AI-assisted workflow

Generate a detailed case report

Create a source-constrained internal report that stress-tests theory and audits every proof chain.

The Generate detailed report button submits a bounded case packet to the selected model. The report is internal work product for attorney review. It is not a pleading, a prediction, or a substitute for checking the record and law.

The report’s required method

Lattice directs the model to follow a fixed audit sequence:

  1. Stress-test the attorney theory instead of adopting it as true.
  2. Trace authority → claim → element → fact → evidence.
  3. Separate supported, contested, inferred, missing, and attorney-authored propositions.
  4. Audit chronology, including undated events and conflicting dates.
  5. Give adverse authority and contrary facts full weight.
  6. Cite exact Lattice record IDs for each factual finding.
  7. Disclose budget omissions and dataset limits.
  8. Identify the specific verification, discovery, research, or attorney decision needed next.

Generic advocacy, invented facts, invented holdings, invented record IDs, and unsupported confidence are prohibited by the prompt and checked by the server. A report that cites an unknown record is rejected instead of displayed.

Read the classifications correctly

  • Supported means the cited dataset supports the statement; it does not mean the proposition is judicially established.
  • Contested means the dataset contains a dispute or meaningful contrary material.
  • Inference identifies reasoning drawn from cited premises, not a recorded fact.
  • Gap means a chain or necessary record is incomplete.
  • Attorney theory identifies counsel’s strategic position and must not be confused with evidence.

The confidence percentage measures support within the submitted dataset. It is not a probability of winning.

Choose report depth

  • Fast is best for a quick orientation or a smaller, well-organized matter.
  • Balanced is the default for ordinary cross-record analysis.
  • Deep is for large or ambiguous matters where long chains, conflicts, and adverse material warrant more reasoning.

The model name, combined input/output tokens, and cost appear at the bottom of every report. Regeneration creates a new analysis from the current graph; it does not silently alter the prior graph.

A disciplined review workflow

  1. Read limitations before the executive summary.
  2. Check each high-impact source ID against the underlying record.
  3. Confirm that contrary sources were characterized fairly.
  4. Convert open questions into assignments, discovery, research, or graph updates.
  5. Correct the graph rather than asking the model to write around bad data.
  6. Regenerate only after material context changes.

Use Print report for a review copy. The printed warning remains: verify AI-generated analysis before reliance or filing.

Examples of useful outcomes

  • The report finds that every element of breach is fact-linked, but the causation element has no evidence-linked fact.
  • It separates counsel’s notice theory from a fact whose status is only alleged.
  • It flags two records placing the same meeting on different dates.
  • It identifies a favorable case linked to the claim but not to the element it actually supports.
  • It shows an adverse case with a strong court weight but no fit analysis or verification note.
  • It asks counsel whether a concession should change the theory or merely be distinguished.

Next: review AI-proposed actions.