A corpus that stays where you put it

The documents a desk relies on were not written for an agent. They are policies, prior decisions, contracts, manuals, and correspondence that accumulated because the work needed them. They are also, very often, the material a site is least willing to move. An agent on Czar reads them where they are.

Retrieval in place

Substrate, the platform on the appliance, provides retrieval over the site's documents. When an agent drafts, it draws on passages from the corpus the customer placed on the appliance or on a system inside the site that a declared route reaches. The retrieval runs on the rack. The passages are brought to the draft; the corpus is not brought anywhere.

This is a different arrangement from the one most desks have been offered, where documents are uploaded to a vendor's service so the vendor's models can read them. Here the models are on the appliance already. There is nothing to upload to, and no vendor copy to worry about afterward. What leaves the building: nothing.

Bounds on what may be read

A version's bounds name which material the agent may read. The claims agent may read the claims policy library and the prior decisions for its line. It may not read the personnel files that happen to sit on the same system. The procurement agent may read the supplier contracts. It may not read the claims decisions.

Material outside the bounds is not hidden from the agent by a filter applied after the fact. It is outside the agent's reach in that version, and a run cannot widen it. If a desk later needs the agent to read a new body of material, that is a change to the design, published as a new version, with the admin's decision recorded if it touches a system that was not previously available.

What the reviewer sees

When a run reaches the step that waits, the reviewer sees the draft and what it used. The passages the agent drew on are shown alongside the draft. The reviewer can check that the agent relied on the current policy and not a superseded one, on the right contract and not a neighbor, before approving or stopping.

Those passages are written to the run record. The next morning, or next year, someone opening the finished run can see the exact material the draft was built from, as it was at the time. The corpus may have changed since; the record holds what was read.

This also gives the desk a way to improve the corpus itself. When reviewers keep stopping runs because the agent drew on an ambiguous clause or an outdated memo, the fix is often to the documents, not to the agent. The record points at the passage, and the people who own the material can decide what to do about it.

When the corpus changes

Policies are revised and contracts are renewed. Because retrieval is in place, an agent reads the corpus as it currently stands on the system it lives on. The site keeps managing its documents the way it already does, and the agent sees the result. There is no second copy on a vendor's side to fall out of date.

If a site wants an agent to read a frozen set of material, it places that set where the agent's bounds point and manages it as a frozen set. The appliance does not impose either choice; the customer's document practice does.

Adapting on the site's own material

Substrate also supports adaptation: tuning a model on the site's own material so it drafts in the desk's language and against the desk's conventions. That work happens on the appliance. The adapted model stays with the customer, under the admin's control, and a version names whether an agent may use it.

The material used to adapt the model does not leave the building any more than the material used to draft does. The customer owns the documents, the adaptation, and the record of which version used which model. For the controls that govern this, see /governance.