cloudrop.it

Adaline

She proposes the moves. You hold the veto.

Adaline is the commercial brain of the deck. She is fast, she is opinionated, and she cannot spend a euro that a compiled ceiling has not already allowed. The fences came first; the intelligence was allowed in afterwards.

Standing order

Propose freely.
Spend never.
Explain always.

The fences

Three things she cannot do.

Compiled ceilings

Daily and per-action spend limits live in the binary, not in a settings row. Raising one is a code change, a review and a deploy. Friction proportional to the consequence.

Human veto

Every proposal waits for a person. There is no timeout that converts silence into consent, because silence at 03:00 is not consent.

Dormant ramp

Autonomous execution exists in the codebase, tested, behind flags that are off. It stays off until an operator turns it on deliberately, one lane at a time.

A spend limit you can change from a settings page is not a limit. It is a suggestion with a number next to it.

Try it

The console, with the fence in it.

Three proposals, exactly as the deck files them. Approve one, veto another, then meet the third: the one that no button on this page, and no button in the product, can approve.

Proposal Repricing

PR-2841

Reprice SKU-2841 down 4.2% to recover the buy box.

Competitor undercut at 04:12. Margin after the move stays 3.1 points above the floor for this category.

Current
EUR 148.00
Proposed
EUR 141.78
Floor
EUR 137.40
Buy box
lost 6h 41m
Daily exposure EUR 0 ceiling EUR 400 / day
Decision log
  1. 14:02:07 adaline repriced SKU-2841 · buy box recovered (+3.1%)
  2. 13:47:55 operator vetoed PR-2839 · seasonal, hold until Friday
  3. 13:31:02 system PR-2836 blocked · exceeds compiled daily ceiling

Every proposal, approval, veto and block is written down. The log is the product's memory and the operator's evidence.

What she does

Four verbs, in order.

  1. 01

    Watches the board

    Every offer, on every marketplace, continuously. Competitor moves, buy-box state, stock drift, margin erosion.

  2. 02

    Forms a view

    Not a score. An argument: what changed, what it costs to respond, what happens if nobody does.

  3. 03

    Files a proposal

    One action, one rationale, one stated exposure. Filed into a queue a human reads in the morning.

  4. 04

    Writes it down

    Approved, vetoed or blocked, the decision and its reason enter a log that is never rewritten.

Listing intelligence

One photo in. Five listings out.

Give her a product photo and an EAN. She reads the item, matches it to each channel's category tree, and writes a listing per marketplace: title, attributes, long description, all in the language and the register that channel rewards. One source of truth, five native listings, none of them a translation of the other four.

one photo, one EAN

  • Allegro Parameters read off the photo, title trimmed to the channel limit. PLparametersGPSR block
  • Kaufland Global German long description, unit price and hazard fields resolved. DEunit price
  • OTTO Market Category tree matched, size chart attached, tone tuned to the channel. DEsize chart
  • eMAG Romanian copy written natively, not translated from a template. ROattributes
  • Shopify Long-form product story, metafields populated, alt text written. ENmetafieldsalt text

Compliance-safe by construction. Hazard fields, unit pricing, country of origin and the GPSR block are filled from the catalogue, never invented by a model. A field she cannot evidence is left empty and flagged, not filled with something plausible. That one rule is the difference between generated copy you can publish and generated copy you have to read twice.

Daily wake

The queue is already there when the day starts.

Every morning, before anyone opens the deck, Adaline reviews every lane: signals, buy box, margin, stock, competitor moves. By the time the first coffee is poured the proposals are staged, each with its evidence attached and its exposure stated. Nothing is armed. The queue is simply waiting.

  1. 05:40 Signals swept across every lane.
  2. 05:52 Buy box and margin deltas resolved against the floor.
  3. 06:04 Proposals staged, evidence attached, exposure stated.
  4. 06:05 Queue handed to the operator. Nothing armed.

The ramp

Autonomy, in four stages, two of them switched off.

  1. L0

    Observe

    live

    Reads state, forms a view, writes nothing. Every module is readable to her; none is writable.

  2. L1

    Propose

    live

    Files a proposal with evidence and a stated exposure. A human approves it or it expires.

  3. L2

    Act within a lane

    flag off

    Executes inside one marketplace, under a compiled ceiling, with an operator veto window.

  4. L3

    Act across lanes

    flag off

    Coordinates moves across marketplaces. Surfacing-only in code today, and never auto-armed.

The ramp is built, reviewed and tested up to L3. Two of its four stages are switched off in configuration, and flipping one is an operator decision with a runbook attached. Not a product default, and not a gradual rollout.

Why she is safe by construction

The uncomfortable truth about AI agents in commerce is that the dangerous failure is not a wrong answer. It is a plausible answer executed two hundred times before anybody reads it. Confidence scales; consequences scale with it.

So the design starts from the blast radius and works backwards. Adaline's spend fence is not a prompt instruction. A prompt is a request, and requests are negotiable. It is a set of identifiers bound at call time and a ceiling checked in code paths she does not author and cannot reach.

Prompts are requests. Code is the fence. We only ship fences.

The second layer is evidence. Every proposal carries the observations it was built from, and a proposal whose evidence is thin says so rather than filling the gap with fluency. Absence is reported as absence. It is a small discipline that removes the single most common way these systems mislead the people running them.

The third layer is the log. Approvals, vetoes and blocks are append-only. When a move turns out badly the question is never "what did it do". The answer is already written down, with the reasoning that produced it and the name of whoever let it through.