TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Studio Practice · AGE 18—21 · POSITIONAL · METHOD
Protect the load-bearing words
You write and lock the words, and the model does everything else.
When to use
Any document where a changed word changes what you are promising: a bill of quantities, a fee proposal, a scope note, a specification clause, a letter to a client or a municipal office.
The method
Write the load-bearing text yourself, first. Then mark it: the quantities, the exclusions, the dates, the fee, and the small words that carry weight — "shall", "excluding", "subject to". Brief the model to keep the marked text exactly as written — no paraphrase, no tidying, no synonyms — and to work only on what surrounds it: layout, headings, a covering note, a table. When the model returns, compare the marked text against your original word by word. Anything changed is reverted, however much better it reads.
Watch for this
The helpful paraphrase. "Excluding external works" becomes "external works to be discussed." The sentence reads better. The promise has changed. The model cannot tell which words are load-bearing, because to the model every sentence is just likely text.
The Lab's note
This is our method, from the studio's own documents: write the words, lock the words, let the model do everything else. Others let the model draft and then check. We find that checking a fluent paraphrase is harder than writing the sentence, because fluency hides the change.
Try it
Take one real document — a scope note from a studio, or a college project brief. Mark the load-bearing words. Ask a model to "improve the clarity" with no protection brief. Then run it again with the brief. Lay the three versions side by side and find where the meaning moved.
Prove it
Show one responsible sentence a model paraphrased into a different promise, and the protection brief that would have stopped it.
How it works
A model paraphrases because that is what it does: it continues text in the likeliest way, and the likeliest way is rarely your exact wording. An instruction to "preserve exactly" works most of the time, never every time. A written instruction asks; the system's own controls enforce. The controls here are simple: keep the protected text in a file the model never receives and paste it back yourself, or lock the cells and fields it must not touch, and then run a deterministic diff of the marked passages against your original. That is why the method ends with a word-by-word comparison rather than trust. The skill underneath is older than any tool: knowing which word cannot move. That is what a senior reads a junior's letter for.
What this idea builds on
What this idea opens up
- Nothing yet names this as a foundation.
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.