Playground

Open the playground and sign in. It uses your account’s private Conceiver learning.

Your account and the playground are one app. Moving between pages never reloads it. The sidebar has two sections:

• Workspace: Overview, API keys, Usage, Billing, and Settings for your account.
• Playground: where you query and teach the selected model. While you are in the playground, its menu also shows the model selector, Training sequences, Stored artifacts, the current session, and the guided demos.

Choose Conceiver · Language for text answers or Conceiver · JSON for structured JSON answers. The current session card refreshes after each action. The account icon in the top-right corner opens the overview, billing, settings, documentation, and sign-out.

Modes

• Ask: submit one query. The response shows an answer or abstention, confidence, evidence, and correlation details.
• Correct: submit a query, then the authoritative correction in the same field.
• Teach: submit alternating query and answer lines. Each pair follows the ordinary query and correction path.
• Evaluate: submit alternating query and held-out expected-answer lines. Only queries reach Conceiver; the browser compares returned text exactly.

The confidence selector defaults to Hypothesis. Higher settings filter displayed answers but do not change Conceiver’s result. Select Memory explicitly when a request should include relevant episodic examples. Hover or focus Ask, Correct, Teach, Evaluate, Memory, or Confidence to see a short explanation.

Guided demos

Pick a demo under Guided demos in the playground menu. Each one starts a fresh, isolated tutor; reloading or picking it again starts over. Audit history and the fixed lifetime credit balance remain, and your own account learning is unaffected.

Each step selects the right mode for you, types its example into the input, and explains in the demo card above the input what that input is and why it matters. Submit is highlighted; pressing it is the only thing a step asks of you. The next step appears once the reply arrives.

Each demo is a JSON file in service/src/learner_service/site/demos/. A file names the demo, its answer kind, its steps (mode, confidence, typed text, title and explanation) and follow-up suggestions. Add a file to add a demo. A step may carry an expect block; the real-runtime test plays every file and checks it.

• Teach its first words starts empty and teaches bob and alice as words by correcting each with itself.
• Teach an operation corrects create account bob and create account alice with JSON results, then asks for a name it has never seen.
• Resolve ambiguity teaches two conflicting destinations for one request, shows the abstention and both memories, then adds the detail that separates them.
• Workflow planner uses Correct mode: ask for a plan it has not learned, give it the right plan, repeat with a second person and team, then test with new names in Ask mode.
• Records desk and Invoice extractor use Teach mode: send two question-and-answer examples at once, then test with new values in Ask mode.

Changing mode yourself, or pressing End demo, leaves the walkthrough.

Tool-call answers produced on the model’s turn appear as generic simulated Tool call bubbles. Each bubble shows only the tool name, remains in a Calling… state for one second, and then reports that no action was executed. Corrections and teaching do not trigger tool-call widgets. Plans remain non-executing.

After the walkthrough finishes, sample follow-up inputs appear in the history. Selecting one only fills the input. It does not submit a query or teach an answer. New results automatically keep the history at the latest entry. Select a model response bubble to open or close its evidence.

Plans are not executed and records are not moved. Use sample or non-secret data.

Training sequences and stored artifacts

Open Training sequences in the playground menu. The page is ordered by task:

1. Connect, copy, or save the selected account model.
2. Load a sequence saved for this model back into Teach, or export it.
3. Expand advanced training tools only when importing, preparing, evaluating, or ingesting pair text.
4. Review every saved sequence on the account, with its size and a download link.

Stored artifacts lists the verified artifacts that keep a model's learning after its last live instance retires.

Copying current training text does not require storage. Saving does. Older event documents remain exportable in their original format. Import and ingestion always target the selected private account model, not a guided demo.

Activity and layout

Open Activity from the workspace header to inspect or refresh audit events, clear the visible history, copy the history, or download events. Authentication headers are excluded from browser exports.

The playground has two columns: the guided demo card with the input under it on the left, and the history on the right. On narrow screens they stack, with the history below. The input grows with multi-line text. The desktop layout keeps the same persistent navigation beside playground and account views. On phones, that navigation becomes a drawer. Presentation mode hides playground navigation without removing it.