Conformance

How feelc fares against the official DMN Technology Compatibility Kit (TCK) and against six other rule engines, measured by re-running their own test suites / scenarios through feelc.

The headline: the engine is never wrong on a feature it supports. It computes the correct answer or honestly refuses an out-of-scope construct. The former DMN-import fidelity gaps (OUTPUT ORDER, PRIORITY) are now closed (ADR 0021 — 7/7 hit policies), so every remaining non-pass is a deliberate refusal. feelc's omissions are deliberate exclusions (unbounded lists/iteration, loop-until-fixpoint, string/regex) — consistent with being a total, deterministic, verifiable evaluator.

Official DMN TCK (OMG conformance suite)

Run with feelc's built-in conformance runner — feelc tck --suite <dir> — which imports each .dmn, compiles it, and checks every <testCase> against the TCK's own expected results (exact-decimal equality). Conformance % = passed / (passed + failed); out-of-subset cases are honestly skipped.

• Level 2 (feelc's core): 84.1% at the last full upstream-TCK run — and of the 10 non-passing cases, none is a wrong value for a supported feature. 7 are honest refusals of out-of-scope constructs (string concatenation, ** power [use power(x, n)], full Kleene null logic, a .872 leading-dot literal, a spaced FEEL name).
  • The other 3 were hit-policy import limitations — now closed (ADR 0021): DMN OUTPUT ORDER is a first-class hit policy (hit: output order) and DMN PRIORITY / OUTPUT ORDER import faithfully (reading <outputValues> into a priority: line) instead of degrading to FIRST. feelc now supports 7/7 DMN hit policies. (The headline Level-2 % above predates this and will rise on the next full upstream-TCK run; the fixes are locked by internal/dmnxml/import_test.go and internal/engine/hitpolicy_test.go.)
• Level 3 (full FEEL): feelc is a deliberate subset — for/some/every, lists, string functions, time-of-day, etc. are out of scope, so the runner honestly skips them (3366 skipped) rather than faking conformance. It still never returns a wrong value.
• Non-compliant: feelc correctly rejects the recursion / string-function models.

Cross-engine scenario coverage

71 representative scenarios drawn from six engines' own examples/tests were ported to feelc and proven on the CLI (compile and reproduce the engine's asserted output). 56 of the modelable .rules are committed as a permanent test corpus (packages/engine/test/corpus/x-*.rules).

Every cross-cutting decision primitive ported 1:1 — all hit policies (first/unique/priority/collect), set membership, fact-vs-fact comparison, chained derived facts as a DRG, exact-decimal arithmetic with units, round/floor/ceiling/trunc/modulo, nested if/then/else, marginal brackets, BKM, and applicability gating. feelc was stricter and more correct in two ways the others don't offer: exact decimals (no float drift — 0.1 + 0.2 = 0.3, vs json-logic's 0.30000000000000004 and grule's float32 drift) and a totality/completeness checker that surfaced uncovered-band warnings none of these engines perform.

The gaps (all deliberate)

The 8 cross-engine gaps + the TCK out-of-scope cases reduce to three intentional exclusions — each breaks determinism, totality, or static verification, so feelc rejects them by design (see comparison.md):

1. Unbounded lists / iteration / higher-order — map/reduce/filter/some/every, sum() over a runtime list, list-typed inputs. (Bounded quantifiers over fixed-arity tuples are a candidate add.)
2. Loop-until-fixpoint / recursion / re-feeding outputs as inputs — feelc is a total, single-pass, acyclic DRG.
3. String manipulation & regex — concat, substring, ToUpper, wildcard .match. (starts_with/ contains as cell tests are a candidate add.)

Nested-object/list-path access and async/dynamic/JS-side-effect facts are likewise out of scope, but they only affect data plumbing (flatten to typed scalar inputs first), never the rule logic.

Reproduce

# DMN TCK (clone github.com/dmn-tck/tck): for each TestCases/<level>/<dir>, feelc import + run vs expected.
# Cross-engine: scenarios ported under packages/engine/test/corpus/x-*.rules, swept vs the CLI by
npm -w @feelc-examples/node-smoke test     # WASM == native CLI across every example + corpus decision
npm -w feelc test                  # frozen-output conformance corpus + rejection/tripwire tests