Reproduced October 18, 2026

When “the run ended” became “the lesson helped.”

An administrative stop was being counted as evidence about unfinished work. That could award helpful credit to a lesson without the relevant outcome.

The starting failure

A workflow recalled a lesson and followed its planning instruction. A smoke-check child failed. Another child, still carrying out a repair, was stopped when the parent run ended. Its relevant work outcome was unknown.

The settlement reader could treat the stopped or cancelled child as evidence that the original obstacle had not recurred, then emit a helpful-feedback intent.

The bounded correction

Keep lifecycle settlement separate from the observed work outcome. With no relevant outcome, recurrence remains UNKNOWN and helpful credit cannot be awarded. An actual diagnostic remains evidence; normal completed work retains its bounded feedback behavior.

Five existing methods were replayed on both snapshots. The original had four failing subtests across the first two methods. The corrected version passed all five methods.

Evidence and limits

Original source: 906703c91893ce3e3696070a870a20da6b145341. Corrected source: 945b3a9cd73e6fe9ecacaab37f62440f9337f364. The same regression file was used for both. Exact logs and source digests are retained with the project; this page does not release the private source tree.

This is isolated fixture evidence from our own project. No live actor, customer data or memory store was used. It does not demonstrate general agent reliability, causal benefit from every lesson, current deployment of the fix, or a customer result. No time-saving figure is claimed.

A paid engagement applies this style of work to one agreed customer problem and leaves checks the buyer can rerun. Its feasibility and scope are established separately.