Decisions pilot guide
Train a classifier from labeled examples. Access is by arrangement. There is no self-serve signup or billing.
Before you start
- Ask for access and pricing at https://slancha.ai/contact.
- Remove sensitive information from examples before uploading.
- Never paste a service key into chat, a URL or frontend code.
Service discovery at https://decide.slancha.ai/api/pipelines lists text, image, audio, trace and tabular services. This shows what the service advertises. It does not show that your account can use them.
Two paths
Path A: Console (trace models)
Use this if Slancha has enabled the Decisions pilot for your organization.
- Sign in and open https://app.slancha.ai/app/decisions.
- If the pilot is not enabled, the page says so. Sign-in alone does not give access. Contact Slancha.
- A proposer creates a trace model and adds labeled examples.
- An approver starts training and acknowledges that the first ready version deploys automatically.
- Any member with access (viewer or above) can test a new example.
Roles:
| Action | Role needed |
|---|---|
| Check access, list models, test | Viewer |
| Create a model, add examples, save corrections | Proposer |
| Train, deploy, export dataset | Approver |
Machine and personal skill tokens do not grant Decisions access.
Console access is not generally available. Slancha turns it on for individual pilot organizations only, and account-based access is not part of the current public release. If Slancha has not confirmed that your organization is enabled, use Path B or ask at https://slancha.ai/contact.
Path B: Gateway workspace (approved service key)
Use this if Slancha gave you an approved service key.
- Open https://decide.slancha.ai/static/workspace.html and connect with your key yourself.
- Use the Build, Projects, Try and API tabs.
- Pick a type the service reports as available at https://decide.slancha.ai/api/pipelines. Text models are created by Slancha staff outside the workspace.
- Train, check the result, then test.
Data rules
| Rule | Console | Gateway |
|---|---|---|
| Labels | 2–64 | 2–64 |
| Examples per label | at least 3 | at least 3 |
| Examples per label for held-out check | at least 4 | at least 4 |
| Upload size | 1 MB per batch | up to 1,000 examples and 64 MB per request |
Media goes in as inline base64 or data URLs only. Local paths and URLs are not fetched.
Training and deploy
- The first version that becomes ready deploys automatically. Later versions need approval.
- Check the real status: queued, training, failed or ready. A failed run is not a model.
- Held-out results do not show real-world accuracy.
- Training has costs. Agree them first.
Test a model
Console example (software integration needs approved account authorization):
POST /api/decisions/projects/{id}/decide
{"state": {"trace": "synthetic example"}, "store": false}
The server picks your organization's deployed model. You do not send a model ID or a gateway key. Today, saving corrections is off by default and needs proposer access.
On the gateway, inference inputs are stored by default. Today, sending store:false asks the gateway not to store that input. Retention and deletion are not set out in this guide.
Console API summary
All routes are under the console API and use your session and organization.
| Route | Purpose | Role |
|---|---|---|
GET /api/decisions/access | Enabled status and role | Viewer |
GET /api/decisions/projects | List models | Viewer |
POST /api/decisions/projects | Create, for example {"name":"example","modality":"trace"} | Proposer |
POST /api/decisions/projects/{id}/examples | Add, for example {"examples":[{"input":"synthetic trace","label":"review"}]} | Proposer |
POST /api/decisions/projects/{id}/train | Body {"acknowledge":"first-ready-autodeploy-v1"} | Approver |
POST /api/decisions/projects/{id}/deploy | {"version":1}, ready versions only | Approver |
POST /api/decisions/projects/{id}/decide | Test | Viewer |
GET /api/decisions/projects/{id}/export | Dataset as JSONL | Approver |
The single example in the table shows request shape only. It is not a trainable dataset. Meet the label and examples-per-label minimums (see Data rules) before training.
Gateway API summary
These routes use an approved service key in the Authorization header, read by server software from an approved secret mechanism. They are not the signed-in console routes above.
| Route | Purpose |
|---|---|
GET /api/projects | List projects |
POST /api/projects | Create with a name and supported modality |
GET /api/projects/{id} | Read one project |
POST /api/projects/{id}/examples | Add examples |
POST /api/projects/{id}/train | Train |
POST /api/projects/{id}/deploy | Deploy a version |
POST /v1/systemone | Test with the owned proj-{id} model and trained-label questions. Inputs are stored by default; see the store flag in the API tab. |
GET /api/projects/{id}/export | Dataset as JSONL, not model weights |
For the owned model identifier and the request shape for /v1/systemone, use the API tab in the workspace.
Two access paths, two sets of limits:
| Service-key gateway | Signed-in account wrapper (console) | |
|---|---|---|
| Credential | Approved service key | Your session and organization |
| Upload | Up to 1,000 examples and 64 MB per request | 1 MB per examples batch |
Data and limits
- The Decisions privacy notice lists subprocessors for decide.slancha.ai only. A complete subprocessor list for the console is not published.
- Export is the dataset only. Trained model files cannot be downloaded.
- Setup does not turn on tracing.
- Data-use terms aren't set yet. Retention, deletion, rate limits, uptime and support for the pilot are not set out in this guide. Agree them with Slancha before you upload data: https://slancha.ai/contact.
If something goes wrong
- Sign-in interrupted: reopen the workspace and sign in again.
- No pilot access or rejected credential: contact Slancha. Do not use a shared admin key.
- Access revoked: stop.
- No deployed model: check versions, or ask for approved training.
- Training failed: fix the data, then request a new run.
- Unknown result of an upload, train or deploy: refresh status before retrying by hand. Do not replay automatically.