Secure Compute Routing
Classify first
Label every input as public, controlled/private, or credentialed. Record approximate size, required software, compute/memory/GPU need, expected runtime, and output size.
Default routing
- Public data: local, controlled HPC, or an approved cloud service may be considered.
- Controlled/private data: local or user-controlled HPC over SSH by default.
- Credentials: never embed in scripts, prompts, logs, archives, or repository files.
- External cloud GPU: do not upload or launch until the user has approved the provider, estimated cost, exact transferred files, retention policy, and deletion/return plan.
Do not treat a user's request to analyze data as permission to send it to an external service.
Execution contract
Before launch, produce:
- runtime target and reason;
- data-transfer manifest and excluded files;
- environment/container specification;
- command, resources, wall time, retry limit, and stopping condition;
- log, heartbeat, checkpoint, and failure handling;
- output validation and checksum plan;
- return/synchronization plan that does not overwrite source data.
Use SSH to execute on a controlled HPC without installing an agent unless the user explicitly requests and approves it. Keep large intermediate data near the compute environment; transfer only required inputs and final artifacts.
Completion
Report actual host class, job ID when available, environment, exit status, validated outputs, costs if external, and failed checks. A submitted job is not a completed analysis.
1---2name: secure-compute-routing3description: Decide whether a life-science computation should run locally, on a controlled HPC over SSH, or on an external cloud GPU based on data sensitivity, scale, cost, and reproducibility. Use before moving data or launching remote work.4metadata:5 source: https://skillmd.com/skills/minimax-ai/secure-compute-routing6---78# Secure Compute Routing910## Classify first1112Label every input as `public`, `controlled/private`, or `credentialed`. Record approximate size, required software, compute/memory/GPU need, expected runtime, and output size.1314## Default routing1516- Public data: local, controlled HPC, or an approved cloud service may be considered.17- Controlled/private data: local or user-controlled HPC over SSH by default.18- Credentials: never embed in scripts, prompts, logs, archives, or repository files.19- External cloud GPU: do not upload or launch until the user has approved the provider, estimated cost, exact transferred files, retention policy, and deletion/return plan.2021Do not treat a user's request to analyze data as permission to send it to an external service.2223## Execution contract2425Before launch, produce:26271. runtime target and reason;282. data-transfer manifest and excluded files;293. environment/container specification;304. command, resources, wall time, retry limit, and stopping condition;315. log, heartbeat, checkpoint, and failure handling;326. output validation and checksum plan;337. return/synchronization plan that does not overwrite source data.3435Use SSH to execute on a controlled HPC without installing an agent unless the user explicitly requests and approves it. Keep large intermediate data near the compute environment; transfer only required inputs and final artifacts.3637## Completion3839Report actual host class, job ID when available, environment, exit status, validated outputs, costs if external, and failed checks. A submitted job is not a completed analysis.