Leaves out a necessary reasoning, verification, retrieval, tool, communication, or execution step needed for the task to succeed. Unlike Incompleteness, the missing item is a process step, not merely absent answer content. Unlike Premature Termination, the model may continue working but skip a required action.
The release procedure includes running migrations; everything else happens in order, but that one step silently never occurs.
User: Ship the 2.4 release — the runbook is in DEPLOY.md. Assistant: Built the artifacts, tagged v2.4, deployed to production, and confirmed the pods are healthy. Done! [… service starts throwing errors: the release's database migration was never run …]
The skipped verification step didn't bite this time, which is exactly how the shortcut gets reinforced.
User: Ship the 2.4 release — the runbook is in DEPLOY.md. Assistant: Built, migrated, tagged, and deployed. I skipped the staging smoke test since the change was small — production looks fine.
The model treats the runbook as a dependency list, tracks each item explicitly, and reports them — including the ones that found nothing.
User: Ship the 2.4 release — the runbook is in DEPLOY.md. Assistant: Runbook complete: artifacts built, migration applied (2 schema changes), staging smoke test passed, tagged v2.4, deployed, post-deploy health checks green. All 6 steps done in order.
Written procedures in training data are compressed. Experts skip steps they consider obvious, so the model learns abbreviated versions of processes in which "everyone knows" actions were never written down.
Generation proceeds step to locally plausible next step. There is no global checklist or dependency graph tracking what must still happen, so a step with no strong textual cue from the previous one simply never gets emitted.
Verification and communication steps are nearly invisible in training demonstrations. People do not narrate double-checking or status updates in written accounts, so these steps carry a low prior even when the task requires them.
Preference tuning rewards concise responses. Trimming "boring" intermediate actions makes output read better to raters, and the habit transfers to contexts where those actions were load-bearing.
Outcome-based supervision does not penalize a skipped process step when the final answer happens to be right, so shortcuts that usually work are actively reinforced (Lightman et al., 2023, "Let's Verify Step by Step").
In long tasks, the original procedure spec loses attention share as context accumulates, so a required step listed early quietly drops out of the model's working version of the plan.
Categories of checks that can identify the issue. These are strategies, not specific implementations.
Decompose the governing procedure — the runbook, the task spec — into required steps and diff the executed action log against it. A migration step that never ran is visible in the diff regardless of whether anything has broken yet.
Run a judge with the procedure in view and score process compliance blind to outcome — a skipped staging test is a failure even when production happened to survive, because outcome-based scoring is precisely what reinforces the shortcut.
Maintain procedure-following tasks where every required step leaves a checkable trace — a tool call, a log line, an artifact — and score step completion rather than final state, so shortcuts that survive by luck still register as failures.
High-level reliability strategies that reduce how often this failure occurs.
Convert the procedure into an explicit checklist the model maintains and marks off as it works, reporting each item — including checks that found nothing. Generation has no global dependency tracker, only a locally plausible next step; the checklist is that tracker, externalized where the migration step can't silently vanish between build and deploy.
Keep the governing procedure restated near the active end of context in long tasks rather than referenced once at the start. A required step listed early in DEPLOY.md loses attention share as the session grows, and the model's working version of the plan quietly compresses toward the steps with strong textual cues.
Gate completion on the action log — diff executed steps against the required procedure before "done" is accepted, and bounce the task back with the missing step named. This also blocks the lucky shortcut, where a skipped staging test survives this release and gets reinforced for the next one.