core-labs/core
1
0
Fork 0
Code Issues Pull requests 1 Projects Releases Packages Wiki Activity Actions
core/evals/hebrew_fluency/gaps.md
Shay ad7993e861 feat(phase5): land 5.2–5.7 — six new fluency lanes, parallel sweep 
Completes the Phase 5 curriculum-era lane checklist alongside 5.1.

English-substrate domain lanes (5.4–5.7) — extend the proven
english_fluency_ood pattern with new vocabulary domains. Same
13-construction realizer, same grammatical_coverage rubric, new
triples. All four lanes land at 100% on both splits:

  5.4 elementary_mathematics_ood    117/117 public + 39/39 holdouts
      domains: arithmetic, set, geometry  |  holdout: probability
  5.5 foundational_physics_ood      117/117 + 39/39
      domains: mechanics, electricity, thermodynamics  |  holdout: optics
  5.6 foundational_biology_ood      117/117 + 39/39
      domains: cell, organism, ecosystem  |  holdout: genetics
  5.7 classical_literature_ood      117/117 + 39/39
      domains: epic, tragedy, lyric  |  holdout: comedy

New-language lanes (5.2 Hebrew, 5.3 Koine Greek) — scoped honestly to
v1 = C01 only, script + length rubric. The realizer's
tense/aspect/quantifier/negation logic in generate/templates.py is
English-only; C02-C13 in HE/GRC requires Hebrew/Greek morphology +
rhetorical templates, named explicitly in each lane's gaps.md as the
v2 unblock path. v1 measures what infrastructure exists:

  5.2 hebrew_fluency       3/3  (predicate-subject-object assembly,
                                  Hebrew script gate)
  5.3 koine_greek_fluency  3/3  (subject-object-predicate assembly,
                                  Greek script gate)

Lane scaffolds follow the established pattern: contract.md, runner.py,
__init__.py, gaps.md, public/v1/cases.jsonl, dev/cases.jsonl,
holdouts/v1/cases.jsonl (5.4–5.7 only; HE/GRC holdouts deferred to v2
when vocabulary expands).

Generators + scorers:
  scripts/generate_phase5_domain_lanes.py      — 5.4–5.7 case emit
  scripts/scaffold_phase5_domain_lanes.py      — 5.4–5.7 contracts/runners
  scripts/generate_phase5_language_lanes.py    — 5.2/5.3 case emit
  scripts/score_phase5_holdouts.py             — parallel holdouts scoring
                                                 via multiprocessing.Pool
                                                 (mirrors the parallel-eval
                                                 pattern from evals/parallel.py)

Lanes are wired into core eval --list automatically through the
framework's lane discovery; parallel sweeps via bash background jobs
(one process per lane).

Regression clean: smoke 54, runtime 19, teaching 17, packs 6,
cognition 57, algebra 132. Cognition eval 100% across all metrics.
2026-05-16 20:59:31 -07:00

2.7 KiB
Raw Blame History

Hebrew fluency — gaps

v1 (current)

  • Construction coverage: C01 only (simple declarative).
  • Word order: verb-second (predicate-subject-object), via generate.articulation._assemble for language == "he".
  • Grounding: through ChatRuntime.chat() with frame_pack="he".
  • Rubric: script + length only (surface contains Hebrew script AND length ≤ max_words). Lexeme-level subject/predicate/object slot matching is not gated at v1 — the runtime currently folds multi-token Hebrew input to a single lexeme through articulation (a known limitation, see v2 below).

v2 unblock path — Hebrew construction coverage

To extend to C02–C13 (negation, tense, aspect, quantification, relative clause, etc.), the realizer needs Hebrew analogues of the English-only modules:

  1. Hebrew morphology — analogue of generate/morphology.py. At minimum: verb conjugation (qal, piel, hifil binyanim), past tense (qatal), imperfect (yiqtol), participles, plural agreement.

  2. Hebrew predicate display map — analogue of _PREDICATE_DISPLAY in generate/templates.py mapping seed-pack predicates to surface forms.

  3. Hebrew rhetorical templates — analogue of _MOVE_TEMPLATES. Verb-second order changes the slotting; placement of function words ("furthermore", "in contrast") needs Hebrew renderings.

  4. Hebrew negation marker — לֹא placement before the verb (predicate-first position), not the English does not <verb> pattern.

  5. Hebrew quantifiers — כל (all), יש (some/exists), etc., with appropriate gender + number agreement to the subject.

  6. Lexeme-level slot grounding — the HE runtime pipeline currently produces a single-lexeme articulation (e.g. surface 'דבר דבר' regardless of multi-word Hebrew input). v2 needs the grounding + planning layers to preserve distinct subject/predicate/object slots through articulation so the rubric can check lexeme presence per slot, not just script presence.

Out of scope for this lane

  • Pointed/unpointed text policy (nikkud). Lexicon entries are currently unpointed. Whether to render pointed surfaces is a pack-level decision, not a fluency decision.
  • Right-to-left text rendering in tools/UI. This lane scores the string; rendering is downstream concern.
  • Modern Hebrew vs. Biblical Hebrew register. Seed pack is cognition-tier Biblical Hebrew; modern register expansion is a separate pack.

Related

  • ADR-0020 (Rust parity sequencing) — language-track work is independent of the Rust parity track.
  • Phase 5.1 (english_fluency_ood) — same harness shape; full 13-construction coverage because English realizer infrastructure already exists.