Reference inferences

A small set of well-tested inferences ships with the MCP extra (under machinable.mcp.inferences). They are shareable implementations, not core; each brings its own scipy statistics, and they back the inference tools. You write your own the same way (add-inference), or share one by URL like any interface.

Inference / tool	Requires	Operands	Question
Outperforms (outperforms / differs)	scalar	2	Is A's quantity greater than B's? (Welch / Student / Mann-Whitney; differs is two-sided)
BeatsBaseline (beats_baseline)	scalar	1	Does A's quantity beat a fixed baseline? (one-sample t-test)
Ranks (ranks)	scalar	≥3	Do the operands differ, and how do they rank? (Friedman omnibus + mean ranks)
ConvergesFaster (converges_faster)	series	2	Does A's curve reach its target sooner? (AUC or steps-to-threshold, then a two-sample test)

Verdicts

Every verdict is a JSON-able dict with a claim, a boolean holds, the p_value, the sample sizes, and alpha, plus inference-specific fields (effect size, mean ranks, the summary used). It is saved as verdict.json on a normal interface, so it is reproducible, cached on (method × operands), searchable (kind="Inference"), and can be surfaced as a widget.

Using them

From the MCP, call the tool (outperforms(a=…, b=…, quantity=…)). From Python, resolve the module directly:

from machinable import get

a, b = get("optimizers", ["~sgd"]), get("optimizers", ["~adam"])
a.launch(); b.launch()

verdict = (
    get("machinable.mcp.inferences.outperforms", {"quantity": "loss", "tail": "less"})
    .of(a, b)
    .launch()
    .verdict()
)

Multiple comparisons

Correction is implementation, enabled by the searchable kind. Where the family is the operand set, Ranks runs the omnibus itself. Where comparisons span many calls, the kind="Inference" search makes the family enumerable, so a correction is expressible as another inference over a set of verdicts (e.g. a HolmCorrected(verdicts)), also in userspace. machinable makes the family discoverable; it never corrects silently.