There is a real problem with how language model-based legal research tools are typically evaluated. The evaluation focuses on whether the answer is correct, not on whether the practitioner using the tool can verify that it is correct. These are different standards, and for professional legal work, the second standard is the one that matters.
A language model that correctly identifies the relevant OR provision for a contract dispute and correctly summarizes the applicable BGE doctrine has produced an accurate output. But if that output does not include the specific BGE citation, the current SR provision number, and a link to the authoritative source text, the practitioner cannot use it in a submission. They must independently verify everything before relying on it. The research time saved by the initial generation is partially or fully consumed by verification.
This piece examines why source traceability is a precondition for usable legal research output, not an optional feature, and what it means technically to build a legal research tool around that constraint.
The Confidence-Correctness Problem
Language models generate text that sounds authoritative regardless of whether the underlying content is accurate. In the context of legal research, this creates a specific risk: the model produces a citation to a case or a provision that does not exist, or that exists but says something different from what the model attributes to it, in language that is indistinguishable from a correct citation.
The legal research community refers to this as "hallucination," though the term understates the severity of the problem for professional contexts. A model that confidently cites a non-existent BGE decision is not making a random error; it is producing plausible-format output that requires specific expert verification to detect. A practitioner who receives a "BGE 147 III 312" citation and does not check the Federal Tribunal database is relying on unverified output. If that citation is incorrect, the reliance creates professional risk.
Source traceability addresses this by changing the structure of the output. Instead of "the relevant provision is Art. 97 OR, which has been interpreted in BGE X to mean Y," a source-traced output provides: "the relevant provision is Art. 97 OR (SR 220), current version at [link to admin.ch SR 220 Art. 97]; the Federal Tribunal addressed this in BGE 133 III 421, retrievable at [link to bger.ch entry]; the key statement is in E. 3.2." The practitioner can check each link independently and verify that the citation exists and says what the output claims.
We are not saying that source tracing eliminates all risk of incorrect legal research output. It does not: a model could link to a real decision that does not support the proposition attributed to it. But source tracing changes the verification burden from "I must independently find the source from scratch" to "I must check whether the provided link supports what the output claims," which is a substantially lower verification burden.
The Temporal Accuracy Problem
Swiss law changes. The OR has been amended at intervals. The ZPO has been revised. New Federal Tribunal decisions modify the interpretation of unchanged provision text. A language model trained on data from a cutoff date will not reflect statutory amendments or BGE decisions issued after that date.
For some practice areas, the temporal gap between a model's training cutoff and the current state of the law is not material for routine matters. A basic contract formation question under Art. 1 ff. OR is unlikely to have been fundamentally altered by recent amendments. But for provisions in active legislative evolution, for practice areas where the Federal Tribunal has been active in recent years, and for recently amended statutes, a model that cannot confirm the currency of its output is not safe to rely on without independent verification.
Source traceability addresses the temporal problem by requiring the research tool to link to live sources at the time of the query rather than to static training data. An output that links to admin.ch's current SR text for a provision reflects whatever version is live at that moment. An output that cites a provision without a current-source link may be reflecting a pre-amendment version.
This is the technical reason why building a legal research tool around retrieval-augmented generation (connecting the language model to live, indexed sources rather than relying on training memory) is not a design preference but a correctness requirement. The language model contributes the reasoning and synthesis; the live source index contributes the current, verifiable text. Neither component alone meets the standard for professional legal research output.
What Source Traceability Requires Architecturally
From a technical standpoint, source traceability in a Swiss legal research tool requires several things that are not trivially achieved:
A current, complete index of the SR at admin.ch: the SR is updated continuously as amendments are published in the Amtliche Sammlung. An index that was built three months ago and not refreshed does not reflect amendments published in the interim. Keeping the SR index current is an ongoing operational requirement, not a one-time build task.
A complete index of Federal Tribunal decisions from bger.ch: BGE decisions are published in the official series on a rolling basis. BGer (unpublished) decisions are available at bger.ch with a delay after the decision date. An index that reflects BGE volumes up to 147 but not 148 is missing the most recent decisions, which may be the most relevant for current practice.
Cantonal court decision indexes with coverage documentation: as discussed in our earlier writing on cantonal precedent aggregation, cantonal databases vary substantially in coverage. The source traceability requirement applies to cantonal results as well as federal results, which means the tool must know what it has indexed and communicate that to users.
Citation parsing and linking: the research output must contain machine-readable citations that map to specific documents in the index, not just human-readable citation strings. A citation string like "BGE 133 III 421" must be resolved to a specific URL in the bger.ch database, not left as a string for the practitioner to enter manually.
The Practitioner Workflow Implications
When source traceability is built into the research output, the practitioner workflow changes. Instead of treating the research tool output as a starting point for independent verification, the practitioner can treat it as a starting point for substantive evaluation of the cited materials. The verification that the source exists and is correctly identified has already been done by the tool. The remaining question is whether the cited source actually supports the argument they need to make.
This is a meaningful reduction in the research burden for routine citation tasks. The part of legal research that consists of finding and confirming sources is a precondition for the substantive work. Tools that short-circuit that precondition with plausible-but-unverified citations create hidden verification overhead. Tools that complete the precondition step correctly allow the practitioner to spend research time on the substantive analysis.
We built Silex Legal around this constraint from the start, which affected architectural decisions that would have been harder to change later. The tradeoff is that indexing live sources and maintaining currency requires ongoing operational work that a pure language model deployment does not. That cost is justified because it is the cost of producing output that practitioners can actually rely on without incurring unacceptable professional risk.
A Realistic Boundary Statement
Source traceability does not make legal research output self-verifying. The practitioner still needs to read the cited source and evaluate whether it supports the use being made of it. A BGE decision may be cited correctly, current in its version, and linked to a live URL, and still not support the argument it is being used for if the relevant Erwagung applies a narrow factual context that does not match the current matter.
What source traceability does is move the verification responsibility from existence and currency to substance. That is a higher-value task for a practitioner to perform, because it requires the legal judgment that only a qualified professional can apply. The existence-and-currency check is mechanical; the substantive evaluation is professional work. A legal research tool should handle the mechanical check so the practitioner can focus on the professional work.
