Local revision beats regenerating the whole

What it is

The number in paragraph three is wrong. The first two paragraphs are already usable. Generate again is another draw over the usable parts as well. Local edit over full regeneration means that once an artifact is mostly acceptable, the next act should hit the named piece, not sample the whole document again. It is better first because of cost — people have already paid to read and to keep — and only then because a full redraw throws uncomplained-of parts back to the dice.

What it asks is that a “change this bit” action exist, and be the default. Whether untouched parts actually freeze is the next constraint. Whether history can be clicked back is the one after that.

Why it happens

On the human side, generation’s cost is mostly evaluation: finish reading, judge, decide what to keep. Full regeneration zeros evaluation progress; every round pays from sentence one. Local revision shrinks the evaluation scope to the named span; already-passed parts need not be reviewed. Engineering-wise a full re-decode is also more expensive, but what is expensive in the interaction is the human reread, not the GPU.

There is also a choice structure. Full redo puts people back on accept-or-throw. Local change keeps them in edit. The latter is iteration. If the only legal next step is still another draw, people will accept a “good enough” full text early to avoid reading it all again.

Studying it

Give a draft that is locally wrong and otherwise usable (a bad number, one section off-tone). Two groups: full regeneration only, or select-a-span then regenerate/rewrite. Dependent variables: rounds to a sendable state, share of already-passed paragraphs that moved, reading time, whether they accept an error-bearing version mid-way. Independent variables: salience of the error, draft length, whether the local action requires learning a selection gesture first.

The contrast task should not be tiny — if the draft is two sentences, local advantage is drowned by length. This entry wants artifacts long enough that evaluation cost is already visible.

Where it stops holding

When the first draft is globally unusable (wrong structure, wrong object), local revision is pouring spend after sunk cost, and a full redo is cheaper. Extremely short artifacts (a slogan, an icon) have no meaningful “local.” Images or audio without a selectable object layer give local action nowhere to land. If users cannot or will not make a selection gesture, the local entry equals none. This entry does not treat failed reference (“change that” and the system cannot find that). It treats why the action should be local.

Applying it

• On the artifact, offer select-then-change: rewrite, shorten, replace a number. Do not make “regenerate the whole” the only primary button by default.
• If a full redo must remain, demote it to an explicit destructive act, and preview that already-passed parts will move too.
• The object of a local act must be visible (a highlighted span). Do not rely only on “please point to where.”
• Check: a three-section draft, error only in section three. If no span-level action is available, or it exists but the default is still a full redraw, local-over-whole has not landed. Time to sendable: the local path should reread the two passed sections less. If those two are also rewritten after a so-called local change, the action is named local and is in fact whole.

Related

• Same group: L2.05.2 A revision must leave unmentioned parts unchanged · L2.05.3 Iteration history must be revisitable
• Nearby: L2.06 Granularity of Control · L2.12 Iterative Edits and Local Regeneration · L3.12 Editing and Taking Over Generated Content
• Search terms: local edit over full regeneration · surgical revision · evaluation cost