The model is not the product. The loop is.
Helix is a clinical operating system the sponsor runs itself. Data comes in, the AI drafts, a human signs, and every signature becomes a rule the system cannot forget. The model is replaceable; what the sponsor learns is not.
Per company materials dated September 2026, Helix runs in production at a cell-therapy (ATMP) biotech, supporting registrational trials and multiple IITs. The partner is anonymised under agreement.
A trial submits one thing in the end: an evidence chain.
Today that chain is assembled by hand across eight functions and breaks at every handoff. Worse: it learns nothing. A mistake corrected in the last trial is made again in the next; when the person who decided leaves, the judgment leaves with them.
Industry-reported figures, not Helix measurements.
A system built on a model gets replaced. A system built on a loop gets stronger.
Frontier models get an order of magnitude cheaper and a step stronger every couple of years. Betting a product on one model is betting on a part that will be swapped. Our only design test: if tomorrow's model is ten times better, is Helix more useful or less? It has to be more, because context, decision history, eval sets and execution controls do not live in the model. They live in the loop, and a better model multiplies onto them.
Every action leaves evidence. Every piece of evidence improves the next decision.
Data enters the sponsor's lakehouse and becomes context. The AI drafts inside a policy boundary. A human signs. The signature enters the ledger with its evidence. Corrections become regression cases. Cases enter memory and guard the next change. Each turn of the loop leaves the system one layer thicker, with nobody asked to build it.
- DataData
- ContextContext
- DraftAI drafts
- SignHuman signs
- EvidenceEvidence
- CorrectCorrection
- MemoryMemory
Learning
Corrections become regression cases; a threshold cannot change until it passes them.
Growth
Each new programme and site joins the same context graph instead of opening a silo.
Self-growth
The loop thickens with every cycle, without anyone being asked to build it.
One number from PDF to signature. One signature guarding the next change.
Two synthetic walkthroughs you can click through. First: an ALT value from a central-lab report, through OCR, controlled terminology, SDTM, ADaM and TFL, into the Decision Ledger a human signs. Second: the correction in that signature becomes a regression case and, two weeks later, blocks a proposal to loosen the gate.
A system that can decline is a system you can trust.
Self-improvement without boundaries is just drift. Every improvement in Helix passes a gate first: changes run the eval before release, nothing ships that fails regression, and the AI generates only inside the boundary the policy engine draws. Every tool has three returns: a value; an abstention, naming the missing evidence and the cheapest next experiment; or a veto, when the premise is false and the chain must stop with its reason recorded. A human always signs.
A value
Inside the trusted boundary; with model version, input hash and threshold, recomputable.
An abstention
The question stands, the evidence is missing: which edge, which evidence, the cheapest next experiment. A row is written; the value is null.
A veto
The premise is structurally false and no experiment fixes it: the chain stops, the reason is recorded.
Swap the model or the vendor. The sponsor keeps everything it has learned.
A CRO's work leaves with the contract; people, judgment and context go with it. Helix sits on the sponsor's side of that line: every trial run on it leaves rules, thresholds, eval cases and project context behind, owned by the sponsor, and the next trial starts from there. CROs can still work inside Helix; Helix is the memory those hands work from.
Corpus
The sponsor's lakehouse: source to statistical output, versioned snapshots with field-level lineage.
Ledger
Evidence, rationale, alternatives, policy, approver and outcome for every signature.
Eval sets
Red-line tests and regression cases every change must pass first.
Policy registry
The versioned history of rules, thresholds and gates, and why each changed.
From keeping the trial coherent to letting a human sign a decision rather than assemble one.
Today Helix assists: drafts, answers and records aligned to one context. Each turn of the loop moves it closer to the decisions themselves: how to escalate a dose, how long to hold an observation window, whether a programme proceeds, where a regulator will push back. Clinical judgment and sign-off stay with people; what follows is direction, not status.
- 1CoherenceIn production
- 2Decision intelligenceNext
- 3Controlled executionNext
- 4CompoundingNext
- 5Tiered autonomyHorizon
What is true today, and what is direction, kept apart.
The harness and lakehouse are private to the sponsor; the foundation model is called through external APIs behind a vendor-switch layer. A fully private loop is the goal, not a current claim.
Today it assists: drafts, answers and records aligned to one context. Decision support is the direction; clinical judgment and sign-off stay with people.
ALCOA+ and 21 CFR Part 11 are reference standards; no certification is claimed. Efficiency figures are internal targets, not validated SLAs. All demonstration data on this site is synthetic.