The question
The previous entry found that the hauler given the heavylift class’s quicker response made no bad landing in a steady cross wind at 1.5 m/s. Does that hold when gusts and turbulence are added? And the quicker body changed two things at once: how fast the machine itself responds, and the response time its autopilot plans with. Which one does the work? The two main reviewers ranked these questions differently for the fifth cycle running, so both were registered with fifteen predictions and flown together in one job on the same seeds.
The flight
The CRAFT campaign flew the Seventh Avenue pad world on 200 fresh seeds in eleven cells, scoring about 1,690 scheduled hauler landings per cell. The full package (steady wind, gusts and turbulence) at 1.25 and 1.5 m/s on the assumed slow body (1.5 s to respond, 0.3 s delay) and on the quicker one (0.6 s and 0.15 s, with the planning the simulator ties to it); the station-keeping setting (slT 0.75) at 1.5 m/s on both; the steady wind alone at 1.5 m/s on both. Then the split: the quick machine with the slow planning (steady wind and full package) and the slow machine with the quick planning (steady wind only). The simulator itself was not changed.
What happened
Under the full package the quicker body made 4.8 bad endings per 100 scheduled landings at 1.5 m/s against 47.2 for the slow body, and 2.8 against 27.6 at 1.25 m/s. What gusts and turbulence added on top of the steady wind fell from 22.3 to 4.8 per 100: cut by about four-fifths, but still above the 2-per-100 line set in advance.
Under the steady wind alone, either half of the quicker body was enough on its own: the quick machine with the slow planning made 0.5 per 100 (9 of 1,694) and the slow machine with the quick planning made none (0 of 1,694), against 24.9 for the slow body. The two halves did not add; each alone removed almost everything. Under the full package the quick machine with the slow planning still made 29.8 per 100, so the planning carried about 60 per cent of the cut. The slow machine with the quick planning was not flown under the full package.
The quick planning began the final descent about 8 m lower and much closer to the pad’s line, whichever machine flew it. The station-keeping setting changed nothing on the quicker body, by construction: it sets the same planning time the quicker body already uses. On the slow body it still cut the full-package rate from 47.2 to 8.7. Ten of fifteen predictions held; the misses all underrated the planning.
What we learned later
Follow-up (2026-10-12, CRAFT cycle 38, provisional). The missing cell was flown on 200 fresh seeds: the assumed slow machine with the quick planning under the full package made 23.5 bad endings per 100 scheduled landings at 1.5 m/s against 47.0 with the slow planning, and 16.4 against 26.4 at 1.25 m/s, while the quicker body with the quick planning made 5.2. So planning alone halved the loss but did not carry the gusty package; the slow body kept a demonstrated cost. Measured both ways round, body and planning added up on this scale, the planning giving about 56 to 58 per cent of the cut, so this entry’s ‘about 60 per cent’ (measured along one path) holds in either order. Of the planning, only the part that plans the stopping and the final approach did the work (23.5 per 100 alone); the spacing-margin part alone lost 406 scheduled haulers before they reached the final, so that cell was not scored. Among the 265 arrivals held at start-up (outside the score), 112 ended badly with the slow planning, 212 with all of it quick and 155 with only the stopping part quick. The cycle-38 review also quarantined one telemetry column in this entry’s record, the label of which controller branch was steering at each wall contact, because it was reconstructed with the slow 1.8 s planning time in every cell; no figure in this entry used it. The planning values are an uncalibrated controller-parameter experiment on assumed bodies; nothing was adopted.
research/city/design/M-CRAFT-PAD-PLAN-FULL-record.md
The cost was in how the drone starts. Amendment (2026-10-13, CRAFT cycle 39, provisional). The previous note reported that, with only the stopping part of the planning quick, more of the haulers held at start-up ended badly (155 against 112 of 265). Those held haulers are the only arrivals the simulator releases from a standstill at the pad edge; every other arrival enters at cruise speed. On 200 fresh seeds the same 279 held arrivals were flown both ways. Released from a standstill, the stopping-only planning again did worse: 52.0 against 35.8 bad endings per 100 (+16.1). Released like every other arrival (at cruise, on the normal clock, with no wait for the pad; a harness switch, not a procedure any craft flies), it did better: 26.5 against 47.3 (−20.8), as it does for the scheduled haulers, and the difference between the two releases (+36.9) is demonstrated. So the held-arrival cost belongs to the model’s standstill release, not to arrivals in general. The ordinary release changes speed, timing and the pad wait together, so speed alone is not identified as the cause, and a standstill start in a cross wind is not a real hover. Six of six registered predictions held. The planning values and the slow body remain assumptions; nothing was adopted.
research/city/design/M-CRAFT-PAD-STOP-HELD-record.md
Correction and follow-up (2026-10-13, CRAFT cycle 40, provisional). The cycle-40 reviewers found the previous note too strong where it said the held-arrival cost belongs to the model’s standstill release: the bad endings depend on the release and the planning together, so the planning is not taken out of the question (descriptively, the +36.9 split into about −11 per 100 under today’s planning and +25 under the stopping-only planning). The numbers stand. Cycle 40 then took the release apart on eleven chosen held arrivals: changing only the release speed from rest to cruise changed every ending and reproduced the normal arrival’s ending in 21 of 22 cases; releasing a third of a second later changed one, and the pad clearance check none. Selected cases, not a rate; see the CRAFT / 40 entry.
research/city/design/B-CRAFT-LOW-SPEED-ENTRY-record.md
What this does not establish
One simulated world, wind direction and arrival rate. Both hauler response figures are assumed, and no sourced response for a six-tonne multirotor was found; the wind package is a model convention. So no real landing risk, wind limit or controller setting follows. The planning parameter reaches several controller readers at once, and which of them does the work is not separated. Among arrivals held at start-up, a simulator artefact outside the score, the quick planning did worse. The reviewers of the previous entry split (one flawed, one sound); this entry’s own review has since corrected parts of its record, which a later integration will carry.
The next question
Nothing is adopted: the planning time and the station-keeping setting stay test keys, and the quicker body is a test cell, not new hauler numbers. In this model the hauler’s landing loss now moves mostly with the response time its autopilot plans with, not only with how fast the machine is. Next candidates: the slow machine with the quick planning under the full package, the split of the planning’s readers, the quicker body’s remaining few bad landings, and the held-arrival cost of the quick planning.
Source references for reproduction
Optional repository references, pinned to this snapshot. Repository access may be required.