Filling parsed language back into controls lets users see what the system understood

What it is

Someone says “shorter, as a table.” Before send, the length slider jumps to 80 and format switches to table. The jump is wrong — they wanted 120 — and they can correct the slider on the spot, without gambling another sentence of “not that kind of short.” Parse-to-control fill-back writes the structured reading extracted from language back onto visible controls, so “what the system heard” becomes an object to inspect and fix, rather than something proved only indirectly by a finished artifact.

It is not the override rule itself. Fill-back is visibility: understanding is pinned to controls. Who counts in a conflict can be decided separately. Without fill-back, people cannot even see whether a conflict happened.

Why it happens

Language understanding happens in a parser nobody can see. Unexternalized, the only evaluation is reading the artifact: if the artifact is wrong, the cause may be parse, sampling, model competence, or the user not being clear — four causes in one miss. Controls are already a shared metric for quantities; using them as a projection screen for understanding pulls parse out of the four. A miss is then a miss on a specific tick, and the gulf of evaluation shortens to before send.

Fill-back also gives a cheaper lever. Editing a control is cheaper and more precise than another near-synonym. Once understanding is visible, correcting it need not go through another generation.

Studying it

The same sentence: one group gets fill-back and may edit controls before send; one group generates with no fill-back. Code: share of parse errors caught before send, whether the user can point to “which band it heard ‘short’ as,” match of the final artifact to intended quantity slots. Independent variables: whether fill-back is animated or highlighted, whether low-confidence parses use a ghost state rather than a hard jump, whether fill-back is undoable. Dependent variables: pre-send correction rate, error attribution (parse versus model), calibration of “it understood.”

There must be a parse gold standard (human-labeled slots), or “was fill-back right” has no ground truth. Hard jumps at low confidence create a new false confidence and should be read separately.

Where it stops holding

Propositions with no matching control (“don’t let legal hear this as an admission”) have nowhere to fill back; stuffing them into the nearest tone slider lies about understanding. When the parser is weak, fill-back writes noise as determinate state, more misleading than no fill-back. Token-by-token live fill-back thrashes before the sentence is finished; do it once at a pause or before send, not on every token. This entry does not treat fill-back as an override policy; if after fill-back the control becomes the only source, that is an extra rule.

Applying it

• Before send, write extracted quantities into controls, with a short mark “from this sentence.” Controls that were not extracted keep their previous value; do not zero them.
• Low confidence should be a ghost or a question (“as a table?”), not a hard jump. No undoes that fill-back.
• Fill-back must be undoable item by item. Undo restores the pre-fill value; the sentence stays; do not delete the whole sentence.
• Check: a sentence you know will be heard as the wrong band (“a bit shorter” often mapping to a too-small length in your system). With fill-back, someone should change that tick before send. Without it, the same error crawls into the artifact. Then add a ban no control can express — if it is stuffed into the tone slider, fill-back is lying.

Related

• Same group: L2.04.1 Natural language is suited to expressing intent · L2.04.2 Parameter controls are suited to precise adjustment · L2.04.3 Having both beats choosing one · L2.04.4 Natural language is strong at goals and constraints; parameter controls are strong at degree and numeric values · L2.04.5 Parameter controls provide a revertible determinate state; natural language does not · L2.04.6 When both exist, which one overrides the other must be explicit, or conflict is unpredictable · L2.04.8 A control’s value range is itself a display of the capability boundary
• Nearby: L5.01 Types of Explainability · L2.15 Ambiguity and Clarifying Follow-ups · L5.02 Local and Global Explanations
• Search terms: parse-to-control fill-back · grounded understanding · pre-send parse visibility