What the simulation found, and what it got wrong
Shared Sky is a scenario model of delivery drones (Wing, Zipline, Amazon) and of air taxis between Texas airports and vertiports, built from FAA filings, company announcements and Census data. It is not a forecast and nobody has endorsed it. No person outside the project has reviewed it yet; the reviews so far were done by AI reviewers. Predictions are written down and fingerprinted before the runs; this page shows the ones that came true and the ones that did not.
Read this first. Every result here comes from a simulation, not from real flights, and none has been checked against real operations: the companies do not publish flight-level data, and the air-taxi networks are scenarios built from announced sites. Read each percentage as “what happens in this model under these assumptions”, not as what operators will see. Being reproducible and pre-registered makes the work checkable, not accurate. The assumptions that move the numbers most are listed at the end.
Being re-checked (22 November 2026). A review of the engine found three counting problems. Drones that change their plan in flight (stops, return-home, outages) are undercounted; the random errors given to neighbouring flights were correlated; and some studies counted meetings after the measured window ended. The headline result (72 % / 73 %) is not affected. Results from ERRORS1, ERRORS2, PRELAUNCH, RECOUNT1 and the bad-day and return-home studies are provisional until they are re-run with the fixed engine, with the re-runs pre-registered and any changed verdicts reported here.
Guess first
Before you read the results, try four of the questions this project asked. Each answer also shows what I predicted, in writing, before the simulation ran. Your guesses stay in your browser; nothing is sent anywhere.
1. If the three delivery-drone companies each flew their own slice of the height band they already filed, how much would planned conflicts between them fall?
About 72 % in Houston and 73 % in Dallas–Fort Worth (10 m vertical gap, no plan sharing; if Amazon really flies near 400 ft the cut shrinks to about 29–35 %).My prediction before the runs: 60–80 % in Houston, 55–80 % in DFW. Both held (HEADLINE2).
2. If drones check their flight plans against each other before take-off and wait when they overlap, what share of flights would have to wait in windy hours?
About 1 in 5 or 6: 18.9 % in Houston and 17.5 % in DFW over 10 windy hours (SEEDS2). Resizing the layer gaps to fit the reservations removes more than half of that cost but not all of it (COORD2).My prediction before the runs: at most 5 %. It failed (COORD1, SEEDS2).
3. Flying between twelve announced Texas air-taxi sites, what share of Dallas–Fort Worth jobs would have a commute at least 10 minutes faster at peak?
About 1.2 % (median; 5th–95th percentile 0.02–4.9 %), mostly long cross-metro commutes, and that is at zero fare (UNCERT1).My prediction before the runs: between 0.1 % and 3 %. It held in the main run (DFW 2.57 %, Houston 1.25 %), though DFW sat close to the upper edge and one earlier single run came out at 3.05 % (ACCESS1, UNCERT1).
4. With air taxis flying until midnight at the busiest level, how many people live where the noise reaches a day–night average of 55 dB (a busy street)?
None, in either metro. The loudest squares off the pads are airport land with no residents. About 7,600 people near the DFW network and 2,100 near Houston's would still hear 10 or more loud flights a day (NOISE2).My prediction before the runs: at least some people. It failed (NOISE2).
The headline: sharing out heights cuts planned conflicts by about 72 % (Houston) and 73 % (DFW)
If each delivery company flies its own slice of height inside the band it already filed with the FAA, planned conflicts between the companies fall by 72 % in Houston (95 % range 67–76 %) and 73 % in Dallas–Fort Worth (66–80 %), with Amazon at its best-sourced heights. A planned conflict is two companies’ planned paths within 30 m sideways and under 10 m vertically, with no coordination between them. Real layouts at the filed site counts, 10 runs each.
It depends on: where Amazon really flies home (if near 400 ft, as one secondhand staff remark says, the cut is 29 % / 35 %), and on 10 m between companies being enough (at 15 m it is 55 % / 45 %). The ranges cover random site layouts, not model error. With a realistic plan-sharing check 99–99.6 % of them are removed either way (10 windy hours, SEEDS2), and height sharing roughly halves how many flights have to wait (COORD1). In collision risk at average GPS, though, that check leaves the shared-out heights slightly more residual risk than the best-sourced heights, because Amazon’s and Zipline’s layers sit exactly 20 m apart, at the edge of what the check sees (RISK2); the gaps need a margin. Wing’s and Zipline’s sites are placed at random because their lists are not public. Two later checks: with all three companies on departure corridors the cut is 71 % / 82 % (CORR2); with the filed overlap caps applied the cut is 72 % / 59 % but absolute conflicts fall by 82–90 %, so every absolute count here is an upper bound (CAPS1). The model flies in wind but not in rain: in 2025 about 22 % (Houston) and 13 % (DFW) of daytime hours had measurable rain and 4–5 % had heavy rain or 55 km/h gusts; the data cannot see thunderstorms (WX1).
