Skills 2.0 — the generalization boundary

The four-row table.

Extends BW-2026-006's three-row table by one row. The story now reads: arc of three experiments converged on v0.7 → v0.7 won on growth-accounting → EXP-011 tested v0.7 on two more skills → boundary mapped.

Both falsifier clauses fired, pre-registered, blind-judged with κ_self-consistency = 0.980. The rubric is reliable. The result is real.

Why v0.7 worked on growth-accounting and not these.

Three properties make a skill v0.7-friendly. growth-accounting has all three. The other two each miss at least one.

v0.7 is a runtime contract. It enforces structure on the steps and lets an elaborator add case-conditional prose. When the framework's value is in the structure (growth-accounting: the identity check is the answer, the quick ratio is the diagnostic), the runtime IS the lift. When the framework's value is in the per-instance walk-through (psych: a 10-element scoring trace, abw: case-by-case design judgement), the runtime constraint costs more than it adds.

What we missed about EXP-008.

The psych regression has a specific mechanism. In EXP-008, psych-framework A3 ran through subagent-mediated runners whose @step bodies were stubs returning {}. The subagent's own prose reasoning carried the framework application — per-element scoring, fuel-curve traces, Ask-Value verdicts. The judge scored that 4.00 because the prose surfaced every signal the rubric demands.

EXP-011 v0.7 Design A authored real @step bodies that compute aggregated outputs and an elaborator that composes prose on top. The judge transcript named the failure directly:

The runtime aggregated information the rubric wanted disaggregated. The EXP-008 "v0.5 baseline" for stub-step skills was confounded. The runtime was nominally firing, but the subagent's prose was doing the framework application. We never had a clean v0.5 measurement on psych or abw to begin with.

For growth-accounting, the v0.5 baseline was real. Steps were fully implemented. They computed real numbers. The runtime carried information the prose used. That's why EXP-008/9/10 produced a coherent arc on that skill — and why EXP-011 reveals the arc was always one-skill-wide.

v0.8 candidates.

Three mechanism shifts, ordered by expected leverage:

1. Per-element output schemas. Step bodies return list-of-records rather than aggregates. For psych-framework, every element score becomes a row in step_outputs["score_each_element"] with name + type + numeric score. The elaborator now has the per-element trace the rubric wants. Trade-off: response length rises; the growth-accounting reference benefits from aggregation, so the mechanism would need per-skill output-shape choice.
2. LLM-driven elaborator. Replace deterministic Python conditional blocks with a small LLM call that reads the case + the rubric's key_observable_signals + the runtime's typed outputs, and adapts the elaboration to the specific decision points raised. Cost rises by one LLM call per skill run. Generalization probably improves; per-call latency rises.
3. Accept the boundary as a public claim. Ship Skills 2.0 v0.7 as a quantitative-skill mechanism and document the boundary in the public skill-author docs. No further mechanism work until a real production use case forces it. Honest but narrow.

The next planning session locks the v0.8 direction. EXP-012 (cross-model decoupling) and the Elaborator UX track were dependent on EXP-011 passing. Per the meta-plan, both are now cancelled.

What this issue does not claim

(1) That v0.7 doesn't work. It works on growth-accounting — see BW-2026-006. The boundary is which skills, not whether. (2) That all qualitative skills will fail v0.7. The boundary is mapped on one qualitative skill (psych) with one rubric. Other qualitative frameworks with different scoring shapes may behave differently. (3) That EXP-008 results are invalid. EXP-008 stands for growth-accounting (the only skill that wasn't confounded). The psych + abw measurements in EXP-008 were nominally v0.5 but effectively Design B done via subagent prose. This is honest re-reading, not retraction. (4) That v0.8 will succeed. Three candidate directions, each with trade-offs. The next experiment runs against whichever wins the planning session. (5) Generalization beyond three skills. The corpus of Reforge-derived skills is larger; the boundary may shift with different skill shapes.