Deterministic pedagogical orchestration architecture with the LLM as linguistic executor¶

Design · All designs
Evidence · 2 claims (2 for) · 1 study (1 design), q2 · 0 of 1 report an effect size · 2 claims rest on one study

Description¶

The architecture separates pedagogical orchestration from text generation: instructional decisions (intent recognition, concept selection, chunk typing, prompt structure) are handled by deterministic modules, and the LLM is restricted to controlled disambiguation and surface-form generation. The pipeline proceeds through six stages: intent detection, concept detection, constrained retrieval, didactic resolution, prompt construction, and LLM generation. This separation, in the authors' words, "ensures pedagogical traceability (every decision maps to an explicit rule), enforces correctness through deterministic unit-testable logic, and makes the pipeline model-agnostic". It was designed to operate without commercial LLM budget or GPU infrastructure.

Design Implications¶

Context¶

Requirements¶

A formalisation of the instructor's own pedagogical reasoning: intents, a domain ontology, and an instructor-authored structured editorial corpus
Detection lexicons calibrated to the course vocabulary and language

Constraints¶

Adapting to a new course requires the instructor to formalise their own pedagogical practice, not merely configure a tool
Courses centred on theorem derivation, algorithmic processes, or code production may require different taxonomies and retrieval strategies

Target Learners¶

University students in a French-language dimensional modelling course

Learning Goals¶

Answering conceptual questions in dimensional modelling in line with the instructor's didactic practice

Claims¶

Examples¶

-

Key Sources¶

Laurent Brisson, Maria Teresa Segarra and Gregory Smits. (2026). A didactical-driven teacher assistant for a dimensional modeling course. https://arxiv.org/abs/2607.22598

Page metadata

Field	Value
`type`	`design`
`id`	`deterministic-pedagogical-orchestration-architecture`
`title`	`Deterministic pedagogical orchestration architecture with the LLM as linguistic executor`
`description`	`The architecture separates pedagogical orchestration from text generation: instructional decisions (intent recognition, concept selection, chunk typing, prompt structure) are handled by deterministic modules, and the...`
`status`	`review`
`generated`	by `process:settle-candidates` on `2026-10-10`
`trust`	unverified — no one has confirmed this page's content
`sources`	1 — `laurent-brisson-2026`

Frontmatter is the page's machine-readable record — see the Schema & Guide.