SlimeCOBOL — COBOL → Java/Rust bit-exact transpiler

Headline numbers

Validation corpus — the largest publicly-disclosed COBOL benchmark we are aware of in Japan

Note: the 99.9995% bit-exact figure is measured on the NIST CCVS85 + Portfolio Management subset (501 programs / ~354K LOC). The 2,876 programs / 1.15M LOC figures below describe the total validation corpus; they are not the denominator of the 99.9995% measurement (the additional volume is used for dialect verification, sales demos, and silent-pass true-value measurement).

Major sources include:

• NIST CCVS85 459 progs / 346K LOC — the US Federal compiler validation suite (FIPS 21-3)
• AWS Mainframe Modernization carddemo 106 progs / 41K LOC — Apache 2.0, official AWS demo
• IBM CBSA (CICS Banking) 66 progs / 27K LOC — EPL 2.0, IBM’s official banking sample
• IBM db2-samples cobol_mf 22 progs / 3.5K LOC — Apache 2.0, DB2 EXEC SQL
• Defects Suite (OpenCBS) 43 progs — MIT, real-world defects benchmark
• Rocket BankDemo 164 progs / 35K LOC — commercial banking demo
• VAULT-CBS 73 progs / 32K LOC — production-grade core banking
• UUSIG Public Domain archive of Unisys Users Group (Burroughs MCP)
• Burroughs B5500-software 13 progs — actual 1960s mainframe sources

Comparison vs. vanilla COBOL compilers

When the same corpus is syntax-checked with GnuCOBOL 3.1.2, the value gap is quantitative:

Target / processor	Acceptance
GnuCOBOL -std=cobol2014 (pure ISO) on Defects Suite	35 %
GnuCOBOL -std=ibm / -std=mf on Defects Suite	75 %
GnuCOBOL on enterprise COBOL with EXEC CICS / EXEC SQL (CBSA / db2-samples / carddemo)	0–6 % (preprocessor missing → rejected)
SlimeNENC (same domain, measured on 501 NIST + Portfolio)	99.9995 %

Observation: vanilla GnuCOBOL accepts only 0–6 % of enterprise COBOL with EXEC CICS / EXEC SQL because no preprocessor is available. SlimeNENC processes the same domain at 99.9995 % and produces bit-exact transpiled output. The three-tier value staircase — ISO 35 % / IBM-MF 75 % / SlimeNENC 99.9995 % — is itself a rare feat of measurement.

Supported target languages

A single Slot IR (intermediate representation) feeds a per-language Stage 6 emitter. Stages S1–S5 and S7–S8 are language-agnostic and fully shared; adding a new target language requires only one new emitter file (Rust, Go, C#, and Kotlin emitters were each ported in roughly 30 minutes).

For Java specifically, we have validated cross-version transpilation across 8 ↔ 17 ↔ 21 (bytecode major versions 52 / 61 / 65), producing physical evidence of JVM compatibility across a 13-year version gap.

Pipeline (9 stages)

Each stage corresponds to an independent claim in the patent, and each behaves deterministically. On failure, Mini-PSV (Stage 8) and the audit chain (Stage 7) preserve the evidence.

Audit suitability

• Bit-exactFor identical input, the transpiled code's computed output matches the source down to the last bit. COBOL features such as COMP-3, SIZE ERROR, GIVING, and multi-target are tuned per dialect (14 values detected, 18/18 dialect-verified bit-exact runs primarily on IBM / Unisys / MF families).
• Round-tripRe-converting the output back to the original language produces a token-by-token match (90/90 cases). This is a formal certificate that conversion did not discard information, and it is something an auditor can verify back-and-forth.
• Audit chainEvery transformation step receives a tamper-evident hash chain. After the fact, "who, when, by which rule" is reconstructible and verifiable (FNV-1a-64 + direct struct hashing).
• Hallucination detectionA trigram language model plus Mini-PSV (bigram) catches 95.0% of mistranslations / hallucinated output. Silent-pass (overlooked) rate is below 1.1% on the CORE set.
• DeterminismSame input + same version → bit-identical SHA-256. Output does not drift across parallel or GPU execution, which makes regression testing meaningful.
• Build-time LLM onlyLLMs are used only at rule-construction time. The runtime (actual transpilation) is purely deterministic and rule-based. This is the basis of our claim that "99.9995% is a measured count, not a probability."

Typical use cases

Finance	Migrate regional-bank and insurer COBOL batches (interest accruals, maturity notifications, etc.) to Java/Kotlin platforms without losing a single yen of interest. The 14-dialect auto-detect handles legacy IBM and Unisys implementations from the same engine.
Pension / Healthcare	For computations subject to long-term retention obligations, Round-trip + Audit chain proves that pre- and post-conversion outputs match exactly. Auditors can independently re-verify "no numerical drift introduced by conversion."
Public sector	Modernize 20+ year-old MUMPS / COBOL assets via a domestic transpiler, in incremental phases. Eliminates the indeterminacy of LLM-driven translation by replacing it with a formally guaranteed transformation.
Enterprise IT modernization	For Java 8 → 17 → 21 cross-version migration, we chain bytecode major 52 → 61 → 65 with bit-exact guarantees throughout. "Behavioral invariance" is demonstrable without an exhaustive regression suite.

