In gold mining, a trommel is a rotating screening drum designed to separate valuable material from everything that surrounds it.
A trommel: a rotating screen that keeps what matters and lets the rest fall through.
Troml is the layer underneath. It reads everything your business receives, ties each record to the client, building or matter it is about, and holds that context so every person and every tool you add starts from your own records, with sources.
Desk AI knows what someone pastes into it. A deployed layer already holds every email, document and rule, and knows which client, building or matter each one belongs to before it answers. The model is the same kind either way. What changes is what it stands on.
| AI at the desk | Troml, deployed under the business | |
|---|---|---|
| What it knows | What the person pastes in | Every email, every document, and the law and rules that govern your work |
| Which client or building | Whatever the person types | Worked out before anything is looked up. If it cannot tell, it asks instead of guessing |
| Where it works | A separate chat window | Inside Outlook and the tools your team already opens |
| Grounding | Checked by hand, if at all | Every statement tied to a source; unsupported sentences removed |
| Who acts | A person copies and pastes | The system reads, files, drafts and looks things up. A person decides |
| What can be measured | Time saved, as people remember it | Every message and action, from the records |
From AI in Strata Management: A Field Report from One Firm's First Year, October 2026.
The substrate builds the anchor: your records, tied to what they are about, checked and cited. A domain harness puts your industry on top of it. Tools put the work where your people already are. Build the anchor first, then add the harness.
Where the work lands. Each tool is small because it inherits the context underneath.
Your industry, as configuration on top of the anchor.
The same machinery for every business. It turns raw records into bounded, cited context.
Pick an industry and an incoming message. The left column never changes: that is the substrate. The right column is what the harness supplies and what comes out. Then switch the anchor off to see what the same message does without it.
Without a substrate, every AI tool connects to every source on its own and works out for itself which client an email is about. Each one can decide differently. With the anchor, that decision is made once, and every new tool inherits it. Move the slider.
A new workflow does not rebuild connectors, entity matching or grounding. It asks the anchor. That is why the harness can be small and why the second tool costs less than the first.
A building's or client's history lives in the records, searchable by everyone, not in one person's head. A resignation becomes a handover, not a loss.
Once the anchor exists, more clients and more mail mean more processing, not another hire for every new portfolio.
When an answer is disputed, the run can be replayed stage by stage: what arrived, what it was tied to, which sources were used, what was removed. Desk AI leaves nothing to replay.
Live means running in production for the founding deployment. Built means it exists and is not doing that job yet. We list the gaps in the same type size as the rest. Open Learning and self-improvement for how the system gets better, and where a person decides.
Troml improves only through steps a person can see and approve. Below is each improvement loop, what it does today, and where a person decides. Nothing here changes a live answer, setting or record on its own.
How a correction from your staff becomes a lesson.
How the system looks for its own weaknesses, and why it cannot act on them alone.
The scale runs from B0, an AI that refines one answer, to L5, an AI that changes how its own improvements are judged. It rates loops, not whole products.
| Level | What the AI decides | Troml today |
|---|---|---|
| B0 | How to improve the answer in front of it | Live: unsupported statements are removed before an answer is shown |
| L1 | How to carry out a fixed improvement procedure | Partly: corrections are captured and distilled; re-use is not changing answers yet |
| L2 | Which improvement to try next | The tuner chooses what to test; a person decides what is kept |
| L3 | What to learn from next | Not deployed. Worklists go to people |
| L4 | Which lessons from live work persist | Not deployed, by design. Every lesson needs a person's approval |
| L5 | How future improvements are judged | Not deployed, by design. The evaluator and answer key stay with people |
Levels B0 to L5 from Duan et al., The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement, applied loop by loop.
A new business does not get new machinery. It gets a harness: the parts of its world the substrate needs to know about, written down once and tested against its own people.
What a record can be about, and the numbers that prove it.
Every type of document and message you hold, defined up front, never added ad hoc.
Intents, urgency rules and who owns the response, taken from your system of record.
The acts, regulations, decisions and house rules answers must cite.
What the system may prepare, in your house format, and who approves each one.
A held-out set of real questions with known answers, including traps and ones it must refuse. The pass mark is set before the first run.
Read-only access to one mailbox and its documents.
Probes over your own records: how much resolves, what is scanned, what is missing. You get a written gap list.
Answers checked against questions your people wrote, graded blind, with the pass mark set first.
Drafts and answers inside the tools your team uses, with a person on every decision.
Measured, Oct 2026: a new building's full handover, 2,542 files including the previous manager's mailbox, was loaded the afternoon it arrived. 2,525 files (99.3%) were searchable and linked to the building, and the assistant answered five test questions correctly with links to the sources. One building, one afternoon.
The founding deployment is a Victorian strata manager:
These figures come from locked measurement queries over production records. The failures sit next to them.
Reply-time measurement, re-run on the complete mail record with the firm's own business mail removed; approved 12 Oct 2026. Counts replies by email to the customer's own address, in clock hours. Phone and portal replies are not seen. No control group.
Each figure comes from a locked query that records its code version and a fingerprint of its results. Every figure on this page sits on a claims register and was approved for external use by the general manager. The register retires a figure rather than editing it.
We build the anchor on your own records first and show you what resolves, what does not, and what your people's questions get back. Then you decide whether to add the harness.