Source: planned/HEADLINE2_RESULTS_2026-10-25.md, COORD1_RESULTS_2026-10-26.md, CORR2_RESULTS_2026-10-28.md, CAPS1_RESULTS_2026-10-28.md, WX1_RESULTS_2026-10-29.md
The scorecard
97 registered predictions failed and 4 were not scored (a pre-run check failed, or the case never occurred). They are across 77 studies, drones and air taxis; each is listed in the study record. Some passes were easy predictions; the failures are the informative part. A model whose predictions always come true has not been tested.
1. Thinner reserved blocks plus a different layer order cut drone delays by about three quarters
Drones that fly in the same airspace reserve a block of sky around their planned path and wait if it overlaps another. With the shared-out heights this project proposes and 10 m blocks, about 1 flight in 5 to 6 waits in windy hours. I had predicted at most 5 % delayed (COORD1); it was 17–19 % (confirmed over 10 windy hours each, SEEDS2). Putting Zipline on the top layer does nothing alone, and thinner blocks do most of the work; together they cut the delayed share to about 4 %.
Table view
| Metro | Shared-out, 10 m blocks | Thinner blocks | Zipline on top + thinner blocks |
|---|---|---|---|
| Houston | 18.5 % | 7.9 % | 4.4 % |
| Dallas–Fort Worth | 17.0 % | 6.8 % | 4.1 % |
Source: planned/LAYER2_RESULTS_2026-10-27.md
Not shown: that 5 m blocks are safe. The simulated conflict counts are 0 to 3 per arm, too few to show a difference. GPS-level errors are tested in the next section.
1b. The height-sharing cut shrinks with GPS error, mostly because of height error
Provisional: ERRORS1 and ERRORS2 used correlated random errors between neighbouring flights and are being re-run (see the note at the top).
The headline runs fly every drone exactly on its plan. ERRORS1 (276 runs, both metros) gives each flight a random sideways error, height error and late start, sized to published figures: near the best monitored GPS sites (L1), the GPS standard’s committed average (L2), its committed worst site (L3), and the size the FAA Remote ID rule allows (L4). This study compares the shared-out heights with the heights the companies filed (not with Amazon’s best-sourced heights, which give the 72 % / 73 % headline and were not rerun here). Without errors that cut is 80 % in Houston and 81 % in DFW on these six seeds; with them it is 76 % / 79 % at L1, 63 % / 59 % at L2, 30 % / 26 % at L3 and 20 % / 23 % at L4. Height error alone does nearly all of this; sideways error or late starts alone change the cut by 3 points or less. I predicted the cut would fall further at L2–L4 (3 of 13 predictions failed).
The plan-sharing check is affected too: the minimal reservation (30 m wide, ±10 m high) removes 83 % / 77 % of flown conflicts at L2 but only 18 % / 22 % at L4, while a block taller than the height error (±46 m) removes almost all, at the cost of roughly twice as many flights waiting (37 % / 47 % in windy hours against 18 % / 17 %).
Read with care: errors here are independent between flights, but real GPS and barometer errors are partly shared by nearby drones, so L2–L4 probably overstate the damage. The late-start size is my assumption, and I found no published figure for the height accuracy of delivery drones, so none of these levels is a measurement of any company’s aircraft. Six seeds; intervals are wide (for example DFW at L4: 12–32 %).
The headline case and shared errors (ERRORS2). Against Amazon’s best-sourced heights (the 72 % / 73 % headline), the cut behaves the same way: 71 % / 73 % at best-site GPS, 60 % / 54 % at the average, 22 % / 23 % at the Remote ID size. But nearby drones share most of their error, because they see the same satellites and the same air pressure. Splitting each error into a shared and a personal part shows that only the personal part matters for separation: if 80–95 % is shared, the average-GPS cut is 66–75 %, close to the no-error 72–74 %. How much is shared on real delivery drones is not measured, so the independent-error numbers above are the pessimistic end. All 10 predictions held, some with wide bands.
