Fluency is not evidence. That sentence is the whole argument in miniature, and it is why ACRA exists.
Large language models can sound clinical. They can arrange sentences the way a careful clinician might. They can cite nothing and still feel cited. In medicine, that fluency is not a feature waiting for better prompting. It is a failure mode with institutional consequences. A recommendation that cannot be walked back to verified sources is not decision support. It is a polished guess wearing a white coat of syntax.
ACRA, the Auditable Clinical Reasoning Agent, is our research answer to a narrower and harder question than whether an LLM can answer medical questions. The aim of the paper is to design and evaluate an architecture for clinical decision support in which recommendations are evidence grounded, inspectable after the fact, and shaped to sit honestly beside the transparency and evidence standards that UK governance environments already ask of AI in health. Full clinical deployment sits outside the scope of the project. The work is architecture and evaluation direction for future feasibility, not a medical device, not a trial result, not an NHS rollout story.
That aim breaks into sub problems that matter as much as the headline. How do you retrieve from trusted corpora instead of from the model's prior alone. How do you force reasoning steps to stay bound to what was actually observed in retrieval. How do you leave a structured audit trail that a human can reconstruct. How do you score claim to source alignment instead of settling for surface fluency. Those are engineering questions dressed as research questions. They are also the only kind of questions worth asking if you intend for clinical AI to survive contact with real scrutiny.
## Not sidelined. Occupied.
For a stretch of the year the lab has been deep in Leverage, our proving ground for inspectable, evidence bound agent systems in a hard market domain. That work is public enough to visit: the Floor lives at [leverage.blueprintlabsai.tech](https://leverage.blueprintlabsai.tech/), and the product page sits at [/leverage](/leverage). People who only see the shipping surface sometimes ask whether ACRA was parked.
It was not.
Leverage and ACRA share a thesis family. Different domain. Same refusal: an agent that cannot show its working is not ready for stakes. In trading, the stakes are money and trust under time pressure. In clinical reasoning, the stakes are guidance that must remain reconstructable under governance. We did not abandon the clinical research line to chase a product. We pressure tested the deeper pattern, evidence before narrative, audit before theatre, in a domain where we could iterate daily. That proving ground sharpens what ACRA must become. It does not replace it.
So this post is not a pivot announcement. It is a return to the paper's architecture in public language, and a clear statement of where the build plan starts for the last quarter of 2026.
## The problem the paper refuses to paper over
Vanilla retrieval augmented generation improved the industry's conscience more than it solved the industry's problem. You retrieve documents. You stuff them into a context window. You ask for an answer. Sometimes the model uses what you retrieved. Sometimes it performs a fluent collage that looks grounded while quietly inventing the joint between sources, or inventing past them.
In ordinary chat, that failure is annoying. In clinical decision support, it is disqualifying. Governance framing in the UK, from NHS AI Lab transparency expectations to NICE style evidence standards, does not ask systems to sound wise. It asks them to be accountable: what was used, what was inferred, what can be inspected. ACRA takes that demand as an architectural constraint, not as a slide in a pitch deck.
The research posture is deliberately modest. We are not claiming a clinic ready product. We are asking whether a particular agent architecture can make evidence grounded clinical recommendations that remain auditable end to end, and whether that architecture can be evaluated with metrics that care about grounding rather than only about answer polish.
## Architecture: from retrieval to structured audit
ACRA is a research stack with four planned movements. None of them is optional if the aim is real.
First, evidence retrieval over trusted corpora. The proposal points toward dense retrieval across materials such as PubMed abstracts and NICE style guideline text. The point is not "search the open web and hope." The point is to begin inside corpora that already carry an institutional notion of evidence, then to surface the passages that will become the only legal fuel for later thought.
Second, a structured reasoning engine in a ReAct style loop: Thought, Action, Observation. The agent does not leap from prompt to paragraph. It plans. It acts, typically by retrieving. It observes what came back. It thinks again only in the presence of that observation. This is slower than one shot completion. Slowness is the feature. Multi step structure is how you create places to enforce discipline.
