Reviewable labeling for small AI teams

Abstract

Qualora is a browser-first workspace for preparing small text datasets. It combines structured first-pass annotation, an explicitly simulated second-pass review, deterministic statistics, portable project files, and SHA-256 consistency receipts. The product is designed for independent developers and compact teams that need traceability without adding a hosted data pipeline.

1. The problem

Small labeling jobs often begin in an improvised document. Categories drift, confidence is lost, reviewer changes are hard to reconstruct, and a downloaded file provides no simple way to detect later changes. Large annotation platforms can be disproportionate for a 20-item evaluation set. Qualora addresses that gap with a focused, local-first workflow.

2. Product principles

Context travels with every label. Review remains reversible. Statistics are derived from the current records rather than stored as claims. The application does not imply consensus, model evaluation, expert independence, or objective correctness. Working data stays in the browser until the user chooses an export.

3. Workflow

A contributor creates a batch of up to 100 non-empty text items and defines up to 12 categories. Each item can receive a category, context note, and confidence score. A review pass records a separate opinion, decision, explanation, and confidence. Filters expose unfinished, low-confidence, and review-pending items. The summary reports completion, review coverage, and label disagreement using deterministic rules.

4. Evidence model

Qualora normalizes the project into a canonical JSON representation. Keys and collections are ordered, text whitespace is normalized, and volatile update timestamps are omitted from evidence. The browser hashes that UTF-8 representation with SHA-256. An unchanged project reproduces the digest; a changed item, category, annotation, or review does not.

A receipt proves consistency of normalized content only. It does not prove authorship, truth, intellectual-property rights, reviewer independence, or label quality.

5. Local data and portability

The current project is saved with a versioned schema in local storage. Export creates a human-readable JSON project. Import checks size, structure, schema version, ID uniqueness, category references, numeric ranges, and enum values before the user is asked to replace existing work. Invalid or cancelled imports do not overwrite the active project.

6. Wallet and receipt anchoring

Wallet connection is optional and reads only a public account and chain ID. It never signs or sends during connection. A current receipt may be saved locally with a business ID. A contract call is available only with valid deployment configuration and always shows operation, digest, wallet, contract, chain, and zero BOT value before a second confirmation. Without verified configuration the interface says Local Demo / On-chain unavailable and never invents chain evidence.

7. Security and privacy

Qualora avoids a remote application database and does not invoke a paid model. This reduces routine disclosure but does not make the browser a vault. Anyone with access to the device or browser profile may access local data. Users should avoid sensitive data, export deliberately, verify files before relying on them, and clear browser storage on shared devices.

8. Accessibility and resilience

The interface uses named controls, visible focus, textual status messages, keyboard-reachable actions, and responsive layouts. Empty datasets calculate a zero percent completion rate. Wallet absence, rejection, account or network changes, malformed imports, stale receipts, pending transactions, and receipt lookup failures remain recoverable states.

9. Roadmap

Future work may include optional team synchronization, signed reviewer identities, configurable adjudication, provenance adapters, and public receipt anchoring. These are directions, not commitments. Any remote or on-chain feature would require new consent, security review, and clear cost disclosure.

10. Governance and limitations

Users remain responsible for lawful data collection, label definitions, reviewer instructions, exports, wallet actions, and decisions based on the resulting dataset. Qualora is a preparation and consistency tool, not a certification authority or source of professional advice.