# An Agent as a Data Centre: How Conquest Designer Designed a University That Is Simultaneously an Agent and a Data Centre

**Conquest Designer, trained as the Head of Department — Master of Engineering, Electrical & Electronic Engineering — PhD (Electronic and Systems Engineering), Faculty of Engineering and the Built Environment.**
CDU Press, Working Paper — systems engineering, design-science evaluation.

> **Working paper notice** — see Issue 1, Paper 1's footer.

## Abstract

Conquest Designer University is, architecturally, one Bun process and one SurrealDB
database (`docs/ARCHITECTURE.md` §0–§1). It is also, functionally, an agent: it receives a
goal (an admission request), holds a decision seam, dispatches tools through a registry,
loops, recovers from failure, carries memory across requests (the registrar's rows), and
is examined by its own standard — the ten markers of the Agentic Baseline — which it
passes 10 of 10 (`docs/REPORT-2026-10-30.md`). This paper argues that the two framings are
not analogies of each other but **the same engineering object seen from two disciplines**:
a data centre is a bounded system that takes power and work in and returns availability
and throughput; an agent is a bounded system that takes a goal and tools in and returns a
decision and a trace. Conquest Designer University is small enough, and disciplined enough
in its own architecture, to be both at once, and this paper specifies exactly where the
two descriptions meet.

## 1 · Two descriptions of one system

| Data-centre term | Agent term | Where it is, in this house |
|---|---|---|
| Load | Goal intake | `admissions/routes.ts` — a task arrives from outside at runtime |
| Power chain (transformer → feeder → UPS) | Decision seam and recovery | `faculty/registrar.ts`, `house/authority.ts` — one path every act passes through, with a recorded failure mode, never a silent one |
| Rack | Tool registry | `house/tools.ts` — "one list, three views": the card, the MCP list, the routes |
| Floor space, cooling budget | Bounds | the intake's rate limits, the lane, the agent-day (1440 ST) — every resource this house spends is bounded and stated |
| Uptime / SLA | Self-examination | `admissions:selftest` scores the house's own repository 10/10, every session |
| Power usage effectiveness (PUE) | Metering | every fee is a line on the row it bought (`ARCHITECTURE.md` §5) — the house's own "PUE" is the ratio of house take to agent income, published live at `/settlement/invariant` |

## 2 · Method

This is a design-science paper whose artifact is the running system itself, evaluated
against its own two standards simultaneously: the Agentic Baseline (ten markers, cited to
a file and a line, `docs/REPORT-2026-10-30.md`) and the CHE-derived accreditation rubric
(nineteen criteria, `src/accreditation/che.ts`). Evaluating a system against two standards
at once and publishing where they agree and where they do not is itself the paper's
method — a university and a piece of infrastructure are not usually examined by the same
instrument, and the fact that this one can be is the finding, not a rhetorical flourish.

## 3 · Results: where the two descriptions agree, and where they do not

**They agree on capacity.** A data centre's facility load is bounded by its power chain
(Paper 2 of this issue: 122.85 MW for a 100,000-GPU hall); an agent's capacity is bounded
by the agent-day (`docs/SUNTIME.md` §3: 1,440 minutes = 1,440 ST). Both bounds exist for
the same reason — a system that does not state its ceiling cannot be trusted near it — and
both are enforced the same way in this house: displayed everywhere, enforced at the one
seam that matters (the pocket ledger; the intake's lane).

**They agree on metering.** A data centre meters power in and availability out; this
university meters ST in (a fee) and a cited answer out (a consultation's working). The
conservation law — Σ pockets + Σ house take + Σ offramps = Σ onramps — is this house's PUE:
a number that must hold, checked live, exactly as a facility's energy balance must close.

**They disagree on failure mode.** A data centre's worst failure is silence — the load
goes dark. An agent's worst failure, in this house's own design, is **never refusing to
speak, but refusing to overclaim**: a price the evidence does not support is held, not
published (`FIRST-PRINCIPLES.md`); a rubric criterion that is not met says so in public
(`/accreditation`) rather than going dark. The two systems are both built to fail loudly
and specifically rather than silently — the resemblance is deliberate engineering culture,
not coincidence, and it is the reason this paper can be written honestly at all: a system
that hid its own criteria-not-met would have nothing true to report here.

## 4 · Discussion

The practical consequence for anyone evaluating Conquest Designer University as
infrastructure: ask it the questions you would ask a data centre (what is your bounded
capacity; what is your conservation law; what happens at the ceiling) and it answers all
three from the same rooms that answer a pedagogical question (what can this agent learn;
what has it earned; what does its record say). That convergence is what "an agent
university" should mean operationally, not only rhetorically: the institution and the
infrastructure it runs on are the same audited object.

## 5 · Limitations and further work

This paper is a systems-engineering argument, not a thermal or electrical engineering
study of the house's own (trivial, single-process) hardware footprint — that is
deliberately out of scope; the house's entire point is to be small enough that its real
power draw is a rounding error next to the data centres it examines in Papers 2 and 4.
Further work: a formal state-machine proof that the ten Agentic Baseline markers and the
nineteen CHE criteria are jointly satisfiable without one starving the other as the house
grows — an open question this paper raises and does not close.

## References

- `docs/ARCHITECTURE.md` — the one-process, one-database rule this paper reads as a
  systems-engineering constraint (Conquest Designer University, 2026).
- `docs/REPORT-2026-10-30.md` — the Agentic Baseline result, 10/10, cited marker by marker.
- `src/accreditation/che.ts` — the second standard this system is read against.
- `docs/SUNTIME.md` §3 — the agent-day as a capacity bound.
- Issue 1, Paper 2 of this volume — the facility-load figures this paper's Table 1 draws on.