World-building field guide

A coherent world is built before the first task is written.

The expensive work is not generating files. It is keeping facts, dates, entities, traps, sources and deliverables consistent across an entire fictional corpus.

Firsthand workflow from one legal-AI contributor · Reviewed September 30, 2026

What is legal AI world building? On some legal AI projects, a contributor designs the fictional matter and creates the document set a model must reason across. The world must be realistic, internally consistent, machine-readable, difficult for the intended reasons, and free from accidental answer clues.
Use it now

Find your phase. Protect the right thing.

Select the stage you are working in. Moonlux will show the control points and handoff evidence that matter at that stage. This planner runs entirely in your browser and does not ask for project files or confidential facts.

Plan the world before describing the files.

Define the precipitating event, parties, task shapes, anchor date and intended difficulty before document production begins.

Protect now

  • One clear precipitating event connects the matter.
  • Every proposed task has a reason it should challenge a model.
  • The anchor date makes the world temporally coherent.

Carry forward

  • Approved parties and entity names.
  • Task intent and planned traps.
  • Unresolved assumptions that require project guidance.
Observed workflow

Generation is fast. Coherence is expensive.

6observed workflow stages
5–7hreported specification stage
43independent gates in the reviewed kit
8repair sessions in one build initially planned for five

These figures come from one contributor and one independently developed legal-world kit. They are useful evidence of where work can accumulate, not market-wide averages or promises about another project.

Failure surface

The recurring problems are broader than proofreading.

Answer leakage

Emphasis, notes, comments or conspicuously broken drafts can reveal the intended answer or turn the world into a checklist.

Cross-file disagreement

A renamed party, new file or changed fact can remain stale elsewhere. One upstream change may affect every task whose scope includes it.

Task-design failure

A trap can disappear during a rebuild, a rubric can double-count one issue, or the corpus can reward a shortcut instead of legal reasoning.

Unreadable evidence

A file may look correct to a person while failing under the extraction tools a model actually uses. A clean visual review does not prove machine readability.

Design direction

Merit should remember the project’s laws.

The future World Builder should not be an unbounded chatbot. It should maintain a local, phase-aware project map: required artifacts, approved facts, dependencies, intended traps, permitted tools, known failures and proof that each required input was inspected.

The first product test is narrower: can Merit help a contributor identify the current phase, inventory the required handoff, and prevent a repeated mechanical or workflow failure without receiving confidential work? That is the threshold Moonlux will test before making a larger claim.

Operating principle

A gate the builder has to remember is not a gate. A dependable control runs at the correct phase, proves it can detect a planted failure, and reports when it could not inspect the relevant material. A skipped check never becomes a pass.

World-building boundaries

Does every legal AI project use this workflow?

No. This workflow was documented from one contributor’s projects. Project instructions, terminology, tool permissions and review procedures differ. The live project’s rules control.

Should I upload a confidential project to Moonlux?

No. This public planner needs no files. Merit’s current browser checker processes selected files in the tab, but contributors must still follow their project’s confidentiality and tool-use rules.

Can Merit build the world for me today?

No. Merit 1.0 is a private pre-submission checker. Phase-aware project mapping and world-building support are a product direction Moonlux is testing, not a current production promise.

Already preparing a submission?

Use the pre-submission checklist and private fictional sample to catch preventable scope, consistency and file-hygiene problems before review.

Open the task checklistJoin the founder beta