We do not ask AI to guess an answer. We preserve domain facts, dependencies, constraints and grounds as inspectable objects. A violation writes nothing and explains why.
| Dialect | Input | Machine criterion | Evidence output |
|---|---|---|---|
| Narrative / White Drum | Long-form serial | Five gates + foreshadowing state machine | Each chapter’s world state is Node-verifiable. Example: a thread marked resolved without payoff is rejected with zero writes. |
| Judgment grounds chain | Judgment text | Statutory sentencing ranges | Zero writes on violation + grounds. Example: a 55% voluntary-surrender reduction exceeds the 40% cap and is rejected immediately. |
| Spreadsheet dependency chain | Excel formulas | relations dependency graph | “Which ten manual inputs determine this number?” A broken formula reference does not produce a plausible-looking trace. |
| CAD review | DWG / DXF openings | Object-level constraints | Audit list + location anchors. An opening count without source-object anchors is not written. |
| Textbook check | DOCX illustration rules | Figure-number / citation consistency | Health report with red flags. A reference to a missing figure cannot be marked passed. |
| Project state | Chat history | Idempotent state transactions | MCP restores in seconds. A stale expected_state_version returns a conflict and writes nothing. |
The point of the law dialect is not to generate a judgment, but to reject an unsupported state change. The local browser demo deliberately pushes a surrender reduction past its ceiling and shows the statutory range and refusal.