mindXtrain Coach probing…

Every mindX artefact — training manifest, MEI receipt, evolution proposal — is stamped with chronos.agent's promised time: a UTC moment plus a measured confidence interval. The interval comes from the standard deviation of recent transaction anchors (the chain confirmed our event at T_block while our local clock saw T_local). When the spread widens, the consensus tier drops from correlated to degraded to drifted.

Which training lanes are reachable from this host. CPU is always available (the trl_cpu lane); AMD GPUs surface when rocm-smi is on PATH; NVIDIA GPUs surface when nvidia-smi is on PATH. The "recommended lane" is the most capable lane detected and matches the recipe's train.backend field.

The training data is mindX's dream-cycle JSONL output. This step confirms the corpus path is reachable and has examples to train on. Set MINDXTRAIN_MINDX_HOME to point at your mindX checkout.

Author a small script (the training examples) for an actor (the model) in its persona voice. Saved as source: local JSONL under out/datasets/<name>/script.jsonl — point the mindx_persona_imprint_local recipe’s data.path at it, train, then measure the imprint (recall before vs after).

The differentiator. A short on-device micro-benchmark picks Composable Kernel vs Triton SDPA, hipBLASLt heuristics, and the RCCL collective config — static plan, no JIT autotune in production.

Translate the canonical mindXtrain config plus the autotune plan into the backend-specific YAML the trainer subprocess consumes.

Spawn the trainer subprocess and watch loss, learning rate, and the raw train.log stream in over Server-Sent Events. On a CPU-only laptop the launch returns a 503 with an install hint; on the MI300X box the loss curve fills in as accelerate launch progresses.

Waiting to start…

The AOT-only discipline is a verification primitive: binding the frozen AutotunePlan hash to the checkpoint hash proves which compiled backend, GEMM heuristic and RCCL config produced these weights — the bitwise-reproducibility that decentralized-training verifiers (Verde/RepOps) require. This re-hashes the run’s manifest.json against the artifacts on disk.

waiting for a completed run…

The model is an actor; the classroom graduates it; the boardroom (any number of role-based members) decides whether to promote it. A tie escalates to a dojo — a panel of a prime number of judges, so the dispute always settles with no tie. Members deliberate with real models via the configured chat backend.

Composite score over five sub-indices — quality, decode throughput, prefill throughput, memory footprint, and energy per useful token — combined via weighted geometric mean (spec §5). A checkpoint must clear MEI ≥ 0.55 with every sub-index ≥ 0.30 to be promotable to AgenticPlace.

One-click bridge from this laptop to github.com/professor-codephreak/mindXtrain and onto an AMD Dev Cloud MI300X droplet. GitHub first so the cloud-init bootstrap clones from a deterministic commit; then either provision a fresh droplet (cloud-init runs mindxtrain bench as it boots) or rsync to one you already own. All output streams over SSE.

Chat with a local model — responses stream token-by-token (the AI-SDK text-stream pattern). Pick a model, or start the ollama server if it isn’t running.

No model available — start ollama (above) or run serve --to ollama, then ↻ models.