Stitched Ledger
A small, deterministic record-answering engine — a librarian, not a brain. It answers questions by exact lookup over one loaded table, shows the exact cell every answer came from, and honestly abstains when a fact is not in the table. No language model, no guessing, no network calls: all computation runs in the visitor's own browser.
Live app: https://stitched-ledger.pages.dev/
Built for Kiroween 2026, category Frankenstein — four parts stitched into one app: a table parser, a deterministic retrieval formula, an abstention gate, and an evidence-linking interface.
How it works
- Records are stacked as a table (the "ledger"). The built-in ledger is a 48-row synthetic Frankenstein laboratory ledger (reagents, equipment, safety data; 8 columns, no personal data). The same engine answers over any table loaded in its place.
- Questions are matched against nine frozen question templates (T1–T9: value lookup, count, comparison, and similar exact forms).
- Every answer carries its evidence as table coordinates, [Row, Col], and the interface lights up the actual evidence cell — the "stitch" between answer and source.
- If a fact is absent from the table, or a question falls outside the templates, the engine abstains with "기록에 없음" ("no record") instead of inventing an answer.
Features
- Split view: question input and answer card on the left, the full ledger grid on the right.
- Stitch highlight: the evidence cell glows and the answer card shows evidence → [Row, Col].
- Run 66-Point Live Verification button: re-runs the 66-question verification in the browser through the real core path and displays the pass count with the actually measured time.
- Ledger loader: drag and drop (or select) a CSV/TSV file to replace the ledger and ask questions over your own table. Malformed files are rejected with the reason stated, and the previous ledger stays in place.
- Dark laboratory theme; no external fonts, images, or network requests.
Verification (measured, not estimated)
The exam is 81 questions: 66 generated mechanically from the table plus 15 trap questions built only from items and attributes verifiably absent from the table. Core exam, Node build: retrieval 66/66, trap abstention 15/15, evidence coordinates 66/66. Core exam, web build modules: retrieval 66/66, trap abstention 15/15, evidence coordinates 66/66. Live verification button, local page (real Chrome): 66/66 passed, measured in 1.4 ms. Live verification button, live page (real Chrome via CDP): 66/66 passed, measured in 1.4 ms. Loader unit tests: 10/10. Type checks (tsc, core and web configurations): clean, exit 0.
Honest boundaries, by design: questions outside the nine frozen templates are all abstained (that boundary is the point of the piece), and the loader does not accept quoted CSV fields containing line breaks — such files are rejected with the reason shown.
Built With
- cli
- cloudflare
- css
- html
- javascript
- kiro
- node.js
- pages
- typescript