Orchestration graphs for an AI-native classroom

1. The model

Dillenbourg (Orchestration Graphs, EPFL Press) puts time on the x-axis and social planes on the y-axis: individual, group, class, plus periphery, community and world. Each activity occupies exactly one plane. Edges carry operators, controls (branch, loop) and a pedagogical label. FROG made operators first-class nodes, because an operator on an edge cannot split, cannot run without a source, and produces three different random groupings when it feeds three activities. “Orchestration can be described as manipulating data structures. Each operator implements an orchestration act” (p. 101).

2. Compared with Minerva's Forum

Forum is the closest existing system, and the comparison is based on Stian's time at Minerva. Both start from the same premise: a live online class should be designed in detail beforehand, not improvised around a video call. Forum supports detailed lesson plans with timed interaction segments, breakouts with some configuration (such as students per room), a few student tools (polls, quizzes, workbooks) and teacher dashboards. The professor sends students to breakouts, brings them back and highlights the written work of different groups. What it lacks is the dataflow: each step stands alone, and nothing one activity produces decides who works with whom in the next.

3. Data model worth keeping

// output of a social operator
{ group: { "1": ["ana","ben"], "2": ["cy","dee"] },
  role:  { skeptic: ["ana","cy"], scribe: [...] },
  home:  { "A": ["ana","dee"], "B": ["ben","cy"] } }
// each group activity declares: groupBy: "group"

Activity contract: a JSON-Schema config, an initial document, a merge function that seeds each instance (per student, group or class) from incoming data, and a runner that changes state only through four verbs (listAppend, objInsert, numIncr, specialize), plus a product formatter, a dashboard and example data. The sync backend is therefore swappable (ShareDB → Yjs or Loro, one room per instance), and AI-generated activity types can be sandboxed behind a postMessage version of the verbs. Dataflow is lazy: an operator runs when its downstream activity opens, on a snapshot of what students had produced. Preview runs any graph with N simulated students.

4. The pedagogical layer FROG dropped

FROG dropped the labels because the engine had no use for them. An AI designer can use them as machine-readable design rationale. Elasticity: an edge's value decays with the lag between its two activities; the ArgueGraph debrief “miserably failed” when it was postponed by a week.

5. What in-class activity is for

“Add a breakout” is not a purpose. Each script below targets a different learning process. KLI (Koedinger, Corbett & Perfetti 2012) distinguishes memory and fluency, induction and refinement, and understanding and sense-making. Its asymmetry hypothesis holds that simple knowledge components (KCs) gain most from practice and feedback, while complex ones with rationale gain from sense-making; its evidence map lists collaboration and accountable talk on the sense-making side. ICAP (Chi & Wylie 2014) ranks engagement passive < active < constructive < interactive, where interactive means both partners construct and build on each other.

6. Joining the tutor and the classroom

Intelligent tutors and CSCL have rarely been combined. They have different units (the knowledge component vs the group's discourse), methods (learning curves and A/B tests vs design-based research and interaction analysis) and venues (AIED/EDM/L@S vs CSCL/ICLS). The exceptions are instructive. Olsen, Aleven and Rummel built collaborative versions of a fractions Cognitive Tutor and found collaboration helped conceptual items while individual work helped procedural fluency, which is KLI's asymmetry applied to social planes; Jennifer Olsen later co-authored the FROG papers. Holstein, McLaren and Aleven's Lumilo gave teachers real-time tutor analytics in the room. Campus's own engineering post (From Sizzle AI to Campus, March 2026) names KLI and ICAP as guides, describes extracting KCs from course materials and composing a practice feed, lists instructor information and classroom interaction as directions, and notes that interactive learning remains unresolved. The orchestration graph is the missing piece: it decides which KCs are worked on in which plane, and it routes data in both directions.

The carrier: Learning Items and plane-scoped pages. To move work between async and live, FROG (2018–19) used Learning Items: typed, addressable content units (an idea, an image, a rich-text document, a spreadsheet, even a whole embedded activity), each with its own edit history, that any activity can display and any operator can move. They can be dragged between activities and embedded inline in rich text. The wiki built on them gave pages a plane. A class page is one shared document; a group page is instantiated once per group; an individual page once per student, so the same “weekly reflection” link opens a different page for each student. A page can be pulled into a live session, its contents routed by operators, and the session's products written back. This is the missing async ↔ class link from the Forum comparison. In a rebuild it maps onto ProseMirror/Tiptap node views over Yjs or Loro.

Precedents from Stian's recent projects. Both are single-learner systems built around the same before-and-after loop. Alif, an Arabic reading trainer, keeps a word-level FSRS model and generates sentences and stories from words the learner already knows. Reading then exposes difficulties that become the next practice, so practice prepares a sustained encounter with a text and the encounter feeds practice. Petrarca, a reading tool, models what a reader knows at the level of claims, surfaces texts at about 70% novelty, and tracks a book chapter by chapter. Moved to a classroom, the claim-level model is what makes “unique points of view” computable: which claims a student holds that their classmates don't, and therefore whom they should talk to.

7. Scripts for specific Campus courses

Built from the 2026 catalogue descriptions. Notation: π1 individual, π2 group, π3 class; [S] social operator, [A] artefact operator. These are design sketches, not observed practice.

8. Authoring as compilation

The FROG record on authoring is unambiguous. In an external usability test (Jan 2018) no participant managed to create a chat activity. The most committed outside teacher sent a ten-step list and concluded: “Building successful graphs is difficult.” The working pipeline was teacher prose → Stian compiles the graph → walkthrough video. That compile step is now the model's job, and the 2016 functional spec already listed “AI modules to distil educational scenaria… from informal teacher lesson plans”.

9. The analytics layer nobody built

• Why now. Most of these states show up only in language (free-riding, groupthink, a misunderstanding repaired in dialogue), and LLM classifiers make them observable. KC mastery from the tutor gives the matrices a prior state on entry. Estimating the matrices needs the same graph run on many students, which an institution with shared course designs and quarterly terms has.
• Experiments embedded in graphs (book point 39). Swap one node or operator, keep the rest, and compare transitions and downstream KC learning curves: random vs disagreement pairing, AI facilitator vs none, split vs shared prompt.
• Timing. Faucon, Olsen, Håklev & Dillenbourg, JLA 2020: real-time prediction of activity completion. “Do I close for 20% of the learners? Should 80% wait?”
• Repair economy (ch. 6). Drift (time), heterogeneity (social) and uncertainty (diagnosis) each trigger graph modification above a threshold. An assistant proposes the repair and faculty accept or bypass: “design for what cannot be anticipated”.

10. Where AI sits

Engine requirements from FROG's failures: operators must rerun when a student joins late (FROG's ran once); activities need per-group pacing (lockstep was never solved); the teacher always controls advancing (teachers rejected a timeline that advanced itself); mobile first; one phone per group can turn a transcribed discussion or a photographed worksheet into a product.

11. Open questions

1. Can a course design carry typed activities with data passing between them, plus edge labels and rationale, rather than timed segments?
2. Which KCs belong in class at all? KLI's asymmetry suggests sense-making KCs in class and procedural KCs in the tutor. How would that allocation be tested?
3. What counts as evidence that a session worked: its products, the transitions into the next activity, or downstream KC learning curves and assessment?
4. How much routing should be automatic and how much proposed to faculty (the balance of control)?

Current test bed: Tadpole, a live layer for Slidev talks (QR join, polls, pairs, quizzes, AI views). The next step is activities that read each other's products, starting with pairing by disagreement.