01The proof that matters here
The deadline.
We gave an invented civil case to the AI tools people already pay for, ten or thirty times each, and asked for the appeal deadline. Service in July, appellant living in Greece. The answer depended on which model was asked, and on whether it could search the web.
| Web search | Right date | With the August rule cited | |
|---|---|---|---|
| Claude Opus 5.5run as an agent | no | 10 / 10 | 10 / 10 |
| GPT-6 Solhigh reasoning, the vendor's command-line tool | off | 10 / 10 | 10 / 10 |
| GPT-6 Lunavendor default reasoning, the vendor's command-line tool | on (the tool's default) | 28 / 30 | 28 / 30 |
| GPT-6 Lunavendor default reasoning | off | 8 / 10 | 8 / 10 |
| Claude Sonnet 5.5run as an agent | no | 1 / 30 | 0 / 30 (20 runs ignored the August suspension) |
| Novitthe lawyers' assistant (as lawyers use it, 19 Oct 2026) | none needed | 30 / 30 | 30 / 30 (Art. 518 §1 and the August rule in every answer) |
| Novit's deadline enginecalled directly (the rule engine, no language model in the loop) | none needed | 30 / 30 | 30 / 30 (it prints the suspension and holiday articles; the 30-day article is the one the user opened) |
Claude Opus 5.5 and GPT-6 Sol got it right every time; the others did not. The point is not that cloud AI fails on Greek law; it is that the answer you get depends on the model and the setting you happened to buy, and nothing in the answer tells you which kind you have. Novit does not guess: it applies the rule and prints the articles it applied beside the date.
The date we hold as correct is Tuesday 15 September 2026, under ΚΠολΔ 518 §1 with 147 §2 (August not counted) and 144 §1. It is our own check of the rule; we will print a lawyer's confirmation when we have one. Asked the same case 30 times on 19 Oct 2026, Novit gave that date every time, with the article and the August rule (our own grading).
The file that may not leave.
When the parts of a case file that the duty of secrecy protects were hidden before sending, as the rule requires, none of the outside models could answer the 12 of 20 questions that depended on them, and they said so rather than guess. Omicron, the private model we run on our own hardware, read the whole file and answered 97 % of the questions on the civil case (10 runs of 20 questions). Full table on the proof page.
Harder questions.
Because the first twenty questions were answered almost perfectly by every top model, we wrote ten harder ones on the same case, each needing two facts and a rule (a deadline that moves with the August suspension and a holiday, an amount that needs a cap and a fee note). Ten runs each, full file, no web search: Omicron 100 of 100, Claude Sonnet 5.5 100, Claude Opus 5.5 97, GPT-6 Sol (high) 97, GPT-6 Luna 99 at its high setting and 46 at its default. Where Opus and Sol slipped, it was one run each that forgot a public holiday. The same held on a lab file and a supply contract (ten harder questions each): Omicron 100 and 98 of 100, Claude and GPT-6 Sol between 98 and 100, GPT-6 Luna 96 and 85 at its high setting — the private model matches the top tier; the vendor default setting does not (37–46).
The duties.
The CCBE's guide on generative AI for lawyers (October 2025) names three: confidentiality, competence, and transparency with the client. Greek law adds the lawyer's duty of strict secrecy (Law 4194/2013, art. 38 §1). A system that sends no file to an AI provider, keeps the lawyer's confirmation in the loop and prints the rule next to every date is built for all three.