Five major mainframe vendors covered by dialect fingerprints

SlimeNENC’s dialect-detect engine carries identifier-marker fingerprints for the three major Japanese mainframe vendors plus two key US-origin platforms. For Coast Guard, municipal, regional-bank, international payment, and SMB-ERP engagements where multiple vendors coexist, “one engine that identifies all of them” is a clear differentiator.

Japanese mainframe vendors (3)

Vendor	Primary markers	Calibration state
Hitachi (OpenTP1 / VOS3)	CBLEELOG / CBLEETRN / CBLEEMEM / CBLEERPC / CBLEEDBQ / CBLEETIM CBLEE\* API family, ADDRESSED BY clause, KFSB*****-{E,I,W} message-ID convention, P-1M64 product code	✅ Calibrated against actual samples (OpenTP1 manual, 8 UAPs)
Fujitsu (NetCOBOL / GS21 / PowerCOBOL)	PFD-V*L* product code, AIM transaction base, Symfoware RDB, JEF kanji code, CALL 'PFM*' performance API	△ Public-spec based, calibration via PoC in progress
NEC (ACOS-4 / ACOS-6)	ACOS-4 / ACOS-6 OS, NCRP transaction, ADBS / RIQS II RDB, NX7700 / Express5800 hardware, S2900 / S3900	△ Public-spec based, calibration via PoC in progress

US-origin platforms (2) — international payment, SMB-ERP reach

Vendor	Primary markers	Calibration state
HP NonStop / Tandem (Visa / Mastercard payment cores)	GUARDIAN OS, Enscribe file system, PATHWAY / PATHCOM transaction, IPCF IPC, ENTER TAL '...' clause, NOWAITED / WAITED I/O, S/H/J/NB-Series hardware, SQL/MP / SQL/MX	△ Public-spec based, awaiting US PoC for calibration
IBM i / AS/400 ILE (SMB ERP / public-sector residual)	CRTBNDCBL / CRTCBLPGM CL command, DDS / DSPF integration, CALL PROCEDURE, CALL 'Q*' system calls, QGPL / QSYS library, RTNVAL modifier, COBOL/400 (legacy name)	△ Initial validation against martinfx-cobol (28 files, GPL)

Note: the above describes dialect-detect (identifier-marker classification). dialect-transpile (actual conversion) coverage centers on GnuCOBOL / opensource-cobol / IBM z/OS / MicroFocus / Burroughs; the five vendor families above are at the fingerprint stage, with conversion PoCs proceeding via customer-supplied code. Please contact us for details.

Subset structure

SlimeNENC consists of two subsets:

Subset A	Cross-language transpilation (this page) Bit-exact migration to a different target language. The 90/90 evidence on this page belongs here.
Subset B	PSDP (Phase-Synchronous Deterministic Parallelism) Bit-exact within-language parallelization. A sequential Java batch becomes a parallel Java batch with no change in output. PSDP detail page →
SlimeJCL (companion)	JCL → Spring Batch transpiler (separate product, separate license) SlimeNENC-user-only distribution, targeted GA 2026-Q3. Bit-exactly carries the JCL counterpart of COBOL batches into Spring Batch, completing the migration project in a single toolchain. SlimeJCL detail page →

Technical specifications

Patent	JP Patent App. 2026-046620 (Subset A) / JP Patent App. 2026-046625 (Subset B = PSDP)
Claims mapped	Stage 1 (claim 1f) / Stages 2–3 (claim 3) / Stage 4 (claim 6) / Stage 5 (claims 2, 7, 8) / Stage 7 (claim 9) / Per-dialect Slot 14b/c/d
Standard tests	NIST CCVS85 501 programs / Internal 18 cases × 5 languages × 2 dialects = 90 cells
Performance	median 62 ms / p95 145 ms / WASM build within 2× overhead / O(N) scaling
Dialect coverage	dialect-detect (identifier-marker): 14 values (IBM Enterprise / Micro Focus / GnuCOBOL / Burroughs(Unisys) / Hitachi OpenTP1 etc.), dialect-verify bit-exact 18/18 (primarily on IBM/Unisys/MF families). dialect-transpile (actual conversion) coverage centers on GnuCOBOL / opensource-cobol; Hitachi etc. at fingerprint stage.
License model	WASM converter is licensed; converted output is unlicensed. Ed25519 3-hop activation.

Related documentation

Detailed materials are available on request:

• Technical overviewSlimeNENC Technical Overview (A5, 21 pages, available as PDF / Markdown / HTML)
• PSDP paperPSDP Paper JP v5d (Subset B, 910 KB PDF)
• Patent specificationsJP Patent App. 2026-046620 (Subset A) / 2026-046625 (Subset B = PSDP)
• Reference implementations5 languages × 23 converters (with 90/90 bit-exact validation)
• Benchmarks501-program × 8-stage performance profile (median / p95 / scaling)

Contact us See PSDP (Subset B) See other products