Third, a source binding rule that is almost embarrassingly simple and almost always missing. Thought steps should not introduce claims that are not supported by a prior retrieved observation. The model is allowed to interpret evidence. It is not allowed to smuggle new medical assertions out of parametric memory and dress them as retrieval. Binding is the difference between chain of thought as theatre and chain of thought as a controlled instrument.
Fourth, structured audit as a first class artifact. ACRA plans JSON audit records that carry the query, the retrieved documents, the reasoning chain with its bindings, the final text, timestamps, and session identity. The final response is then generated conditioned on that chain and on citations, not conjured beside an invisible scratchpad. The audit log is not a debug dump for engineers who already know the system. It is the product of the architecture: a reconstructable trail from question to recommendation.
If you compress the stack into one sentence: retrieval feeds grounded reasoning, grounded reasoning feeds a recommendation, and every step leaves a paper trail that can be scored.
## What we will measure when we measure
The proposal's evaluation direction matters because it refuses the vanity metric of "sounds good to a demo audience."
ACRA is meant to be compared against a vanilla LLM baseline and against standard RAG. The comparison dimensions include accuracy, hallucination style failure modes, a proposed Evidence Grounding Score (EGS) for claim to source alignment, and audit log completeness. EGS is the philosophical center of that list. An answer can be locally correct and still fail if its claims do not bind cleanly to retrieved evidence. An answer can be incomplete and still succeed as research if the system knows what it does not know and leaves an honest trail.
We will not invent results here. There are no clinic deployments to announce. There is no quiet NHS pilot hiding behind this post. There is a measurement agenda that matches the aim: if the architecture cannot improve grounding and auditability under controlled evaluation, it does not deserve a prettier interface.
## Q4 2026: where plans start
The last quarter of the year is when ACRA moves from proposal gravity into laboratory motion. Honest scope looks like this.
We start with the corpus and retrieval substrate. Choose and prepare a trusted evidence base in the spirit of the proposal, PubMed abstract material and guideline style text, then make dense retrieval boringly reliable: indexes, chunking contracts, citation identifiers that survive into the audit record. If retrieval identity is mushy, every later binding is fiction.
We implement the reasoning loop next, with source binding enforced as a structural constraint rather than as a polite system prompt. Thought, Action, Observation must be machine checkable enough that an unbound claim is a failed step, not a rhetorical flourish we forgive because the final paragraph was elegant.
We freeze an audit schema early. Queries, retrieved documents, bound reasoning steps, final text, timestamps, session ids. Completeness is an evaluation target. Schemas that drift every week cannot support EGS or audit completeness scoring. The log format is part of the science.
We stand up the evaluation harness against the two baselines named in the proposal: vanilla LLM and standard RAG. Early runs will be small, instrumented, and public in their humility. The question for Q4 is not "did we ship clinical AI." The question is "can we show, on a defined task set, that the auditable loop improves grounding and trail quality in the ways the paper claims matter."
We keep deployment out of scope on purpose. Feasibility work can prepare the road for later clinical collaboration. It cannot pretend the road has already been walked. That boundary is how we stay consistent with the proposal and with the lab's habit of refusing theatre.
If Leverage taught us anything useful for this quarter, it is cultural rather than mechanical: evidence trails beat folklore, shadow evaluation before behavior change, and reconstructability as a design object. Those instincts transfer. The clinical stack still has to be built as its own system, against its own corpora, scored with its own grounding metrics.
## The literary fact under the technical one
There is a quiet poetry in auditability, if you let yourself see it. A clinical recommendation is a sentence that asks to be believed at the worst possible moment. Belief without a trail is charisma. Belief with a trail is work: retrieval that can be named, thoughts that can be checked, observations that can be replayed, a final claim that still points at the page that authorized it.
ACRA is that work, stated as architecture.
We are a lab that builds where evidence matters. For months the proving ground was a trading desk you can open in a browser. For the last quarter of 2026, the research desk turns back toward the clinical proposal with the same stubbornness: no fluent void, no unbound thought, no recommendation that cannot survive being asked "show me why."
The paper's aim remains the compass. Design an auditable, evidence grounded clinical reasoning agent. Evaluate whether grounding and audit quality actually move. Leave deployment to a future that has earned it.
That is where we start.