Legal AI evaluation has turned into a leaderboard. A table with model names down the left side, legal tasks across the top, and scores in the cells. Best for drafting. Best for research. Best for contract review. A new model drops, the table updates, the rankings shift, and the cycle repeats. It's a recognizable pattern by now.
Ranking models like this only works if the model is the single lever determining output quality. It assumes you can hold the rest of the pipeline constant, swap the inference layer, and read the difference as a clean signal of capability. Systems don't work that way.
A lot happens before a model generates text. A user asks a question, and the system has to figure out what documents matter. Search involves decisions about indexing, chunking, metadata, and what "relevant" actually means to the retrieval layer. Source documents have structure that carries meaning. Headings create hierarchy, definitions create scope, cross-references connect clauses. When a document gets processed for retrieval, that structure either survives or gets destroyed. Poor chunking strategies turn a coherent legal instrument into a pile of detached paragraphs, and the retrieval layer searches them as if they were independent units.
What gets retrieved has to be assembled into something the model can use. Order matters. Selection matters. Whether the retrieved chunks include enough surrounding context to make sense matters. Whether the system found the right passage at all matters more than how well the model summarizes the passage it was handed.
If the system has an agent layer, the variables multiply. A multi-step agent decides what to search for first, whether the results are sufficient, what to look for next, and when to stop. It decides whether to verify a citation against the underlying source or trust the model's recollection. It decides whether to check its own intermediate reasoning or just pass it along. These decisions dictate the final output far more than raw model capability, because they determine what the model works with.
The model sits at the end of this chain. A strong model given the wrong documents will produce a confident, well-written answer to a question that wasn't really asked. A weaker model given the right documents, assembled with enough context to interpret them, will produce something useful. The model generates language. It doesn't determine whether that language is grounded in the right source material, or whether the right material was found to begin with.
The common defense against this is to point out that a system isn't tied to one model. Vendors will say they can swap the model underneath. That's technically true and substantively empty. Flexibility at the inference layer says nothing about whether the search found the right passage. It says nothing about whether document processing held structure together or destroyed it. It says nothing about whether the agent's multi-step reasoning actually verifies anything or just produces more text. Swapping models is a statement about the model layer. The actual problems are upstream.
Start with search. Habeas retrieves before it generates. It runs natural language search across a closed, Australian-specific corpus of case law, legislation, and commentary, understanding legal meaning rather than matching keywords. Jurisdiction, court level, and date filters apply before the query runs, not after. The authority is found and checked before a sentence gets written, not attached afterward to justify one.
Document processing is where structure either holds together or falls apart. Every proposition Habeas returns carries a clickable citation traceable to the subparagraph it came from, not the case generally. Confirming an authority actually supports a claim becomes a read, not a re-investigation. Document Stores extend this to a practitioner's own materials, matter-specific repositories the system reasons over exclusively when asked, so the same retrieval discipline applies to uploaded briefs and evidence as it does to the public corpus.
The RAG layer is where retrieval either grounds the output or sits next to it unused. Habeas reasons over what it retrieved rather than generating an answer and back-filling sources. That ordering is the actual hallucination mitigation. It isn't a claimed side effect of a bigger model. And it holds regardless of which model sits at the generation step.
The agent layer is where the system decides what to verify. Habeas's research assistants break one query into multiple searches, call multiple tools and sources in sequence, and can run an adversarial pass, surfacing the strongest authority against a position, not just for it, before opposing counsel does. That's a system-level decision about what to check and re-check, sitting well upstream of which model phrases the final answer.
None of this produces a single comparable score, and none of it fits neatly in a table. Which is why the industry defaults to model comparisons. They are easy to read and easy to rank. They give the impression of a moving field, of progress you can track by watching numbers shift. The harder layers don't work that way. There is no benchmark for whether a retrieval layer respects document structure. There is no leaderboard for whether an agent's verification step actually catches errors. The work that determines output quality is harder to measure, harder to rank, and harder to talk about in public. That's why it gets skipped in benchmarking, and why it matters more in practice.
The model is the part everyone can see. It has a name, a version number, a release date. It gets benchmarked, ranked, and compared. In a well-built system, it's the least important part to get right. The upstream pipeline matters more, even if it gets less attention. The model receives what the system gives it. If the system feeds it the wrong material, the best model on the market will still produce the wrong answer. The model is the last step, not the first. Evaluating it as the only step is the easiest thing to do, and the least informative thing to believe.
