Smithery is now a part of Arcade.dev! Read more in our announcement here
Scores the quality of an AI legal output before it goes out, in two layers: objective (mechanically checkable - do the citations exist, are the provisions and dockets real) and subjective...
Before you send it, score it - and keep what is checkable apart from what must be judged. An AI output can read smoothly and still be wrong. This skill scores it in two layers: objective (facts you can verify mechanically) and subjective (quality you have to judge against an explicit rubric). Mixing the two is where false confidence comes from - "sounds right" is not "the citations check out".
It scores; it does not write or send. The decision is the lawyer's; the scorecard is the basis.
Delegate, do not guess:
The objective result is binary per item: matches / does not. Even one mismatched citation means the output is not ready, whatever the subjective score.
Score each dimension 1-5 with a one-sentence reason:
| Dimension | 1 | 5 |
|---|---|---|
| Legal correctness | thesis wrong or unsupported | thesis sound and grounded in the rule |
| Completeness | omits material issues | covers everything material to the question |
| Clarity | muddled, inconsistent | precise, readable for the reader |
| Jurisdiction fit | confuses systems/rules | correct for the governing law |
| Grounding (anti-hallucination) | claims with no basis | every claim has a basis |
When the judge has no basis for a number, it does not invent one. Instead of a forced 3 it writes UNCERTAIN, with a subcategory and a note on what evidence is missing:
Hard rules:
"AI that knows what it does not know" as a field on the scorecard, not a claim about it.
A score card is often issued twice: once on the first pass, once after revisions. The second pass is where scores quietly rise, because the author now knows the charges and can argue around them. Two hard rules.
Monotonicity. A re-score may lower the decision at any time, for any reason. Raising it requires new evidence, not new argument. New evidence means Layer 1 re-run with a different result, or material that was not in the session before (the judgment finally supplied, the contract finally attached). "On reflection this is good enough" is not evidence and does not move the score.
Anchors. While any of the following is open, the decision cannot rise above Full verification, whatever the subjective average says:
Only re-running the tool that set an anchor can clear it. A judge's opinion cannot. Same rule
agentic-risk-asi-pl applies to finding severity: direct evidence is not negotiable.
Record every blocked attempt to raise the score. Seeing who tried to move the verdict, and on what argument, is often worth more than the verdict.
LAYER 1 (objective): citations OK/X | provisions OK/X | dockets OK/X | completeness OK/X
LAYER 2 (1-5 rubric):
Legal correctness: 4 - <reason>
Completeness: 3 - <...>
Clarity: 5 - <...>
Jurisdiction fit: UNCERTAIN (INSUFFICIENT_EVIDENCE) - the output rests on a judgment whose
text is not in the session; missing: the reasoning as handed down
Grounding: 4 - <...>
Average: 4.0 (from 4 numeric dimensions; UNCERTAIN outside the average)
ANCHORS: none (or: citation [2019] UKSC 12 unmatched - decision capped at Full verification)
DECISION: Revise (Completeness 3 - limitation period not addressed)
On a re-score, add one line:
RE-SCORE #2: Revise -> Revise. Raise to Send rejected: argument only, Layer 1 not re-run.
Two-layer method (objective matching + a subjective LLM-as-judge rubric on a 1-5 scale) based on DISC-Law-Eval from DISC-LawLLM (Fudan DISC Lab), Apache-2.0. Exam data and judge model dropped; rubric dimensions, thresholds and framing are MateMatic's own. MateMatic interpretation, not the position of any bar or regulator.
These rules apply to every skill in this plugin, including where the skill itself is silent. They are copied into each skill, so they hold whether you install the whole plugin or a single skill.
If an error would pass without being stopped by layers 1-5, the fault is in the skill. Fix the tool, do not add a note. "This is not legal advice" stops no error and transfers no responsibility.
(GDPR Art. 6), (Case C-311/18, CJEU).[check on EUR-Lex].The tag sits next to the line it concerns. A case number that exists is not enough - check the content.
Nothing is sent, filed, signed or published before a qualified person checks and approves it. A skill output is a draft, not a finished document.
The verification foundation is jurisdiction-neutral - the method, not the substance of any one legal system. It connects to no external source (no MCP connectors) and makes no outbound calls of its own; what you type still goes to your configured model (see https://github.com/matematicsolutions/awesome-matematic-skills-en/blob/main/TRUST.md). Substance comes from separate plugins (eu-law-sources). Deep national law (e.g. Polish case law) lives in the Polish marketplace, by design.
