Registry / 009SystemProductionLLM

Scene production.

A deployed managed product that turns a reference image and product packshots into brand imagery. Every model step is gated by an operator, and both gates exist because their absence already caused a failure.

Deployed system — this configuration is running in production. Client identities are withheld.

Classified as a system: an arrangement of automations, capabilities and human judgment responsible for an outcome. The outcome belongs to the arrangement, not to any one component in it.

A model does part of the work here, under a fixed prompt. It interprets or generates; it is not what grants this entry its authority.

Problem it stands behind

Publishing at volume without publishing mistakesEvery figure an author declares is recomputed before a lesson can build. Narration is checked before a voice is made from it. Image generation stops at a human gate wherever a wrong result would otherwise ship.In the owner's words, on /problems

Objective

Let a brand operator produce publishable scenes without touching infrastructure, while keeping a human judgment step at each point where a model's output would otherwise ship unexamined.

Runtime

01
Describe
A reference image is converted to prose. Nothing is copied from the image itself.
02
Approve
The operator reads and edits the description before anything renders from it.
03
Render
Scenes are generated from the approved prose, with a generic placeholder in the product slot.
04
Place
The real product is composited into selected scenes.
05
Compare
The packshot sits beside every result so product errors are visible feature by feature.

Anatomy

Context

Operator reference image
Product packshots
An editable scene description
Per-job output history

Capabilities

Describe a reference as prose
Generate scenes from a description
Place a real product into a scene
Record estimated cost per call
Fail over between serving pools

Activation

Operator action at each of the three steps

Authority

What can it change?

Autonomy: Recommend
Horizon: Operator-invoked

Read

Uploaded reference and packshots
The operator's own job history

Write

Generated images in the operator's workspace
An append-only cost ledger

Conditional

Render only from a description the operator has seen and approved

Prohibited

Render before the description is shown
Reproduce a face or identity from the reference
Name a brand in a generated description
Publish anything — output leaves only when the operator takes it

Escalate

Description wrong or incomplete — the operator edits it
Product detail wrong in a result — the packshot comparison surfaces it
Transient model outage — noted to the operator when a fallback served the call

Composition and verification

Made from explicit parts.

Composition

Hosted operator interface
Language and image models
Two human approval gates
Append-only cost ledger
Access control in front of the application

Verification

Estimated cost recorded per call against dated list prices
Fallback engages only after the provider's own retry policy is exhausted
Identity excluded at the prompt level, not filtered after
Release-based deploys with the previous release kept for rollback
Both gates added after the failure each prevents

No evidence is published for this entry. The items above are our own account of its source, read before publishing — not something you can check without taking our word for it.

Next step

What you can do with this.

This configuration runs in production, client identity withheld. To discuss the same for your operation, start with the problem rather than the system: a first call is thirty minutes and costs nothing.