15% of all profit is donated to Heal Palestine and the Palestine Children's Relief Fund (PCRF)
Platform Research Validation Safety Pricing Company Demo Request access
Validation

How we know it works, and where it doesn't.

This page reports cohort-level evidence for the Provenance-2 site-of-origin core. The same page that states the measured results states the failures and distribution limits.

Scope boundary. These metrics validate anatomical site-of-origin classification and its uncertainty controls. They do not validate subtype, immune-context, target-context, or other exploratory research heads to the same standard.
The numbers

The numbers, and what they measure

0.940macro-F1 on the held-out test set (balanced accuracy 0.942), every site weighted equally
3 in 5off-pipeline single tumors it abstains on rather than guess; on the ones it does call, it is right about 98% of the time
8xcalibration error cut (0.091 to 0.011)
90%conformal coverage in-distribution; about 90% resolve to a single confident call

Macro-F1 weights all 25 sites equally, so the common cancers cannot carry the rare ones. We report it, and balanced accuracy, rather than top-line accuracy, which would look higher for the wrong reasons. The 0.940 is measured on a held-out set the model never trained on. On independent external tumors the honest number is lower, a 7-site top-1 of 0.906 on 381 samples, and we report that separately rather than blend it in to look better.

Evidence files

Download what is reproducible, label what is not

The machine-readable summary below is exported from the repository's passing cohort-level verification output. It is Provotics-generated evidence, not an independent audit.

Validation summary

Metrics, cohort sizes, scope, source hash, and limitations for the current served site-of-origin core.

Download JSON

Public-sample status

No accession-linked bundle pairing a full publishable input with the current served-core output is public yet. The evidence ledger explains the publication gate instead of substituting the canned demo.

Open evidence ledger

Interface walkthrough

The guided demo teaches the output structure and abstention behavior. Its profiles and results are illustrative and are not evidence.

Open guided demo
Interactive evidence

Compare the cohorts directly

Change the evaluation cohort to see what improves, what falls, and which per-site results are sufficiently powered to interpret.

Powered per-site results

Method

How we validate

The validation was built to try to break our own results, which is why the limitations further down are stated plainly rather than hedged.

  • Held out by study. An external study never appears in both training and test, so the model cannot memorize a cohort and call it generalization.
  • Batch-origin probe. We train a classifier to tell our data apart from an external cohort. If it can still separate them after correction, the apparent "signal" is a batch artifact and the cohort is rejected.
  • Shuffled-label control. We re-run with the labels shuffled. If a recovery survives that, it was never real. This caught and excluded a false improvement on one site that was a majority-class shortcut.
  • Bootstrap intervals and deduplication. Headline numbers carry resampled confidence intervals, and near-duplicate samples are removed so a profile cannot sit on both sides of a split.
Uncertainty

It knows when it does not know

A site prediction is only useful if its confidence is meaningful and the model can decline when it should. Three independent mechanisms make that true.

Calibrated confidence

Raw scores are calibrated so a stated probability matches reality. Calibration error drops from 0.091 to 0.011 on held-out data, roughly an 8x reduction, so "80% confident" means about 80% in practice (in-distribution).

Conformal sets and abstention

Instead of forcing a single answer, the model returns a candidate set with a 90% coverage target. About 90% of cases resolve to one confident site; when the evidence is genuinely ambiguous it returns the short list it cannot rule out, and when nothing clears the bar it abstains rather than guess.

Novelty and input validity

A distance check flags profiles unlike anything in training, and an input-validity gate catches profiles that are not tumor RNA-seq at all. The gate flagged 100% of held-out normal-tissue samples with no tumor false positives. These are guardrails, not a clinical detector.

Calibrationreliability
predicted confidence → observed accuracy →

Schematic. After temperature scaling, calibration error falls from 0.091 to 0.011 (roughly 8x), in-distribution. The dashed line is perfect calibration.

Coverage and abstentionin-distribution
Of 100 in-distribution calls about 90 resolve to one confident site about 10 abstain or return a short candidate set

The 90% conformal coverage target holds in-distribution only. Off-pipeline, it abstains on about three in five single tumors and is right about 98% of the ones it does call.

The generalization gap

The cross-platform gap, in full

The hardest case is a single tumor sequenced on a different pipeline than the model trained on. We publish that gap rather than hide it. On those off-pipeline single tumors, the model abstains on roughly three in five cases, and on the ones it does call it is right about 98% of the time. So it declines rather than confidently misroutes. We do not quote the 90% coverage figure for off-pipeline inputs, because conformal coverage is an in-distribution property and does not hold under platform shift. The honest framing is selective: it answers less, but it is right when it answers.
Limitations

Where it is weak

The same rigor that produced the numbers above produced these. They are part of the model.

It does not recognize real mesothelioma

On real mesothelioma cases the model scores 0% recall and confidently sends them elsewhere. An earlier per-site figure for Pleura and Mediastinum reflected one external batch's signature rather than the biology; we found it with the per-source split and corrected it. Treat Pleura and Mediastinum output as unreliable.

Rare sites are overconfident and data-starved

On a handful of rare sites the stated confidence runs well above measured accuracy (for example Soft Tissue and Esophagus). Those sites have very few examples, close to the entire public universe of their kind, so the data is exhausted and stronger models do not move the number. It is a data ceiling, not a model ceiling.

The guarantees are in-distribution only

Calibration and conformal coverage are measured on a held-out split from the same distribution and do not hold under platform or batch shift. The novelty and input-validity checks exist for exactly that reason, and are themselves reference-only.

No fairness audit, no clinical validation

All evaluation is retrospective on public cohorts. There is no prospective study, no independent clinical-site validation, and no subgroup-equity audit. Demographic representativeness is uncharacterized, and performance on underrepresented groups is unmeasured and may be worse.

Read the rest

See the full model card, the input and output contract in Docs, and the responsible-use notes on Safety.