substrate/tui-corpus-producer

The TUI-as-Corpus-Producer pattern designates the operator terminal interface (slm-cli) as a primary source of high-quality training data for the per-tenant model adapter. Every interaction with the Doorman through this interface is a curated corpus contribution.

[edit]Why terminal interactions are high-quality training data

Three properties distinguish system administration and IT-support interactions from general training data:

Verifiable ground truth. When an operator follows AI advice — running a suggested command, applying a proposed configuration change — the system either recovers or it does not. Other domains such as creative writing or strategic reasoning lack this immediate-feedback property. IT-support has it by default. The operator knows immediately whether the response was correct.

Narrow domain. Archive operations, system conventions, and customer-specific workflow vocabulary form a bounded command set and failure-mode space. Models train more efficiently on bounded domains than on general corpora because the signal-to-noise ratio is higher.

Domain-expert feedback. The operator issuing a verdict is the person who knows whether the response was correct — not a proxy labeler separated from the actual work. Published reinforcement-learning-from-human-feedback literature consistently reports that high-quality verdict-signed interaction tuples train an order of magnitude more efficiently than observation-only tuples. [^1]

[edit]The /feedback mechanism

After every assistant response in the TUI, the operator is offered three explicit verdicts:

Good. The response was correct and useful. The tuple is flagged as a positive direct preference optimisation example.

Refine. The response was close but needed adjustment. The operator provides a correction inline; the tuple captures the response-and-refinement pair as training signal.

Bad. The response was wrong. The tuple is flagged as a negative direct preference optimisation example.

If the operator dismisses without providing a verdict, the tuple is captured as unsigned and contributes to supervised fine-tuning but not to direct preference optimisation.

[edit]Adapter quality budget

Published fine-tuning literature suggests 200 to 500 high-quality verdict-signed interactions are sufficient for a first adapter training cycle in a narrow domain. [^2] The platform's intended sequence for each tenant is: accumulate signed interactions from dogfood operations, train the first per-tenant adapter via the LoRA training pipeline, apply a validation quality gate, and promote the adapter to the deployment. Each subsequent training cycle incorporates additional interactions, progressively tuning the adapter to the customer's specific environment — their systemd units, their seed taxonomy, their workflow vocabulary.

[edit]Per-tenant adapter ownership

The corpus produced by a customer's operators trains that customer's adapter, not a general adapter. Per the customer-owned-graph-ip convention, the trained adapter weights are the customer's property. The platform distributes the model architecture and the training pipeline; the customer retains the trained adapter that results.

[edit]Verdict capture discipline

Some terminal sessions should not contribute to the training corpus: test sessions initiated with a no-corpus flag, sessions interrupted by unavailable tiers before completion, and sessions using forced-tier debug mode are audit-logged but excluded from normal training data. The boundary between operational corpus and test corpus is enforced at the Doorman's verdict intake endpoint.

[edit]See also

• single-boundary-compute-discipline — the TUI never calls inference tiers directly; all calls route through the Doorman
• customer-owned-graph-ip — per-tenant adapter weights are the customer's intellectual property
• knowledge-graph-grounded-apprenticeship — training tuples carry graph context when the Doorman grounds the request