Skip to the content.

ISAR: Invariant Kernel for Closed Computational Dialects

ISAR is a Lean 4 formalization of a four-operator combinatory calculus (norm, konst, comp, dup) and its observational quotient, the Invariant Layer (InvariantLayer := ISKSubtype / OperEq). The central result is morphism_uniqueness: the Invariant Layer is the terminal object in the category of admissible semantic kernels, so every closed dialect — lambda calculus, term rewriting, stack bytecode, hereditarily finite sets, linear interaction nets — admits a unique structure-preserving morphism into it. All theorems are machine-checked in Lean 4 with no sorry.


Project


Four Verified Results

1. Confluence and Normal Forms

The ITerm calculus with operators norm, konst, comp, dup is proven confluent: every term has at most one normal form. S is constructively derived from the basis. A bracket-abstraction compiler maps de Bruijn lambda terms to ITerm and is proven to preserve beta-reduction steps.

2. Terminality of the Invariant Layer

InvariantLayer := ISKSubtype / OperEq is proven to be the terminal object in the category of admissible semantic kernels (morphism_uniqueness). Lambda, TRS, bytecode, HF-set, and quantity-kernel interpretations each yield a unique factorization morphism. The layer is not postulated as a universal structure; it is derived as the quotient of a concrete calculus and proven terminal within a formally stated category.

3. Futamura projections (mix formulation)

Substitution soundness (futamura_first) is proved at the meta specializer. Object-level 2nd/3rd projections are mix instantiations over PESetup (specTerm + selfApp). TrivialPE shows mix+selfApp alone need not optimize; fragment OptimizingPE (identity/konstβ folds) proves Nontrivial. Full Jones–Gomard–Sestoft BTA for all of ISAR remains open.

4. Linear Reduction, Matrix Geometry, Named Completion Axioms

LinearIKTerm admits a bounded-fuel normalization certificate (sufficient_fuel_correct), structurally isomorphic to HVM2 interaction net reduction. The $4 \times 4$ ISAR operator matrices satisfy $I^2 = I$ (idempotency) and $(I \cdot R \cdot A \cdot S)^2 = 0$ (nilpotency). Gauge equivalence $P K_1 P^{-1} = K_2$ is proven, unifying two matrix representations of the same kernel. ISARApproximation.lean keeps ISAR_UAT and completion/embedding as named axioms; there is no metric on KernelAddress and UAT is not proved.


Module Status

Lean Module Content Status
ISAR.lean Syntax, reduction, confluence, basis completeness Verified
InvariantLayer.lean Quotient construction, linear fragment, fuel certificate Verified
LambdaFragment.lean Bracket abstraction, beta-simulation Verified
TensorSemantics.lean Tensor denotation, compositionality, separation Verified
KernelCategory.lean Terminal object, morphism_uniqueness Verified
HFSet.lean, ZFCInterpretation.lean HF set axioms, Ackermann bijection, HF kernel morphism Verified
TRSView.lean, BytecodeView.lean, QuantityKernel.lean Dialect morphisms, universal factorization Verified
ViewIndependence.lean, ViewUnification.lean Syntax-independence theorem, isomorphism unification Verified
ReverseRosetta.lean Forward invariance, referential openness Verified
Futamura.lean Three Futamura projections Verified
ISARMatrices.lean Idempotency, nilpotency, gauge equivalence Verified
ISARApproximation.lean Named ISAR_UAT + completion axioms (frozen; no metric) Named axioms

Build Locally

git clone https://github.com/cypoe/isar-proofs.git
cd isar-proofs
lake build

Blueprint PDFs (run twice each for cross-references):

latexmk -pdflatex=pdflatex -pdf blueprint/src/print_monograph.tex
latexmk -pdflatex=pdflatex -pdf blueprint/src/print_paper_a.tex
latexmk -pdflatex=pdflatex -pdf blueprint/src/print_paper_b.tex
latexmk -pdflatex=pdflatex -pdf blueprint/src/print_paper_c.tex

HTML blueprint:

plastex -c blueprint/src/plastex.cfg blueprint/src/web.tex

Check all blueprint declarations against Lean source:

lake exe checkdecls blueprint/lean_decls