Capabilities and limits

This page lists what Studium does today, for each kind of user, with the practical limit of each capability. Studium is a research prototype of adaptive teaching: its unit is an explicit path of concepts and tasks, not only a conversation with a model. A capability listed here exists in the code; it is not necessarily available on every surface.

For learners and self-learners

Capability What works Practical limit
Learn a topic in a conversation Your assistant turns sources you chose into a playable lesson with get_authoring_guide, prepare_sources and create_lesson; each graded item cites a verified quote Agent-drafted, not teacher-validated; up to 20 sources and 200 000 characters; no reproduction scene
Adaptive path Navigation follows prerequisites, measured mastery, hints and remediation Only as good as the lesson and the judge
Study over days get_study_plan says what is due across lessons; start_review_session plays it; rate_recall records a self-rating A concept no lesson can replay is listed but never proposed
Honest feedback Graded answers move mastery; without a judge you get the reference and rate your recall Without a judge, free text is not assessed
Metacognition Declared confidence compared with measured performance Prototype indicators, not a diagnosis or an institutional grade
Experiment journal Prediction, confidence, observation and conclusion kept through record_experiment and get_experiment_journal Private, opt-in, not graded
Self-study from references experience_packet.py passages, prepare, check, bundle --self-study Open book for quizzes and flaw audits; no teacher validation, no release
Reproductions The harness runs a capsule; Studium checks the run against its manifest; VS Code asks before running code Studium never executes; a run check is not a proof of execution
Peer review Peer rooms (open_peer_room, submit_peer_rating) aggregate ratings Rooms live in memory; a target stays hidden below 3 raters
Language set_session_language and activity dictionaries Dictionaries translate the texts provided; older interfaces are not all localised
Two-lane lesson Panels and panes, or prototype forms, handle reading, choices and verdicts without a model turn; the conversation remains available for written answers and questions, with get_companion_brief keeping the assistant informed The two-lane interface has automated checks but no learner trial yet; native MCP Apps and Claude Code host trials are deferred
Concept map and contest The map shows concept states and evidence; export_concept_map and tools/concept_map_report.py produce an inspectable local page, and a learner can contest a reached concept through check questions A map records session facts, not certified learning; the claim alone changes no mastery and skips no scene
Learner report tools/learner_report.py export builds a local ASTRA/MySTRA copy of actions you chose A partial trace: proves neither learning, nor the absence of help, nor authorship

Start with Learn with your assistant, the student guide and Review over time.

For teachers

Capability What works Practical limit
Literature to course Passage capture, a canonical plan and exact links to evidence No automatic scientific approval
Review and approve review_lesson shows a draft without its keys; experience_packet.py approve writes approval.json bound to the exact lesson bytes; dashboard library and review pages The record documents a review; it is not a signature
Follow a class class_report: per scene and per concept accuracy where a key or judge graded, hint reliance, calibration, completion Consented telemetry only; floor of 5 learners for each figure; covers the whole configured data folder
Learner contributions Proposal, review and release of learner bundles; discovered with list_community_bundles A course must have its bundle at <BUNDLES_ROOT>/<course>/lesson.json; release follows the teacher's decision
Activity templates Five portable activities: linked representations, parameter comparison, evidence and claim, analysis diagnosis, conclusion robustness Local viewers without an embedded agent; they do not advance the session
Media and visual support Mermaid, KaTeX, 2D SMILES, images, audio and video with provenance Coverage differs by viewer; generation is external and optional
Provenance view of a draft tools/experience_review.py renders a checked draft through ASTRA and MySTRA, so a reviewer can tell verified, declared and still-open elements apart A provenance view, not an approval nor a judgement of scientific validity
Exploration record tools/experience_exploration.py captures what a harness explored: sources, versions, named files, checks Hashes detect later changes; they prove neither the science nor an execution
Contextual help ? controls explain goals, prediction, confidence, evidence and adaptation, in English and French They state pedagogical intent, not proven benefits
Observatory Course inspection and review views, including MySTRA Local service, not an authenticated class portal

See the teacher guide, Review and approve and Follow a class.

For developers and harness builders

Capability What works Practical limit
MCP server 40 tools over stdio (25 lesson, 5 study, 3 authoring, 2 teacher, 1 form tool in prototype, 3 panel tools, 1 companion tool), prompts learn and review Each client has its own tools, account and session
Local HTTP API The 25 lesson operations plus rate_recall, schema discovery, bearer token, protection against duplicate submissions Loopback only; other study, authoring and teacher tools are MCP-only; active sessions end with the process
File channel A task and answer loop in .studium-session/ for a harness without MCP Formats and atomic writes must be respected
VS Code extension Lesson pane, chat participant, notebook cell, restore on reload, reviews in the status bar Its sessions are separate from an assistant's own MCP server
Shared harness panels apps/lesson-panel-ui supplies VS Code and MCP Apps renderers; apps/claude-code-mod supplies a native Claude Code pane, with orientation, voices, help, settings and concept-map views Prototype surfaces: automated contract checks, no native host trial and no learner use yet; capsule reproduction requires the VS Code terminal
ASB capsule Task card and reproduction protocol A capsule's tier does not attest a learner's run

See Harnesses, the MCP tools reference and Architecture.

