[ 03 · BUILDING THE EVIDENCE ]

The evidence map before your next experiment

What do I actually need to have proved, and am I about to spend money on the wrong study?

Most early programmes generate evidence in the order the science suggests, which is rarely the order a submission needs. Mapping the required evidence against what you hold turns a sequence of experiments into a plan, and usually reveals one study that should happen sooner and one that should not happen at all.

WHAT THIS IS NOT

Orientation for translating research toward a regulated product — not a determination of device status, regulatory pathway, submission type or clinical adequacy, and not a substitute for qualified regulatory advice. SPEQ is independent: linked institutions, incubators and agencies are third parties, and nothing here implies affiliation, endorsement or partnership with any of them.

ARE YOU HERE?

  • You are designing the next study and the justification is momentum rather than a gap
  • Nobody has written down what the eventual submission has to demonstrate
  • You can describe your results but not which requirement each one satisfies

WHAT TO DECIDE NOW

  • What the eventual evidence package must contain, at the level of "characterisation, safety, performance" rather than protocol detail
  • Which existing results already satisfy part of it, and which will need repeating under different conditions to count
  • Which study is on the critical path, which is usually not the one the team is most interested in

START KEEPING THESE RECORDS NOW

  • The mapping itself — requirement against result against gap — because it is the artefact that makes a diligence conversation short
  • Any study run before formal standards applied, with its conditions recorded honestly, so a later decision about whether it can be relied on is possible at all
  • Reference standards, control material and their characterisation, which age out quietly

THE PROBLEM THIS ANSWERS

Development work generates records whose evidential value depends on properties captured at the time — contemporaneity, material provenance, instrument state — and none of them can be added afterwards.

What it costs: The loss is silent. Nothing fails, no alarm sounds, and the result simply stops being reproducible evidence at the moment somebody needs it to be.

NOT YET MEASURED — Repeat use of the evidence inventory by one organisation, which is the signal that it is maintained rather than read once.

An AI-enabled product depends on dataset provenance, model lineage and a frozen evaluation, and each of those is routinely reconstructed from memory rather than recorded when it was true.

What it costs: A performance claim that cannot be reproduced cannot be defended, and retraining without recorded lineage makes the earlier evidence unattributable to anything shippable.

NOT YET MEASURED — Governance preparation runs originating from a startup account rather than an enterprise one.

WHERE THIS GOES WRONG

  • Running the pivotal study before the analytical method that measures its endpoint is validated. The result then measures the method as much as the product.
  • Assuming a study can be retro-fitted into compliance. Conditions that were not controlled cannot be documented into having been.
  • Building the evidence map from the technology rather than from the claim, which produces a thorough package answering a question nobody asked.

WHAT THIS DOES NOT ANSWER

  • Which specific studies a regulator will require, which depends on the pathway and is a question for scientific advice rather than for a guide
  • Whether existing non-GLP work can support a submission — sometimes yes, sometimes as supporting rather than pivotal evidence, and it is fact-specific