Anterior generates medical records by reversing its inference workflow, starting with a label and reasoning trace before creating the data that would produce them.
2
Decision trees let clinicians sample diverse policy paths and test rare cases that customer production data may not contain.
3
A coarse-to-fine generation pipeline produces patient journeys and documents in text, with consistency checks and clinician-controlled skills refining the result.
Summary
Anterior works with scanned medical records that arrive mostly by fax. These records contain handwriting, tables, checkboxes, images, and information spanning a patient's clinical history. They are useful for evaluations, but contracts prevent Anterior from retaining them or deriving redacted, anonymized, or other persistent copies. Anuj Iravane explains how the team generates synthetic records instead. It reverses the normal workflow: sample an outcome, derive a policy reasoning trace, then generate the patient record that would lead to it. Explicit policy decision trees provide more varied conditioning inputs than an LLM inventing cases from scratch. The pipeline builds records from patient invariants to encounters to documents, then checks consistency and agreement with the original task labels. Clinicians control the process through human review and editable skills. Around 90% of Anterior's datasets are synthetic, and clinicians distinguished synthetic records from real ones about 60% of the time in a blind review.
Medical records are rich, varied inputs for policy decisions
Anterior builds AI systems for healthcare administrative workflows such as prior authorization, payment integrity, and HEDIS measures. Iravane describes the shared task as applying a policy to highly unstructured data. The input often arrives as scanned fax bundles, with bad handwriting, tables, checkboxes, key-value pairs, and images. These records can cover an entire clinical trajectory, and every patient's journey differs. He describes them as an observation through a fuzzy lens over a person's lifespan. The data has a long tail of rare cases and nuanced scenarios, which makes it useful for testing workflows.
The data needed for evaluation is often forbidden from persistence
Iravane says healthcare accuracy standards are unusually high, and 95% is not good enough for Anterior's work. Medical records are therefore a regular source for datasets and evaluations across its workflows. The problem is that Anterior cannot keep the records under its contracts. The restrictions also cover deriving information, redacting or anonymizing the records, and keeping derivative copies. As a result, nothing survives in a persistent dataset. The team needs a way to create evaluation data without retaining the customer records that would otherwise provide it.
Reversing inference makes synthetic labels correct by construction
The normal task starts with unstructured data and a policy, follows a reasoning trace, and produces an outcome or label. Anterior reverses that process. It samples a label, finds a reasoning trace for the label, and generates the medical data backwards from those inputs. This conditions generation on a diverse set of outcomes and traces, which helps avoid the repeated cases that can result when an LLM is asked to invent records directly. Since the generation starts with the task labels, the resulting records can be checked against those labels through a round trip. Iravane says this can remove the expensive ground-truthing step.
Symbolic policy trees provide a more useful source of diversity
Anterior models policies explicitly as symbolic structures similar to decision trees. Iravane uses a CPAP medical-necessity policy as an example, with branches describing conditions under which the device should be approved or rejected. These structures help Anterior execute policies with better accuracy and consistency in LLM-based workflows. They also allow the system to sample reasoning traces for a chosen outcome. Iravane argues that this distribution is more uniform and effective than asking an LLM to invent diverse cases. Sampling from the policy can expose rare scenarios that a customer's production sample may not contain.
A coarse-to-fine pipeline builds records around encounters
The synthetic record pipeline starts with patient invariants such as biological sex, birth date, and blood group. The system combines those invariants with a reasoning trace to generate an ordered list of events and provider encounters, called the patient journey. It then creates a document plan for each encounter and generates the documents using the plan and the patient's preceding history. This mirrors the way documentation is produced during real provider encounters. The layered process keeps prompts efficient and can scale to longer patient journeys because the system fans out into documents instead of placing the whole record in one context window.
Evaluation loops catch contradictions across parallel documents
Documents are generated independently in a parallel fan-out process, so Anterior runs a refinement loop over the result. One evaluation checks consistency across all documents for contradictions, inaccurate details, and conflicting information. The system also compares the generated record with the labels and task inputs that began the pipeline. This round-trip check tests whether the record remains in agreement with the intended workflow outcome. Because the label was sampled at the start, the data can retain correct labels by construction while the evaluation loop improves particular parts of the documents.
Text generation is enough because PDFs add little value
Anterior generates and evaluates records as plain text and Markdown rather than rendering them as PDFs. Iravane says it is possible to create a PDF, but the team sees little value in doing so because modern PDF parsers can convert complex PDFs into Markdown. Keeping the pipeline in the text domain makes the generation and evaluation process simpler. The synthetic record does not need to reproduce the original file format to test the policy workflow. What matters is the information, its organization across encounters and documents, and its consistency with the intended task.
Clinicians control the pipeline through human review and skills
Anterior gives clinicians a chance to steer generation at each stage. They can inspect production cases, use failures or interesting examples as ideas, and guide the pipeline toward similar scenarios. Clinicians also own the pipeline logic through skills that run on Anterior's internal agent harness. Patient journeys, document generation, document enrichment, and evaluations are all modeled as skills. Adding a new document type can involve creating and attaching a skill file rather than changing engineering code. Iravane presents skills as an interface that lets AI engineers and domain experts work on the same workflow.
Synthetic datasets let teams test customer edge cases before launch
Around 90% of Anterior's datasets are synthetic, and the team uses them for evaluation. In a blind review, clinicians distinguished synthetic records from real ones about 60% of the time. Iravane calls this an area for improvement. The practical benefit is that datasets can be created just in time for customer deployments instead of waiting for customer data. Teams can model edge cases, simulate them, and test workflows before going live. His advice is to reverse the inference workflow, sample diversity from a distribution that fits the use case, imitate how the original data was generated, and give domain experts control of the pipeline.
"It is really important to give your domain experts the keys because these are the people who know about your data."16:00
Who should watch
You work with sensitive healthcare records that cannot be retained or turned into evaluation datasets.
Your generated cases repeat common patterns and miss rare policy branches or production edge cases.
Clinicians understand the data better than the engineers maintaining the pipeline, and you need them to edit workflows without code changes.
