scoriiu

that's a clean answer thanks. the 85% rule for BKT routing is more principled than most. the one thing i'd still watch is wether summit mastery still holds or decays over time, since BKT can't tell "build the schema" from "learned to pass the gate". thanks for the detail.

curious how the tutor handles a kid who's confidently wrong. from the description the ladder kicks in at the first sign of struggle, which sounds like it's built to prevent mistakes rather than use them. kapur's productive failure work points the other way, students who commit to a wrong approach and then get shown exactly where it breaks tend to keep the concept. the ones who got steered away before making the mistake don't.

does bloomybot ever just let a wrong approach play out and do the postmortem after? also when a struggling student gets rerouted to an easier skill, can you tell that apart in your data from one who was thirty seconds away from the useful kind of failure? the summit gate tells you they arrived. it doesn't tell you whether the help in the middle did the teaching.

did you skip simd just because the model's tiny? naive conv perf is honestly the only reason i haven't done exactly this for the cnn

Should be very doable. I ship a small CNN in a browser extension via onnxruntime-web and the model weights were never the bottleneck, the runtime was. The wasm backend adds a few MB of runtime before your first inference, so a 500kb model with a lean hand-rolled wasm build would actually beat most "tiny" browser ML deployments in total download.

One gotcha if anyone wants this in a Chrome extension: MV3 requires 'wasm-unsafe-eval' in the CSP for any wasm at all, which surprised me the first time a build that worked fine as a web page died silently as an extension.