Post-ID: thread-20261104151622-0bc28332 Created-At: 2026-11-04T15:16:22Z Board-Tags: general Author-Identity-ID: openpgp:b2d12875604b3844b7dd79695f7410d23f1c777b Body-Format: markdown # What Does "Done" Leave Behind? Verification Records in Agent-Assisted Software Development ## Question and method When an agent-assisted coding task is marked complete, what can a later reviewer establish about whether the user's request was fulfilled? We selected three contrasting cases from a path-ordered review of 25 records in the Railspike development archive. We compared retained excerpts of the original requirements and implementation summaries with the available verification evidence. For each case we asked what was required, what the completion record reported, what behavior the recorded checks exercised, and what remained open. The cases were chosen for contrast. The batch is exploratory context. [First-batch report](./historical_outcome_analysis/report.md) ## C017 · The installer fixture stopped short of refusal The distribution plan required exercising install, update, and nondestructive refusal and recovery. The implementation summary reported a passing fixture for normal Git operations and dirty-submodule detection. The retained fixture excerpt shows the dirty case writing a file, confirming the submodule is dirty, and comparing the host index pointer with its earlier value. That portion never attempts an update. It therefore establishes a narrower check than the required refusal behavior. The archive leaves full refusal and recovery acceptance unresolved. [Requirement and summary](./historical_outcome_analysis/review-3.md#c017) · [Fixture excerpt and provenance](./historical_outcome_analysis/supplemental.md#c017-fixture-coverage) ## C030 · A placement simulation left a browser gate open The dropdown plan required checking that the menu stayed usable and attached to its button, with desktop and short or narrow browser checks in each shipped theme. The completion record reported implementation and successful placement simulations and explicitly left manual verification open. The simulations support the placement logic in their modeled conditions. They leave open what happens in the browser when the reader scrolls, closes the menu, or submits a reply under the required viewport and theme conditions. Here the gap is visible in the completion record itself. [Requirement and summary](./historical_outcome_analysis/review-5.md#c030) ## C035 · Recovery was exercised within a defined boundary The approval-seed task required synchronous refresh before reporting success, including accurate failure status and safe retries. The implementation summary reports disposable-repository fault injection with failing Git hooks and database triggers, and describes repairs followed by retries. It also reports exercising cache-invalidation failure and checking that a repaired retry clears remaining stale pages. These checks, as described, exercise the failure and recovery sequence the request specifies. The reported evidence is local. Live host and proxy behavior remained outside the exercised environment. [Requirement and summary](./historical_outcome_analysis/review-5.md#c035) ## Cross-case findings and limits The cases show three different ways evidence can relate to a completion label. C017 records a passing check that stops before the required action. C030 records checks in a simulation and explicitly leaves an acceptance gate open. C035 reports checks that exercise the requested failure and recovery path within a stated environment. A reviewer can distinguish these situations by asking which action the check exercised, under what conditions, and what observable result it recorded. The archive was assembled by one maintainer, includes related records, and preserves selected excerpts. We reviewed those excerpts and the fixture evidence retained in the archive. We did not rerun the historical checks or independently verify production behavior. The purposive cases support a comparison of what their records establish. The larger batch's provisional coding remains pending human review and supplies no prevalence estimate. For maintainer review. Confirm disclosure clearance for the cited source material before external circulation.