Richard asked a narrow question with a wide shadow. Our scarcest resource for Maureen is held-out test data, and RGCC, the best we have, is not great. Olink kept emailing him. What did we already know, and what could we buy?

The notes were there, in pieces

More prior work existed than either of us remembered.

• paan_planning.md had a table of supervision-producing tests and a ranked acquisition list. It named an organoid drug screen the single highest-payoff purchase.
• The protein-expression writeup and two June vector notes covered Olink.
• The population-supervised planning doc argued for cell lines and PDX models as proxy patients.

None of them had been joined up around the question “what counts as held-out test data?”.

The test on the to-do list was already running

The mission spec listed “order tumor-informed ctDNA (Signatera)” as an open Tier-1 action, and two vector notes called the assay “still outstanding”. GH #131 said nine Signatera reports sat unparsed.

A plain keyword search of the document store returned 21. Nineteen are distinct monthly draws from January 2025 to March 2026, two are duplicates, and one more draw appears only in the reports’ history tables.

The series reads:

• 7.26 MTM/mL in January 2025, rising off treatment to 20.74 by October. The very first draw was already positive, about seven months before the August CT showed the nodes.
• 0.15 on October 27, after FOLFOX, cetuximab and encorafenib began.
• Negative every month since December.

I first called this the best held-out data we had. Richard corrected that the same day, and he was right. Each report is one number, and the number has been zero since December. As validation it is worth one regimen-level label, a combination whose credit can’t be split between its drugs. Its future value is as a relapse tripwire.

Two smaller lessons came with it. The semantic document search had returned nothing for “Signatera”; a plain ILIKE found all of them. And the OCR block stored beside each report’s text misreads the history table: one draw came out as five different numbers across reports, while the text layer agreed with itself in all eighteen that carry it.

What RGCC cannot see

Checking RGCC against that one in-vivo response exposed a structural blind spot rather than noise. The panel has no encorafenib or binimetinib. The BRAF and MEK inhibitors it does test score zero on both draws, and cetuximab barely registers.

The regimen that worked is a combination whose whole point is that BRAF inhibition fails in colorectal cancer unless EGFR feedback is blocked too. A single-agent assay cannot see that. So a ranker that correctly puts those drugs high for her is penalized by the validation set.

RGCC’s own co-authored outcome study (BMC Cancer 2022) reported a positive predictive value of 47% in 36 refractory patients. That design is too selection-biased to condemn the assay, but it is all the outcome validation the assay has.

Olink answers a different question

Olink is the PDR2 paper’s native biomarker input: plasma protein levels, measured as a deviation from a reference. That makes it a candidate for the model’s input channel, not drug-response ground truth.

The practical picture is also worse than the June note said:

• Core labs sell it by the plate: an 86-sample minimum at Vanderbilt.
• NPX values are relative to their plate, so a lone sample cannot be placed on the UK Biobank scale.
• The promotions in Richard’s inbox are reagent kits for labs with their own sequencer.

Where held-out labels can actually come from

More sequencing adds inputs, and we are already rich in those. Only response measurements add labels. In vivo, that means drugs she actually receives, which is one regimen so far. That leaves two honest routes.

1. Proxy patients, now and free:

   • CTRPv2, GDSC and the PRISM Repurposing screen, whose compounds are 77% non-oncology.
   • A 2026 Oslo biobank of 213 organoids from 102 MSS patients, with its drug data open.
   • The Novartis PDX encyclopedia’s colorectal arm.

   Personalizing to each proxy and testing against its measured responses turns an n-of-1 check into a benchmark of the personalization mechanism itself. PRISM also screens outside compounds, which would let a frozen candidate list be tested prospectively.

2. Her own tumor, at the next biopsy. The PARIS organoid test, now run by the nonprofit Cure First, tests up to 47 drugs with dose-response, including combinations. Nothing like it can run while she is NED. So the useful work today is logistics, arranged in advance, so that the next biopsy does not go straight into formalin.

Richard’s follow-ups

Richard’s reply moved the work on in four directions.

• Keeping the record current. The Kaiser token had expired, and Kaiser never grants a refresh token. No Cerbo connection is configured at all, and the connectors still carry silent-failure bugs. That became #512: a freshness monitor, a re-auth reminder, and an audit.
• The proxy benchmark and his response-space embedding. He saw that both are the TabPFN spike’s first track. They went into #327, together with the sources and UK Biobank.
• UK Biobank. Applications are paused until late 2026, and the fees are modest. The real barrier is that an individual can’t apply without an institutional affiliation.
• A harder question. He asked whether a system that ranks drugs against a disease through “treats”-style edges can ever discover anything, when the regimen that worked for Maureen is a combination the graph cannot express.

On the ranking, the answer is reassuring: every drug in the vocabulary is scored, so the design does not limit candidates to known indications. On granularity, it is not. The graph has no notion of a drug combination, and her variants are still gene-level. He wants the end-to-end run first, to see what is really possible before redesigning, and that is the right order.

The general lesson is an old one in a new place: before recommending that something be ordered, search the record for its name.