# Run 01.5 selected reference plant protocol

Observation and design date: 2026-10-05. Plant: `PLANT-GB300-1GW-2026Q4-A`.
Compute parent: `CFG-GB300-2026Q4`. Governing Accounting Standard v0.1 remains unchanged.

Run 01.5 constructs one conditional engineering design. It does not establish an
actual commissioned project, qualified procurement, or a universal equipment/GW
coefficient. Its inputs retain OBSERVED, DERIVED, SELECTED DESIGN ASSUMPTION,
PROXY, or UNKNOWN classification in canonical objects. Researcher selections
carry no empirical evidence grade and remain conditional sensitivities in the
historical common ledger. A Grade B source supporting a choice does not turn the
choice into an observed facility inventory.

The fixed Run 00 central choices (140 kW rack input, 5% extra-rack IT, 1 GW IT)
are selected explicitly for this plant, with links to the preserved reference.
No old coefficient, observation, conflict, snapshot, or supersession is edited.
Normalized compute equivalents and integer provisioned plant capacity remain
separate. Full repeated blocks are installed, including the residual block;
empty rack positions are not additional commissioned accelerators.

Vertiv RD030 is a candidate principal basis, assessed against other inspected
contemporary OEM architectures. Adaptations require explicit engineering
reconciliation, rating and failure-capacity checks. The Supermicro shared 4+4
power-shelf conflict remains unresolved. An upstream transfer design cannot
prove rack-level failure tolerance or accepted rack usable output.

Electrical and thermal numerical design inputs are first-class assumptions.
Equipment counts are named formulas over module counts and per-module factors.
Nameplate and usable capacity are calculated separately at every redundancy
layer. Losses, COP, hydraulic head, coolant properties, operating fan power,
reserve margins, routes and climate duty are selected assumptions where exact
curves or geometry are unavailable. They are not empirical conversion constants.
Product proxies retain voltage, duty, vintage and material-transfer mismatches.

Power accounting distinguishes B2 input at campus MV delivery from incremental
B3 HV transformation losses. Rack AC/DC conversion and PSU fans remain inside
the 1 GW IT input. Cooling electricity is inside the calculated B2 overhead.
The design-day bottom-up PUE is compared with the preserved 1.20 scalar without
forcing agreement. An annual modeled PUE requires weather/control/load profiles;
it cannot be obtained from a peak COP or a manufacturer peak-efficiency claim.

At steady state, parallel IT heat partitions and all electrical parasitics must
ultimately reconcile with heat released to the environment. Serial CDU and
chiller heat transport is not added repeatedly. Air-cooled heat rejection has
zero evaporative spray demand for the selected circuit. This does not establish
zero annual maintenance make-up, leaks, source withdrawal or basin consumption.
Circulation, inventory, make-up, withdrawal, consumption and discharge remain
distinct. Missing empirical water or material coefficients remain null.

Design closure is conditional on this plant; empirical unknowns remain unknown.
Unselected product counts are not resolved by selecting a different product.
Historical unknown counts and question counts use separate denominators.

Every new object is append-only. A later correction requires a new ID and
supersession or an explicit resolution relation. Offline reproduction verifies
Run 00 and Run 01 entity preservation, source hashes, formula results, plant
capacities, power and heat conservation. Engineering acceptance, detailed
protection/hydraulics, supplier allocation and construction design remain explicit
requirements beyond this accounting reference design.
