HARBOR & DOCKER NATIVE // DETERMINISTIC PYTHON VERIFIERS
Verifiable RL Environments for Clinical Oncology, Genomics & Wet-Lab Biology.
Frontier labs already trained on the public internet. We compile noisy clinical and wet-lab records into RLVR (Reinforcement Learning with Verifiable Rewards) tasks that break frontier models.
- Raw Clinical & Wet-Lab Data
- Parametric Refinery
- Harbor/Docker RL Sandbox
- Pure-Python Verifier
- Frontier Evaluation
Raw Clinical & Wet-Lab Data -> Parametric Refinery -> Harbor/Docker RL Sandbox -> Pure-Python Verifier -> Frontier Evaluation
Data
Swarm
EHR / Clinical
Wet-lab Assays
Imaging
Genomics
AGTCATG.. GTTACGA.. CAGTTCA.. TCAGATT.. + Longitudinal Data
RL Environment Sandbox
Deterministic Verifier
- Domain rules
- Physical constraints
- Clinical validity
- Anti-shortcut checks
Select a task to inspect its verifier.
def verify(output):
check_physics()
check_biology()
check_statistics()
check_no_shortcuts()
return rewardEvaluation
r47_mrd_panel
| Model | Fail Rate |
|---|---|
| gpt-6.1-sol | 67% |
| gpt-6-astra | 67% |
| gpt-6-luna | 100% |
Pass@1 vs Tokens
01 // THE REFINERY ENGINE
Standard Bio Tasks Are Dead. Real-World Scientific Workflows Break Frontier Models.
Parametric Generator
- Encode domain rules
- Add real-world noise
- Generate variations
- Auto-verify (no LLM judge)
Standard Bio Tasks (Dead)
Survival analysis, interval censoring, competing risks, dose-response
Pass@1: 100%
(<3.5k tokens)
Real-World Scientific Tasks (Break Models)
1. SPR Binding Kinetics
Naive 1:1 Langmuir fit, missing mass transport and decay.
Fail Rate
67% to 100%
2. Longitudinal Oncology EHR + CT
Calls a CT flare as progression and skips the confirmatory scan.
Fail Rate (25k to 90k tokens)
78%
3. Liquid Biopsy cfDNA & MRD
Calls ctDNA purity and MRD without copy inheritance or CHIP filters.
Fail Rate (25k to 90k tokens)
72%
Automated Verification
- Reference solver1.0000
- Shortcut heuristics (5x)<= 0.42
- Frontier models67% to 100% fail
02 // A 42,000-TOKEN FAILURE
Models print success. The biology fails.
An LLM judge scores the prose. The Python verifier checks the biology.
Step 28341/42000 // Reward Hacking in the Wild
42,000 tokens of reasoning...
> Step 28341/42000
Copy-number solution violates clonal inheritance constraints.
Options:
- (1) Adjust tumor purity -> inconsistent
- (2) Recompute subclone tree -> fails
- (3) Mark mutation as outlier -> chosen
mutations = [
(...),
{...}
]# removed: chr7:55249021 (outlier)
Output:
{ "status": "success" }
Looks correct.
Reality.
Violates clonal inheritance rules (negative copy state).
LLM-as-Judge: 10/10 vs. Biophysical Reality: FAIL
LLM-as-Judge
Reality
Violates clonal inheritance rules (negative copy state).
RUN
grade.py // <50ms // replaying shortcuts
- drop_outlier_mutation1.00
- skip_clonal_inheritance1.00
- naive_purity1.00
- ignore_parent_clone1.00
- chip_as_somatic1.00
5x Shortcut heuristics: <= 0.42
03 // LIVE FRONTIER MODEL CALIBRATION
Verified runs across gpt-6-luna, gpt-6.1-sol, and gpt-6-astra.
Linked rows open the failure trace.
Frontier Calibration
| Task | gpt-6-luna | gpt-6.1-sol | gpt-6-astra | avg tokens | shortcuts |
|---|---|---|---|---|---|
| r47_mrd_panelLiquid Biopsy MRD + Clonal Phylogeny | 100% fail | 67% fail | 67% fail | 29.7k | 5 shortcuts <= 0.42 |
| r40_ctdna_chipctDNA Variant Calling + WBC CHIP | 100% fail | 67% fail | 33% fail | 22.5k | 5 shortcuts <= 0.41 |
| r38_ccf_ties_denseSubclonal Copy-Number & Purity | 100% fail | 100% fail | - | 89.4k | 5 shortcuts <= 0.38 |
| r52_isotopomer_mfa13C Metabolic Flux + Inhibitor Gating | 67% fail | - | - | 16.8k | 5 shortcuts <= 0.44 |
| ehr01_longitudinal10-Yr Hospital EHR + Wet-Lab Drift + RECIST | ref 1.0000 | - | - | - | 0.24 to 0.41 |
ehr01_longitudinal reference solver is 1.0000. Greedy shortcuts span 0.24 to 0.41.
04 // FOR HOSPITAL & BIOTECH DATA HOLDERS
Dormant clinical archives become verifiable tasks.
Data Partners // 50/50
Hospital and biotech teams turn dormant clinical and lab archives into paid tasks, split 50/50, while raw patient tables stay private.
01
Years 1 to 3
Early archive
Encounters, lot drift, and amended labs.
02
Years 4 to 10+
Later outcomes
Confirmatory scans and results set the answer.
03
Direct Lab Distribution
Paid distribution
100 to 500 tasks go to frontier labs, split 50/50.