The Private Corpus
Corpus specification, monthly rotation, and why the private data asset is Signet's core defensibility.
Core defensibility. The rotating corpus is not a supporting asset. It is the mechanism's defensibility layer and one of Signet's primary long-term assets.
Initial corpus
The planned production corpus begins with 20 hours of private ASL footage, including a 3-hour canary subset.
The corpus specification requires:
| Requirement | Specification |
|---|---|
| Signers | At least 12 deaf native signers who do not appear in public benchmark corpora |
| Conditions | Studio, broadcast-lit, conversational, and off-axis |
| Register | Instructional, narrative, and conversational |
| References | Certified English reference translations |
| Verification | Independent second-interpreter verification |
| Clip length | Sentence-level, between 2 and 30 seconds |
| Interpreter diversity | No single translator's wording may dominate a slice |
The last row matters more than it looks. If one interpreter wrote most of a slice, models would be rewarded for learning that person's phrasing rather than the meaning of the signing.
Rotation
- Four hours of new private footage are added monthly.
- The oldest four hours are retired from active scoring.
- Retired footage would then be released as a public self-evaluation archive, once the corpus exists and rotation has begun.
Rotation does three things at once. It keeps the active evaluation set beyond reach. It builds a longitudinal record that a competitor cannot assemble quickly. And it eventually gives the research community real evaluation data, once that data can no longer be used to game the network.
Active clips also obey a 30-day reuse cooldown inside the draw. See For Validators.
Why this is the moat
A protocol can be copied in a weekend. Twenty hours of certified, second-verified, condition-diverse footage from twelve deaf native signers cannot.
The corpus is funded as recurring core infrastructure, at roughly $3,000 per month for rotation on top of the initial acquisition. It is not treated as a one-time benchmark purchase. See Economics.
Community participation
Deaf community participation is required in corpus creation and in product evaluation. This is a stated project requirement, not an aspiration.
The project targets unserved content volume. It does not position automated output as a replacement for certified human interpretation in high-stakes contexts.