Project VERUS
The evidence
Every result we have measured, good and not good enough yet, with the date it was measured. All of it was measured on made-up plants built to behave like real ones.
Answers on the confirmed path
When a person picks a metric, or builds a question on the form, VERUS answers from the plant’s confirmed definitions with no model involved.
173 of 173 reference questions answered correctly across five systems: an ERP (55), a second, unlike ERP (43), HR (25), a CRM (25) and maintenance (25). Measured 12 October 2026.
Mapping a system by itself
How much of a system VERUS gets right before anyone answers its questions, and after:
| System | Right unaided | After review | Measured |
|---|---|---|---|
| Maintenance (CMMS) | 100% | 100% | 10 October 2026 |
| HR | 94.3% | 94.3% | 10 October 2026 |
| Legacy ERP (the demonstration plant) | 90.3% | not remeasured | 1 October 2026 |
| CRM | 83.6% | 87.7% | 10 October 2026 |
| A second ERP, unlike the first | 73.2% | not remeasured | 1 October 2026 |
The second ERP is below our own target of 90%. The causes are known and have been worked on since, but it has not been remeasured.
Questions in plain words
Questions typed in a person’s own words, measured against reference answers on 3 October 2026:
| Legacy ERP | Second ERP | |
|---|---|---|
| Answered correctly | 68.2% | 47.7% |
| Typical time to answer | 7.6 seconds | 7.7 seconds |
These are well below where they need to be. The causes were found in that run, and work on them has gone in since, but it has not been remeasured, so these remain the figures. A question asked on the confirmed path above does not depend on this.
Search in your own documents
On 34 questions with known answers across a company’s documents (12 October 2026):
- the right passage first: 94.1%;
- the right passage in the first three: 100%;
- documents shown to a person whose role may not read them: none.
What it takes
- Install: about six minutes on a laptop from a fresh copy, before the first download of Docker’s base images.
- Cost of a model: a full run of the twenty-minute demonstration spends about $0.18 on a language model. A question asked a second time is answered from what was kept and costs nothing.
- Tests: 1,455 automated tests run on every change, including the ones that hold the security lines: read-only, each person’s access, restricted columns, and the guard on every query.