-
-
Pitch workspace with a clearly labeled fictional example.
-
Evidence audit flags numeric gaps and planned results; source links do not verify claims.
Inspiration
What it does
How we built it
Challenges we ran into
Accomplishments that we're proud of
What we learned
What's next for DefendMyDeck
Inspiration
A polished presentation can hide a fragile claim. A student builder may say “94% accuracy” while the actual test note says “45 out of 50.” DefendMyDeck makes that gap visible before the follow-up question arrives.
What it does
Paste pitch or slide text, review editable claim candidates, and attach relevant evidence excerpts. The audit highlights numbers missing from the excerpt and distinguishes planned experiments from measured results. Rehearse a question anchored to a particular claim and its source, revise an answer, and compare explicit evidence-reference, measurement-context and limitation checks. Optional browser saving, JSON backup import/export and Markdown reports keep the rehearsal portable.
How it was built
This new proof of concept was built on 18 October 2026 using Codex with the official Devpost Learn Skill Pack. The repository includes scope.md, prd.md and spec.md in devpost/. The implementation uses JavaScript modules, HTML, CSS and Node.js built-in HTTP and test libraries, with no runtime package dependencies or API keys. The MIT-licensed source is public. AI assisted the planning, implementation, tests, technical documentation and this submission copy under the entrant’s direction.
A deliberate design choice
The running app uses deterministic text patterns, numeric comparisons and question templates—not a language model. It processes the pitch locally and never fetches the supplied source URLs. A linked source is not proof, and finding checklist mentions is not a measure of correctness or a predicted jury score. Missed claims can be added manually. This small scope makes the feedback inspectable and the complete workflow usable without a paid API.
Challenges and verification
Editing a claim initially left its audit state stale; the input handling was corrected and checked in the browser. Changed claims or evidence also invalidate incompatible rehearsal attempts. Backup imports are bounded and rebuilt rather than trusting stored feedback. Sixteen automated tests passed, covering numeric mismatches, planned evidence, import validation, stale history and local-server boundaries. Browser checks exercised the audit, two attempts, saving/reloading and backup export/import. These are development checks, not a user study.
Demo and current limits
The short captioned demo uses actual captured app states and a clearly fictional CampusSort example. Its results and agent-entered rehearsal answers are demonstration data, not a real experiment or user-impact finding. Keyword checks can be fooled by incorrect or negated statements; users must judge the evidence and meaning themselves.
What comes next
Test usefulness with real presenters before expanding into slide parsing, audio rehearsal or optional grounded language-model questions. No personal learning reflection or real-user validation is claimed by this agent-prepared account.
Run it
Install Node.js 24 or newer, clone the repository, run npm start, and open http://127.0.0.1:4173. Run npm test for the test suite. No installation of packages, login or API key is needed.
Log in or sign up for Devpost to join the conversation.