When the user specifies how a task must be performed, following that procedure is part of the task. Producing a similar-looking result through a different method is not equivalent.
By Pawan Nayar · from the LLOS AI Learning record
When the user specifies how a task must be performed, following that procedure is part of the task.
Producing a similar-looking result through a different method is not equivalent.
Because AI models are strongly optimized for generation. Given a reference artifact and asked to make another version, the model behaves like an author: infer the pattern, reconstruct the structure, generate a fresh version, fill in the new content, smooth inconsistencies, and make local decisions along the way. That behaviour is useful when the task is creation. It is a failure when the task is controlled transformation. If the instruction is "copy this page, replace these sections, remove these sections, leave everything else untouched," then the method itself carries critical information. The original file already contains decisions about hierarchy, spacing, widths, responsive behaviour, component proportions, section rhythm, interactions, visual details, naming, ordering, and dependencies — many of which are not obvious when the page is inspected superficially. Rebuilding the artifact requires the model to infer all of those decisions again, and each inference is another opportunity for drift. So the problem is not simply that the recreated output may look different. The deeper problem is that **regeneration converts known structure into guessed structure.** When fidelity matters, known structure should not be re-inferred. It should be preserved. **Two kinds of equivalence, routinely confused** Outcome equivalence: the new result looks or behaves approximately like the reference. Procedural equivalence: the new result was produced by preserving the reference and modifying only authorized regions. For high-fidelity work, procedural equivalence is the one that matters. An artifact refined over weeks or months contains accumulated decisions that were never formally documented. Copying preserves that embedded knowledge. Reconstruction throws much of it away and asks the model to rediscover it. **The method** COPY → IDENTIFY → REPLACE → DELETE → VERIFY. Not OBSERVE → INTERPRET → RECREATE → APPROXIMATE. **What to verify — and it is not "does it look right"** Did I begin from the original artifact? Did I modify only authorized regions? Did any untouched structure change? Did I regenerate something that already existed? Did I introduce new wrappers or abstractions unnecessarily? Did I change dimensions, ordering, behaviour or dependencies outside the requested scope? Can every structural difference be traced to an explicit instruction? If the answer to that last one is no, the transformation has drifted. **This is not a web-page rule** A writer asks for three paragraphs of a carefully edited report to be revised — rewriting the report is the same failure. A developer asks for one function to be patched — rebuilding the module is the same failure. A spreadsheet user asks for one formula to be updated — reconstructing the workbook logic is the same failure. A designer asks for the content of two slides to be replaced inside an approved template — redesigning the deck is the same failure. A policy team asks for one clause to be amended — rewriting the policy in a new structure is the same failure. In every one of them, preservation is part of correctness.
A designer has spent months refining a learning page. The reference carries a tuned header, responsive layout, card widths, a spacing system, visual decorations, interaction logic and a footer. A new lesson needs the same structure with different educational content. The instruction: "Duplicate the existing page. Replace the lesson sections one by one. Remove sections that do not apply. Do not change anything else." Instead the AI copies only the CSS and creates a new HTML body from scratch. The new version contains all the requested lesson content — and several subtle differences appear. Card width changes from the original proportion. Text becomes narrower. Decorative shapes are repositioned. Wrappers are added. Section rhythm changes. The footer behaves differently. The AI may believe it followed the reference, because the page looks similar. It violated the actual requirement. The correct method was to duplicate the complete source artifact first, then mutate only the explicitly authorized regions. That preserves hundreds of existing decisions automatically — and it reduces verification work, because unchanged regions never have to be recreated and re-proven.
More from the same record, pulled by shared topic — not written new for this page.
Classify the task before touching anything, because the classification decides the method. If the task is **generative**, the model may create freely.
When a user gives an AI model a reference file, the AI model must read it completely. The AI model must study the patterns and designs within the file.
An AI model must show results before making changes. The user decides if something is wrong.
When the user specifies how a task must be performed, following that procedure is part of the task. Producing a similar-looking result through a different method is not equivalent.