LLM generator mini
Prompt pack, model route, evaluator rubric, and reviewer owner stay attached before the job enters the queue.
follow-up-access-v4
eligibility-copy-variants
finance-hardship-language
Synthetic Data Lab demo
A dataset you can defend before you train on it.
78% of teams cannot validate training data before training, and 77% cannot trace where it came from. This demo walks one loop: a coverage brief becomes a reviewed dataset with the schema signed, privacy bounded, and every field change recorded. Read-only here. Scope it for your own work in a pilot.
Quality strip · illustrative seed data
Jobs last 24h37
LLM, physics, privacy-profile, and coverage-planning jobs visible in the seeded preview queue.
Avg quality score92%
Latest approved slices average realism, fidelity, privacy, and coverage checks before export.
In review11
Dataset slices waiting for approve or reject decisions in the review queue.
Open alerts3
Quality alerts still attached to generated slices before release.
Schema builder
Decide what exports before a row is generated.
Schema, row preview, privacy meter, and export manifest sit in one workbench. A reviewer sees exactly what changes and signs off, so the dataset arrives with its rights and its bounds already settled.
Privacy risk meter
k=12epsilon=1.6low risk
Suppression and generalization are active for sparse region buckets before export.
eligibility-response-v4
5 fields · 3 generated rows · reviewer required
Fields
household_iduuid
synthetic
region_bucketcategory
generalized
eligibility_reasontext
reviewed
risk_scoredecimal
bounded
follow_up_daysinteger
clipped
household id
region bucket
eligibility reason
risk score
follow up_days
syn-10482
rural-north
transport gap
0.42
14
syn-10483
urban-west
benefit renewal
0.31
7
syn-10484
suburban-east
appeal review
0.58
21
Export manifest
schema hash signedprivacy profile attachedreviewer approval in review
Approval diff
Requested dataset compared with governed output.
Reviewers see which fields changed during privacy and quality approval, why the governed output differs, and what still exports in the export manifest.
Field
Requested
Governed output
Decision
region_bucket
full ZIP-level geography
regional bucket, sparse areas merged
generalizedeligibility_reason
free-text policy narrative
reviewed reason taxonomy with examples
reviewedfollow_up_days
raw long-tail values
clipped at approved range
boundedprivacy profile attached
quality reviewer approved
manifest records field changes
Workbench preview
Three checked-in lab previews.
The same screens you can scope in a pilot, scaled down here as a checked-in preview with review and privacy controls.
Physics generator mini
Simulation slices carry scenario parameters, quality checks, and edge-case alerts into the same review path.
warehouse-collision-edge-cases
robotics-grasp-lighting
sensor-noise-replay
Privacy editor mini
k-anonymity, l-diversity, t-closeness, epsilon, and delta are visible before dataset export.
k = 12
l = 4
t = 0.18
epsilon = 1.6
Illustrative seed data · scoped workspace metrics in a pilot
Jobs last 24h
37
Avg quality score
92%
In review
11
Open alerts
3
Review queue
Nothing exports until a reviewer says so.
Every slice carries its dataset, source, quality score, and the reviewer's decision. That record is what survives an audit. The rows below are illustrative seed data; in a pilot they read from your scoped review queue.
follow-up-access-v4
LLM generator
94%
Approve after privacy note
warehouse-collision-edge-cases
Physics generator
89%
Reject until balance improves
eligibility-copy-variants
Coverage plan
93%
Approve for export
The loop
Five steps from a coverage plan to a dataset you can defend.
1
Coverage plan
Define target scenarios, quantities, privacy profile, and reviewer owner before generation starts.
Evidence captured
2
Generate
Launch LLM or physics jobs and keep queue state visible while slices are produced.
Evidence captured
3
Review
Approve, reject, or annotate each slice before it can move into an export.
Evidence captured
4
Quality check
Inspect realism, fidelity, privacy, alerts, and coverage trends on one dashboard.
Evidence captured
5
Export
Release a governed dataset packet with settings, reviewer notes, and signed manifest context.
Evidence captured
EU AI Act provenance enforcement begins August 2026. Run this loop on your own data in a pilot and walk away with a reviewed, rights-cleared dataset you can put in front of an examiner.
Start a pilot