Synthcore

Inductive program synthesis with library learning, for JS/TS. No LLM, no GPU.

Give it input→output examples and synthcore returns a verified function (or {ok:false}). It never hallucinates: it only returns a program that passes all your examples. Zero marginal cost, deterministic, auditable, offline.

import { synthesize } from "synthcore";

const r = await synthesize([
  { input: '{"email":"[email protected]"}', output: "[email protected]" },
  { input: '{"email":"[email protected]"}', output: "[email protected]" },
  { input: '{"email":"[email protected]"}', output: "[email protected]" },
]);

if (r.ok) {
  console.log(r.recipe); // "get(parseJSON(arg0),\"email\")"  ← real recipe, verified
  console.log(r.code);   // a standalone `solve` function that passes the 3 examples
}

r is { ok:true, code, recipe, size, entry, ast } or { ok:false, reason }. code is a standalone JS function ready to run; recipe is its readable form; ast is the recipe as a tree (consumed by learn()); reason ∈ "bad_input" | "not_found" | "unverified".

What it solves / what it doesn't

Enumerative search over a bounded DSL: it shines at small, composable transformations over numbers, strings, arrays and JSON. It is not general-purpose code generation.

| ✅ Solves today (verified in demo/tests) | ❌ Out of scope today | |---|---| | Scalar arithmetic — double → add(arg0,arg0) | Large programs / application logic (it won't write an app) | | Strings — reverse → rev(arg0), UPPERCASE → upper(arg0) | Filtering / dedup: the filter/unique combinators are missing | | JSON parsing — JSON field → get(parseJSON(arg0),"email") | Arbitrary recursion (e.g. Fibonacci) / deep structures | | List reduction — sum → sum(arg0), average → idiv(sum(arg0),len(arg0)) | Algorithms with deep recursion or backtracking | | Lists — sort → sort(arg0), max → max(arg0), map/zipWith | Transformations depending on external state or I/O | | Dates/numbers/regex via the opt-in std bundle — year → year(arg0), "$1,234.50" → num(arg0) | Large multi-step programs (enumerative budget) |

For the right column, the path is to seed a primitive (extraPrims, e.g. proposed by an LLM) or raise the search budget — it's not a wall, it's a frontier that moves.

Use cases

Synthcore shines wherever you have examples of a small, repeatable transform and need it guaranteed correct, cheap and offline — not a plausible guess you still have to test:

• Data cleaning / ETL at scale. Normalize a messy column ("$1,234.50" → 1234.5) once, then run it over millions of rows at zero per-call cost, offline (no data leaves your machine). Transpile to Python for your pipeline.
• Kill the regex you were about to google. Extract a field by example instead of by pattern ("ERR-404: not found" → "404"). You describe the what, not the how.
• Schema mapping / API glue. Turn "their payload → your payload" examples into a verified extractor; commit it, run it in CI, no external API in the loop.
• Recover an exact formula from data. Verified symbolic regression (½mv², pH) for unit conversions, pricing rules and discrete laws — exact, not approximated.
• A verified tool for your agents. Via MCP, an LLM can call Synthcore for a guaranteed transform instead of hallucinating one.

How it works

• Typed bottom-up search over a DSL of combinators (n-ary application, map, fold, zipWith, constants), with observational-equivalence pruning (dedup by the output vector) → the exponential pruning that makes it tractable.
• Deterministic verification in a hardened subprocess (no env, memory cap, temporary cwd, hard kill on timeout). Only code that passes all examples is returned.
• Library learning (wake/sleep, DreamCoder/LILO style): mines recurring sub-programs, generalizes them into abstractions ranked by compression (MDL) and reuses them → solves deeper with the same budget.
• Optional LLM hybrid: when pure synthesis fails, an LLM can propose one new primitive (extraPrims); the engine verifies it and, if it unlocks the solution, recomposes it for free forever. The LLM introduces the rare, expensive bit; the engine recomposes deterministically and at $0.

API

synthesize(examples, opts?) → Promise<{ ok:true, code, recipe, size, entry, ast } | { ok:false, reason }>

• examples: a list of { input, output } | { in, out } | { args:[...], expect }. An array input is one list argument (not varargs). At least 3–5 examples recommended.
• opts: { entry?, tools?, rounds?, maxEvals?, extraPrims?, std?, verify? }.

Rich types (opt-in): dates, numbers in text, regex

The standard primitives are not in the base DSL — they enlarge the search and raise the overfitting risk — so they are enabled with { std: true } (or by passing extraPrims: stdPrims):

await synthesize([{ input: "2024-03-15", output: 2024 }, /* … */], { std: true });  // → year(arg0)
await synthesize([{ input: "$1,234.50", output: 1234.5 }, /* … */], { std: true });  // → num(arg0)

stdPrims = datePrims (year/month/day/weekday, in UTC) + numberPrims (num/digits) + regexPrims (regexExtract/regexMatch).

