Product roadmap

R0: release foundation

• Resolve package name, ownership, publication access, and the GPLv3 versus MIT metadata mismatch.
• Publish the existing jethc executable as a versioned package, then provide checksummed standalone binaries.
• Define semver, compiler/language versioning, changelog, compatibility window, and deprecation policy.
• Add CI across supported operating systems and Node versions.
• Generate reproducible artifacts, checksums, SBOM, and signed provenance.
• Add SECURITY.md, a private reporting channel, and an incident process.
• Freeze the first artifact schema and EVM-target policy.

R1: professional CLI and developer experience

• Extend the current project configuration, multi-file builds, contract selection, artifact profiles, structured diagnostics, stable exit codes, and standard JSON mode without breaking their versioned contracts.
• A deterministic local EVM runner for deploy/call scenarios, fixture state, snapshots, traces, revert decoding, and reproducible fork configuration.
• Source maps, source-level traces, revert decoding, gas reports, and storage layout diffs.
• Incremental/watch builds and a validated content-addressed cache.
• LSP, editor extensions, formatter integration, and generated docs on hover.
• Foundry/Hardhat adapters, deployment/linking helpers, and explorer verification.

R2: compiler correctness and conformance

• Convert the supported matrix into machine-readable capabilities and generated conformance cases.
• Expand grammar, type, location, aliasing, inheritance, call, ABI, storage, proxy, and optimizer fuzzing.
• Preserve deploy/run/decode comparisons with solc 0.8.35, then define a process for adding future solc baselines.
• Add metamorphic tests, invariant tests, test-order shuffling, deterministic seeds, timeouts, and corpus minimization.
• Audit optimizer transformations with semantic equivalence and gas baselines.
• Prioritize OR lifts by user impact, workaround quality, implementation risk, and proof coverage. Do not optimize for the raw number of accepted constructs.

R3: language and runtime completeness

• Work through the current "Still gated" list in SUPPORTED.md.
• Prioritize common contract and ABI shapes, arbitrary contract construction beyond the existing EIP-1167 CREATE/CREATE2 clone surface, function values where safely expressible, and storage-reference library workflows.
• Improve import/package boundaries and deterministic dependency resolution.
• Add analyzer checks that prevent proxy/storage misuse and common application hazards where the compiler has enough information to prove them.
• Add opt-in optimizer passes only with before/after semantic and gas evidence.

R4: offline AI verification lab

Architecture#

1. A local LLM, SLM, or reasoning model proposes tests, mutations, invariants, or audit hypotheses.
2. A grammar-aware generator keeps candidates inside a chosen language region.
3. JETH and solc compile equivalent programs.
4. A local EVM deploys and executes both with generated calls and state.
5. An oracle compares bytes, logs, slots, gas classes, and diagnostics.
6. A reducer produces a minimal permanent regression.

Model strategy#

• Begin with retrieval, constrained generation, and tool feedback around a strong local base model.
• Build a normalized corpus of fixes, rejects, parity mirrors, diagnostics, and minimized failures.
• Fine-tune only after evaluation is stable and leakage-resistant.
• Maintain held-out bug families and score reproduction rate, unique findings, false-positive burden, reduction quality, and compute cost.
• Let models propose. Let deterministic execution and compiler invariants decide.

R5: audited numerical packages

First package#

• full-precision multiplication/division;
• signed and unsigned fixed-point types;
• explicit floor, ceil, toward-zero, and nearest rounding;
• saturating and checked variants where the distinction is visible in the name;
• units and range-constrained branded types;
• roots, powers, logarithms, and exponentials with documented error bounds;
• vector and statistics helpers only after gas and overflow analysis.

Calculus-like functionality#

Integrals, derivatives, and differential methods should not emulate real numbers implicitly. Choose one of two explicit models:

• compile-time symbolic transformation for supported expressions; or
• runtime numerical approximation with fixed-point input, a stated algorithm, domain restrictions, convergence conditions, maximum error, iteration cap, rounding mode, and worst-case gas.

Keep numerical packages outside the compiler core, version them separately, and commission specialized audits. The compiler may later add verified intrinsics for proven hot paths without changing package-level semantics.

R6: production trust and ecosystem

• Independent compiler, CLI, standard-library, and generated-proxy audits.
• Public bug bounty and transparent advisories.
• Long-running fuzzing infrastructure with retained corpora and seeds.
• Reproducible release ceremony and signed release index.
• Stable documentation website with version switching and searchable diagnostics.
• Reference projects, migration guides, training material, and maintainer policy.
• Governance for language proposals and backwards compatibility.