Prove two codebases do the same thing — or hand you the input where they don't.
Isomorph takes two implementations of the same logic — different languages, different repos, different authors — and adversarially tries to make them disagree. It ends one of two ways:
DIVERGENCE FOUND —
computeRSIdisagrees when the input contains a NaN gap ≥ period length. Python returns 50.0 (neutral seed); JS returns NaN and poisons every downstream value. Minimal repro: 15-element array attached.
NO DIVERGENCE FOUND — 4 properties proven by exhaustive small-input enumeration, 12 properties survived 50k adversarial inputs, 3 behaviors UNVERIFIABLE (I/O boundary, listed below). Confidence: HIGH on pure functions, NONE on the network layer.
It never says "equivalent." It says exactly what was proven, how, and what
remains unverified. UNVERIFIABLE is a first-class verdict.
AI made rewrites cheap. Python→Rust ports, monolith→service extractions, duplicated validation across mobile and backend, protocol reimplementations — all of them ship on the strength of golden tests that both authors wrote, which means both authors' blind spots are baked into the safety net. The question every migration team avoids asking out loud is "do these actually behave the same?" Isomorph is the adversarial third party whose only job is to answer it.
A seven-stage pipeline, run by a frontier model inside your Claude Code session (your code never leaves your machine):
- Ingest — read both codebases, build an alignment map pairing comparable functions by semantic role, not name. You confirm the map — the one required human step.
- Model — write a shared behavioral contract per pair: domain, edge cases, numeric semantics, statefulness. Contract disagreements are findings before any code runs.
- Synthesize properties — decide, explicitly and in print, what "equal" means per pair (exact, float-tolerance, order-insensitive…).
- Harness — generate a thin sandboxed differential runner per pair.
- Adversarial search — structured fuzz, coverage-guided mutation, and reasoned attacks: the model reads both implementations' logic differences and constructs inputs targeting each suspected divergence, with a stated hypothesis per attack. The reasoned attacks are the point.
- Triage + shrink — every divergence gets input-shrunk to a minimal repro, root-caused to a line pair, and severity-graded.
- Report — one artifact: divergences with repros, proven properties, the UNVERIFIABLE boundary, and the alignment map used.
Full detail: docs/ARCHITECTURE.md.
# as a Claude Code plugin (recommended)
claude plugin install /path/to/isomorph # see docs/INSTALL.md
# then, inside any Claude Code session:
/isomorph path/to/implA path/to/implBSee docs/INSTALL.md for setup and docs/USAGE.md for worked use cases.
- Pure-ish deterministic logic only: parsers, calculators, validators, encoders, indicator math, pricing/tax rules. Concurrency, networked services, and UI are honestly reported as UNVERIFIABLE, not guessed at.
- Language pairs: Python ↔ JavaScript/TypeScript first; next pair chosen by demand.
- Local-only: all execution happens on your machine, in sandboxed runners (no network, scratch cwd, timeouts). Nothing is uploaded anywhere.
v0 — pre-benchmark. The pipeline runs as a Claude Code skill; the tool has not yet passed its own Gate 1 blind benchmark (see docs/PLAN.md). Do not trust a NO-DIVERGENCE verdict from a v0 run for anything with money on it — DIVERGENCE findings, however, come with executable repros you can verify yourself.
| Doc | What |
|---|---|
| docs/INSTALL.md | Install and setup |
| docs/USAGE.md | Use cases and a full walkthrough |
| docs/ARCHITECTURE.md | Pipeline internals, formats, sandboxing |
| docs/PLAN.md | Execution plan, kill/continue gates, guardrails |
| CLAUDE.md | Rules for agents working in this repo |
MIT. See LICENSE.