Kernel band
L0 – L1home of the performance engineer
The innermost accelerator code and how it launches. Occupancy, register pressure, grid/block shape, streams. Speedups here are measured in kernel microseconds.
The signature
One stack. Eleven layers. Every problem diagnosed where it lives.
Portent reasons from the innermost compute kernel up to a self-improving knowledge base. It holds every layer at once instead of working one level and guessing about the rest, so a fault is named, fixed, and verified at the layer where it lives.
Middle band
L2 – L7home of the data-center operator
Framework, collectives, transport, the build environment that pins it all, fabric, and node. This is where a cluster is brought up and where most bring-up traps live. Kubernetes orchestration spans L2–L7.
Brain band
L8 – L10home of the self-improving loop
The benchmark harness that measures what actually happened, the autonomous loop that plans and repairs itself, and the knowledge base that feeds each lesson into the next run.
The signature
The 11-layer stack
Portent reasons across every layer, hardware to knowledge base. A problem at any layer is diagnosed at that layer.
Substrate
Red lines
What the agent has learned never to do
Each red line is a codified rule in the corpus: earned from a real failure, attributed to the event that taught it, enforced by a deterministic check. Not a prompt-time suggestion. One per layer.
- L10
Knowledge base
Never promote an unattributed lesson. Folklore with no source event is rejected at the gate.
- L9
Autonomous loop
Never auto-retry an irreversible or billable step. Stop, prepare it, and hand a human exactly one command.
- L8
Benchmark harness
Never report a throughput number unless a correctness gate passed on the same run that produced it.
- L7
OS / infra / driver
Never infer NIC count from the instance type. Enumerate and verify on the node before any workload starts.
- L6
Network fabric
Never enable RDMA on a secondary subnet without validating multi-rail routing first.
- L5
CI / build environment
Never promote a container image whose CUDA / driver ABI wasn't verified against every target node's driver.
- L4
Transport plugin
Never load a transport plugin whose ABI wasn't built against the running NCCL. A silent fallback to TCP shows up only as a 'slow' run.
- L3
Communication library
Never pin one NCCL algorithm across cluster sizes. A ring tuned for 8 nodes is the wrong choice at 64.
- L2
Framework
Never stack full-shard FSDP and activation checkpointing on a model that already fits; you pay recompute for memory you didn't need.
- L1
Launch config
Never raise occupancy by cutting registers when it forces spills to local memory. Measure the launch, don't assume the win.
- L0
Compute kernel
Never accept a kernel rewrite without a bit-exact (or ULP-bounded) check against the reference output. Faster and wrong is still wrong.
Every rule above is retrievable, actionable, and attributed. When one fires, the agent stops at that layer and cites the lesson. It never silently works around it.