Surfaces and maturity

Surface Status Meaning
Any MCP harness (Codex, Claude Code, Copilot, OpenCode, Cursor) and the file channel Supported Tested in CI, documented, the recommended way in
VS Code extension (lesson pane, chat participant) Supported Tested in CI, built as a VSIX by CI
lesson-server (HTTP, answer keys redacted), teacher dashboard, Observatory Maintained Kept working and tested; fewer new features
Tauri student app (apps/student-runtime) Frozen No new work; built only on manual request
apps/author-web Legacy Kept for the M1 demonstration; use the teacher dashboard instead
MCP Apps panel (apps/lesson-panel-ui) and Claude Code pane (apps/claude-code-mod) Prototype Shared lesson panels with automated contract checks; native host trials and the planned formative learner trial have not happened

The release workflow can publish to the VS Code Marketplace, Open VSX and npm once their tokens are configured; until then, a release is a GitHub draft with the VSIX and the server file. See Release.

Who holds the key

Class Scene types In a self-study lesson
ungraded intro, concept, meta_checkpoint Nothing is graded
author_keyed quiz, spot_the_flaw The key was written by the person who answers (open book), or by the agent that drafted the lesson
production explain_back, comparison, summary_audit, prompt_practice, communication_sim, discussion, ai_critique Without a judge model, the reference is shown and nothing is assessed; a judge (LITELLM_BASE_URL) makes the assessment external
externally_grounded reproduction Checked against an executed run

The notices

A self-study lesson shows a notice on its first screen. It depends on who drafted it and on whether a judge is configured.

Drafted by, judge Notice (English lessons)
learner, none Self-authored lesson: quiz and spot-the-flaw keys were written by the learner (open book). Without a judge model, production scenes are not assessed: compare your answer with the provided reference and rate your recall; that rating schedules your reviews and never changes your mastery.
learner, configured Self-authored lesson: quiz and spot-the-flaw keys were written by the learner (open book). Production scenes are assessed by the model.
agent, none AI-drafted lesson from your sources, not checked by a teacher: each item cites its passage; tell the assistant about a key that looks wrong. Without a judge model, production scenes are not assessed: compare your answer with the provided reference and rate your recall; that rating schedules your reviews and never changes your mastery.
agent, configured AI-drafted lesson from your sources, not checked by a teacher: each item cites its passage; tell the assistant about a key that looks wrong. Production scenes are assessed by the model.

A lesson that holds a reproduction adds: "Executed reproductions carry an external assessment."

More in Honesty and provenance.

Assets and display

The surfaces render images, SVG, LaTeX equations, code, CSV, protocols, chart specifications and executed results; scientific assets include orbitals, 3D molecules and spectra. Rendering libraries are vendored for local use (studium.assets = vendor, the default); a cdn option exists. When a renderer is missing, a fallback keeps the content readable without promising the same interaction. Modality preferences can reorder and emphasise assets without hiding content; they answer a stated need and do not diagnose a cognitive profile. See Language and supports and Media and visual support.

Auditability

Adaptive decisions carry their rule and modulator, and the graph and decisions can be exported as indicium claims and ASTRA records. This makes an inference inspectable; it does not make it correct.

Who does what

  • Studium keeps the concepts, scenes, expected answers, progression and run protocol.
  • The harness owns its tools and model, reads the skills, searches and runs in its own environment.
  • The learner predicts, chooses a variation, interprets and answers.
  • The teacher sets the scope, reviews and decides what is released.

Web search, Perspicacité, libraries and skills are not available because a lesson mentions them: the harness has to check its real connections. The agentic experiment tutorial shows how to use them within the lesson contract.

Open boundaries

  • Deferred, by decision: item variants and misconceptions, native MCP Apps and Claude Code host trials, HTTP delivery of class telemetry, concept identity by namespace, npm and Marketplace publication.
  • No authenticated multi-user teacher portal; the dashboard and the Observatory are local.
  • The VS Code pane and an independent MCP server do not share a session.
  • Passing software tests does not establish institutional compliance, multi-user deployment or teaching effectiveness. Start with an accompanied pilot and the rules in Privacy and limits.

View the source of this page on GitHub