Guided mode and integrations v0.4¶

Status: available in v0.4 (milestone M3), the first public launch

v0.4 builds on the multi-tenant core of v0.3. With v0.4, analysts go from a spreadsheet to a trustworthy report and a what-if scenario. They do not write YAML. Other applications can call fitted models through stable endpoints. v0.4 is the first public launch, cut line A. All persons can sign up and run declarative specs and the built-in models on their own data. Only trusted tiers can upload code. For instructions, read the Guided tutorial, the Integrations hub and Tutorials.

Guided mode: organised around tasks¶

The app has two modes over the same ModelSpec. A toggle in the header changes between Guided and Expert. Each Guided screen has "View as YAML". Edits in Expert mode go back to Guided mode.

journey
    title An analyst's first marketing-mix model
    section Set up
      Choose "Geo marketing mix": 5: Analyst
      Upload a CSV, answer "which column is revenue?": 4: Analyst
      Enter beliefs as ROI ranges per channel: 4: Analyst
    section Run
      Watch plain-language progress, with an ETA: 5: Analyst
    section Decide
      Read the report and the sanity scorecard: 5: Analyst
      Drag what-if sliders, see bands update instantly: 5: Analyst
      Share a link or connect a spreadsheet: 4: Analyst

Shipped	Where to read about it
Four templates (geo marketing mix, weekly forecast, demand by store, customer choice) and a question flow with role detection	Your first analysis in Guided mode
Sandboxed ingest of CSV, Parquet, MFF and Meridian-shaped tables	From a spreadsheet to a Guided draft
Business priors (ROI or share ranges) translated in a sandbox, lift tests as calibration	Guided modelling
The sanity scorecard and the promotion gate	Guided mode, Integrations
What-if sliders over per-draw response grids, "Run exact", forecast and counterfactual views	What if…
The persona lock: projects that stay Guided-only	The persona lock

Integrations¶

Shipped	Where to read about it
Endpoints and aliases with history, rollback and a frozen input contract; `score:<alias>` tokens	Endpoints and webhooks
Generated snippets: Python, TypeScript, curl, R, Apps Script, a notebook cell	The Integrations hub, Connect Google Sheets
Signed webhooks with an outbox, retries, a delivery log, replays and an SSRF guard	Receive webhooks
Saved plots, PNG/SVG/PDF renders in the sandbox and expiring embed links	Embed a chart, Sharing plots
The Python client's endpoints; the CLI's `endpoints`, `webhooks`, `plots`, `guided`, `priors` and `tenant` commands	Optimise a budget, The Python client

Trust and the data lifecycle¶

Shipped	Where to read about it
Self-service sign-up with e-mail verification, disposable-domain blocking, rate limits and Sybil flags for operators to review	Sign-up and abuse controls
Identity verification and cut line A at run admission	Sign-up and abuse controls
Tenant deletion that purges every store, and the GDPR export	Deleting and exporting data
A Cloud Armor policy on every load-balancer backend	Security hardening

The exit criteria¶

Criterion	Where it stands
Five non-technical testers go from upload to report to scenario without help in less than 30 minutes	Prepared. An automated run takes 135 s of system time. We repaired the six usability defects that it found. We wrote the protocol for the testers. The sessions are not done yet.
An external optimiser works against an endpoint	Met on a laptop: +0.22 % revenue at the same budget against the live endpoint (details). From v0.5, the platform optimises budgets next to the model, in a sandboxed `optimize` run on mmm-framework planning (Optimise a budget). We removed the external-optimiser adapter.
Cut line A: public launch of declarative specs and built-ins	Met for admission. Sign-up workspaces are `public`, and the platform refuses uploaded code. Public runs need an executor that enforces egress and DNS (`docker-gvisor` or Cloud Run Jobs). This laptop could not run these executors.

We verified all items locally. M3 verification gives the numbers. It also tells which items still need real persons, Docker or Google Cloud.