Activity templates

Five ready starters turn the current scene of a lesson into an interactive activity your students open in a browser: a self-contained page, the instructions for their assistant and the format of the action trace they may export. You adapt the data, units, sources and criteria in a small configuration file. Preparing an activity runs no retrieved code, installs nothing, approves no lesson and submits no answer.

Template The learner
linked-representations Follows stable identifiers across an edge table, a peak plot and a candidate table
parameter-comparison Holds an input fixed, varies one parameter and compares baseline and variation
evidence-claim Reads a supplied source statement and picks the narrowest supported claim
diagnose-analysis Inspects analysis steps, picks a suspected flaw and compares supplied repairs
conclusion-robustness Compares assumption scenarios against an explicit decision threshold

The studium-activity-templates skill lets your assistant do these steps for you (see skills). A full walk-through is in the activity to report tutorial.

Prepare an activity

.venv/bin/python tools/experience_templates.py list
.venv/bin/python tools/experience_templates.py prepare \
  --template linked-representations --context operator-context.json --out activity-linked

--context is the public scene context, the JSON your assistant gets from the get_operator_context MCP tool (or from task.md in the file channel). --out must not exist. list --json prints the full catalogue; an entry marked proposed is not runnable.

Option Use
--config FILE.json Adapt the template; examples in examples/activity-templates/<template>/config.json
--language, --translations Presentation language and translation dictionaries
--supports FILE.json Figures, equations, chemistry, media (see media and visual support)

A configuration is closed: unknown fields, non-finite numbers, duplicate identifiers, mismatched template ids and incomplete sources are refused before anything is written. Slider baselines and variations must sit on the declared step grid. The parameter, evidence, diagnosis and robustness templates take either synthetic data (data_status: "synthetic teaching fixture", source_kind: "synthetic_fixture") or supplied evidence ("supplied source-backed data", "supplied_evidence"). The linked-spectrum starter stays synthetic: it reuses a checked example core and refuses a network whose numbers differ from its oracle.

The output directory holds:

File Content
activity.json The validated definition and its scene binding
index.html The self-contained local page
task.md Editable instructions for the assistant, with the exact report commands
learner-actions.schema.json The schema of the exported trace
provenance.json Input and generator hashes, and the stated validation limits

What the student sees and keeps

The page first asks whether to keep an in-memory trace. Without a trace the activity still works, and nothing is reconstructed later. With a trace, Export activity trace can download a partial trace before results are revealed; Complete and export final trace adds the completion event and locks the controls. The export carries the SHA-256 of activity.json, so a later definition with the same identifiers cannot reinterpret the events.

.venv/bin/python tools/learner_trace.py validate LEARNER_ACTIONS.json
.venv/bin/python tools/learner_report.py export REPORT_DIR \
  --actions LEARNER_ACTIONS.json --activity activity-linked/activity.json
.venv/bin/python tools/learner_report.py verify REPORT_DIR

The page has no network sink and no automatic storage, holds no learner identity or answer key, and keeps the learner's optional text only in an explicit download. The report is described in learner report.

Diagnose an analysis

In the diagnosis config, steps have an id, label and description; repairs have an id, label, explanation and a numeric result. Set flawed_result, result_unit, supported_repair_id and oracle.repaired_result, which must equal the result of the referenced repair. That is a consistency check, not a scientific verification. Selections are recorded as step_<id> and repair_<id>; the selected step is not graded automatically.

Test conclusion robustness

In the robustness config, each scenario has an id, label, explicit assumption and numeric value. Set baseline_scenario_id, metric.label, metric.unit and a decision with operator (gte or lte), threshold and a bounded claim. oracle.supporting_scenario_ids must agree with the values, in order, including equality at the threshold. A selection is recorded as scenario_<id>.

Identifiers, with their prefixes, are limited to 120 characters. The templates compare captured results; no equation or code is evaluated. Run real analyses with your assistant, keep their provenance, then regenerate the activity from those outputs.

Across disciplines

The diagnosis pattern fits a wrong denominator in ecology, a unit conversion in astronomy, an omitted blank correction in chemistry or an aggregation error in survey analysis. The robustness pattern fits alternative background estimates, missing-data assumptions or calibration choices. Keep domain terms and values in the config files, with units and source locators.

Limits

  • Three supplied scenarios do not estimate a probability of truth, and a threshold crossing does not establish causation.
  • The checks shown after reveal are instructions unless a separate workflow records that they ran.
  • Neither the page nor its report measures learning or approves anything for students: review the objective and interpretation of source-backed content before class.

Code: tools/experience_templates.py.

View the source of this page on GitHub