lin-0178

14.2 The workflow learning sketch

Let \(W\) be a partially ordered workflow and let

\[ F:W\longrightarrow \mathsf{Md} \]

assign a typed Markdown artifact to each stage. Objects may include a task contract, plan graph, tool manifest, observation schema, evidence ledger, answer template, validator contract, and handoff memory.

A skill section \(s_i\) reads selected anchors and writes selected anchors. The workflow declares:

  • prerequisite order between repairs;

  • invariants of the tool and evidence contracts;

  • restriction maps between local artifacts;

  • validator paths that must agree; and

  • the final artifact and trace semantics to preserve.

Definition 14.2 LASKO residual profile

For ordered sections \(s_i,s_j\), define

\begin{align*} R_{\mathrm{vis}}(i,j) & =d_{\mathrm{art}}(F_{ij},F_{ji}),\\ R_{\mathrm{DB}}(i,j) & =d_{\mathrm{blame}}(\eta _{ij},\eta _{ji}),\\ R_{\mathrm{ker}}(i,j) & =d_{\mathrm{trace}}(\tau _{ij},\tau _{ji}) \quad \text{when }R_{\mathrm{vis}}(i,j)\approx 0. \end{align*}

The first observer measures visible order sensitivity. The second measures whether validators assign blame differently. The third is a practical kernel proxy: the artifact may look unchanged while the route, evidence, or future repair behavior differs.