01 / record
Optimistic concurrency
Every draft write carries an expected revision. A stale editor gets a conflict, never a silent overwrite.
Annotation workflow infrastructure
Raw records in.
Reviewed evidence out.
LabelForge is a collaborative workbench for text and structured data. Import a file, pin an immutable schema, assign real work, resolve disagreement, and publish a dataset with a manifest you can verify.
Open sourcePostgreSQLNo model provider required
01CSV and JSONL import
02Server-enforced assignment
03Revision-checked autosave
04Schema-pinned answers
05Checksummed releases
The workflow
An auditable path from file to release
Every step is stored, scoped to a workspace, and visible in the audit trail. There is no hidden AI step and no fabricated productivity score.
01
Import and validate
Parse real bytes, preview row errors, quarantine duplicates, and commit only when the result is explicit.
02
Pin a schema
Publish an immutable annotation schema. Tasks keep the version they started with.
03
Assign and annotate
Managers assign eligible annotators. Drafts autosave with an expected revision.
04
Review or adjudicate
Accept, return, reject, or send a disagreement to a reasoned final answer.
05
Release with proof
Freeze accepted or adjudicated work into a versioned export with SHA-256 checksums.
Built for the awkward parts
The form is easy. The record integrity is not.
LabelForge treats concurrency, schema drift, and disagreement as first-class product states instead of edge cases.
02 / record
Immutable schema versions
Answers are interpreted under the schema pinned to the task. New schemas apply only through explicit, audited actions.
03 / record
Honest agreement measures
Cohen's kappa, average Jaccard, and span overlap are shown with sample size, numerator, denominator, and limitations.
04 / record
Reproducible releases
A release is a stored snapshot. The manifest records the policy, counts, schema, and checksums used to verify it.
Stated limits
Small enough to verify. Explicit by design.
The first release is deliberately constrained. Limits are enforced on the server and explained before upload, not hidden behind a truncated file.
Upload per file4 MB
Rows per import10,000
Records per project25,000
Preview staging lifetime1 hour
Import formatsCSV / JSONL
External model callsNone
Vercel Functions cap request bodies at 4.5 MB. LabelForge sets a 4 MB application limit to leave room for multipart overhead. Host limits can change; the app states the limit it enforces.
Start with real workflow
Open the demo, follow one record, publish one release.
The seeded workspace contains synthetic product reviews and support records, multiple roles, drafts, a correction, a disagreement, and a completed release.