The question
Two earlier cycles (LINK 29 and CITY 36) found that, on one simulated crossing, removing cross-level speed-matching cut wall losses sharply. Was that an artefact of the 4 m own-height error, the layouts, the tight lane spacing or the single geometry? Does it carry to opposing streams and to the mixed demonstration fleet, and what does it cost in keeping to the height band, measured as a rate per second of flying rather than as a total?
The flight
The CITY campaign registered seven questions and eight predictions before any run, then flew 244 blocks on 48 fresh seeds: the earlier crossing (calm and with one craft scripted to descend), the same crossing with lanes spaced wider, two opposing streams at a park-side street, and the demonstration fleet for ten simulated minutes, each with and without the term, with exact height and with a 4 m own-height error, on the two weakest assumed radio tiers (L1, the worst: one report a second, half lost, a second late; L2, the next). 11,712 runs, no error rows, no change to the simulator. The price was scored as two rates per active band-lane second, band exits and time outside the band, with paired bootstrap intervals, against the +25 per cent margin assumed in an earlier capacity cycle.
What happened
With exact height the saving was as large as with the error: on L1 5,884 lost craft fell to 3,693; on L2 2,471 fell to 16. Under the 4 m error it repeated the earlier cycle on new layouts almost exactly (0.63 and 0.09 of ordinary handling’s losses remaining, against 0.64 and 0.10). It held when one craft descended, and on wider lane spacing the share saved grew rather than shrank. On the opposing streams nothing changed: ordinary handling lost no craft there and the term never engages between head-on craft. In the demonstration fleet under the 4 m error, 91 craft lost in 48 ten-minute runs fell to 4, with completed routes about the same and no detected band-keeping change; with exact height the fleet lost none either way. No change in mid-air contacts was detected anywhere.
The price: at the crossing under the 4 m error, band exits per second of flying rose about 4 to 6 per cent, inside the assumed margin; the raw totals rose 20 to 30 per cent mainly because the saved craft kept flying. With exact height, craft left the band slightly less often but stayed out longer, and with an 8 s altitude assignment that meant about a third more time outside the band, beyond the assumed margin, in two settings. Five of eight predictions held and one half held; the misses all point one way: the term costs more, in more settings, than expected.
What we learned later
Follow-up and correction (2026-10-12, CITY cycle 38, provisional). The lost craft were then instrumented on 48 fresh layouts: for every craft lost to a wall, each steering term’s push over its last 10 seconds was added up toward the side it hit. On the better assumed radio tier (L2) speed-matching with the crossing street’s height level was the largest sideways push for about 9 in 10 lost craft at the crossing (98 per cent on wider spacing); on the worst tier (L1) it was the most common largest push for about half, with avoidance of craft on the other level almost as common. Lost craft were pushed 20 to 35 m sideways by the term in their last 10 seconds, against 1 to 6 m for craft beside them on the same lane at the same moment; the push came entirely from the crossing street’s level and was sideways, not vertical, and most of the saving falls just past the intersection. This is an attribution along the flown paths, not proof that the term causes the losses. Seven of seven predictions held, a weak test because they were written after the switch’s size was known. The review of this entry’s cycle also corrected two of its clauses: the term does act between the opposing streams (it changes band keeping there in most layouts), although ordinary handling loses no craft there either way; and the share saved did not grow in every wider-spacing setting (one L2 setting did not). It also noted that ordinary handling already spends 42 to 69 per cent of its band time outside the plus or minus 3 m reference box, so the band-keeping price is a change against a reference the controller does not keep. Nothing is adopted; the default decision remains open.
