Displayed steps must be the real execution path; a fabricated process misleads when to intervene

What it is

If the steps on screen are an after-the-fact “I thought… then I decided…” while the actual tool calls are another chain, people will put their hands down at the wrong moment. Displayed path must be the real path: each step joins a real scheduled event — which tool, which object, what returned. A fabricated process description is another kind of hallucination.

A handsome reasoning narrative is not observation. It is copy. Intervention aimed at copy does not hit execution.

Why it happens

When to intervene depends on “what will happen next beat.” Fake steps mis-say what will happen: people may abort on a safe narrative, or stay while the narrative still says “analysing,” and the outbound send has already gone. That is worse than no window: with no window people at least know they cannot see; a fake window gives certainty of having seen. If plan approval also reads this description, what is approved is a script and what runs is another play — that is plan consistency; this entry is that the in-execution process view must be wired to the scheduler, not to a language model’s self-narration.

A real path may be abstract (node level), but the abstraction must be an aggregation of real events, not another story.

Studying it

Compare three process views: a real tool trace wired to the scheduler, a model-generated reasoning narrative, both side by side with the narrative deliberately forked from the trace. Dependent variables: whether the moment of intervention falls before the real crossing, whether the narrative causes a miss or a false stop, whether people can later point at the fork. Independent variables: whether the narrative is labelled “not execution,” where the fork sits.

The primary endpoint is where the intervention sits relative to the real crossing, not how readable the narrative is.

Where it stops holding

When there is no internal reasoning to show, show the tool trace only; do not pad a story. A model soliloquy needed for debug should be labelled non-execution, and must not drive the semantics of the abort button. Position and remainder built on fake nodes make the skeleton precisely wrong. Whether plan and execution agree is the approval layer; this is what the view is wired to once it is running.

Applying it

• The process view’s data source can only be scheduler events. If a model narrative is shown, it is a separate column titled “what the system says, not execution.”
• Abort, change-object, skip bind only to real events. The narrative column cannot be clicked into execution control.
• Check: leave the narrative on “analysing the recipient” while the scheduler sends the mail. If people therefore do not stop, the fake path has already misled timing. Take the narrative column away; intervention should return to before the real crossing — if it does not, control is still bound to copy.

Related

• Same group: L4.12.1 Observable intermediate steps turn after-the-fact correction into in-process intervention · L4.12.2 Steps that are too dense exceed the user's ability to follow; visibility degrades into a scrolling log · L4.12.3 Users need to tell where they are in the plan and how much remains · L4.12.5 Long phases with no output need their own explanation, or they will be judged as stuck
• Nearby: L4.14 Plan Visibility and Revision for Multi-step Tasks · L4.08 Visibility of Task Progress · L5.01 Types of Explainability
• Search terms: faithful trace · chain of thought · observability