Onkydra
CHANGELOG

What we shipped.

Engineering, methodology, product, business. Mostly the results that came back badly: an axis read the wrong way round for five days, a lineage result that turned out to be arithmetic, a cost line wrong by 4.4 times. Every figure traces to a file in the repository, with the command that reproduces it.

16 August 2026·7 min read

The axis was the wrong way round for five days, and every test agreed.

Our in-silico knockout table is nested [target][programme][cell state]. The serving layer read [target][cell state][programme], and a commit whose subject line announced it was fixing an inverted claim broke the one code path that was right. Three of the four cell-state names are spelled exactly like the three programme names, so every wrong read returned a plausible number.

Read post →

16 August 2026·6 min read

Take the gene out of its own readout: one result keeps 99%, the other keeps none.

Our headline knockout readout is a mean over three marker genes, and the gene we knock out is one of them. Storing the per-marker atoms let us re-score every result with the target removed from its own average. OLIG2 keeps 99% of its effect. OLIG1 keeps exactly zero, so the whole of its apparent effect was arithmetic, and every OLIG1 result on that programme is retracted.

Read post →

16 August 2026·8 min read

Biology or handling? The deposit already contained the control arm we said it did not have.

The strongest non-lineage comparator in our knockout results is an immediate-early gene, which the literature calls a dissociation artefact. That was an argument, not a measurement. The reference turned out to carry both arms of the experiment: 11 single-cell libraries against 36 nuclei libraries. The immediate-early panel falls in nuclei at Cliff's delta 0.96, and all three lineage programmes fail to separate the same arms.

Read post →

16 August 2026·7 min read

Two thirds of our network is generic. We deleted it and ran the sweep again.

The base network under our in-silico knockouts is a union of DMG chromatin data with a published human promoter network, and the generic half is 65.33% of the rows. Removing it leaves the ordering intact at Spearman 0.9803 and the top four unchanged. It also makes FOS stronger, so the test does nothing for the specificity problem and slightly worsens it.

Read post →

16 August 2026·5 min read

Two ways to be wrong while every check you ran came back clean.

A cost line was filed 4.4 times too high because seven days of a month were assumed to be a full month, and the arithmetic around it was faultless. An effect size was published a third too small under the wrong name, and the two figures beside it were right. Neither error was a mistake in the working. Both were an unsourced number sitting next to sourced ones.

Read post →

16 August 2026·6 min read

One thousand draws is a resolution parameter. Every interval comes from the sixty.

Our sampler emits a thousand simulated co-mutation profiles from a copula fitted to 60 real H3 K27M cases. They are not patients and they are not extra evidence: a thousand draws and ten thousand draws from the same fit carry the same information. So every confidence interval is resampled over the sixty, the fine joint is deliberately pulled toward the prior, and we are barred from ever pricing per draw.

Read post →

29 June 2026·4 min read

Ewing sarcoma: the cohort sampler run on a second cancer, and a published finding that did not reproduce.

Cohort sampling run on a second cancer with sharply different molecular biology, anchored to two published Ewing cohorts fetched through the cBioPortal API. The sampler carried. One published co-occurrence finding, STAG2 with TP53, did not reproduce. Read with the note at the top: the portability claim this post was written to demonstrate was retracted on 2026-08-13, and Ewing is not selectable.

Read post →

29 June 2026·4 min read

Leave-one-out cross-validation: where the joint model helps, and where it hurts.

We tested our own claim that a joint model beats a per-gene baseline. On the DMG anchor as it stood in June it helps on Brier and hurts on log-loss; on the larger Ewing anchor it wins on both. The DMG anchor has since been corrected upward and this has not been re-run against it.

Read post →

29 June 2026·3 min read

Research-use-only · the line goes on every report header, every footer, every JSON export.

We added a single-source-of-truth disclaimer module, wired it into the workspace report (top and bottom), and embedded the full text into every JSON export envelope. Customers, grant officers, and regulators get the same line.

Read post →