What kind of system this is. The
Bench Book is a retrieval-augmented generation ("RAG") system. The term
describes a specific architecture: rather than asking an artificial
intelligence model to answer from whatever it absorbed during training,
the system first retrieves the governing text — here, the
provisions of the Michigan Court Rules relevant to your question — and
then instructs the model to compose its answer from that text alone. The
model functions less like an expert reciting from memory and more like a
clerk directed to answer only from the record placed in front of them.
Why the architecture matters.
General-purpose AI systems generate language by statistical prediction,
and when asked about law they can produce authority that does not exist —
confident, well-formatted, and wrong. The professional consequences of
relying on such fabrications are by now well documented. This system is
built to foreclose that failure mode structurally rather than by
instruction: the model is never asked what it remembers about Michigan
procedure. It is shown the pertinent rule text, retrieved verbatim from
the official publication, and confined to it.
The verification layer. Confinement is
then checked rather than assumed. Before any answer reaches you, every
citation it contains is tested against the parsed rules: does the cited
provision exist, and was it among the passages the model was actually
given? The results of that audit are displayed with the answer, and each
cited provision links to its verbatim text and to the page of the source
PDF, so the answer can be verified against the rule itself rather than
taken on trust. Where the rules do not address a question, the system is
designed to say so and to indicate where the answer likely resides —
a statute, court precedent, or local order — rather than to guess.
The measured record. In evaluation
against more than a thousand benchmark questions written from the rules
themselves, the system produced no fabricated citations, and it declined
all thirty-two questions deliberately designed to have no answer in the
court rules. Its stored copy of the rules was verified word for word
against the official PDF — all 625 rules accounted for, nothing added or
paraphrased.
What this does not claim. The
architecture minimises fabrication; it does not abolish error. Retrieval
can miss a pertinent provision, and composed language can state a rule
more broadly than its text supports. That is why every answer carries its
sources, why the audit is shown rather than merely performed, and why the
banner above asks you to verify against the official rules before
relying on anything here.
Where the work happens. Search and
verification run entirely on this computer. Only the final composition
step is sent to an external language-model service; your searches and the
documents never leave this machine.
Reference identifiers (for technical staff)
source michigan-court-rules.pdf
(
this tool's verified local copy)
sha-256 3267004f83eacdc8baa71adffadbb46b598c1ed6b4e1dbc9ccd163833e2d202a
provisions 11,860 citable, from 12,298 parsed blocks
passages 3,580 (rule-scoped, 256-token budget)
retrieval Qwen/Qwen3-Embedding-4B, dense, citation router
answers glm-5.2, restricted to retrieved passages
verification parse reconciled against the document's own contents
(625/625 rules); word-count identity 384,507 = 384,507
Michigan Court Rules
Bench Book · search with provenance