Author a world from a question

Ask what you want to compare in plain words, for example "Would a fifth packing station cut order waiting time at about 60 orders an hour?". Alive builds a synthetic world from its model library, labels every assumption, and asks only questions whose answers change feasibility, interpretation or spend. The same session continues in the web app, the CLI, the SDKs and coding agents.

Authoring is in beta. The model library is a set of draft, unvalidated models with synthetic parameters; a session produces an approved, compiled world you can review. Running authored worlds through the hosted API comes in a later release. Results describe the synthetic world, never your real setting.

Start a session

A project needs a confirmed intake first. Then:

alive author start --project <project_id> --question "Would a fifth packing station cut order waiting time?"

--mode auto_draft (default) fills gaps with labeled defaults; --mode guided asks about the numbers that drive the comparison; --mode expert keeps the values you set exactly. Values you give in other units (minutes, percent, per hour) are converted and the conversion is recorded.

A session ends in one of three ways:

• A world: the models, the comparison, what is measured and every assumption, ready to approve.
• A question: the smallest set of answers that changes the result.
• A narrowed question or a refusal, with the reason: for example, mortality is not something workflow models produce, live systems need connected execution, and real patient data is out of scope.

A drafted world also shows how each result is measured: the unit, what it is averaged or counted over, the time window, and what is still waiting at the end, reported beside waiting times so a backlog cannot make them look better than they are.

Every answer also lists what you can do next: for example, Alive's concept review for an outcome a workflow simulation cannot produce, or the evidence that would answer your real question. Suggestions marked as AI are hypotheses to consider, not findings.

Answer and resume anywhere

Each answer creates a new revision. In a terminal, alive author start asks the questions in place. Scripts, CI and agents never wait for input: with --json, --no-input or no terminal they print the checkpoint and exit with code 4. Continue from any client:

alive author show <session_id> --json
alive author answer <session_id> --set change="add one more packing station"

The web app shows the same session at /projects/<project_id>/author?session=<session_id>. Agents relay a person's answers with --relayed (answered_by: agent_relayed is recorded); they never answer on the person's behalf.

Approve the exact revision

alive author approve <session_id> --revision <n>

An approval names one revision. Any later answer or change creates a new revision without it, so a changed world always needs a new approval. The HTTP endpoints are POST /v1/projects/{project_id}/authoring-sessions, GET /v1/authoring-sessions/{session_id}, POST .../answers, POST .../revise and POST .../approve; see the HTTP API reference.

AI stages

By default no AI provider is called: a deterministic baseline authors the world. With --ai hosted (metered) or --ai openai|anthropic|google (your organization's own key), AI stages read the question, choose models and propose values. Every value an AI stage reads must quote your exact words, or it is not used; AI review notes are labeled hypotheses, never findings.