Practising oncologists reviewed the engine's plans for the first time

Until this point, the project measured itself. There is an internal scoring script that rates a generated plan out of 100, and it is a reasonable proxy for whether the machinery is behaving. It is not evidence that the output is any good.

In April 2026 the engine generated treatment plans for a real patient — the two-track output the charter requires, standard and aggressive, with sources attached to each recommendation. Practising oncologists read them and judged them strong.

What this changes

The success metric. Clinician judgement now sits above the internal rubric: when a change would raise the score but a reviewing oncologist would think less of the output, the score is wrong. That reordering is the point of this post.

What it does not establish

• One case is not validation. It is a single patient, one disease area, and a handful of readers. Nothing here supports a claim about accuracy in general.
• The reviewers assessed plans, not outcomes. They judged whether the reasoning and the cited options were sound — not whether a patient did better. Those are different questions, and only the second one is clinical evidence.
• Nothing about the disclaimer changes. This remains a research and support tool, not a medical device, and every plan it produces still has to be verified by the treating oncologist.

On publishing the case itself

The charter (§9.3) sets three conditions before any real case can become public: informed consent from the patient, removal of all identifying information, and ethics committee approval from the treating institution. The project owner reports all three are now in place.

Publishing the case as "Reference Case #0" is a separate step and has not been done yet. This post deliberately carries no clinical detail about the patient — it does not need any to make its point.