Language models read.
They are not allowed to decide.
Dasyante is built on one architectural rule: the LLM does the only thing it is uniquely good at — understanding clinical language — and deterministic code owns everything that must be exact: identity, verification, linking. The parts that need to be right every time are written in rules, not weights.
A strict division of labor.
The model turns free-text prose into candidate frames — findings, sites, sizes, changes. From that point on, no model touches the record. Whether a fact survives, which lesion it belongs to, and how it connects across years is decided by inspectable, versioned code that behaves identically on every run.
Understanding language.
Parses prose into candidate finding frames — the one task where probabilistic models genuinely excel. Proposals only.
Identity, verification, linking.
Deterministic, versioned code decides what survives, what it belongs to, and how it connects — the same way, every run.
Three stages. Only one is probabilistic.
Propose frames.
The model reads each report and proposes finding frames: structured slots for finding, site, laterality, size, change, and assertion — each with the sentence it claims as evidence. Proposals, nothing more.
Pass the gate.
Every proposed frame passes through the evidence gate: deterministic checks that the quoted sentence exists in the report, verbatim, and supports each slot. A frame that fails is dropped, not repaired.
Build the tracks.
Surviving frames are linked into tracks by a deterministic composite key — finding type, anatomy, laterality — not by asking a model "is this the same lesion?" Same reports, same tracks, every time.
The gate does not negotiate.
The evidence gate is a set of hard checks, not a confidence score. The anchor sentence must appear in the source report character-for-character. The slots must be recoverable from that sentence. A fact that cannot cite its source does not get a lower rating — it gets deleted. This is why every green chip in Dasyante means the same thing.
Hallucination isn't filtered downstream. It is structurally unable to enter the record.
It improves the way software improves — not the way models drift.
When a clinician corrects a fact in the workspace, that correction doesn't vanish into a fine-tuning run. It becomes a proposed rule: a small, human-readable change to the deterministic layer.
- 1 · Correction. A clinician marks a fact wrong or amends a slot. It's captured with its full context — report, frame, track.
- 2 · Proposed rule. The correction is generalized into a candidate rule — explicit, inspectable, written down. You can read exactly what the system wants to learn.
- 3 · The frozen gate. The candidate runs against a frozen regression suite: every previously verified case, untouched. If fixing this case breaks any earlier one, the rule is rejected.
- 4 · Human-gated merge. A rule that passes still ships nothing on its own. A human approves the merge. Every rule is atomic and revertable — one commit, one diff, one undo.
No forgetting. No drift. No silent updates.
The system on Tuesday is the system from Monday, plus exactly the rules someone approved in between.
Determinism is a clinical feature.
In oncology, "the system changed its mind" is not an upgrade — it's a liability. Dasyante's deterministic core means its behavior can be audited, its changes reviewed like code, and its answers reproduced years later, exactly. That is what it takes for software to sit inside a medical record.
Watch the mechanism run.
The demo shows real frames, real tracks, and real anchors — on synthetic patients.