Measure resistance in the lab.
Essential biological evidence, but experiments require specialist teams, materials and lab capacity. Timelines can run to weeks or longer.
Planning example: 4+ weeks, depending on the assay.THE PROBLEM / CANCER RESISTANCE
Anticipate cancer’s
next move.
Imagine a treatment kills most of the cancer cells. But a few resistant cells survive. With fewer cells competing against them, they can multiply, and the cancer can return, harder to treat.
● ● ● ● ●
● ● ● ● ●A diverse tumor
→● ● ● ● ●A diverse tumor
○ ○ ○ ○ ○
○ ○ ● ○ ○Treatment leaves survivors
→○ ○ ● ○ ○Treatment leaves survivors
● ● ●
● ● ● ●Resistant cells can regrow
● ● ● ●Resistant cells can regrow
ESCAPE ASKSBuild a research lab
Build a research lab
that learns from every experiment.
Train a model for the question.
Gathering data, preparing labels, training and validating a specialized model also takes time, expertise and compute resources.
Reliable predictions depend on relevant data and validation.Our ambition is to map cancer’s escape routes before resistance emerges. Escape Labs brings agents, biological models and verified experiments into one research loop, starting with real FGFR assays.
Biological context: National Cancer Institute ↗ · Illustration is conceptual. Research prioritization; independent biological validation remains necessary.
ESCAPE LAB / BUILT IN SIX HOURS
A paper. An experiment.
A measured finding.
Digest the study data, replay the research loop, and inspect the results.
Recorded replay
Interactive replay of the recorded study ingestion and experiments. Published assays and aggregate results; no new scientific run.
SIX HOURS / A WORKING RESEARCH LOOP
From a proposal
to measured progress.
Real FGFR data. Bounded agent proposals. Numerical tools.
Independent verification and an auditable experiment history.
Historical replayNemotron on Crusoe · Recorded Oct 17
AUTONOMOUS MODEL REVISION / REAL RUN
NEMOTRON / CRUSOECan the agent improve resistance prediction?
FROZEN EVALUATION CONTRACT01 / 05
Summary playback · original run 100.74s
BUILT IN APPROXIMATELY SIX HOURS
Data ingestion → Crusoe proposals → numerical experiments → verified decisions.
RESEARCH + NUMERICAL TOOLS
Nemotron plans experiments. GLM-5.3 is the tested alternative. ESM2 + logistic models score activation variants.
MEASURED FGFR2 HIGH-ACTIVATION IMPROVEMENT
14.4% better
ranking.
Three additional highly activating variants found within the same 68-pick budget.
3 fixed development foldsSame fitted classifier · Raw-margin ranking repair
FGFR2 HIGH ACTIVATION / DIFFERENT ENDPOINT
668 VARIANTS · 3 FOLDSMean-fold average precision
SIGMOID PROBABILITY—
1 / 26 high variants retrieved
RAW DECISION MARGIN—
4 / 26 high variants retrieved
68 picks in either ranking0 / 7 strong variants in bothSame fitted coefficients
What does average precision mean?
AP measures whether highly activating variants appear near the top of the ranked list. Higher is better. An AP of 0.11135 is not 11.1% accuracy. Here, the same classifier recovered 4 of 26 high variants in 68 nominations, compared with 1 before the scoring repair.
| Fold | Probability AP | Margin AP | High hits |
|---|
2.92s verification with cached features594 protected artifacts unchanged+0.01402 absolute AP gain
WHAT CHANGED
In the first fold, 136 scores rounded to exactly 1.0. Raw margins preserved 275 distinct scores. The other two folds were unchanged.
WHAT THIS ESTABLISHES
Raw-margin ranking recovers information from the fitted classifier. Evidence: three development folds; descriptive gain interval includes zero. The next milestone is independent biological validation.
THE RECORD IS PART OF THE RESULT
Build on every
experiment.
One evidence ledger connects each research question to its measured outcome.
No experiments match this search.
Evidence has a boundary.
Original final + S13Consumed. Later S13 reuse is retrospective.
Activation pilot foldsRepeated development evidence.
RLY + MANOReserved for a separately frozen protocol.
A SEPARATE CHECK OF THE REPLAY VAULT
A claim cannot
skip the evidence.
Run the existing commit/reveal API and inspect the verified results.
Live API checkSynthetic fixtures · Live verification of vault controls
REAL API / SYNTHETIC FIXTURE
Commit first. Reveal second. Verify last.
This executes the repository’s actual vault and HTTP handlers in an isolated temporary fixture. It is separate from the historical model-training workflow.
Ready. No check has run in this session.
Service status
Checking local adapter…
The public API runs isolated synthetic checks. Scientific experiment results remain the recorded aggregate evidence.
IMPLEMENTED COMPONENTS / CLEAR MODEL ROLES
Nemotron plans.
The evaluator measures.
The planner proposes a hypothesis and configuration. Numerical tools fit models; independent rules decide what is retained.
RESEARCH WORKFLOW
Open full diagram ↗Where Nemotron works
This maps components across experiments. Nemotron-3.5-Super-VL is the experiment planner on Crusoe. GLM-5.3 was tested as an alternative in comparison and research runs; reviewed You.com cards feed that research branch.
SEPARATE DEVELOPMENT PIPELINE
Open diagram ↗Where the measured activation gain comes from
Frozen ESM2 representations and a logistic classifier score variants. Raw-margin ranking preserves their ordering. Nemotron’s newer diagnostic loop now inspects these aggregate results and retains tested research rules.
POST-SUBMISSION / CONTINUING THE BUILD
A research loop
that keeps learning.
The six-hour build produced the +14.4% ranking repair.
Since then: persistent research policies, representation tests, and public reproduction.
Retain better research habits.
—→—
Correct diagnostic decisions on 48 held-out generated cases across four predefined failure families.
Nemotron proposed two persistent rules: check score saturation before ranking, and feature shift before promotion. A frozen referee accepted the policy.
Use the improved procedure again.
- Load the promoted policy.
- Inspect saturation and feature scaling.
- Compare rankings and sequence ablations.
- Record the decision and stop.
5 Nemotron calls · 10,024 tokens · 0 new fits.
Recorded Crusoe execution over published FGFR development aggregates.
MODEL DIAGNOSTICS / SAME 668 VARIANTS AND 3 FIXED FOLDS
Test the next modeling hypothesis.
| Representation | Mean AP | High hits / picks | Decision |
|---|
The single-branch representations show promising development scores. These are diagnostic ablations with different component counts; the combined model remains the default. The variance floor failed its acceptance gate.
PUBLIC REPRODUCTION / FRESH INPUTS + EMBEDDINGS
—Rebuild the six-hour headline.
The same AP and retrieval results reproduced from public sources in a clean workspace. Three model fits. Dependency installation excluded.
The next milestone: prospective biology.
Expand the research loop from archival assays to independently frozen biological tests, then connect it to real laboratory experiments. RLY and MANO remain reserved for a separate protocol.