Learning actions and learning a voice#

The latest training work explores two separate 9B adapters: one for autonomous planning and one for conversational personality. Both small remote diagnostics completed, and both were compared through the actual llama.cpp serving path. Neither candidate qualified for production.

These experiments extend the earlier 0.8B Intent/Task work. A learned persona would express a consistent identity while still receiving current observations, memory and instructions. It would not acquire live experience simply by having been trained.

What the diagnostics tested#

Candidate	Reviewed training / evaluation samples	Main result
Autonomy	4 / 2	Native cancellation outputs improved, but serving comparisons still lost lighting objectives and regressed one unconstrained cancellation format
Persona	5 / 1	Responses became more concise, but retained invented experiences, source confusion and an unsupported completed-movement claim

These are diagnostic seed sets, not broad coverage. Evaluation families were separated from training, exact messages and authorship were retained, and both adapters started independently from the same pinned Qwen3.5 9B base. They were not trained on top of one another or qualified as a combined adapter.

The existing remote training path now installs the uploaded pinned requirements before validation. Reloading also explicitly loads the pinned base before attaching the adapter. That matters because dependency behavior could otherwise discard the adapter’s recorded base revision.

Lower loss is not enough#

Autonomy’s native held-out loss fell from about 0.416 to 0.022, and reloaded native outputs produced correct cancellation handoffs on the two evaluation cases. The production base already succeeded at cancellation under its runtime schema, however. A native improvement does not automatically improve the served system.

For the serving comparison, the complete GPU projector configuration did not fit. Text tests moved only the unused projector to CPU; 22 unchanged-base outputs matched the production configuration. Each adapter was then sampled independently with full context. That controlled adjustment is part of the comparison’s limits, not a production configuration change.

The persona experiment also tested less static persona context. Fewer prompt tokens did not demonstrate reliable identity learning. The resulting replies still confused sources and made unsupported claims. Reducing production persona context or deploying either candidate is therefore not supported by this result.

Training data follows the current interface#

The single Persona response owner changes the Task training contract. Speech-only requests now have no Task program; mixed requests retain physical or sensing actions, with the reply routed through Intent and Persona. Version 16 contains 6,358 development, 1,518 evaluation and 187 regression records with no Task speech-step targets.

Those are revised comparison references, not an additional completed specialist training run. Earlier adapters and run snapshots are retained. Scores from the old speech-step interface must not be compared as though they used the same contract, and reviewed cases are no longer newly unseen evaluation examples.

Production and next work#

The diagnostic record confirms both remote jobs finished and their pods were terminated. The production 9B service was restored with its original startup arguments and no loaded candidate adapter; model assignments matched the saved pre-comparison state. The existing 0.8B process was unchanged. No physical actions were dispatched by these comparisons.

The next requirement is broader, correctly attributed planner/persona data and fresh final evaluation families. A larger fit on this tiny seed set would not resolve the coverage gap. The companion guide explains how reviewed learning differs from memory and personality context.

This public summary uses the October 15 integration investigation and training handoff records, identified in the source snapshot manifest. Private conversations, account details and raw training messages are not part of the website report.