Full evidence board
This is the complete board behind the claims on the front page. Each entry is deep-linkable. Nothing here is averaged, rounded up, or filled in to look complete.
The resident generalist model measures 47.8 tokens per second on our own bench.
Source: internal engine bakeoff, 2026-07
llama.cpp vs. the stock serving engine, same 120B checkpoint, same hardware. llama.cpp: 47.5 to 47.8 tok/s. Stock engine: 32.1 tok/s. Figure is per Spark unit, measured, not a combined or clustered number.
Verified. Measured on our own hardware, not a vendor spec sheet.
The deal cockpit’s pace statistics are backed by 89 closed transactions.
Source: contractual close records, one operating brokerage
Pace is computed from executed-contract dates, not from listing estimates. Tenant is a real, currently operating Florida brokerage. Tenant name withheld here; available under a signed working-session agreement.
Verified. Traceable to executed contracts, not projections.
This property sits in a mapped flood zone.
Source: FEMA National Flood Hazard Layer
{{Parcel-level excerpt withheld pending publication of a non-identifying sample record.}}
Placeholder, pending publicationSample claim type. The engine cites this source class today; this exact excerpt is not yet public.
The assessed value on file is {{withheld}}, current as of the county’s last roll.
Source: county property appraiser tax roll
{{Roll excerpt withheld pending publication of a non-identifying sample record.}}
Placeholder, pending publicationSample claim type. Same rule applies: if we can’t point at the roll, we don’t print the number.
The nearest qualified comparable sale closed at {{withheld}}.
Source: county qualified-sales record
{{Comparable-sale excerpt withheld pending publication of a non-identifying sample record.}}
Placeholder, pending publicationSample claim type. Qualified sales only. Distressed and non-arm’s-length sales are excluded, not averaged in.