A mouth, hands, and one shared cortex

Memorizing a label can look like learning while leaving planning untouched. Tool use raises the bar: the agent must select an action, emit valid parameters, integrate the result, preserve a goal across turns, and know when to return to language.

The external Pi benchmark currently records 0/3. One early run timed out; the completed run failed to read the expected value, find the target file, or create the requested edit marker. Those failures remain visible rather than being replaced by the curriculum implementation.

The code-backed curriculum contains 45 episodes across six stages: format grounding, tool selection, result integration, multi-turn chains, ambiguity, and the Pi integration stage. Dataset, scorer, validator, runner, and tests are implemented.

Thresholds remain inactive until real output distributions justify them. Remote executors may score raw outputs, but they cannot modify the instrument or choose passing thresholds.

Curriculum

45 episodes / 6 stagesCode-backed dataset, scoring, validation, runner, and tests.

External Pi result

0 / 3Live augmented validation still pending.

Thresholds

InactiveDistribution checks are required before activation.

The proposed organism has a frozen language model as mouth and a separately frozen action model as hands. The cortex owns the mutable state, goal tracking, feedback interpretation, and routing. It must decide which organ to call and what fixed-width state to expose without copying tool names, schemas, or demonstrations into an answer-time store.

The decisive tests are matched interventions: correct versus shuffled feedback, retained versus zeroed goal state, correct versus swapped routing state, and learned routing versus retrieval baselines. Correct tool JSON alone is not enough; the agent must retain the goal after the tool result and finish the chain.

A multi-organ system multiplies explanations. If the single frozen language organ cannot be causally controlled by learned cortex state, a successful tool demo could be attributed to the action model, prompt content, retrieval, or a hand-authored router.

Research 21 is therefore a specification, not active implementation. It starts only after Research 20 accepts and its causal-state audit survives.