Shelly guide

Decisions: you stay in charge

Shelly is designed as a decision-support system, not an autopilot. Anything that moves money or affects compliance waits for a human.

At a glance

Governance policy in config.yaml: confirm anything over $250/week, plus always-confirm areas (recall, budget, range, replenishment, markdown, credit, customer terms, slotting, production, quality)
Every proposal has an id (D-1a2b3c4d), evidence, impact, confidence and a 7-day expiry
Answer in the web app (Confirm/Reject → Download → import), by CLI, or by typing 'confirm D-…' in the chat
The log is append-only: nothing is overwritten, so the full history can be audited

After each run, actions pass through the governance policy. Major ones are stored as 'pending'; minor ones are 'auto-approved' but still logged.
Your answers are imported with python -m shelly decisions import shelly-decisions.json. The file is validated (schema, id format, size) before anything is applied.
On the next run, the learning loop reads confirmed and rejected decisions and updates each rule's precision.

Precision estimate

Beta(1 + confirmed, 1 + rejected) → precision = α ÷ (α + β)

Tuning rule

≥5 answers and precision < 40% → stricter threshold; > 85% → slightly earlier alerts, always within safe bounds

Rejecting a decision is useful feedback, not a failure. It teaches Shelly what counts as a false alarm in your store.
Add a short note when you reject (e.g. 'promo week'). Notes are kept in the audit trail.

The static website can't write back to GitHub, which is why answers are exported and imported. A future version can do this by email or through the Claude task.

What needs my confirmation?confirm D-1a2b3c4dWhat have you learned?

Type or say these in the chat on the main page.