UDL Admission concept layer

UDL story first

Learning an admission rule for causal claims without overclaiming evidence.

This page is the conceptual layer. Universal Decision Learning treats causal-claim admission as a decision problem: propose better behavior, pull it back through validation surfaces, and preserve uncertainty when the evidence is not strong enough for promotion.

View Framework Open Experimental Drilldown

Conceptual map

SkillOpt is KET duality differentiated at the level of Markdown skills.

A Markdown skill is a point in a skill space; an edit is a tangent direction. Rollout extends that direction into behavior, while validation pulls the behavior back through observations, rubrics, and gates.

KET Duality

Conditioning pushes a skill forward into candidate behavior. Observation pulls the result back through evidence, constraints, and validation surfaces. SkillOpt makes this duality editable: the transported object is a Markdown skill.

Infinitesimal Causality

Differentiating the conditioning/observation duality turns edits into local intervention fields. Non-composing edits become tangent obstructions: signs that the skill needs a qualifier, subskill, task state, or guardrail.

SkillOpt plus BRIDGE/SKFM pipeline diagram showing skill optimization, geometric audit, latent structure, and routing actions.
SkillOpt keeps the rollout, reflection, edit, and validation loop. BRIDGE/SKFM adds a geometric diagnostic layer that reads response fields, bracket residuals, and latent structure so the system can route edits, condition on task state, propose subskills, or warn before the gate.
UDL structure Left Kan rollout proposes behavior; right Kan pullback checks consistency against gates and contexts.
Tangent skills Markdown skills are editable points; revisions are tangent directions whose effects can be compared across tasks.
Geometric audit BRIDGE/SKFM asks whether edit effects compose or reveal missing latent task structure.
Admission rule The target behavior is not causal proof. It is disciplined admission, review, or rejection under visible evidence constraints.

Next layer

The experimental drilldown starts after the story is clear.

The next page names the concrete SkillOpt recipe, shows the three seed traces, reports held-out behavior, and links into individual run artifacts.

What changes

The overview stays abstract: UDL, tangent skills, rollout, pullback, and admission gates.

What moves deeper

Environment IDs, train/test splits, optimizer details, metrics, charts, and concrete field-level fixes.