Reference
A challenge's Type fixes the rules the agent runs the conversation by. This page covers what every type has in common, then each of the five in turn.
Applies to every type
The type's rules below, then the published version's Scenario, Constraints, and Goal. If the assignment sets a Theme, that narrative wrapper is read too, on every turn. The template's Description is shown to people in lists and on cards; the agent never reads it.
After every student turn a judge checks the conversation against the type's checkpoints, below, and the session ends once every required one is met. If they never are, the session ends after 12 student turns. A version published with a time limit adds a third ending: the agent works a warning into the conversation, in character, once the student has used about four fifths of the time, and at the limit the session closes and is graded on the work done, showing as Timed out. Only time spent between turns counts, capped at fifteen minutes per gap, so a student who leaves and comes back the next day does not lose the time they were away. Either way the agent gives a closing reply before the grading starts. A student can leave and resume a session, or abandon it, but cannot force it to finish early. No further turns are possible once the assignment's closing date passes.
Once the conversation ends, a judge reads the student's messages against the type's rules, the Scenario, Constraints and Goal, and the teacher's Evaluation criteria. It decides pass or fail, gives a score from 0 to 100, and writes feedback. A scoring step turns that into the session's final score, and, where shown, four fixed dimensions: conceptual accuracy, communication clarity, problem-solving approach, and completeness, each with its own rationale. A session shows as Completed when every required checkpoint was met; otherwise it shows as Completed (partial), and still keeps its transcript and feedback. The Theme never reaches the judge or the scoring step, so grading stays the same whichever narrative was used.
Each type has a fixed list of checkpoints, the observable moves the judge looks for. When the student meets one, the conversation shows a milestone marker and the report records the turn. When the student makes no progress for three turns, the agent works a hint into the conversation, in character, at up to three levels of directness; the report shows the hint level reached.
Type 1
The student works through a scenario by questioning; the agent answers, pushes back, and never hands over the answer.
The agent answers what the student asks, then returns a question that opens the inquiry further: surfacing the assumptions behind the student's view, asking what would have to be true for it to hold, and putting the strongest counter-argument to them. It never states the conclusion for the student and never argues a side of its own.
Work toward a position you can defend: state it, answer the strongest objection to it, and address the people your choice affects.
The student has stated a defensible position, answered the strongest objection to it, and addressed the people their choice harms.
Goal: Reach a defensible position through Socratic questioning.
Describe a situation with a genuine trade-off and no single correct answer, so there is room for the student to be pushed on their assumptions.
Type 2
The agent argues a position the student disagrees with; the student must build its strongest form before they may attack it.
The agent presents its position and opening arguments, then requires the student to strengthen its case, a steelman, before any rebuttal. A straw version is rejected and met with a harder argument. Once the steelman is stronger than the position as first presented, the agent opens phase two: the student rebuts, and the agent defends by targeting the weakest point in each rebuttal, calling it out if the student slips back to attacking a weaker version than the one they built. It concedes a point that is genuinely good.
Build a steelman, then rebut that strongest version, for the configured number of rounds, or two rounds when none is configured.
After the configured rebuttal rounds, or two rounds if none is configured. The grade is the quality of the steelman and whether every rebuttal engaged that strongest version, not whether the student won.
Goal: Build the strongest form of an opposing position before rebutting it.
Name the position the agent will argue and hold. Pick one the student is likely to disagree with, so the steelman is real work rather than a restatement of their own view.
Type 3
The student writes an argument; each round the agent attacks its single weakest point and the student revises, explaining what changed and why.
The agent asks for the student's argument, then each round identifies and attacks the single weakest point, never several at once, so the student repairs each problem before the next is exposed. A revision that papers over the weakness with a surface change gets the same point re-attacked, more specifically.
Revise your argument after each round and label every revision "What changed: X. Why: Y." A revision with no annotation, or "I made it better", is not accepted.
After the configured number of rounds, or three when none is configured, ending with a short verdict on how the argument moved from its opening to its final form. Grading looks at the annotations and whether each revision addressed the attacked weakness, not the final argument alone.
Goal: Repair an argument round by round under pressure, naming what changed and why.
Pose a question the student can take a clear initial position on, with enough real weaknesses in any first answer to sustain several rounds of attack.
Type 4
Given "X did not happen," the student traces first-, second- and third-order consequences while the agent tests every causal link.
The agent presents a factual baseline and the counterfactual premise; the student traces consequences rather than arguing for or against the premise. It asks for first-order effects, then tests causal necessity: would that effect have followed from the premise, or would a compensating factor have produced it anyway? It then asks what follows from that chain, again for a third order, or accepts an explicit statement that the chain is too uncertain to trace further. It introduces one real complicating factor the student's chain did not account for and requires them to integrate it.
Trace consequences to the configured depth, weigh compensating factors, integrate the complication the agent introduces, and name at least one genuine uncertainty.
The chain reaches the configured depth, or three orders if none is configured, with the complication integrated and an uncertainty named. Grading looks at the coherence of the chain and whether compensating factors were weighed, not whether the outcome is one the agent agrees with.
Goal: Trace a coherent causal chain from an altered premise, weighing compensating factors and naming uncertainty.
State the factual baseline and the altered premise ("X did not happen") clearly enough that first-order effects are obvious, leaving the harder second- and third-order effects for the student to work out.
Type 5
The student builds an analogy that explains a concept to a named audience; the agent finds where it breaks and the student refines it.
The agent asks for an analogy that explains the concept to the audience the challenge names, and checks it fits that audience. It then stress-tests the analogy in two rounds: first pointing out where it is imprecise, then finding a place where it would give that audience a wrong intuition about the real concept. After each break the student refines the analogy or replaces it.
Build an analogy for the named audience, refine it through both stress tests, then name at least one limitation it cannot capture and say what the audience should be told instead.
A refined analogy has survived both breaks and the limitation is mapped. Grading looks at the fit to the audience, the quality of each refinement, and the honesty of the limitation, not the cleverness of the first attempt.
Goal: Build an analogy for a named audience, refine it under stress-testing, and map where it breaks down.
Name the concept to be explained and the audience it is explained to. The gap between the concept's real structure and that audience's background is what gives the stress-test something to find.
