Capabilities and remaining work
Available workflows
Zils provides an authenticated decision API using shared JevK5 4B, with model listing, key management, typed predictions, and durable bulk jobs. Hosted checks have exercised authentication, real GPU inference, private uploads, retries, and account isolation. These establish workflow behavior, not a production-capacity or customer-quality guarantee.
The hosted training queue supports the pinned JevK5 4B model, approved miners, separate calibration and test data, and accepted-artifact delivery. Jobs preserve their own model identity; older Kev jobs remain readable. A completed run may produce no_qualifying_model.
The automatic training workflow connects capacity-gated assignment, evaluation, publication, runtime verification, and private model registration. Version selection compares first versions with the calibrated base and upgrades with the customer's previous accepted model. A stale upgrade cannot replace a newer active version. An API key grants access to an account's models; ordinary predictions do not train them.
The local bundled fleet and guarded Bittensor integration retain their Kev 0.8B contract. The first closed training-to-chain round completed on testnet subnet 579 with verified revealed weights. That historical result does not establish an open JevK5 competition or public-miner inference.
What still needs evidence or implementation
| Area | Remaining work |
|---|---|
| Customer quality | Representative customer-data pilots and independent final evaluation; synthetic checks do not establish improvement |
| Capacity | Sustained concurrency and reliability measurements, capacity planning, and any service-level commitments |
| Model upgrades | Customer-facing upgrade selection; the backend already supports explicit predecessor selection |
| Commercial operations | Billing and operating policies for a broader customer launch |
| Open competition | Public discovery, isolated untrusted-model evaluation, benchmark refresh, and rewards across different jobs |
Training exports remain readable by approved miner operators. Confidential compute and automatic training-data retention controls are not implemented. The decision API's bulk retention policy is separate from training storage and downloaded copies.
Read research within its scope
The product-matching experiment compared shared JevK5 with an additional adapter. The adapter reduced held-out accuracy and worsened probability error; its macro-F1 gain was uncertain. It was not promoted. The recorded queue pilot missed its frozen acceptance threshold and delivered no accepted model. Preserve these outcomes when describing model progress.
The public research page and experiment reports record specific models, data, hardware, and timing conditions. Unrelated benchmark percentages do not establish a model ranking or customer-task reliability.