1c. In collision risk, not just conflicts, height sharing gives about the same cut
A planned conflict at 25 m is not the same as one at 1 m. RISK1 turns every planned encounter between companies (36,150 over 40 runs) into a collision probability with the Reich model used by ICAO: the drones collide if their errors bring them closer than their combined size. Against Amazon at its best-sourced heights, sharing out heights cuts modelled collision risk by 79 % (Houston) and 82 % (DFW) at best-site GPS and 61 % / 58 % at the GPS standard’s average; if most of the error is shared by nearby drones, the average-GPS cut is 78–83 %. At the Remote ID size it is at least 16–18 %. All 5 predictions held.
Read with care: these runs have no plan-sharing check and no avoidance, so the rates (about 1 modelled collision per 10,000 flight hours) describe plans as filed, not real operations, where the plan check removes 98–99 % of conflicts at average-or-better GPS. Zipline’s and Amazon’s sizes are not published and are assumed; risk roughly doubles with double the size. Without errors the counts are whole events (3 planned collisions in about 31,500 Houston flight hours with best-sourced heights, 0 with shared-out heights; 2 and 1 in DFW). This is a model quantity, not a safety assessment.
1d. Drones dodging each other over Remote ID barely helps once GPS has its normal error
AVOID1 gives every drone a simple avoidance rule, with figures from an FAA-funded Remote ID avoidance test (BAA004: detection at 100–500 m, message gaps of 5 s or more, 36 of 51 test encounters avoided): detect the other drone at 300 m, react after 6 s, climb or descend at 2 m/s towards a 10 m gap. With exact GPS this removes 71 % of the modelled collision risk. At the GPS standard’s average error it removes only 3–5 %, even if every alert is acted on (4–7 %). Most of that risk comes from drones planned more than 10 m apart in height that their errors bring together, and each drone broadcasts where it thinks it is, so the broadcasts cannot show it; and a 10 m gap is inside the height error. Height sharing still cuts the remaining risk by 57–61 % and the number of manoeuvres by three quarters. I predicted a 50–71 % cut; 2 of 5 predictions failed.
Read with care: one manoeuvre per encounter, the other drone does not react, climb rate and latency are assumptions, and no plan check (as RISK1). Avoidance with sensors that see true positions (radar, cameras) is not modelled and could do better.
1e. With plans checked before take-off, the modelled risk mostly disappears, and the shared-out heights show a weak spot
RISK2 switches on the plan-sharing check before computing collision risk. The minimal check (blocks 30 m half-width, ±10 m high) removes 81–96 % of the modelled risk at the GPS standard’s average error and effectively all of it at best-site GPS; a larger check (±46 m) removes all of it but delays 40–48 % of flights and cancels about a quarter. With the minimal check, though, the shared-out heights keep more residual risk than Amazon’s best-sourced heights (Houston 0.93 vs 0.82, DFW 1.11 vs 0.51 modelled collisions per 100,000 flight hours). The reason is the layout: Amazon’s outbound height (81 m) and Zipline’s (101 m) are exactly 20 m apart, the edge of what ±10 m blocks can see, so the check never flags the pair while average GPS errors can still bring them together; they cause about 90 % of what is left. Height sharing still halves the flights the check delays (19 % vs 40 % in Houston, 17 % vs 37 % in DFW). Layer gaps and check sizes have to be set together, with a margin for height error. 3 of 5 predictions failed.
Read with care: still air, block sizes are my assumptions (the ASTM standard is paywalled), and the residual rates rest on few encounters. A model quantity, not a safety assessment.
1f. Bad GPS among tall buildings hardly moves the metro-wide figure
ERRORS3 gives downtown the urban-canyon error measured by NASA Langley in Columbus, Ohio (95th percentile 47.8 m sideways raw, 4.4 m with their processing) and keeps average GPS elsewhere. Metro-wide modelled collision risk changes by +0.1 %, because downtown is about 0.2 % of each counted area and holds under 0.4 % of the close encounters. If every place had canyon-sized sideways error, risk without a plan check changes by at most 6 %. I predicted that with the minimal plan check the residual risk would rise by at least 50 % under canyon error everywhere; it changed by −11 % to +17 %, so that prediction failed. A likely reason, not tested here, is that the check's weak spot is height gaps (RISK2), which sideways error does not touch. Errors that wander during a flight were settled by argument (same size at any moment, same expected rate), not by runs. 1 of 4 predictions failed.
Read with care: downtown is a box, not a building-by-building signal model; one city, one study; height error assumed unaffected; still air; lower bound with canyon error everywhere (90 m search). A model quantity, not a safety assessment.
