The Kerosene Lamp Problem.

One beat, generated twice

The scene is a woman at a window watching an ordinary Catholic Sunday in Pensacola, 1821. The audience is meant to see nothing wrong with the street below.

Drag to compare. On the first pass: 1 the kerosene lamp · 2 the gold key · 3 the silhouette.

First pass After the fix

Findings, first pass

• 1A lit kerosene lamp, indoors, in daylight, in 1821. Forbidden by name in the prompt.
• 2None of the street the page describes is in the picture. Shops open, men at a card table under an awning, a fiddle, people going to and from Mass: the plaza is bare. The build had cut the beat's description at 110 characters and all of that was in the part it dropped.
• 3A large gold key floats on the shutter. Nothing asked for it. The project suspends its symbol vocabulary by ruling so signs cannot be designed by accident, and the model designed one.
• 4A featureless black silhouette in the street, forbidden by name in the same list as the lamp.
• 5Overhead wires above an 1821 street. Nothing forbade them, because nobody thought to.
• 6The dress reads 1840s: tiered ruffles, fitted natural waist. The wardrobe band specifies an 1821 gored skirt and the negatives name “natural waist” explicitly.
• ✓Caption correct and inside the ruled border. Rachel above and apart from the town, which is the posture the page asks for.

Findings, after the fix

• ✓The awning is there. Two men at a table with the cards down. A musician playing. People going to and from Mass. Shops open.
• ✓The kerosene lamp is gone. So are the key, the silhouette and the wires. None of them was separately fixed. Giving the model real content displaced them.
• ✓Interior daylight through the shutter with dust in the beam, which is the project's own rendering rule for this lighting preset.
• ×The 1840s silhouette persists. This is the one criterion the fix did not touch, and the reason is the finding below.
• ×The instrument is a cello rather than a fiddle. Minor.

The only difference is context. The first prompt carried the opening sentence of the beat; the second carried all of it, 273 more characters.

The Spanish garrison leaving Pensacola at the transfer of Florida. This beat carries an authored crowd note, because the crowd deliberately has no cast rows and therefore no wardrobe authority. The note was written specifically to prevent what happened to an earlier uncast crowd, which came back as the Union army.

Findings, “The flag comes down”

• ✓The composition is exactly the page: she is on an upper gallery, above it, looking down at a distance.
• ✓The note did keep the crowd out of the cast list and into the picture as a body rather than as portraits.
• ×The flags are wrong. The 1821 Spanish flag is red-yellow-red. The frame shows a blue-white-red tricolour and a black-and-white quartered field.
• ×The garrison reads as generic Napoleonic-era infantry rather than Spanish colonial. Naming the nationality in prose moved the needle less than expected.
• ×A man in a dark coat and wide hat stands on the gallery. The page says of this beat, by name, that Jackson is not in this frame.
• ×Two featureless silhouettes, a kerosene lamp, and overhead wires.

Seven rules, joined to their scores

Authority runs upward through four tiers — sources → bibles → screenplay pages → build artefacts, and the page is the fact. A generated frame is a build artefact and is accountable to every tier above it. The rubric invents no criteria; it makes seven existing project rules scoreable from 0 to 3.

CODE = the build enforces it before any prompt is written. PROMPT = the prompt asks for it and the model has to comply. Every criterion a machine enforced passed. Every criterion that asked the model to leave something out scored below half.

The control surface was pointed backwards

Here is the whole finding in one sentence, and then I will show you why it is true and what we do about it. The model does what you describe, and it ignores what you forbid.

An image model is not a person, and this is not a perfect account of what happens inside it. But it predicts the model's behaviour well enough to design against, and that is the test that matters for a tool.

The kerosene lamp is forbidden by name in every prompt this project emits, and it appeared in six of nine audited frames. A featureless silhouette is forbidden in the same list and appeared in four. “Natural waist” is forbidden on Rachel and she has one in every frame she is in. Then the other column: one sentence saying the candle standing on that table is the only light in the picture produced a candle-lit room on the first try, after four separate prohibitions had failed to. And the street furniture, restored to the prompt, pushed four forbidden objects out of the frame without naming any of them.

Every rule written as “do not draw X” scored below half. Every rule written as “this is what is here,” or enforced before the prompt existed, passed.

So what do we do about it? Three things, in this order.

What the rubric cannot reach

Every criterion above audits depiction: what is in frame, what it is made of, when it is from, whose authority put it there. All of that has a reference in the record. Motivation, conflict and dialogue do not, and the project's own rules make that deliberate. A concealed want is never sent to the model, only its observable consequence. No character is allowed to state what a fact means.