Define the output first
An application may need to classify an enquiry, assign an ordered score or answer a binary question. These tasks have a structured output that can be defined before inference.
Titrit is Base Workers’ research program for this kind of decision intelligence. It is not a released general-purpose chatbot or a qualified production model. Its typed decision contract includes Choice, Score and Noul.
Follow the inference path
The architectural direction is structured text or JSON → tokenizer / encoder → Titrit Decision Core → logits → softmax / probability distribution → typed decision. The intended model returns logits; the runtime constructs the typed result.
The direction is non-autoregressive decision inference rather than generating a final response token by token. Whether this yields a useful efficiency advantage must be measured against appropriate baselines. No latency, cost or accuracy advantage is claimed here.
Calibrate the confidence
A probability is useful when its meaning is explicit and its reliability has been evaluated. A confident prediction is not proof of correctness. Calibration research asks whether stated confidence corresponds to observed behavior.
Evaluation also needs to test changes in option order, context, language and script. Grouped data splits and recorded provenance help distinguish generalization from repeated examples. TCDT—Titrit Custom Decision Training—remains a proposed research program, not a validated training method.
Study Moroccan context deliberately
The long-term aim is global multilingual capability with deep Moroccan specialization. Darija in Arabic and Latin scripts, transliteration, spelling variation and mixed-language text are central research questions. Amazigh varieties should be tracked and evaluated separately.
These are targets, not supported language claims. No language is production-qualified today. Moroccan benchmark work needs to distinguish general language understanding from the narrower task of PII classification.