MEDICAL AI RESEARCH · PARIS
Agentic workflowsfor medicine.
The traceability, the audit and the results of the whole chain stay in one signed .oia file.
One worked case at week 37, real from end to end: a fetal MRI, two segmentation providers, a canal computed from the bone, and an audit that refused two regions. It is a single example of what the format is meant to carry, not the extent of it.
One case file.All its layers.All its proof.
A .oia is one file. It holds the scan, the shapes drawn from it, what each shape is and who drew it, the answers computed from them, and a seal over the whole. Opening it re-checks every claim it makes, so nothing inside has to be taken on trust.
PROOF
The seal over all of it.
Provenance, the audit verdict for every region, and a signature. Opening the file runs the checks again.
TIME
The questions asked and the answers computed.
States from a named solver, kept apart from what was observed, and replayable exactly.
SIGNAL
The scan the shapes came from.
The original volume on its own grid, bit for bit, so the evidence never has to be fetched from elsewhere.
MEANING
What each shape is, and who drew it.
A name, the model that produced it and its version, linked by hash. Label 3 stops being a number.
STRUCTURE
The shapes drawn from the scan.
Closed surfaces in millimetres, so the file can re-measure them instead of taking them on trust.
An orchestrator of providers.
A workflow is declared, not programmed. Every step names a provider, or is code an agent wrote. This is the one that built the case above, top to bottom.
IMPORT · DICOM connector
The scan arrives untouched.
On its own grid, bit for bit, with a declaration of what the import keeps and what it loses.
SEGMENT · nnU-Net · fetal
A model draws the fetus.
Audited the moment it lands: one component, euler 2. It passes, and the file keeps who drew it.
SEGMENT · TotalSegmentator
Another draws the pelvis. Two regions are refused.
hip_left declares one component and measures four. The refusal is not a crash: it stays in the file as a finding.
DERIVE · code · sha256
No provider computes a birth canal. The agent writes one.
A distance transform over the bone, narrowest at 99 mm. The code travels in the file with its environment, hashed.
# the canal, from the bonewall = distance_transform(~bone)axis = ridge(wall)canal = 2 * wall[axis]# narrowest 99.4 mm at z 83SIMULATE · MuJoCo 3.12
The question leaves the file. States come back.
The solver program is preserved. Replaying rereads the answer exactly; rerunning it is a new answer, and says so.
MEASURE · code · sha256
At term, the head clears by 0.6 mm.
Smaller than a voxel and smaller than the segmentation error. The file says undetermined, because that is the honest answer.
THE CASE FILE · one signed .oia
Every result written back, by hash.
THE WORKFLOW IS DISPOSABLE.
THE OBJECT IS THE DELIVERABLE.
Chaining models stopped being hard. What stays hard is that someone who ran none of it can still check what came out.