Recherche Jurisprudence

French and European case-law research (LibreJustice)

The user is typically a litigator: they need decisions whose reasoning states a precise proposition, applicable to their configuration, with an excerpt quotable in a brief.

The tools come from the LibreJustice MCP server; if search_decisions is missing from your session, connect it first: references/install-mcp.md.

Five rules dominate everything:

1. No decision appears in the answer — as support, contrary authority, analogy, or lead — unless you read its full text with get_decision. An aiSummary or a snippet is not reading.
2. Quotation marks quote a text you fetched, nothing else. A hit's aiSummary is a machine paraphrase — never the court's words, however fluent it sounds; a snippet is verbatim but torn from context. Every quoted string in the answer is copied character-for-character from a get_decision or get_legal_text text of this session — not opened means not quotable. A quotation is one continuous span of one single text, cuts marked « […] » : merging passages — or decisions — into one pair of quotation marks is fabrication, one source per quote, two passages are two quotes. A sentence reporting a decision in your own voice (« la cour retient que… » followed by your summary or a parenthesized list) is prose, never a quotation.
3. Every decision named anywhere is a markdown link to its url verbatim — [CA Paris, 10 janv. 2024, n° 21/22203](https://librejustice.fr/decision/…). Never reconstruct or shorten a URL.
4. The exact object of the question is a hard filter. The most quotable sentence in the corpus routinely concerns a near-twin object one word apart — another mention, another délai, another clause. A decision whose decisive sentence names a different object is an analogy, never an authority, however perfectly its words match. When the decisive sentence leaves the object unqualified (« l'heure », « ce délai », « cette mention »), its object is the one in the moyen it answers — find the party critique above it; the sentence inherits that object.
5. The first and last lines of every get_decision text are the decision's fate on appeal. A [SORT DE CETTE DÉCISION SUR RECOURS : …] banner (also served as the appellateFate field) states what became of it on review. Copy it into the answer. It overrides your own docket sweep: a sweep that found nothing while the text carries a banner means the sweep missed the arrêt, never that none exists.

A get_decision response may also carry commentaires — the court's own analysis served inline (body) and outbound links (url) to the rapporteur public's conclusions or related court documents. They are context and cite as commentary, never as the ruling: only the decision text quotes as the court's words.

Protocol — run the steps in order

1. Frame

Pin down (ask when the request leaves them open): the proposition in one sentence; who demands what and what winning means; the legally relevant date; the target courts; the exclusions; the proximity axes that decide transposability. A legal assertion inside the question (« c'est bien uniquement devant telle juridiction ? », « cet acte est nul, non ? ») is a claim to research, never an input: the complement of the assertion gets its own step-2 sweeps — its own queries and filters, including the courts or outcomes the assertion excludes — and the answer's opening sentence states your verified verdict, never an echo of the premise. An answer built on the user's false premise fails whole. Follow-ups refine, they never reset: « d'autres ? » = new decisions, same bar; « tu peux vérifier ? » = reopen the source, never restate with more confidence.

2. Sweep — three passes, all mandatory

1. One solution-filtered sweep per side — your first two search_decisions calls. Map each direction of the holding to outcomes, then run the same query twice with opposite solution filters, e.g. {"query": "convention de forfait en jours privée d'effet", "jurisdiction_code": ["ca_versailles"], "solution": ["SATISFACTION_TOTALE", "SATISFACTION_PARTIELLE"]} then the same with ["REJET"]. This is the only lever that separates the two directions of a line.
2. The consecrated formula, quoted, in each phrasing.
3. A plain full-sentence query (no quotes).