1g. Weather stops and dinner-time peaks add conflicts, because conflicts grow with the square of traffic
Doubling traffic roughly quadruples planned conflicts (exponent 1.98–2.09, WX2). So anything that bunches flights adds conflicts even when the day's total is the same. If drones stop in 2025's bad-weather hours (3.9 % of daytime hours in Houston, 4.6 % in DFW) and catch up afterwards at up to 1.5 times the normal rate, a year has about 6–7 % more planned conflicts. If orders peak at lunch and dinner with the busiest hour twice the average, a day has 40–48 % more, and the busiest hour about 4 times an even hour (DEMAND1). Height sharing keeps its 72–75 % cut in both. Every absolute count on this site assumes even demand, so it is low by about that much. At double traffic the plan check's delayed share barely moved while cancellations rose (I predicted the opposite). 2 of 8 predictions failed.
Read with care: stop rule, catch-up rate and peak size are assumptions; the demand shape is from food delivery in Beijing; thunderstorms are not in the weather data; still air.
1h. Real helicopters fly lower than I expected, but meet few drones
HELI1 uses 14 days of 2025 public flight-tracking data (adsb.lol, ODbL): 1,199 helicopter-days over Houston and Dallas–Fort Worth. 30 % of helicopter time over the counted areas in Houston and 46 % in DFW is below 150 m above ground, and 14–17 % is inside the drone height band (46–121 m). I had predicted under 15 % and under 8 %, so three of five predictions failed. The shares move a lot with the ground height I assumed (roughly 20–60 % below 150 m), and volunteer receivers miss low flights, so they are floors. Against the drone traffic of the simulation, a low-flying helicopter has on average 0.03 drones (Houston) or 0.009 (DFW) in a 450 × 450 × 60 m box around it, and height sharing makes no difference to that. This counts exposure, not encounters or collisions.
Read with care: ground elevation is an assumption, heights are pressure altitude, helicopters and drones are not matched in time of day, and rules or avoidance that would separate them are not modelled. No aircraft identifiers are published.
2. About 1% of jobs or fewer would see a commute that flying shortens, and the number is very uncertain
Using Census blocks and the Census commute-flow data, I counted existing commuters whose door-to-door time falls by 10 minutes or more if they flew between two of the twelve Texas sites Archer announced on 23 October 2026 (a different, larger network than the seven sites on the air-taxi page). The single run I first published gave 3.1 % of jobs in Dallas–Fort Worth and 1.3 % in Houston. Letting the uncertain inputs vary shows that was an optimistic reading.
Table view
| Metro | 5th | 25th | median | 75th | 95th | single point |
|---|---|---|---|---|---|---|
| Dallas–Fort Worth | 0.02 % | 0.32 % | 1.20 % | 2.80 % | 4.93 % | 3.05 % |
| Houston | 0.22 % | 0.34 % | 0.48 % | 0.94 % | 2.43 % | 1.28 % |
Source: planned/UNCERT1_RESULTS_2026-10-27.md, planned/ACCESS1B_ADDENDUM_2026-10-27.md
Not shown: riders. This counts people with a time saving at zero fare, ignoring queues, weather, aircraft limits and whether anyone would pay. The input ranges are my assumptions. The biggest driver is how fast the road is that people would otherwise drive. Beneficiaries commute about 67–77 km in a straight line against a typical 20 km, and they do not hold a higher share of high-wage jobs than average: that prediction failed.
3. How a vertiport lets aircraft land matters more than how many stands it has
When parking stands are limited, an arriving aircraft may have to wait in the air. With two landing pads per site, the simple first-come rule never met the service and battery-reserve test at 3 requests an hour, however many stands there were. Booking the landing slot when the aircraft takes off needed 8 stands per site; holding the aircraft on the ground until a stand is reserved needed 5, at the price of longer waits for passengers (about 225–285 seconds against 46–130).
Table view
| Case | First come | Slot at take-off | Hold on ground |
|---|---|---|---|
| DFW, 2 pads | never (>16) | 8 | 5 |
| Houston, 2 pads | never (>16) | 8 | 5 |
| DFW, 4 pads | 7 | 7 | 5 |
| Houston, 4 pads | 6 | 6 | 5 |
Source: planned/PADRULE1_RESULTS_2026-10-27.md
Not shown: a certified procedure. Real arrival priority, go-arounds and weather are not modelled, and stand counts at real airports are not public. One of the six predictions for this study failed.
