Validation Ledger

Turn customer conversations into traceable product decisions.

Built by David Turner · atomicdjt

Live demo · Architecture · Release validation · Security · Full portfolio

More by David Turner: BuildWorld AI · WeaveStudio · GitHub profile

Validation Ledger is a polished, local-first research workspace for capturing customer evidence, testing product hypotheses, and recording decisions without losing the source material behind them.

Source → Evidence signal → Hypothesis → Decision

Why it exists

Product discovery often decays into scattered notes and conclusions based on repetition rather than evidence. Validation Ledger keeps the full reasoning chain inspectable: verified excerpts remain attached to sources, supporting and contradicting evidence stay distinct, and hypothesis support strength is calculated with a deterministic, explainable model.

Highlights

• Traceable evidence: connect exact source excerpts to hypotheses as supporting, contradicting, or neutral signals.
• Explainable support strength: score independent sources, segment diversity, behavioral evidence, verified citations, and contradictions without a black box.
• Decision history: record what was decided, why, and which evidence informed the call.
• Local-first privacy: research records remain in the browser through IndexedDB; there is no application backend or cloud database.
• Integrity-checked backups: export versioned JSON and atomically restore only after complete schema, size, duplicate-ID, and relationship validation succeeds.
• Optional AI assistance: generate untrusted structured suggestions with Google Gemini, verify quoted provenance, and require explicit human acceptance before saving.
• Responsive workflow: polished desktop navigation and purpose-built mobile layouts.

Product tour

The workspace includes project and hypothesis management, a source library, source-level evidence extraction, an evidence matrix, decision tracking, reporting, backup controls, and guided demo data.

Technology

• React 19, React Router 7, TypeScript 6, and Vite 8
• Tailwind CSS 4 with a custom token-based design system
• Dexie 4 and IndexedDB for browser-local persistence
• Zustand for small persisted interface preferences
• Google Gemini as an optional, user-configured extraction assistant
• Vitest and Oxlint for deterministic tests and static analysis
• Vercel for the static production deployment

Run locally

Requirements: Node.js 24 and npm 11.

git clone https://github.com/atomicdjt/validation-ledger.git
cd validation-ledger
npm ci
npm run dev

Open the local URL shown by Vite. No environment variable is required for the core product. To use AI extraction, add a restricted Gemini API key in Settings; never commit a real key.

Quality gate

npm test
npm run lint
npm run build
npm audit

The scoring suite covers source independence, segment diversity, behavioral evidence, direct citations, neutral/missing relationships, counterevidence semantics, and source-text revalidation. GitHub Actions repeats unit tests, lint, production build, dependency audit, Chromium workflow, and automated accessibility checks on every push and pull request.

Data ownership and backups

The IndexedDB database is named ValidationLedgerDatabase. Data is tied to the current browser profile and site origin. Use Settings → Export backup regularly; clearing browser storage can remove local records. Backup import validates the supported schema before replacing current data in a single transaction.

The live demo does not receive or retain application data on a server. Optional AI extraction sends the selected text and request directly from the browser to Google's API using the key supplied by the user.

Scoring model

Validation Ledger reports support and counterevidence separately. Supporting strength combines:

• independent supporting sources, capped at 60 points;
• segment diversity, capped at 15 points;
• behavioral or willingness-to-pay evidence, worth 15 points;
• direct source citations, capped at 10 points.

Counterevidence receives its own independently capped score. Credible support plus contradiction is labeled mixed; contradiction without support is contradicted. This keeps positive evidence from erasing material counterevidence.

Only excerpts that match the saved source exactly or through conservative whitespace/quotation normalization may be treated as direct evidence. Missing or unverified provenance is displayed as an inference, never as a quote. The score is an aid to judgment, not proof of market demand. AI-generated suggestions also require human review.

Documentation

• Architecture and system boundaries
• Release checks and current limitations
• Security and responsible key handling
• Contribution workflow

License

Copyright © 2026 atomicdjt. All rights reserved. The repository is public for portfolio review; no license to redistribute or create derivative works is granted unless provided separately in writing.