TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · The Brief · AGE 17—19 · POSITIONAL · HELD

Criteria, or only plausibility

Criteria go in the brief before generation, and the verification stays outside the model.

The dilemma

The brief is ready to run: type, constraints, materials, all of it. One thing is missing — how you will judge the answer. Writing criteria takes five more minutes and forces you to say what "good" means before you have seen anything. Leaving them out is faster, and the model will return something that looks right.

The choices

Run it now and judge the outputs by first impression. Or stop, write the criteria — "must keep the view axis from living room to court", "must not add a wet area", "must improve cross-ventilation" — and run it with the criteria inside the brief.

The consequence

Without criteria, the model aims for plausibility, because that is the only target it has. You get a description and grade it on feel. With criteria, you get an assessment: the model checks its answer against the lines you drew, and so can you. That assessment is a first pass, not a verdict. The model that made the scheme shares the blind spots of the model now checking it, so you confirm each check yourself, on the drawing. The second path is slower. It is also the only one that gives you a result you can defend when someone asks "why this one?"

The case

Two schemes for a school on a tight urban plot. Scheme A looks better on the screen. The criteria said: every classroom cross-ventilated, no dead-end corridors, the playground visible from the staff room. Scheme B meets all three — confirmed on the plan, not on the model's say-so. Scheme A meets one. Without the written criteria, A goes to the jury — and the jury finds the dead-end corridor in four minutes.

Our position

Criteria belong in the brief, written before generation. Naming how you will judge the answer gives the model a target other than plausibility, and gives you a test you can run on what comes back. A written criterion asks; it does not verify. The check that counts is made by something independent of the generator — you on the drawing, a measurement, the CAD geometry.

Why we hold it

It is the layer students skip most, and the one that changes the reply most: from description to assessment. It is also the layer that carries the designer's judgment into the machine's work, and that is where we think authorship lives — provided the judgment stays with the designer and is not handed back to the machine as "it checked itself."

The strongest objection

A model assessing its own output against your criteria is not an independent tester. It can misread the output, misapply the criterion, or rationalise a failure it cannot see, because the evaluator shares the generator's ways of being wrong. Explicit criteria can therefore create the look of verification without any. "Does the kitchen open onto the courtyard?" may be answered by the same model that misdrew the kitchen. If students learn that criteria make the model a tester of its own work, the card has taught the opposite of what it meant.

What would make us revise it

If self-assessment against stated criteria turned out, on student work, to catch the errors an independent check catches — measured, not felt — we would let the model's check stand as verification. Until then the criteria go in the brief and the verification stays outside the model. We also watch the older objection: if criteria-first briefs measurably narrow student work without a matching gain in defensibility, we would teach criteria as the second pass, not the first. Reviewed against studio output each edition.

Prove it

Add clear success criteria to a brief you have already run. Run it again, and show how the reply shifted from describing the scheme to assessing it — then show one criterion where you checked the model's assessment against the drawing, and say whether it was right.

Take it to crit

Before generating, could the student state the criteria that would make the result usable or not — did they judge the outputs against those criteria rather than first impression — and did they check the model's own assessment against the drawing?

How it works

A reasoning model plans before it answers, and its plan is built from whatever the brief supplies. Criteria give it a target to check against while planning. That is why the answer changes shape and not only content. What the criteria do not supply is an independent eye: a written instruction asks; the system's own controls enforce, and the control here is a check the generator did not make — yours, a measurement, a separate tool. The six-layer brief on this map gives criteria a layer of its own for exactly this reason.

What this idea builds on

What this idea opens up

Sources

Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.

Age grows from 11 at the centre to 22 at the edge, and six sectors show the learning strands. Tab into the map and the arrow keys step from idea to idea, following the links where there is one. Enter opens the idea under the cursor, and E reads out its links and the reason recorded on each. Press slash for Search, question mark for the full key list, and Escape to leave. Open Ideas for the complete readable list, including what each idea builds on and what it opens up.

LOGIKA · RBDS AI LAB INDIA
ON-RAMP · AGE 11 · FIRST ENCOUNTERS, NOT GATES — IDEAS · — DEPENDENCIES
DONE
OPENS NEXT
SOLID — STANDS ON