Compatibility

cljgo does not self-grade compatibility. It is measured two ways: an internal conformance suite frozen against real JVM Clojure, and an external suite the project does not control.

The conformance discipline

Every semantic behavior in cljgo is a plain .clj file under conformance/tests/ with a frozen ;; expect: output. Each expectation is verified against real JVM Clojure 1.12.5 (the clojure CLI is the semantic oracle, needed at authoring time only) and cited in a comment.

Two properties make this stricter than a typical test suite:

• Dual harness. Every file runs through both execution paths — the tree-walk evaluator (the REPL) and the AOT-compiled binary — and must produce identical output on both. The README counts 416 oracle-cited files enforced on every commit.
• Divergence is a release blocker. REPL-vs-binary divergence is treated as the one unforgivable failure mode. A file that cannot run both ways needs a written waiver in the file itself.

The file format is simple: any number of forms evaluated in order in a fresh user namespace, with exactly one expectation comment — ;; expect: <text> (the pr-str of the last form’s value must match exactly) or ;; expect-error: <text> (the error message must contain the text).

Why divergence is structurally hard in the first place is covered in Architecture: one reader, one analyzer, one AST feed both backends.

The external yardstick: jank’s clojure-test-suite

cljgo is also scored against the jank clojure-test-suite (upstream @164a4b3, unmodified) — 242 real clojure.core test files — as a single ratcheting number in CI. The passing count may only rise.

Measure	Result
clojure.core vars resolved	242 / 242 (100%)
Suite files fully passing	238 / 242 (98.3%)
Failures	0
Errors	4

Reproduce it yourself: cljgo suite. The suite runs interpreted; cljgo’s own dual-mode conformance suite stays fully green alongside it.

The 4 outstanding files, honestly

The four errors (abs, add-watch, short, reduce) are dialect registration, not broken semantics. Those files carry reader conditionals with no :default branch (e.g. #?(:cljr System.Int16 :clj java.lang.Short)), so a runtime the suite has never heard of reads them as nothing — and (instance? (short 0)) then fails with “wrong number of args (1)”.

Adding a :cljgo branch is the same mechanism :cljr / :lpy / :phel already use, and cljgo’s spellings are truthful ((instance? java.lang.Short (short 0)) is genuinely true here, as on the JVM). With those four branches applied the suite reads 242/242 (100%) — but they are not upstreamed yet, so the published number is the one you get from the suite as it ships. Full analysis: docs/suite-upstream.md.

What is and isn’t complete

clojure.core itself is not yet complete — the honest per-namespace ledger is docs/fundamentals-audit-2026-07.md. Complete against the 1.12.5 oracle (per the README, satellite rows re-verified 2026-07-23):

• Satellite namespaces: clojure.string, set, edn, walk, zip, data, repl, pprint; clojure.test complete (39 oracle vars).
• clojure.core.async: 55 publics — every non-deprecated, non-internal var of JVM core.async 1.6.681, over real goroutines.
• JVM-compatible hashing (ADR 0051), the numeric tower (bigint/bigdec/ratios/promotion), transients, tagged literals and reader conditionals.

Where Go forces a deviation (typed nil, no STM, RE2 regex), the design doctrine is to document it loudly rather than approximate silently — see design/00-architecture.md.