cobel

Demand-side existence proof for RFL: demonstrate that a real foundation model can emit valid, retargetable RFL Skill ISA.

cobel is the matched half of Milchick (the supply-side proof, a real embodiment). Milchick holds the embodiment fixed and the planner trivial; cobel holds the planner real (Claude) and the embodiment a set of descriptors. Together they bracket the RFL roadmap's "(VLA, embodiment) pair" existence-proof milestone.

What it proves

A frontier foundation model, given the Skill ISA spec + a task + a scene (but not the embodiment), emits valid, schema-conformant, retargetable Skill ISA across novel tasks — at a measured pass rate (pass@1 / pass@k, spec-only vs few-shot). It does not claim that action-token VLAs (π0 / OpenVLA / Octo) emit RFL — those operate below the Skill ISA; the ISA's natural emitter is the planner / VLM tier.

How it works

flowchart TD
    TS["Task + Scene<br/>(no embodiment; Principle 1)"]
    P["Planner.plan()<br/>mock / claude-opus-4-8"]
    SK["skill_yaml<br/>(Skill ISA)"]
    G1["[1] Schema gate<br/>full 50-primitive skill-isa.schema.json"]
    OK["schema_ok (PRIMARY metric)"]
    G2["[2] rfl.retarget(skill, descriptor)<br/>published binding, per embodiment"]
    O["retarget_ok / capability_rejected /<br/>beyond_engine / malformed"]

    TS --> P --> SK
    SK -->|primary| G1 --> OK
    SK -->|secondary| G2 --> O

Loading

The same Planner interface backs both the offline mock (the CI gate) and the real claude-opus-4-8 emitter, so the proof is a one-flag swap (--planner mock|claude). The model never sees the embodiment; the descriptor enters only at retarget. Schema-validity is the headline metric (it does not depend on engine coverage); retarget is a secondary, four-valued check bounded by how many of the 50 primitives the reference engine implements.

Layout

• planner/ — the planner-adapter contract + a deterministic mock + a real Claude emitter
• harness/ — schema + retarget validation; pass-rate reporting
• tasks/ — task + scene inputs (novel; not the RFL examples)
• run.py — --planner mock|claude --mode spec-only|few-shot
• results/ — generated pass-rate tables (the published artifact)

Authoritative design: rfl/docs/design/2026-06-01-demand-side-existence-proof-design.md.

Results (2026-06-01, claude-opus-4-8, samples = 5 per task)

Headline — schema validity (does the model emit valid full-spec Skill ISA?). The model is given the task + scene + the Skill ISA spec digest + the full skill-isa.schema.json contract (spec-only is blind: no worked example skill; few-shot adds one example anchor).

task	spec-only pass@1 / pass@k	few-shot pass@1 / pass@k
relocate-part	1 · 5/5	1 · 5/5
seat-fuse	1 · 5/5	1 · 5/5
latch-buckle	1 · 5/5	1 · 5/5
pour	1 · 5/5	1 · 5/5
stack-blocks	1 · 5/5	1 · 5/5
sort-by-weight	1 · 5/5	1 · 5/5

A real frontier foundation model emits valid RFL Skill ISA on every novel task, every sample, in both modes. This is the demand-side existence proof.

Contract fidelity matters: given only a prose digest of the spec (no JSON schema), spec-only emission dropped to ~1/6 tasks valid — the model wrote plausible skills with slightly wrong parameter names/shapes. Supplying the full skill-isa.schema.json as the contract is what makes spec-only fair and reliable. (This is "give the model the spec," not few-shot example copying — still blind.)

Retarget outcomes (secondary; an honest map of reference-engine coverage across four descriptors — a tendon-driven hand, a direct-drive hand, a 6-finger pneumatic gripper, and the real 4-DOF no-F/T PincherX-100 — not a model metric: the engine implements 18 of the 50 primitives). Claude's emissions are valid full-spec Skill ISA that often use primitives beyond that 18-set (beyond_engine = an engine coverage gap, not an emission error). In few-shot mode a single real-model emission exhibits all four outcomes: seat-fuse retargets cleanly to canonical actions on the three force-capable hands (retarget_ok) and is correctly capability_rejected on the no-F/T PincherX; latch-buckle (force.snap_engage) is capability_rejected everywhere; the rest are beyond_engine. The mock baseline (results/mock-spec-only.md) additionally shows the same embodiment-agnostic relocate-part skill retargeting onto both a dexterous hand and the $600 PincherX. Full tables: results/claude-spec-only.md, results/claude-few-shot.md.

Honest caveats. Claude is a general frontier model — the planner / VLM archetype, not a deployed motor-control VLA; the claim is that the planner tier can target the ISA (action-token VLAs operate below it). Retarget coverage is bounded by the 18/50 reference engine. pass@k is a measurement, not a binary (here saturated at 100% for k = 5).

Private during development. To be published Apache-2.0 when the proof runs.