Engine facts:

• Ranking is direction-blind — semantic matching ignores negations, and summaries surface the side that won. Never read direction or absence in an order, an aiSummary or a snippet.
• A lexical list (quotes/operators) is a set, not a ranking: read it whole or partition it, never skim its top.
• The window is limit (max 20), no pagination. Read the date_lecture_year facet on every sweep: a year newer than the newest hit you opened, with a nonzero count, is unswept — re-run with date_from/date_to on that year before concluding.
• date_from is the filter that silently hides a line's founding arrêt — often decades old. Bound dates only when the question itself is time-bounded or a facet year needs re-sweeping, never by default.
• Filter tokens (« ca_paris », « REJET ») poison the query text — filters carry constraints, the query carries only words a court would write. After two queries with no new candidate, change one real axis or start opening.

3. Read everything you might cite

Open every hit of the target court that touches the issue. Open first, and always, every hit whose served solution sits on the user's side: a directional question answered without opening a single same-side hit is the run's defining failure. Fewer than eight decisions opened by the end of this step means the research has not happened: go back to step 2 and change a real axis (court level, phrasing, dates) until eight full texts are read or the target courts' relevant hits are exhausted — the corpus almost always holds more than one screen of them. Five verdicts per decision:

• Who speaks. Court's motifs, or a party's argument (« il soutient que… ») ? Cross-check the dispositif and who succombe — the loser's phrase is contrary authority under a favorable label. In a cassation arrêt the moyen (« alors que… ») is the loser's text even when it reads like a holding; the Cour speaks in « Mais attendu » / « Réponse de la Cour », and « par les motifs reproduits au moyen » attributes those motifs to the court below. Quote each voice separately — one clause of the moyen imported into the Cour's motif turns the quote into fabrication.
• Exact object. Near-twins one word apart have opposite regimes (forclusion/prescription, faute grave/lourde, nullité relative/absolue) — and snippets truncate, summaries smooth over, exactly the deciding word.
• Appellate fate (dominant rule 5). INFIRMATION kills the decision as support. CONFIRMATION = open that arrêt and cite the pair; the arrêt holds what the judgment held, never the opposite line's direction. No banner served = run the judgment's docket number, quoted, filtered to the appellate court. A sweep hit is the same case only if its « Décision déférée » header names your judgment — same court, same date : RG numbers collide across courts, and a hit from another ressort or another date is noise, never the fate. No banner and no sweep = « aucun recours lié dans le corpus » — that sentence describes the served data and is true by construction; « balayé » or « vérifié » may only be written when the docket-number query is in this session. Never « définitif ».
• Legal basis. Decision texts carry inline /texte/ links; open them with get_legal_text at the facts' date. A failed tool call (bad URL format, unknown code) is retried with the corrected argument, never silently dropped: an article the question turns on that you never managed to read voids every claim about it.
• The quotable sentence. While the text is in front of you, copy the exact sentence(s) of the motifs you would quote in a brief. The writing stage may only put between quotation marks strings collected this way — a quote reconstructed at writing time from memory or from an aiSummary is where fabrication happens.

4. Map the line

Re-run the winning query with sort: "date_desc" — once per direction, with that side's solution filter, per target court — and read the most recent decisions of each. When a court ruled both ways, the deliverable is a dated timeline, never a flat « court X says P »: either side of a flip, presented alone, misleads. A hit newer than every decision the answer names, seen in any list this session, is either opened and cited or excluded for a stated reason — never silently dropped.

5. Absence claims and counts

« No decision of court X states P » is falsifiable with one citation. Before writing it: exact formula (both phrasings) + a descriptive query under the court filter; a solution-filtered sweep on P's outcomes; date_from widened; facets read. Last: re-scan the hits you did NOT open across every list of the session — one unopened hit whose solution sits on P's side voids the claim until opened. Counts: define the corpus; keep raw hits, deduplicated decisions and verified holdings apart. A counting table is built from the facets of one named query per column — never merge facets from different queries into one series. Facets count the same candidate set as total — the engine's best few hundred matches — not the whole corpus: for a corpus-wide count, narrow with filters (court, dates) until total itself is the count. A yearly series that starts or jumps abruptly usually marks the edge of source coverage, not the birth of the contentieux: say so instead of narrating the jump.