EraLock — Period & Culture Accuracy for AI Video

The problem

Ask any current video or image model for 1985 Seoul and you get Chinese temple roofs and red paper lanterns. Ask for a Korean coal briquette and you get a barbecue firelighter — right colour, right size, right function, wrong object. Ask for a street and every shop sign comes back covered in letterforms that look like a language and are not.

These are not random glitches. They are three specific, repeatable mechanisms:

• Adjacent-culture bleed. A thinly-photographed culture gets overwritten by a heavily-photographed neighbour. The model is regressing to the regional mean, not guessing.
• Signature-prop approximation. For a culturally specific object, the model substitutes the nearest thing sharing silhouette and function — confidently, and without any visible sign of failure.
• Text as texture. Any surface that should carry writing gets filled with pseudo-language.

Each has a countermeasure. Generic instructions are not among them. "Make it culturally accurate" does nothing measurable; naming red lanterns and temple roofs works.

What EraLock does

Give it a country, a year, and your scenes. It returns:

1. An era anchor across six production-design categories — architecture, street, vehicles, interiors, clothing, hair and grooming. Six, because each is a separate thing the model gets wrong independently. Vehicles and hair carry the most information per word; hair is where models most reliably default to the present day.

2. A culture-exclusion negative set naming the specific elements that bleed in for your target, not a generic disclaimer. Ships with maps for East Asia, Southeast Asia, Central Asia and the Caucasus, Eastern Europe, Latin America, and Africa — plus a method for deriving one when your target is not covered.

3. Signature-prop form locks that specify the distinguishing geometry and then ban the wrong object by name. Naming the correct object is not enough. You have to close the door on the substitute.

4. Composition steering away from text surfaces. No prompt reliably stops fake lettering. The countermeasure is to not frame the sign — shift the angle, throw the plane out of focus, or composite real text in post.

5. An inspection checklist for what comes back. Because verifying your settings is not inspection. Looking at the frames is inspection.

Sample output

Input: Seoul, 1985, night alley

Positive

authentic South Korea, 1985 — mixed low-rise brick and concrete housing, fully paved roads, heavy overhead wiring; a narrow alley at night, wet asphalt, single bare bulb over a doorway, a man in a wide-shouldered jacket walking away, cool fluorescent spill from a window, 35mm, shallow depth of field

Negative

STRICTLY NOT Chinese and NOT Japanese: no Chinese architecture, no red lanterns, no hanfu, no qipao, no Chinese temple roofs, no kimono, no torii; no text, no signage, no lettering, no watermark

Plus prop locks where the scene needs them, and the inspection checklist for the batch.

Negative delivery adapts to your model: --no for Midjourney, the negative field for Stable Diffusion, Seedance and Kling, and — for Veo and Sora, which have no negative field — prohibitions folded into the positive prompt as statements about the scene rather than commands, which those models tend to ignore or invert.

Included references

Who this is for

Anyone generating period or culturally-specific footage: short-form history channels, documentary reconstruction, period drama previs, museum and heritage work, ad concepts with a retro setting.

Most useful when your subject is a culture that is not the most-photographed one in its region — that is exactly where the bleed is worst, and exactly where audiences notice fastest.

What this does not do

Stated plainly so you can decide before you subscribe:

• It runs on your own model key, and it does not render anything. Requests execute against your own LLM API key — you need one, and you keep direct control of what you spend. What you get back is prompts, locks and checklists; generation happens in your own video tool on your own account. Nothing is bundled, and there is no quota to exhaust.
• It does not fix text generation. Nothing does. It routes your composition around the problem.
• It is not a historical research service. The Korea presets are deep because they came from repeated production. For other places it gives you a rigorous method, not a finished answer.
• It cannot guarantee a clean batch. No prompt technique can. It cuts the defect rate hard and tells you exactly what to look for in what survives.

Where it came from

A year of running an automated pipeline that produced period pieces daily — short-form and long-form, mostly mid-century Korea.

Every rule here traces to something that shipped wrong first. Chinese roof curvature that reached publication and came back as viewer complaints. A fake title card in the first second of a clip, missed because inspection frames were sampled from the middle. Episodes set in Germany and England rendered with Korean architecture, because the country was named in the script but not in the opening image prompt. A coal briquette that came back as a firelighter and nearly passed review.

Every one of those passed a settings check. The prompt was right and the picture was wrong.

That is the habit this skill is really selling: look at the frames.