World-model · koru

Where koru is, right now

1151/1358
tests passing
207
gap to goal
84.8%
pass rate

Goal is 1358/1358 — gap 0. That gap is the surprise the world-model holds, and the breath that inhale/exhale moves.

Signals — what breath sees

koru's pass-rate across its whole snapshot history — fires when the breath goes wrong: a drawdown held ≥8pp below the running peak for longer than anything the in-sample bring-up ever showed (a long inhale that never exhaled).

21
out-of-sample fires
84.7%
rate at first fire (snapshot 250)
0
in-sample false fires
0
oracle mismatches
24995
pass rate % out-of-sample (from 2026-05-02) alarm fired

Performance — world vs model

Host-parity — "at parity with the host language (Zig) — any gap is measurement noise": UNMEASURED — 0 / 8 workloads have a Zig baseline. Asserted, not measured.

Against the field, koru is fastest in 7 / 8 workloads (koru = green, fastest rival = grey):

  • filter_map_reducekoru fastest
    koru186.2ms
    rust187.6ms
  • multi_kind_dispatchkoru fastest
    koru154.5ms
    rust158.2ms
  • regex_matchkoru fastest
    koru46.4ms
    rust95.1ms
  • regex_match_boundkoru fastest
    koru47.7ms
  • regex_match_groupskoru fastest
    koru111.9ms
    rust303.9ms
  • regex_redoskoru fastest
    koru29.5ms
    rust30.0ms
  • sum_range_foldable
    koru24.1ms
    rust24.1ms
  • sum_range_xorkoru fastest
    koru136.8ms
    rust139.7ms

What it watches 3 watchers

  • ● wm-adapter adapter · wm/run.sh
    koru regression suite — the wall that writes each snapshot; fires (red) when a green test flips red
  • ● breath native · models/breath/model.wmfx
    koru's pass-rate across its whole snapshot history — fires when the breath goes wrong: a drawdown held ≥8pp below the running peak for longer than anything the in-sample bring-up ever showed (a long inhale that never exhaled).
  • ● commit_cadence native · models/commit_cadence/model.wmfx

What it's missing 19 open · 1 rejected

Candidates surfaced by automated hunts — not confirmed defects. ⊙⊙ = independent agents converged · ✓ = already built · the rest are unverified hypotheses, not yet re-checked against source.

  • med·high Diagnostic-code registry coherence: emit ⊆ declared, declared ⊆ emitted, pinned ⊆ declared ⊙⊙ ✓ built
    2026-06-07 · test
  • med·high Product version is one number across all five sites
    2026-06-07 · test
  • high·high canonicalize post-condition: every DottedPath gets a module_qualifier, no silent null-skip downstream
    2026-06-07 · test
  • high·high No orphaned expected.txt that degrades to a false-green compile-only test
    2026-06-07 · test
  • high·high Hand-authored doc code blocks must compile against the live compiler (stderr-gated, not exit-code) ⊙⊙
    2026-06-07 · test
  • med·med A behavior-changing fix(...) under src/ must land with a regression test
    2026-06-07 · test
  • high·high proc bodies skip cross-module private-event visibility (KORU044 blind spot)
    src/main.zig:5415 (.proc_decl => {})
    2026-06-18 · test
  • high·high 4+ glob matchers with divergent semantics (tap vs transform dispatch disagree) ⊙⊙
    src/runtime_registry.zig:260; src/shape_checker.zig:410; src/ast.zig:1758
    2026-06-18 · test
  • med·med [aspirational] stdlib markers have no compiler enforcement gate ⊙⊙
    koru_std/fs.kz:19; OUTSTANDING_DESIGN_DECISIONS.md:1
    2026-06-18 · test
  • med·high compiler_passes modules compiled but never wired in (dead_strip, purity_analyzer, effect_analyzer, metadata_aggregator) ⊙⊙
    src/koru_lib_facade.zig:20; build.zig:800; absent from src/main.zig
    2026-06-18 · test
  • med·high MUST_FAIL negative tests that pass without pinning WHICH error fires ⊙⊙
    tests/regression/500_INTEGRATION_TESTING/510_NEGATIVE_TESTS/510_011_invalid_module_qualifier (no expected_error.txt); 510_052 (FRONTEND_COMPILE_ERROR only)
    2026-06-18 · test
  • med·high TapRegistry emits a no-op placeholder into PRODUCTION output (taps silently do nothing)
    src/emitter_helpers.zig:830; src/visitor_emitter.zig:658
    2026-06-18 · test
  • med·high duplicate ~proc silently last-wins (no diagnostic, unlike registerEvent)
    src/type_registry.zig:231
    2026-06-18 · test
  • med·high wildcard tap matching zero events warns only via log.debug (invisible unless DEBUG=true)
    src/shape_checker.zig:611
    2026-06-18 · test
  • med·high koruc init stamps hardcoded version 0.1.0 (drifts from compiler version 0.1.7)
    src/main.zig:5734 vs src/main.zig:36
    2026-06-18 · test
  • med·high CCP 'ast' command documented but not implemented
    docs/CCP.md:59 vs src/ccp.zig:14
    2026-06-18 · test
  • med·high obligation-matrix CELL key typo silently drops a cell to 'no cell yet'
    scripts/generate-obligation-matrix.js:113
    2026-06-18 · test
  • med·med branch-constructor field names never validated against declared fields
    src/shape_checker.zig:1523 (TODO)
    2026-06-18 · test
  • low·high effect-branch ordering (!-before-|) is convention-only, unenforced
    docs/EFFECT_BRANCHES.md:47
    2026-06-18 · test

Rejected

  • ✗ annotation typo (e.g. [comptme]) compiles silently
    koru annotations are opaque by design (intended behavior, not a bug)

Is this really "world-modeling"?

Of 19 open gaps: 19 are deterministic checks (tests/probes) and 0 are true temporal watchers. For software, the gaps mostly want a check, not a watcher — the force-multiplier is the hunt that finds them and this one place that shows them.