For researchers checking patient-level generalization
Clean by slide.
Shared by patient.
New images can still come from familiar patients. Asterisk checks the levels behind your split—and shows the rows behind the answer.
NVIDIA Nemotron on Nebius Token Factory.
Fresh checker execution with Nebius Sandboxes.
Same test set.
Two different answers.
One retained witness · no live call
One patient. Two different slides.
TCGA-2G-AAFZ- Training slide
- TCGA-TGCT/TCGA-2G-AAFZ-01Z-00-DX2.69C768D9-0B6B-4AE7-8CFA-04A26DDC9193.pt
- wsibench/Report_all_type_all_train_single.json /5108
- Evaluation slide
- TCGA-TGCT/TCGA-2G-AAFZ-01Z-00-DX1.A782BB9C-2121-4E36-878D-475540CE5E8E.pt
- wsibench/Report_all_type_all_test_single.json /110
The participant is identified by the first three barcode components, Project-TSS-Participant, for example TCGA-02-0001 (12 characters).
Recorded documentation: docs/tcga-barcode.md, line 2. TCGA barcode definition ↗. Identifier references, not image contents.
A result you can question
“No overlap”
needs a definition.
In the published WSI-Bench report-task files, Asterisk measured 42 patient identifiers shared between training and evaluation, affecting 43 of 208 evaluation rows. A slide-only check misses that distinction.
This is a saved local measurement of pinned public files, not a live model response or a new Sandbox run. Shared patients matter when the claim requires unseen patients; they do not, by themselves, quantify an effect on model accuracy.
What the measurement establishes
Exact identifier overlap under the recorded derivations. It does not establish duplicate image content, preprocessing leakage, a performance penalty, or independence at an undocumented level. The downloadable JSON retains source provenance and witness rows.
From the claim to the source rows
The model proposes.
The engine decides. You confirm.
- 01 / Define
Start with what must be unseen.
Add your split files and documentation. Preview the information selected for the model before sending it.
- 02 / Inspect
Review the proposed identities.
Nemotron proposes extraction rules with citations. Application code validates and measures them. Missing definitions stay explicit.
- 03 / Retain
Leave a receipt someone can inspect.
Confirm the identity meaning and evaluation partition, inspect witness rows, and request a fresh Sandbox replay. Save the result and export readable HTML or JSON.
Research preview
A precise check.
An explicit boundary.
Asterisk checks recorded identities, not whether an entire experiment is scientifically valid. Zero overlap stays pending when the identity meaning needs confirmation.
In 41 evaluated cases across tested configurations, guarded model runs produced no false reassurance. Nemotron Super proposed 21 ladders correctly; the Sonnet comparison proposed 36. The shipping Super v4 prompt scored 5/10 on set D and failed its preparation gate. A later denial control elicited an unsupported binding. These results do not guarantee reliability on new datasets. Researcher usefulness remains unmeasured.
The workspace is a separate application. Uploads and execution happen there. This public page has no upload form, model calls, or analytics.
No account or provider credentials needed
Run a local measurement.
From a trusted Asterisk source checkout with Python 3.14:
python3.14 -m venv .venv .venv/bin/python -m pip install -r requirements.lock.txt .venv/bin/python -m uvicorn app.main:app --host 127.0.0.1 --port 8791
Open http://127.0.0.1:8791/static/ladder.html#check. Expand Advanced, load the subject-disjoint example, then measure on server. The fictional example needs no model or Sandbox call. Zero overlap stays pending until you confirm the identity meaning.
The real benchmark loader requires an optional pinned identifier cache. The source README explains installation. The hosted workspace requires reviewer credentials and has a finite live-call reserve; manual measurements and saved results do not consume it.
Make the assumption inspectable
Before you call it independent,
check what “it” means.
Open Asterisk workspace ↗Research preview · workspace access may require reviewer credentials