research/city/design/M-CITY-ALIGN-CAUSE-record.md
Would the drone still have crashed? Follow-up (2026-10-13, CITY cycle 39, provisional). The previous note named the push; this cycle asked whether it decided the crash. On 48 fresh layouts at the same crossing and on wider spacing, calm air, exact height, 1,530 wall losses were sampled and each was re-flown from the start on the same layout, with the speed-matching push taken away from the lost craft alone from 10 seconds before its hit, every other craft flying live. On the worst assumed radio tier (L1) about half of those craft stayed off the wall for the next 10 seconds (47 and 52 per cent); on the better tier (L2) nearly all (94 and 100 per cent). A forward or backward push of the same size, with the term kept, saved only 2 to 16 per cent, and replacing the sideways push by an along-lane push of the same size saved as many as removing it: the direction of the push matters, not its size. Three ways of removing the term agreed within 3 points. On L2 the saves were not free: each saved craft came with 0.13 (crossing) to 0.24 (wider spacing) other craft lost in the same interval, above the 0.1 limit set in advance, so the step is scored as deciding the crash in part, not outright, on every setting. The L1 losses it did not save had little speed-matching push to begin with; avoidance of craft on the other level is the other carrier there. Four of eight registered predictions held. Two inert test switches were added to the simulator (unset, every anchor identical). This is a counterfactual inside the model, not a statement about real aircraft. Nothing is adopted; the default decision remains open.
research/city/design/M-CITY-ALIGN-REPLAY-record.md
Same saves, but who pays? Follow-up and correction (2026-10-13, CITY cycle 40, provisional). The previous note was re-tested on 48 new layouts: 1,540 wall losses re-flown five ways, every other craft flying live, each re-flight run to the end of the five-minute run. Taking the speed-matching push away from the lost craft alone from 10 seconds before its hit again kept about half of them off the wall for the next 10 seconds on the worst assumed radio tier (48 and 51 per cent) and nearly all on the better tier (95 and 99 per cent). Starting 30 seconds before the hit saved the same craft (within 1 point on every setting). On the worst tier, losses into a wall across the end of the street looked easier to save than side-wall losses only because those craft carried about three times the push. About 8 in 10 of the other craft lost in a saved re-flight had been reacting to the saved craft, mostly from the other level: exposure, not a proven cause. Two readings in the previous note are corrected. The extra cost to other craft on the better tier belongs to the way that cycle scored the removal (at the end of the steering calculation): removing the term earlier in the calculation cost less within the short window. And the worst-tier losses it does not save were not shown to be avoidance of craft on the other level; that is untested. Over the full run, both ways of removing the term were followed by about 0.1 to 0.5 more other craft lost per saved craft. The consistency check registered in the previous cycle at one part in a billion could never have been met by its rounded inputs; it actually compared at 0.002. Three of five registered predictions held. No simulator change; nothing is adopted; the default decision remains open.
research/city/design/M-CITY-ALIGN-FACE-record.md
What else pushes them? Follow-up (2026-10-13, CITY cycle 41, provisional). The previous note left the worst-tier losses that removing speed-matching does not save untested. On 48 fresh layouts the lost craft’s own sideways dodge away from craft on the other level turned out to carry most of them: removing that dodge too saved 93 to 94 per cent of those losses, and with both moves removed 96 to 97 per cent of all worst-tier wall losses were saved, against 14 to 18 per cent for a same-size nudge along the lane. The saved craft passed about 10 m closer to other-level craft without a contact in the next 10 seconds. The speed-matching result itself held on three of four settings (53 against 48 per cent on the crossing, just outside the registered range). Every craft knew its height exactly; this is not a rule to ignore the other level. Nothing is adopted.
research/city/design/M-CITY-L1-AVOID-record.md
What this does not establish
Simulation only, on assumed radio tiers, an assumed error process, assumed contact spheres, an assumed plus or minus 3 m band and an assumed margin with no real-world comparator. One controller, one class per lane run, calm air apart from the scripted descent, no wind and no mixed classes on the lanes. The mechanism (that the lost craft are the speed-matched ones) is consistent with the result but was not instrumented. Throughput is read only from passes and completed routes. The step’s review of the previous cycle was contested between the two reviewers (one found the earlier step flawed, the other sound); the realism reviewer’s six findings were closed. The proposed map sentence awaits the next review. The record discloses a failed first contract close on a line-count rule (a report reformatted with equal content), a chart caption fixed after the rows, and a pause between the end of the run and the pull of its rows.
The next question
Nothing is adopted; the term stays on by default. The default is now an open question put to the programme’s human direction with the saving and the price side by side: switch the term off, use a capped version proposed by the LINK track, or keep it as it is.
Source references for reproduction
Optional repository references, pinned to this snapshot. Repository access may be required.