Generate a response
- Input
- Model
- Token
- Token
- Token
- Response
Autoregressive generation builds a response token by token.
03 / Decision intelligence · Research & Development
Decision intelligence without token-by-token generation.
Titrit is Base Workers’ research program for efficient structured decisions: typed outputs and probability distributions, rather than a generated conversation.
A different inference direction
A decision is not
always a sentence.
Many application tasks need a choice, a score or a classification. Our research asks what a model designed for that output should look like.
Return a typed decision
- Structured input
- Tokenizer / Encoder
- Titrit Decision Core
- Logits
- Softmax / Probability distribution
- Typed decision
The direction is non-autoregressive decision inference. The intended model returns logits; the runtime converts probabilities into a typed result.
Architectural direction, not a released model. Efficiency advantages must be measured against appropriate baselines.
Decision primitives
Small outputs.
Explicit meaning.
The typed decision contract keeps application logic structured and separates model predictions from what an application does next.
Choice
Select among specified alternatives with a probability distribution over those alternatives.
Score
Return a decision on an ordered scale, with defined semantics and probabilities.
Noul
A binary decision primitive for yes/no-style questions, rather than a free-form answer.
Classification & risk
Research into typed classification and risk decisions with explicit categories and evaluation.
Contextual PII classification
Study whether context can distinguish sensitive spans and classes. PrivacyGuard owns the policy and transformation.
Probability calibration
Evaluate whether confidence corresponds to observed reliability across tasks, languages and conditions.
First strategic specialization
Understand the context.
Keep the boundary.
Contextual sensitive-data classification is the first strategically important specialization for PrivacyGuard.
A future private Titrit integration could classify sensitive information without sending that context to a general-purpose external AI service.
Titrit detects and classifies. PrivacyGuard decides and enforces. Research predictions do not replace application policy or grant permission to reveal data.
Understand the privacy boundaryMoroccan depth, broader ambition
Language is more
than a language tag.
Darija, Arabic and Latin scripts, transliteration, spelling variation and mixed-language inputs are central research questions.
The long-term aim is global multilingual capability with unusually deep Moroccan specialization, including separately evaluated Amazigh varieties. Language support is a target, not a qualified capability today.
Benchmark work must evaluate each language and variety separately. A synthetic PII experiment does not establish general language understanding.
Research status
Experiments.
Not product claims.
Titrit is currently an R&D program. No Titrit-native production model has been released, and no language is production-qualified.
TCDT research
Titrit Custom Decision Training is a proposed research program. Its components are hypotheses, not a proven training method.
Native model experiments
Plan and evaluate independent decision-model baselines. Historical reference experiments are not released Titrit models.
Calibration
Study confidence quality, decision invariance and selective behavior with explicit evaluation criteria.
Structured decision evaluation
Measure Choice, Score and Noul behavior using controlled comparisons, grouped splits and reproducible records.
Moroccan benchmark work
Develop measurement protocols for scripts, language mixtures, local context and generalization, with provenance and leakage controls.
Contextual PII detection
Research candidate discovery, contextual classification and exact span boundaries without claiming unverified accuracy.
Let’s build what comes next
Make decisions measurable.
We welcome focused conversations about structured decision research, Moroccan language evaluation and contextual privacy detection.