4. The aircraft fleet needed is fairly robust; the charging rate is what moves it
Across the stated ranges for pad time, turnaround, charging rate and flown distance, the required fleet moves by one or two steps. A faster charger shrinks it most. It rises quickly with demand: each doubling of requests multiplies the aircraft by about 1.6 to 2.
Table view
| Requests/h per route direction | DFW | Houston |
|---|---|---|
| 1.5 | 8 | 5 |
| 3 | 14 (12–16) | 8 (8–12) |
| 6 | 24 (24–28) | 14 (12–20) |
| 12 | 48 | 28 |
Source: planned/AIRTAXI_NET3_RESULTS_2026-10-27.md, planned/UNCERT1_RESULTS_2026-10-27.md
Not shown: real demand. The request rates are scenario settings, not forecasts; the charging rate comes from a manufacturer’s marketing figure and aircraft range is a cautious assumption checked against Wisk’s and Joby’s stated figures (planned/AIRTAXI_INPUTS_ADDENDUM_2026-10-28.md).
5. How well my own predictions held up
Every study here was registered with predictions before it ran, each written as “this will hold”. Of the 353 that were scored before this check, 274 held (78 %, 95 % interval 73–82 %). I was overconfident by about 22 points, and I found no kind of prediction I was better at.
Table view
| Group | Held / scored | Rate |
|---|---|---|
| All scored predictions | 274 / 353 | 77.6 % |
| Ranges and limits | 251 / 323 | 77.7 % |
| Comparisons and claims | 23 / 30 | 76.7 % |
| Delivery-drone studies | 162 / 217 | 74.7 % |
| Air-taxi studies | 101 / 125 | 80.8 % |
| First half of the project | 138 / 177 | 78.0 % |
| Second half | 136 / 176 | 77.3 % |
Source: planned/CALIB1_RESULTS_2026-10-31.md, planned/calib1_results_table.md
- No improvement over the project. The first half of the predictions held 78 %, the second half 77 %.
- Numeric targets were no harder than comparisons. Ranges and limits held 78 %, comparisons and claims 77 %.
- The newer topic did slightly better, not worse. Air-taxi questions held 81 %, delivery-drone questions 75 % (intervals overlap).
- Misses leaned one way. When a numeric prediction failed and the miss could be placed, the result was worse for the system than I had predicted 37 times and better 19 times (16 could not be placed). Only 8 of the 79 misses were close calls.
Not shown: that these are independent tests. Predictions inside one study share data and a model, so the intervals are too narrow. I also chose the questions. I predicted five things about my own forecasting and only one held (the lean toward worse results); the other four were wrong (registered in planned/CALIB1_PREREG.md). Coding by three independent coders; a blind re-check of 10 % of rows agreed on every verdict.
Other results, in one line each
- Outside check. The air-taxi engine reproduces a published MIT vertiport-capacity study (Vascik & Hansman, 2019) in 5 of 5 predictions: pad use within 10 points in 15 of 16 cases and the same saturation points.
- Shared flight plans. Reserved airspace blocks inspired by ASTM F3548 (the standard is paywalled and was not read; block sizes are my assumptions) cut flown conflicts by 99–99.6 % in windy hours (SEEDS2, 10 runs); about 1 flight in 5 to 6 waits, about 6 minutes on average.
- Air taxis near drone sites. At full drone build-out about 2 in 5 take-offs and landings at the Texas Medical Center helipad, and 1 in 3 at Toyota’s Plano HQ, would have a delivery drone inside the “well clear” volume on planned paths; drones yielding to booked air-taxi movements removes all of them (AIRTAXI1).
- Air-taxi noise. The highest day-night level off-site is 54 dB over 12 daytime hours, under the FAA’s 65; loud (55 dBA) events reach about 2.4 km from a site (AIRTAXI_NOISE1; flying on to 22:00 adds under 1 dB and to midnight 2.4–3.3 dB, AIRTAXI_NOISE2).
- What I got wrong. My explanation that Amazon’s delays come from its own climbs and descents held with 20 m gaps between layers but not with the 10–14 m gaps of the shared-out heights. An early “the DFW cut is about half” claim was wrong: the corrected figure is 75–77 %. Both are on the pages they appeared on.
