Sentient-X
Sentient-X Docs

Choose your path

Deep learning scientist

Run fast experiments without detaching models from governed data and promotion evidence

Your golden path is immutable corpus → training run → comparison → exact application → Worlds gate → canary.

The path

Freeze the data question

Name an immutable corpus snapshot, task revision, embodiment domains, and recipe. Keep native coordinates and per-domain normalization authoritative until the model boundary.

Run and observe

Autonomy owns training; Waddle owns experiment evidence. Record the snapshot, recipe, code, metrics, artifacts, and failure together so comparisons remain explainable.

Evaluate the exact candidate

Held-out Worlds evaluation is promotion evidence, not training input. Never relax or train on the held-out gate.

Assemble the deployable graph

Combine the trained VLA, optional inner agent, runtime images, artifacts, resources, and typed connections into one exact RobotApplication.

Promote evidence to canary

Whole-graph qualification and task-specific capability approval precede Fleet deployment. Keep the previous healthy application one rollback away.

Give this to your agent

Compare the candidate runs for task <task-ref>. Verify that corpus snapshots, recipes,
embodiment domains, and evaluation suites are comparable. Rank hypotheses from governed
evidence, not a single metric. Do not train on held-out evaluation data or promote anything.
Return exact run, artifact, and evidence references plus the next bounded experiment.

Fast without becoming vague

  • Change one scientific variable per comparison when practical.
  • Prefer a small exact snapshot over an implicit “latest” dataset.
  • Keep arrays and padded tensors at the terminal model boundary.
  • Inspect per-domain and failure-slice metrics, not only the global mean.
  • Promote an application identity, never a local checkpoint path.
  • Reattach to durable operations; a disconnected client is not a failed experiment.
OPERATION=$(sx --project "$PROJECT_ID" worlds evaluate policy:pick-mug-v12)
sx watch "$OPERATION"

Let auto-perfect do bounded repetition

When repeated experiments are the work, use an autonomy.improve campaign with a task and round budget. The agent chooses hypotheses; Experience, Worlds, Autonomy, and Fleet still own execution and evidence.

Drill down