Guide, practise, transfer
Connected guided explanations to adaptive skill practice and separately recorded baseline, transfer and delayed checks, with validated versioned browser-local progress.
Marcus Chess
Learn to see the next decision.
An interactive chess education platform connecting guided lessons, adaptive practice, playable Stockfish games and reflective game review.
Chess education / independent engineering project
Next.jsTypeScriptchess.jsStockfish 19Web WorkersVitestPlaywright
Who it is for
Beginning chess players learning to recognize threats and make considered decisions. The pedagogy is a design hypothesis awaiting educator and learner review.
The problem it solves
A correct move alone cannot show whether a learner understood the threat or memorized a board. Practice, feedback and assessment need distinct roles, with a path back from a real game to a focused exercise.
What I personally built
Built the lesson and adaptive-practice experience, browser-local progress, Stockfish worker integration, playable games, coached review and source-separated evaluation exports.
Play and review locally
Integrated the pinned Stockfish WASM worker for games, cancellation/restart and bounded analysis. Game reconstruction and selected-decision review lead into reflective retries.
Qualify coaching evidence
Added bounded tactical proofs and explicit uncertainty for alternatives, sacrifices and special moves. Opt-in timing and exports preserve the source of each observation.
How the product works
5 inspectable stepsFresh screenshots and recording from the public Marcus deployment verified on 27 October 2026. Lesson and real Stockfish play use a new browser-local session. Technical references and earlier suite results remain pinned to Phase 2C; no learner outcomes are claimed.
Choose a detail step to use zoom and dimming.
The public deployment now uses Marcus branding. This fresh capture shows its actual lesson; progress is confined to a new browser context.
Guided inspection
Step 1 of 5
Notice a threat before moving
The guided lesson explains a rook threat visually, then asks the learner to practise and check a new position.
Authentic application captures. These controls explore screenshot evidence; they do not operate the app. Use Zoom to evidence, dim the surroundings, or open the full-resolution image. The text transcript below works without JavaScript.
Read every walkthrough step as text
1. Notice a threat before moving
The guided lesson explains a rook threat visually, then asks the learner to practise and check a new position.
The public deployment now uses Marcus branding. This fresh capture shows its actual lesson; progress is confined to a new browser context.
Inspect corresponding screenshot (opens in a new tab)2. Practise a decision, not a click sequence
Adaptive practice separates baseline, guided exercises, unfamiliar transfer checks and delayed review.
Hint-assisted retries and familiar boards are not treated as independent evidence of mastery.
Inspect corresponding screenshot (opens in a new tab)3. Play an actual Stockfish game
A local WASM worker provides legal engine responses through the UCI adapter. Players can choose a side and a descriptive difficulty preset.
Presets bound engine resources and native skill. They are not calibrated human Elo levels.
Inspect corresponding screenshot (opens in a new tab)4. Reflect on a game decision
Coached review reconstructs a game, selects decisions and offers reflection, candidate moves and qualified tactical feedback.
This fresh capture reviews the actual browser-local demonstration game after a real Stockfish reply and resignation. Bounded proof does not establish strategic understanding or human coaching quality.
Inspect corresponding screenshot (opens in a new tab)5. Keep learning evidence interpretable
The fresh Local results screen requires a completed baseline before showing results. Source-separated assessment export is supported by the earlier attributed verification.
This fresh session has no completed assessment; the screen truthfully asks the learner to finish the check. No evaluation totals, real pilot or learning-effectiveness result is invented.
Inspect corresponding screenshot (opens in a new tab)
Find a threat, then play Stockfish
Read-only public deployment verified on 27 October 2026. Lesson and game progress stay in a fresh browser context; all non-read HTTP methods were blocked. No measured learning improvement is claimed.
Recording transcript
- Read Is Your Piece Safe?, start the lesson and identify the rook on d8.
- Read the real feedback and inspect the available learning practice.
- Start a new game against the real Stockfish worker, play e2–e4 and wait for its legal reply.
- Resign the demonstration game, open decision review and inspect local learning results and their limitations.
Read the Phase 5D capture and verification report
Inspect the mobile captures (390 × 844)
Difficult engineering decisions
Practice success is not assessment mastery
Hints, familiar positions and gameplay review stay separate from unfamiliar checks. Earlier exposure is excluded conservatively; a simulated delayed test is not a retention study.
A worker result needs a lifecycle
UCI lacks request IDs. Cancellation destroys the worker; generation checks reject late results. Bounded searches and explicit restart trade resources and speed against analysis depth.
Coaching must admit uncertainty
Mechanical attack facts cannot evaluate every sacrifice. Bounded mate proofs support qualified feedback; deeper tactical and strategic value remains unassessed. Descriptive difficulty needs actual beginner calibration.
Architecture behind the interface
Deterministic learning domain
Exercise selection and evidence rules live outside React. Versioned progress validation prevents malformed data or repeat attempts from becoming unsupported mastery.
src/domain/adaptive/session.ts (opens in a new tab)Typed UCI worker adapter
A replaceable worker port bounds resources, validates legal principal variations and handles timeout, cancellation, disposal and restart.
src/engine/uci-engine.ts (opens in a new tab)Selected-decision review
Full-game replay and review decisions remain distinct from assessment. Familiar-position observations do not enter unfamiliar evaluation denominators.
src/domain/review/decisions.ts (opens in a new tab)Evidence, with boundaries
Read Phase 5 verification reportUpstream results below are attributed to their recorded checks. This portfolio change does not rerun every application suite.
Phase 2B unit and real-engine checks
The upstream Phase 2B report records 177 passing unit tests across 13 files and one default engine skip. A separate actual pinned-engine integration passed four presets and three benchmark searches. These are prior upstream checks, not rerun by this portfolio change.
Production browser checks
Phase 2B records 34 Chromium scenarios; the later Phase 2C report records 46 while preserving learning/engine logic. Lint, strict TypeScript and the isolated Webpack build passed in those recorded upstream checks. None establishes physical-device readiness or educational effectiveness.
Authentic current experience evidence
The retained Phase 2C captures show the real lesson, learning-practice entry, playable game, reflection review and local evaluation. Synthetic assessment totals and the 14-position benchmark are test material. No real learner pilot, retention improvement or 10× effectiveness has been measured.
What remains unverified
Working implementation and release readiness have different evidence requirements.
- Educator review and a real learner pilot have not been completed. No measured learning improvement, calibrated Elo or validated 10× effectiveness is claimed.
- Physical Android/iPhone, Firefox/WebKit, spoken screen-reader and battery/thermal validation remain pending. Chromium viewport checks do not substitute for those checks.
- Bounded tactical proofs cover a constrained set of outcomes. Strategic exchange value, deeper tactics and beginner difficulty calibration need further review.
- Earlier release evidence left Stockfish corresponding-source packaging and application licensing review unfinished. Public lesson and Stockfish playback now work at the verified URL; this capture does not establish closure of those release-compliance gaps.
- Phase 2C adds a local source/build receipt and release configuration. Complete recipient-accessible Stockfish source/build inputs and an approved application licensing treatment remain unresolved; those preparation artifacts are not a completed public release.
Follow the source
Open repository (opens in a new tab)Links point to the current committed Marcus source. Phase 2B and 2C reports are historical technical verification, not learner studies. The latest source contains release preparation, while licensing/corresponding-source and educational-validation gates remain open.
Phase 2B technical verification (opens in a new tab)Later experience checks and pending learner/device work (opens in a new tab)Source-separated learning evidence (opens in a new tab)Stockfish integration and licensing obligations (opens in a new tab)Release blockers and corresponding-source review (opens in a new tab)