Technology roadmap

Research candidates, in sequence.

Every technology below is marked “Research Candidate — not implemented or validated by CellSight.” No dates, milestones, performance gains or delivery promises are stated.

Abstract layered render of geometric foundation blocks, curve planes and feature-map grids in violet, red and off-white. — Abstract progression illustration from product foundation to model research.

Track 1

Product Foundation

Re-establishing what exists before anything is claimed.

Under Verification

Repository and model reconstruction

Objective: Recover a reproducible build of the previous workflow and its model artefacts.
Present status: Not started in this concept release.
Required evidence or dependency: Access to the original repository, environment and weights.
Indicative sequence: Indicative sequence: first

Research Candidate — not implemented or validated by CellSight.

Under Verification

Model and dataset cards

Objective: Document model identity, intended use, data origin, splits and limitations.
Present status: Template not yet drafted.
Required evidence or dependency: Reconstructed build and confirmed data rights.
Indicative sequence: Indicative sequence: second

Research Candidate — not implemented or validated by CellSight.

Track 2

Evaluation Readiness

Building the conditions under which a result could be believed.

Research Candidate

Subject-disjoint evaluation

Objective: Evaluate on splits that prevent subject leakage between train and test data.
Present status: Protocol not defined.
Required evidence or dependency: Documented subject identifiers within the source data.
Indicative sequence: Indicative sequence: third

Research Candidate — not implemented or validated by CellSight.

Research Candidate

Technical input-quality checks

Objective: Detect out-of-specification fields before any downstream step.
Present status: Specification not written.
Required evidence or dependency: An agreed technical acceptance specification.
Indicative sequence: Indicative sequence: third

Research Candidate — not implemented or validated by CellSight.

Research Candidate

Calibration and selective-prediction experiments

Objective: Study whether output dispersion can be made interpretable and whether abstention helps.
Present status: Not implemented or validated by CellSight.
Required evidence or dependency: A reproduced baseline and a held-out evaluation set.
Indicative sequence: Indicative sequence: fourth

Research Candidate — not implemented or validated by CellSight.

Track 3

Model Research

Candidate techniques under consideration only.

Research Candidate

EfficientNet benchmark

Objective: Benchmark a convolutional baseline against the reconstructed reference.
Present status: Not implemented or validated by CellSight.
Required evidence or dependency: Reconstructed baseline and evaluation protocol.
Indicative sequence: Indicative sequence: fifth

Research Candidate — not implemented or validated by CellSight.

Research Candidate

DinoBloom feature benchmark

Objective: Assess whether domain feature representations are useful for this task.
Present status: Not implemented or validated by CellSight.
Required evidence or dependency: Licensing review and evaluation protocol.
Indicative sequence: Indicative sequence: fifth

Research Candidate — not implemented or validated by CellSight.

Research Candidate

Learned OOD evaluation

Objective: Investigate detection of inputs outside the modelled distribution.
Present status: Not implemented or validated by CellSight.
Required evidence or dependency: A characterised in-distribution reference set.
Indicative sequence: Indicative sequence: sixth

Research Candidate — not implemented or validated by CellSight.

Research Candidate

Caveated explainability experiment

Objective: Explore region attribution as a review aid, with explicit caveats on interpretation.
Present status: Not implemented or validated by CellSight.
Required evidence or dependency: A stable baseline and a written interpretation boundary.
Indicative sequence: Indicative sequence: sixth

Research Candidate — not implemented or validated by CellSight.

Track 4

Future Integrations

Described for direction only; nothing is built, connected or planned for delivery here.

Research Candidate

Workflow interoperability study

Objective: Understand what an integration surface would need to satisfy before it could be considered.
Present status: Not implemented or validated by CellSight.
Required evidence or dependency: Evaluation readiness completed first.
Indicative sequence: Indicative sequence: later

Research Candidate — not implemented or validated by CellSight.

Research Candidate

Evidence register automation

Objective: Explore keeping capability states, owners and review dates in one maintained register.
Present status: Not implemented or validated by CellSight.
Required evidence or dependency: Named owners and an agreed review cadence.
Indicative sequence: Indicative sequence: later

Research Candidate — not implemented or validated by CellSight.