Roadmap
A staged path, with each step gated by evidence, from a validated first model toward dependable radiological support.
Our intent is deliberately narrow: to earn trust through evidence before asking for it. We report what the model gets right and where it falls short, we close a gap before we claim a capability, and we will not place a model in clinical use before it has passed validation and appropriate regulatory review.
Each stage opens only when the previous one passes checks that are fixed before the evaluation is run. Version 1 followed this rule when its checkpoint was selected, and every later stage will follow it too.
Tuberculosis, done properly
Version 1 is delivered and validated on small internal sets. Tuberculosis remains our sole focus through Version 5.
- Held-out test round, run once on the final candidate Planned
- Comparison against our earlier version under identical evaluation Planned
- Closing the sensitivity gap across image sources Targeted by the next training round
- Correct outline placement on non-square images Fix identified
Next-generation models and workflows
Versions 2 to 4 are iterative training rounds, each using broader and more varied data obtained through research partnerships, and each measured against the same checks. Each round targets the gaps measured in the last, which is intended to make the model more reliable and capable and the workflow around it better.
- Serial imaging: tracking treatment response by comparing a patient's radiographs across visits.
- Structured reports generated alongside each outline.
- Research and validation partnerships: we are seeking research institutions and data partners to evaluate the model on independent data.
Version 5: a dependable tuberculosis model
Version 5 completes the tuberculosis foundation: a model that is safe, consistent across image sources and presentations of the disease, accurate on held-out data, and efficient. Each quality has a measurable target that is fixed before evaluation.
- Safe: a low false-flag rate on held-out normal radiographs, and no clinical use before appropriate regulatory review.
- Consistent: comparable sensitivity across image sources, with the current source gap closed.
- Accurate: outline quality at or above the previous version on a held-out test set, evaluated once.
- Efficient: inference speed and compute cost measured and reported.
Beyond tuberculosis
Once Version 5 passes its checks, we will extend to additional conditions, one at a time. Each new condition begins only after the previous model passes its own pre-set validation rule. The long-term aim is a family of validated medical imaging models on a shared foundation, built with the same discipline.
What support unlocks
| Support | What it unlocks |
|---|---|
| Data partnership | Broader, more varied training data, which targets the source gap directly. |
| External validation | Independent evaluation on data we have not seen. |
| Funding | Compute, annotation and the held-out test rounds that each gate requires. |
| Research collaboration | Joint protocols for serial imaging and treatment monitoring. |
Trust and data stewardship
Evidence before claims.
Results are reported with their limitations, including the cases the model misses.
Examples chosen by rule.
The examples shown follow a selection rule that was fixed before any results were generated.
Source and attribution.
Images shown on this site are credited to their sources, and any rights holder can request removal.
No clinical use without review.
Elavris Lite remains a research prototype until it has passed validation and appropriate regulatory review.
A site that collects nothing.
This site does not use analytics, cookies or forms.
Interested in supporting or collaborating?
Write to [CONTACT EMAIL] or see the evidence for version 1.
No clinical use before appropriate regulatory review.
Research prototype. Not a medical device. Not for diagnosis.