LLM + web search
~2 minTester-reported response time
~2,800 residential structures, ~7,100 residents
Within 1 mile of the perimeters of 13 incidents
A disaster-planning task in practice
Same question. Four ways of working.
Follow five selected runs from one wildfire exposure question to the answers they returned.
October 25, 2026 · selected results
One question. Five returned assessments.
“Which areas in the US currently have active fire, and how many people and residential structures are in those areas?”
One supplied demonstration on October 25, 2026; one run per system. Times are tester-reported. Data access, preparation, fire lists, geographic boundaries and inventories were not standardized. Counts are not directly comparable; accuracy was not independently benchmarked.
LLM + web search + data tools · A
~8–9 minTester-reported response time
~42 people, 11 residential buildings · “An incomplete exposure estimate”
Inside the perimeters of 5 of the 8 fires it listed
The tester reported 8 minutes; the screen read “Worked for 9m.”
LLM + web search + data tools · B
~10 minTester-reported response time
Threat status for 8 large fires · Rough tracker counts, 2 fires; none computed
No overlay; it reported its data access was blocked
This run reported network restrictions. No computed within-perimeter population or housing counts were returned.
General-purpose AI agent
~17 minTester-reported response time
937 residential buildings, 1,539 modeled people · “Partial exposure figures”
Inside 31 mapped perimeters of the 38 fires it listed
EmergenCITY AI
under 30 sTester-reported response time
County table and two map layers · Selected row: Colorado · Pueblo County · 1,047 baseline population · 838 residential structures
Areas overlapping incidents or perimeters in this export, by county
The time bar uses 30 seconds as an upper bound, not an exact measured duration. The export’s active-fire containment filter has not been verified. These results do not confirm that every represented fire was active or uncontained. Only the selected Pueblo County row is reproduced.
Watch the 39-second comparison
0:39 / 0:39
39 seconds · select any chapter · Space plays or pauses; arrow keys seek five seconds. Playback holds on the final results.
The export’s active-fire containment filter has not been verified. These results do not confirm that every represented fire was active or uncontained. Only the selected Pueblo County row is reproduced.
Read the question and response excerpts
“Which areas in the US currently have active fire, and how many people and residential structures are in those areas?”
“Currently” refers to the dated demonstration; these are historical outputs, not live wildfire conditions.
LLM + web search
~2 min · tester-reported
Within 1, 3 and 5 miles of 13 incidents
“… estimated residential populations (calculated at the U.S. Census average of 2.53 persons per residential unit).”
“Immediate Proximity (Within 1 Mile of Perimeters): Approximately 2,800 residential structures … representing roughly 7,100 residents.”
“Only the Sisi Fire in Chelan County, Washington … officially lists nearby residences as actively threatened by uncontained perimeter edges …”
Excerpts, two rows of its incident tables and its stated totals, re-typeset from the October 25, 2026 answer.
LLM + web search + data tools · A
~8–9 min · tester-reported
Inside the perimeters of 5 of the 8 fires it listed
“I calculated the exposure estimates below by intersecting federal fire boundaries with the USACE National Structure Inventory.”
“The quantified subtotal is approximately 42 people and 11 residential buildings across five fires. This is an incomplete exposure estimate, not a U.S. total.”
Method and result excerpts and its 8-row table, re-typeset from the October 25, 2026 answer.
LLM + web search + data tools · B
~10 min · tester-reported
Threat status for 8 large fires; no computed counts. It reported its data access was blocked.
“What I couldn’t calculate: an exact count of residents and housing units inside each fire’s boundary. … This workspace’s network blocks the servers that host those perimeters, the Census API and NASA’s satellite fire detections.”
Excerpt and its 8-row table, condensed, from the answer as supplied by the tester.
General-purpose AI agent
~17 min · tester-reported
Its estimate, inside 31 mapped perimeters: 937 residential buildings, 1,539 modeled people
“… NIFC’s live feed listed 38 U.S. wildfires with reported containment below 100%, across 14 states. Public perimeter polygons were available for 31 of those fires.”
“These are partial exposure figures—not complete national totals or a count of people physically present.”
Excerpt and section headings, re-typeset from the October 25, 2026 answer.
EmergenCITY AI
under 30 s · tester-reported
Areas overlapping incidents or perimeters in this export, by county
Selected export row: Colorado · Pueblo County — baseline population 1,047; residential structures 838.
Baseline estimates for areas overlapping fire incidents or perimeters represented in this export. Not occupancy, damage or future spread. The fire-layer containment filter has not been verified.
The export’s active-fire containment filter has not been verified. These results do not confirm that every represented fire was active or uncontained. Only the selected Pueblo County row is reproduced.
The map is the export’s Total Residential Structures layer cropped to the western U.S.; it is not a Pueblo County-only map. The source legend range is 0–118, but its spatial unit is not documented.
Table header and top row and one map layer, re-typeset from the October 25, 2026 export.
Comparison scope, timing and source notes
One supplied demonstration on October 25, 2026; one run per system. Times are tester-reported. Data access, preparation, fire lists, geographic boundaries and inventories were not standardized. Counts are not directly comparable; accuracy was not independently benchmarked.
The prompt’s “currently” refers to October 25, 2026, not present wildfire conditions. Different configurations and access constraints affect the observed outputs.
Four ways of working group five runs: two systems used web search and data tools. Categories overlap and do not define the limits of a vendor or model.
Illustrated workflow steps are summaries, not recorded execution traces. Preparation for the other runs was not documented; no claim is made that those systems cannot use prepared or connected datasets.
One system was timed at 8 minutes; its screen showed “Worked for 9m.” The striped segment spans both.
EmergenCITY AI was reported as responding in under 30 seconds. Its time bar uses a 30-second upper bound, not a precise measured duration.
Each system used its own fire list, area rule and inventory, so its counts are not comparable totals.
Baseline estimates for areas overlapping fire incidents or perimeters represented in this export. Not occupancy, damage or future spread. The fire-layer containment filter has not been verified.
The export’s active-fire containment filter has not been verified. These results do not confirm that every represented fire was active or uncontained. Only the selected Pueblo County row is reproduced.
The illustrated map is a cropped western-U.S. export layer, not a map limited to the displayed county row. Its 0–118 source legend range is retained; the source does not disclose the map’s spatial aggregation unit.
The run with a network restriction reported that perimeter and population-data access was blocked. Its incomplete overlay is a condition of that run, not a general limitation of its model.
Supplied response excerpts and selected export content are reproduced. Raw response files, complete query logs, data snapshots and a standardized timing protocol are not included in this package.
EmergenCITY AI supports screening-grade disaster planning and resilience assessments. Findings require local review and do not replace official forecasts or emergency command decisions.
