_r:           1
  max_path_r:           1
  eml_depth:            2
  structural_overhead:  3
  corrections:          Corrections(c_osc=0, c_composite=0, delta_fused=0)
  predicted_depth:      4
  is_pfaffian_not_eml:  False

Suggested rewrites (1):
  [sigmoid] 1/(1 + exp(-x))  (-1 cost units)

Provable non-regression

The library's headline guarantee:

for expr in any_collection:
    for s in suggest(expr, only_improvements=True):
        assert s.score_after < s.score_before  # always

Every rewrite returned with only_improvements=True strictly reduces the EML predicted depth. The test suite validates this on a hand-checked panel; the engine filters non-improvements before returning them.

Installation (pre-release, local only)

pip install -e /path/to/eml-rewrite-pkg

Requires eml-cost (also pre-release).

License

PROPRIETARY-PRE-RELEASE. See LICENSE.