SplitCheck Team

What we’re proposing to build

• Enforced split policy in CI (proposed)The same three detection levels as the free checker and the toolkit, run as a required check on every pull request that touches your training or eval data, not just when someone remembers to run it locally.
• Approved exceptions (proposed)When a flagged overlap is a false positive or an accepted risk, a named reviewer approves it once and the policy remembers that decision instead of failing the same PR forever.
• PR annotation (proposed)A comment on the pull request with the contamination rate and which records triggered it, so a reviewer doesn't have to open a separate report.
• History (proposed)The contamination rate over time, so a team can see whether a data pipeline change made things better or worse.
• 5 or 20 repos (proposed)One subscription covers CI runs across up to 5 repositories at $199/mo, or up to 20 at $499/mo. "Repo" means a GitHub repository with its own CI workflow calling the check — not a seat count, not a per-run charge.

How the data stays local (proposed architecture)

Same principle as the toolkit: the check would run inside your CI runner, using the same open detection code as the $29 toolkit. Your training and eval records never get uploaded anywhere — SplitCheck’s servers would only ever see the report metadata your CI job posts back (pass/fail, the contamination rate, which policy rule fired, which exception was approved by whom) so that the PR annotation and history features can work. The raw text of your records would never leave your infrastructure. This is the same claim the toolkit already keeps today — read /toolkit and the toolkit’s own code if you want to verify it before a hosted version exists.

Illustrative example

# illustrative — SplitCheck has not built this; no PR ever produced this comment

SplitCheck found 3 contaminated records in this PR (0.6% of eval, above your 0.5% policy)

| Level      | Eval matches | Train matches |
|------------|--------------|---------------|
| exact      | 1            | 1             |
| normalized | 0            | 0             |
| near (≥0.85) | 2          | 2             |

2 of 3 matches were pre-approved under exception "synthetic-augmentation-2026-08".
1 new match requires review: eval#4021 ~ train#88213 (similarity 0.91).

Merge blocked until the new match is approved or the training file is fixed.

Price

These are the prices we’re testing, not a final price list. Pilot customers help set what actually ships and what it costs.

What it does not do

• It does not exist as running software yet — reserving a pilot does not give you access to anything today.
• It will not catch genuine paraphrases with no shared 5-gram — same lexical-overlap limitation as the free checker and the toolkit. Embedding-based semantic matching is out of scope.
• It is not a data platform, a labeling tool, or a dataset marketplace. It checks for overlap; it doesn't help you build better data.
• It is not free-tier friendly by design — if your training data changes rarely, the $29 toolkit run by hand or in a manual CI step is very likely the right amount of tooling and cheaper. Buy Team only if you want the check to be a merge-blocking, always-on policy.

The free checker and the $29 toolkit are real, ship today, and don’t require any of this: try the free checker · the $29 CI toolkit.