How a mind learns when
the mission can fail.
A quiz records whether an answer was right. SimLab records a decision committed before the physics answers back: the throttle a pilot chose, the orbit they bet on, the retry after the vehicle broke. The engine reads that stream and keeps a model of how one specific learner thinks. This log publishes what it measures, what the corpus shows so far, and the questions it is being built to answer.
Committed decisions
Every input is an evidence row: a committed answer, a control change, a dwell, a retry. Each one was chosen before the outcome was known.
The cognitive flight map
Mastery per concept by Bayesian knowledge tracing, forgetting-curve decay, misconception flags, and the next reachable step. A live state, not a grade.
Flight-controller agents
A deterministic agent router scores every learner event before any model is called: fast answers cost zero model calls, and only true escalations reach the tutor. FIDO, EECOM, PROP, GNC and CAPCOM read the same telemetry the learner flies and call the mission the way a real console would.
Recorded interventions
Recovery passes, prerequisite reteaches, consolidation holds, step-ups. The router also keeps a live read on the learner's state, in flow, frustrated or disengaged, and the tutor teaches to it. Every adaptive move is stored with the reason that triggered it.
Each entry below is computed from the production database when this page loads. Small numbers stay small on purpose: this is an early corpus, published honestly so it can be watched as it grows.
Reading the live corpus…
Open hypotheses, not findings. They are listed so researchers know what the data is designed to test, and so we can be held to them.
Does consequence make learning stick?
Is first-pass accuracy on a concept higher, weeks later, when it was first learned under a mission that could fail than under a consequence-free drill?
Is a narrated failure worth more than a success?
Do pilots who crash and receive a crew debrief recover faster on that concept than pilots who never failed it?
Can forgetting be predicted per learner?
Does a decay curve fitted to one learner's own flights predict their next miss better than a population curve does?
Does the engine read decisions, not language?
Do Arabic and English flights trigger the same interventions at the same mastery levels, so the model is measuring reasoning and not reading speed?
Every counter here reads the production database, never a slide. Learning-science and neuroscience groups who want to test these questions against the anonymised evidence stream can reach us directly; co-authored work is the goal.