edikt gov grade-compile

Grade the editorial quality of your compiled governance. An LLM acts as judge, reading every rendered .md under the governance directory — ambient core and topic files — and scoring the result 0–10 on six dimensions. The grade is advisory — it never changes your governance.

The six dimensions:

• coherence — related directives grouped into the same topic file
• conciseness — reminders tight and actionable, no redundant noise
• signal-to-noise — high-priority directives surfaced, not buried
• description quality — each topic's registry description in .edikt/topics.yaml tells a reader, mid-task, whether that topic is the one they need — the routing mechanism now that there's no keyword table
• tier assignment — each directive lands on the surface its scope justifies (universal → ambient core, path-scoped → directive index, topical → skill package)
• no double loading — a directive's body appears on exactly one ambient surface, not duplicated across the ambient core, a topic file, and a skill

There is no keyword-matched routing table in the current render (compile_schema_version: 3) for a grade to check the accuracy of — that mechanism was retired and replaced by topic descriptions and paths:-based auto-loading. Compiled governance is the surface your agents read on every task; a degraded surface silently weakens guidance even when every underlying decision is correct. This command gives you a quality signal plus specific, actionable findings.

Synopsis

bin/edikt gov grade-compile                                            # default dir + model
bin/edikt gov grade-compile --dir .claude/rules/governance --model claude-opus-4-7

Run it after gov compile. The grader runs out of the compile loop and never affects compile determinism.

How it works

1. Pre-flight gate. Confirms the edikt tier-2 binary is available. If neither edikt nor bin/edikt resolves, the command refuses and directs you to install — it never falls back to a degraded path.
2. Dispatch the LLM-as-judge via the Go binary, which reads the compiled files and scores them.
3. Write the report to .edikt/state/compile-quality/<timestamp>.json and print the overall score plus the six per-dimension scores.
4. Present the results. Reads the newest report, leads with the overall and per-dimension scores, then lists findings — each names its dimension, a severity (info / warning / critical), a message, and an optional suggested fix. Any dimension below 6/10 is called out as the highest-leverage area to improve, pointing at the specific files and sections named in the findings.

The grade is a signal. Re-authoring decisions or re-running gov compile is your call — the command does not edit governance in response to a grade unless you ask.

Stub mode

Set EDIKT_GRADE_COMPILE_STUB=1 to skip the LLM dispatch entirely and emit a canned fixture report. Use this in CI or when wiring downstream tooling without spending tokens:

EDIKT_GRADE_COMPILE_STUB=1 bin/edikt gov grade-compile

Related

• /edikt:gov:compile — produces the compiled governance tree this command grades.
• /edikt:gov:score — scores overall governance quality (context budget, directive compliance, manual-directive health).
• /edikt:gov:review — reviews governance document language for enforceability.
• Compile — how decisions become compiled governance.