Idiograma
Spanish tense and aspect made visible: any sentence becomes an interactive timeline, built on a homegrown linguistic theory.
Problem
- The only shared language teachers and students have for talking about language is grammar names, and those names carry zero meaning about when to use the thing.
- Students memorize what they can name and never learn the part that matters.
Approach
- A fine-tuned language model returns a structured analysis of every verb, and the browser draws it as an interactive timeline.
- The rendering engine implements a written spec (TimeZones, now at v1.5) with fifteen numbered rulings that work as case law.
- The governing principle is semantics over grammar: morphology is evidence, not criterion.
- Model output is a candidate until I ratify it; a validator enforces the spec in CI with zero errors required.
- An annotation workbench renders every reading as a real diagram before I judge it.
Outcome
- Live, mid-renovation, and the model is the weak link.
- The idea landed with a real teacher. The gap is accuracy and polish.
Try it — this is the real thing
Open full ↗Live production app. The deployed model is the pre-renovation fine-tune, so some analyses will be wrong; the retrain is the next phase.
Idiograma turns a Spanish sentence into a picture of what its verbs do to time. Type any sentence at idiograma.com; a fine-tuned language model returns a structured analysis of every verb (its perspective, direction, timeline, aspect), and the browser draws it as an interactive timeline. Three parallel worlds: Realidad, Realidad alternativa, Atemporal. The objective present sits at the center, pasts and futures on either side, and arrows show the point in time each event is stated from. Verbs only; everything else is stripped.
The problem it attacks is the nomenclature. The only shared language teachers and students have for talking about language is grammar names (“pluscuamperfecto”, “present progressive”), and those names carry zero meaning about when to use the thing. Students memorize what they can name and never learn the part that matters. The positioning I settled on: teachers teach what sentences mean; Idiograma makes it visible. The buyer is the teacher, and the product exists to empower teachers, not replace them.
The app encodes my own linguistic theory, and the theory is the product. The rendering engine implements a written spec (TimeZones, now at v1.5) with fifteen numbered rulings that work as case law. The Vantage Rule: arrows encode the temporal standpoint the sentence states, never causation. La Penumbra: certainty is vertical position, so podría ir sits a quarter of the way toward the other world and the open reading of ojalá pudiera sinks to the midpoint. The governing principle is semantics over grammar: morphology is evidence, not criterion. The wanting in quisiera saber si este plato lleva gluten is real; the subjunctive is courtesy.
Ground truth is a corpus, and nothing enters it without me. 620 sentences and 786 readings live in git as one JSON file per sentence, with a ratification status on every reading. Model output is a candidate until I ratify it; a validator enforces the spec in CI with zero errors required. An annotation workbench renders every reading as a real diagram before I judge it, because an idiograma judgment should never be made without seeing the idiograma. Minimal pairs are the unit of work: que venga and que viniera differ in exactly one thing, certainty, and the diagram shows it.
Status: live, mid-renovation, and the model is the weak link. The deployed analyst is a frozen fine-tune that still speaks the old schema, kept usable by shims, and its accuracy is the thing to fix. Ratification of the readings has only begun. Validation so far is one teacher interview, which went well; the rest is untested assumption, consciously so. The next phase is an analyst agent and a retrain from ratified readings only, then school workshops. The idea landed with a real teacher. The gap is accuracy and polish.