How the work is done
- Each study’s question, settings and predictions are written in a registration file and fingerprinted (SHA-256) before any run. A script checks all 81 fingerprints against the files. The oldest fingerprinted registrations date from 22 October 2026; earlier work was not fingerprinted. Three early studies (VALLOC, VALLOC2, ROBUST, which include the first height-sharing result) were committed together with their results, so their order cannot be checked, and FLEET’s analysis was written after its runs. The fingerprints are not timestamped by an outside party.
- The runner and analysis code are committed before the runs. Failed predictions are reported as failures.
- Changes made after seeing a problem are logged as deviations, with at most two fix attempts. Reruns done to understand a result are labelled exploratory.
- Outside claims are checked against the primary source where one is public (FAA filings, company releases, published papers). Some rest on press reports, and ASTM F3548 was not read; anything I could not confirm is marked unverified or left out.
- Reviews: AI reviewers (fresh sessions of the AI assistant used to build the project, each given a role) have checked the code and pages several times and their findings are logged. No person outside the project has reviewed it.
- The project was built with an AI assistant (Claude), which wrote most of the code and text under the author’s direction.
Method in brief (for regulators and planners)
- What is flown. Each company’s filed flight profile (heights, speeds, hovers, sites and daily volumes from its FAA environmental assessment) is turned into planned 4D paths: straight lines between a site and a customer, with filed climbs, hovers and descents. Nothing reacts to anything else in flight unless a plan check is switched on.
- What is counted. A planned conflict is two different companies’ planned paths within 30 m sideways and less than 10 m vertically at the same moment, checked 4 times a second and counted once per episode. A “close” count uses 15 m and 5 m. There is no published separation standard between delivery drones of different companies, so these values are my choice, not a rule; the 15 m and 20 m vertical columns on the city pages show how the result changes if more room is needed (in Houston the cut is 72% at 10 m, 55% at 15 m and 67% at 20 m).
- What the plan check is. Before take-off, each new flight’s plan is compared with the plans already accepted. The minimal check flags any overlap within ±5 s in 30 m cells; the realistic check (COORD1) reserves blocks of airspace around each plan, block sizes are assumptions inspired by ASTM F3548, which I have not read. A flagged flight waits and tries again. The headline runs have no navigation error or wind drift (wind, WIND20 and SEEDS2, height-keeping error, PRELAUNCH, and GPS-level errors, ERRORS1, were tested separately) and there is no tactical avoidance, so the model measures how often plans would need to be changed, not collision risk.
- What the ranges mean. Ranges are the spread over random placements of Wing and Zipline sites (whose site lists are not public) and random demand, not model error. A few quoted intervals (RECOUNT1, COUNTFIX) are instead counting intervals on the number of events.
What this is not, and what moves the numbers
- Not checked against real operations. No result here has been compared with real drone or air-taxi traffic, because I found no public data at that level. The one outside check is against another model (the MIT vertiport study), and a comparison with the BlueSky simulator gave different counts.
- Not a forecast, a safety case, or a business plan. Demand, pad times, aircraft range and charging rates are assumptions, listed with each study.
- Not affiliated with, or endorsed by, any company named. Coordinates and site lists come from public sources and may be off.
The assumptions that change the results most, from the studies above:
- Where Amazon flies. At its best-sourced heights the cut is 72 % / 73 %; if it cruises near 400 ft it is 29 % / 35 %.
- How much vertical room counts as safe. 10 m gives 72 % / 73 %; 15 m gives 55 % / 45 %. No published standard sets this.
- Where Wing and Zipline sites are. Their site lists are not public, so they are placed at random; the ranges show that spread.
- How many flights there are. Volumes come from the FAA filings; with the filed caps applied absolute conflicts fall by 82–90 %, so every absolute count is an upper bound (CAPS1).
- Flying errors. The headline runs have no GPS or height error, no wind drift and no in-flight avoidance. Wind and height error were tested separately, and with GPS-level errors the cut against the filed heights falls from 80 % / 81 % to 59–79 % at best-site and average GPS and 20–30 % at the worst committed site (ERRORS1; the headline case gives 71–73 % and 54–60 %, and most of the cut returns if nearby drones share most of their error, ERRORS2); rain and thunderstorms are not modelled.
- The size of reserved airspace blocks. My assumption, not the ASTM standard. It sets how many flights wait (about 1 in 5 to 6).
- Road speeds and fares for air taxis. The commute result is driven mostly by how fast the road alternative is, and assumes a zero fare. The fleet size is driven mostly by the charging rate, taken from a manufacturer’s figure.
- The code, registrations and results are in the public repository, github.com/forgottenl5/shared-sky. Questions and corrections: open an issue there.