Science domain bundles (opt-in): physics & chemistry laws

Physical and chemical laws ship as injectable primitives (physicsPrims, chemPrims, or both via sciencePrims). Give it data and it recovers the exact law — verified symbolic regression, like PySR/Eureqa but with an exact guarantee instead of a minimized error:

import { synthesize, physicsPrims, chemPrims } from "synthcore";

await synthesize([{ args: [2, 3], expect: 9 }, { args: [4, 5], expect: 50 }, /* … */],
  { extraPrims: physicsPrims });                      // → kinetic(arg0,arg1)   (½·m·v²)

await synthesize([{ input: 0.01, output: 2 }, { input: 0.001, output: 3 }, /* … */],
  { extraPrims: chemPrims });                          // → pH(arg0)             (-log₁₀[H⁺])

Honest frontier: for noisy experimental data use PySR (it minimizes error); for exact transformations (discrete laws, unit conversions, identities) Synthcore wins (exact verification). Note the engine only composes arity-1/2 primitives in its rounds, so arity-3 laws (e.g. idealGasP = nRT/V) only synthesize when the task itself has that arity.

Multi-language output: transpile to Python

The engine synthesizes a language-independent AST (Recipe); synthesize lowers it to JS, and emitPython lowers the same verified AST to standalone Python — no extra search, just an output pass:

import { synthesize, emitPython } from "synthcore";

const r = await synthesize([{ input: [1, 2, 3], output: 6 }, { input: [10, 20], output: 30 }]);
if (r.ok) emitPython(r.ast);   // → "def solve(*a):\n    return sum(a[0])"

Pass the same tools/extraPrims/std you gave synthesize. Only the base DSL and the bundles we maintain (std, domain) transpile; a learned abstraction or an arbitrary injected extraPrim (raw JS) throws a clear error instead of emitting incorrect code.

Learn a library that grows with use

learn() mines reusable abstractions from solutions you already found (DreamCoder/LILO, MDL ranking). Pass them back as tools and the engine solves deeper with the same budget. serializeLibrary / loadLibrary persist them across sessions (plain JSON):

const r1 = await synthesize(examplesA);
const r2 = await synthesize(examplesB);
const lib = await learn([r1, r2]);                  // mine verified abstractions
const r3 = await synthesize(examplesC, { tools: lib.tools }); // reuse what was learned

const json = serializeLibrary(lib.tools);           // persist
const tools = loadLibrary(json);                     // restore in another session

LLM-seeder hybrid

When pure synthesis fails, an LLM proposes one primitive (extraPrims); the engine verifies it and recomposes it for free. Full copy-paste example: examples/hybrid-llm-seeding.ts.

Advanced surface (composition, independent verification): solveBySynthesis, learnAbstractions, grade (verifier), buildOps, configureSearch, and the types Tool / Recipe / Op / Abstraction.

Use it from an agent (MCP)

synthcore-mcp is a Model Context Protocol server that exposes synthesize as a tool, so agents (Claude Desktop, Cursor, …) can get a deterministic, verified, $0 data transform instead of hallucinating one. The LLM reasons; Synthcore guarantees.

cd mcp && npm install && npm start   # stdio server with a single `synthesize` tool

Then point your client at it and call synthesize with your examples (supports the opt-in bundles and language: "python"). Full wiring for Claude Desktop / Cursor in mcp/README.md.

Limitations (read them — honest selling avoids the hype that burns)

• It is not general code generation. Enumerative search only reaches small programs in a bounded DSL; it does not write an application nor reason over natural-language specs.
• It complements an LLM, it doesn't replace it. The model is the hybrid (LLM proposes rare primitives; Synthcore recomposes them for free and verified), not a head-on competitor.
• With few examples it can overfit. It returns some program that passes what you gave it; with 1–2 examples that may be a coincidence. Real case: for "email domain" with 3 examples the engine finds max(split(capitalize(arg0),"@")) — it passes by accident (the domain wins the lexicographic order), not because it "understands" emails. Give ≥3–5 examples with edge cases so the solution truly generalizes.
• Bounded DSL. Dates, regex, numbers-in-text and the science bundles (physics/chemistry) exist but are opt-in ({ std: true } or extraPrims), because each primitive enlarges the search. The filter/unique combinators are still missing and there is no recursion, so filtering, dedup or Fibonacci return not_found. It's the current DSL frontier (it moves by seeding primitives), not a hidden case.
• Transpilation is template-based, not a full compiler. emitPython covers the base DSL and the bundles we maintain. Learned abstractions and arbitrary injected primitives (raw JS) are not transpiled — it throws rather than emit code it can't guarantee.
• Niche, not mass. The audience is developers/tooling, not end consumers.

Positioning

• vs Microsoft PROSE / FlashFill (mature PBE): Synthcore adds library learning + deep composition, and is embeddable, open TS (PROSE is C#/closed and string/table-centric).
• vs DreamCoder / LILO (academic, Python + GPU): Synthcore is a TS project that runs on a laptop without a GPU.
• vs LLM codegen (Copilot/Cursor): deterministic, verified, $0, offline, auditable — the other half of the stack.

Usage

npm install synthcore   # in your project

Repo development:

npm install        # devDeps only (typescript, tsx, @types/node)
npm run demo       # solves 8/8 example tasks, verified, $0
npm run bench      # honest benchmark: 13/15 tasks, ~150 ms per solved task, $0
npm test           # contract suite (node:test)
npm run typecheck  # tsc --noEmit
npm run build      # compiles the library to dist/ (JS + types) for publishing

Requires Node ≥ 22. Pure ESM, no runtime dependencies.

License

MIT.