When tier-B cost confidence is enough.

Not every action needs the same certainty about cost. Rewriting a product's copy is reversible and cheap to get slightly wrong. Cutting its price is neither. That asymmetry is why Magistry gates the one price action by cost-confidence tier — and why optimize_loser runs fine at Tier B.

The three tiers

• Tier A — verified cost: Shopify cost_per_item or a manual entry someone stands behind.
• Tier B — estimated cost: an AliExpress reverse-image match on the supplier listing.
• Tier C — unknown: no cost data at all.

The tier is stamped on the decision row at the moment of action, so an auditor always sees the confidence the agent actually had — not what we know now.

Why optimize_loser is safe at Tier B

optimize_loser doesn't touch price. It kicks off a copy/image rewrite cycle on a chronic underperformer — and it's fully reversible, because every content change is versioned and the revert restores the previous version. The downside of being wrong is small and recoverable: a SKU carries weaker copy for a while, and one revert restores it. That's exactly the risk profile an estimated cost is good enough for.

{
  "from_state": "ACTIVE",
  "to_state": "OPTIMIZING",
  "trigger": "roas_below_floor",
  "evidence": {
    "action": "optimize_loser",
    "roas_30d": 0.7,
    "reasoning": "Chronic underperformer; rewrite copy before considering a discount."
  },
  "cost_tier_at_decision": "B",
  "status": "applied"
}

Contrast that with discount_test — the only price action — which stays locked until Tier A and a margin floor. A wrong price on an underestimated cost is a margin leak you might not catch for weeks, and reversing it doesn't refund the orders that already sold cheap. Same SKU, same uncertainty — different blast radius, different gate.

The trick isn't being certain about everything. It's matching the reversibility of an action to the confidence you have. Soft and reversible can move on estimates; expensive and sticky has to wait for facts.

— Magistry

Tiering cost confidence is how an agent stays useful when its information is imperfect — which is always. It acts where being wrong is cheap, and it waits where being wrong is expensive. That's not timidity. It's the whole discipline.