AI Tutors and Learning Tools

Trace mastery, ground every card in a source, plan paths and materials, and suggest words from your own corpus.

What each project builds#

Factor	Study coach	Path finder	Autocomplete pad	Materials planner
Builds	Practice page with review dates	Feasible learning path	Local writing pad	Packing list and shortages
Runtime	Node.js 22.18	Python 3.10+	Node 22.18	Python 3.10+
Input	Deck JSON and attempt events	Resources, goals, budget	Corpus of id and text	Participants, materials, inventory
Output	index.html, progress.json	learning-map.html, path.json	report.html, vocabulary-model.json	packing-list.csv, shortages.json
Review loop	Download attempts JSON	`--decisions decisions.json`	Download vocabulary-model.json	`--decisions decisions.json`
Limit	Exact answer match only	Self-reported progress	One next token, two of context	One concurrent session

Command	What it does
`python3 scripts/project_test.py <project> --init <workspace>`	Makes a workspace. The first stage fails
`python3 scripts/project_test.py <project> --stage 1 --path <workspace> --strict`	Grades one stage
`python3 scripts/project_test.py <project> --all --solution --strict`	Checks the reference solution
`node --experimental-strip-types main.ts --deck ./fixtures/deck.json --attempts ./fixtures/attempts.json --initial-date 2026-09-01 --date 2026-09-03 --out ./study-output`	Builds the practice page and schedule
`python3 cli.py samples/input.json --output demo-output`	Runs path finder or materials planner
`node cli.ts sample.json output`	Trains the pad and writes output/report.html
`python main.py`	Runs the simulated tutor efficacy study
`npm run serve`	Tutor web app on :8090

Setting	Default	Note
`p_init`	0.2	Prior knowledge
`p_learn`	0.12	Learn rate per practice
`p_slip`, `p_guess`	0.10	Guess is 0.15
`master_threshold`	0.85	Mastered concepts are skipped
Catalog size	18	Exact planner maximum
`suggest` limit	5	Allowed range 0..50
Planner `compare`	[8, 12]	Participant counts to compare
Study `n_turns`	60	10 learners in each arm

Symptom	Cause	Fix
"Canonical answer must occur in the evidence quote"	Offsets miss the answer	Fix `start` and `end` in UTF-16 units
"Conflicting duplicate attempt ID"	Same id, other content	Give each attempt a new id
"Proposal lacks verbatim answer evidence"	Model quote is not in the source	Reject it. Keep the deck unchanged
"prerequisite cycle"	Resources need each other	Remove one prerequisite edge
"stale progress"	Input changed after export	Rerun, then edit the new decisions
"complete prerequisites first"	Progress skips a step	Mark the prerequisites done first
"inconsistent conversion cycle"	Two unit paths disagree	Fix the conversion factors
"Unsupported model schema"	`schemaVersion` is not 1	Load a model with schema version 1

p = BKTParams()p_init, p_learn, p_slip, p_guess

if correct:

num = mastery * (1 - p.p_slip)

denom = num + (1 - mastery) * p.p_guess

else:

num = mastery * p.p_slip

denom = num + (1 - mastery) * (1 - p.p_guess)

posterior = num / max(denom, 1e-6)

return posterior + (1 - posterior) * p.p_learnlearn transition

Task	Code
Grade one typed answer	`gradeAnswer(deck, cardId, response)`
List due cards	`dueCards(progress, date)`
Order prerequisites	`path_for(index, target, completed=None)`
Rank paths in a budget	`rank_paths(index, goals, budget, completed=None)`
Suggest a next word	`suggest(model: Model, prefix: string, history?: string[], limit?: number)`
Load a saved model	`importModel(value: unknown)`
Scale materials	`scale(model, participants=None)`
Find shortages	`shortages(model, participants=None)`

Answer match: NFKC, lower case, trim, drop end punctuation. Then exact match.
Box: correct adds 1 up to 5. Wrong resets to 0.
Interval: 2 ** (state.box - 1) days when correct. 1 day when wrong.
Path rank: most goals covered, then fewest minutes, then IDs.
Suggestion score: order * 1000000 + count. A longer history wins.
Consumables: quantity*participants.
Shared equipment: quantity*ceil(participants/share).
Unknown stock: unconvertible inventory leaves shortage unknown.

What does the tutor do when asked for the answer?: It asks a leading question.
What proves the tutor works?: A pre/post test delta with 10 learners over two weeks.
Does the study coach judge essays?: No. It uses authored questions and exact matches.
Does a model card go into the deck?: No. It goes to card-proposal.json for review.
Does the planner guess glue volume?: No. It never infers density or package size.