The answer is right. The working is not.
Try Repair Desk · Watch the 2:09 demonstration · Current source and setup
A learner can reach the correct answer through two mistakes that cancel. An interrupted practice session can also lose the distinction between an independent answer and one completed after a hint. Counterstep Relay keeps the work and the help history, not just the final answer.
It is a working repair, practice and resume workflow for learners and tutors. Repair Desk now makes the first correction directly editable and turns the resulting history into a readable study note.
Fix one line. Keep the rest.
The opening example ends at the correct x = 2 but includes a faulty expansion and a second error. Relay identifies the first transition that changes the solution set. It does not approve the derivation merely because the last answer matches.
Instead of retyping the whole solution, select the faulty line, replace that equation and check it. The original problem stays locked in the line editor, later lines remain in place, and the server checks the complete revised chain. Correcting one mistake can expose another. The full workboard remains available when several lines need editing.
The next-step guide follows the checked state: repair, unsupported expression, unfinished work, fresh-number practice or a study note. It does not automatically supply a correct equation. A hint or exact counterexample remains an explicit request.
Practise, leave, and come back without rewriting history
Fresh-number practice tests a requested operation rather than asking for the worked example's final answer again. Independent first attempts, assisted completions and revised answers remain separate.
The demonstration completes one fresh card independently and another after a hint, saves a handoff, reloads the app and reopens that file. Both counts survive. A hinted answer does not become independent after reopening.
A portable session contains the original equations, practice seed and ordered actions. Every tool call validates and replays those inputs. Saved grade fields are not trusted. Repeated identical actions do not double-count; reused action IDs with different work and stale revisions are rejected. Handoff is an explicit file or opt-in device storage, not automatic cloud synchronization.
A study note someone can actually use
The readable Markdown note now includes each original practice question, requested operation, submitted equation and checked outcome. Previously, answer history lacked that question context. The report reconstructs it from the recorded practice seed rather than trusting imported question labels or an answer key.
The same report powers the on-screen preview and downloaded note. Workboard edits and new session actions hide the stale preview. Reopening the handoff reconstructs the earlier questions and help history. A Markdown note is for reading; the separate handoff JSON resumes the session.
Real MCP, explicit-command reference host
The self-hosted service imports and calls the official Model Context Protocol SDK 1.29.0 at runtime. Protocol 2025-11-25 runs over Streamable HTTP. Four tools begin a repair, take a repair/practice turn, resume a session and produce a report. A served Agent Skill documents their use.
The reference web host performs actual initialization, discovery and tool calls. Its activity panel exposes those requests. This is explicitly an Alexa+ experience simulator, not Amazon Alexa, speech recognition or an unrestricted language-model conversation. The primary submission is the functioning MCP service; the web interface demonstrates its workflow. No AWS runtime integration or live-device certification is claimed.
The exact BigInt rational checker controls mathematical verdicts. A previously trained 1,637-parameter model suggests one of five practice families, but cannot decide correctness or set progress. The existing evaluator, learned weights and scoring rules are unchanged in this upgrade. The earlier engine, training generator and evaluation are included with provenance; no retraining or new algorithm is claimed.
What was executed
The complete release passed 105 Node tests and 50 browser checks: 25 inherited checks, 18 Repair Desk checks and 7 question-context checks. The inherited transport tests include an independent official SDK client communicating through TCP.
All 50 browser checks passed again on the anonymous public deployment with its content security policy enforced. Served JavaScript, styles, source archives, documentation and the video matched the verified release. Only the documented hosting rewrites to index/judge HTML links were permitted. The public video fully decoded and has non-silent audio.
Release results · Public verification · Engineering friction log
These are internal checks using synthetic equations, not a learner study or measured learning improvement. The concrete before/after is a targeted repair interaction and a question-aware readable handoff. Teacher review and an authorized usability study remain the next validation steps.
Scope and provenance
The current engine supports 2 to 12 linear equations in x with rational or decimal constants and five authored practice families. Nonlinear products, powers, variable denominators and inequalities are unsupported. Equivalent equations cannot validate unwritten reasoning or establish understanding. Model scores are not calibrated diagnoses.
Sessions are bounded to 48 actions and 52,000 bytes. A fingerprint identifies supplied bytes, not authorship or a certified grade; someone controlling a file can construct another internally valid practice history. The public endpoint receives equations and actions but the application does not persist them. Hosting infrastructure may retain ordinary logs. Use synthetic work, not student personal information, in the public demonstration.
Original core and Relay work were created September 7, 2026; Repair Desk and question-aware notes were added September 8 with substantial AI assistance, during the hackathon window. Original code is MIT licensed and dependencies retain their licenses. The demonstration records actual app actions with disclosed stock Kokoro af_heart synthetic narration at approximately 122 words per minute, not a cloned person's voice.
This updates the existing Alexa+ and Open Source entry, not a new submission. The prior immutable release and the unrelated production portfolio remain unchanged. Current source is on the dedicated repository's relay-desk-final-20260908 branch; its default branch retains the detectable MIT license. The source archive preserves the tested release; later public-verification records are additionally available on the branch.
Download source and tests · Demonstration transcript · Earlier release
Log in or sign up for Devpost to join the conversation.