SolutionsHealth
Health. Research pipelines with an audit trail.
Patient-cohort and clinical-pipeline work run as enforced, evidenced, replayable processes with compliance surfaces built in.
Runs as a service today.
01Capabilities
Research pipelines that keep their receipts.
Health research must answer for itself later. Every cohort, pipeline, and conclusion is recorded as it was made, and replays exactly.
- Patient-cohort construction
- Cohorts are built as runs against defined criteria, kept with the cohort so it can be reconstructed exactly.
- Clinical-pipeline execution
- Research pipelines run as enforced processes: a stage advances only when its gate's criteria are met.
- Evidence in a signed graph
- Evidence is captured as signed, append-only facts: what a run concluded, recorded with what it concluded it from.
- Replayable history
- Every decision reconstructs from the journal; an audit replays what happened.
- Compliance surfaces
- Access, provenance, and forgetting are first-class: facts can be retracted, entities tombstoned, bytes excised, each as a further signed fact.
02The audit trail
A journal that proves what it knew.
Evidence lands as signed, append-only facts; a query answers what was known at any past date, and a compliance erasure is itself a signed, provable act on the record.
03Walkthrough
One decision, replayed for an audit.
The audit question is always the same: why was this decided, and on what evidence. One cohort-inclusion decision answers it here.
decision.selected
An auditor picks one decision: the inclusion of a de-identified record in a study cohort.
replay.started
The run re-executes from the append-only journal. No memory of the original run is needed.
memory.asof
The memory answers with what was known at decision time, whatever has been learned since.
evidence.listed
The signed facts the decision read: the criteria, the source records, their provenance.
gate.checked
The inclusion criteria that held, with the outcome of each check.
replay.completed
Same inputs, same projection, same decision, rebuilt from the record.
The week before an audit looks like any other week.
04Calibration
Calibrated conservatively.
The harness is tuned against evals on task correctness and process adherence. The claim is alignment to the process you defined; nothing promises a clinical outcome.
- Evals on task correctness
- Evals score clinical-task correctness and adherence to the required research process, separately.
- Adjusted until the scores hold
- Process, skills, and memory are tuned until runs hit the target score and hold it when the evals re-run.
- Alignment to intent
- Calibration aligns a run to the process you defined; the console never claims a clinical outcome.
05What this makes possible
The dosage question.
Take the best agent prompt engineering can build: skills, careful instructions, review loops, the works. Would you put it in charge of a patient’s medication dosage? No responsible team would.
That is the point: the trust never comes from better prompting; it comes from enforcement. Clinicians author the decision process, every step is gated, a human breakpoint holds before anything reaches a patient, every decision replays from the journal. Such a workflow could, potentially, be made trustworthy.
We imagine going further: with enough paired data, patient situations and the dosage a clinician chose for each, a harness could be trained to produce the right workflows, an expert validating each before use. An illustrative direction only; nothing here is a shipped capability or a regulatory claim.
06The console
What the console shows.
- A patient-cohort builder working on de-identified, generic data.
- A clinical pipeline with stages and gates.
- An evidence and provenance viewer, backed by signed memory.
- Compliance and authorization surfaces: access, provenance, retraction.
- A replay affordance on any decision.
Every audit-trail property above is open source. What a download cannot give you is the calibration to your protocols and a pipeline operated alongside your clinical team.
A pipeline you can answer for later, with signed evidence behind every decision.
Replay one cohort decision, evidence and provenance in view.
A Health demo replays one pipeline decision end to end, evidence and provenance included.
Or emailhello@a5c.ai
The intro call is 30 minutes and free.