The question
CITY 34 found that, with a 4 m own-height error, altitude assignment traded wall losses for contacts between craft, and suggested that a granted pair skips its avoidance. Does that excess replicate on new layouts, is it on granted pairs, and would withholding exactly those grants remove it?
The flight
The campaign flew the same simulated crossing on 48 fresh layouts with four registered questions and seven predictions (an eighth was withdrawn before any run). It compared ordinary handling, altitude assignment with an 8 s report validity, and a diagnostic oracle: the same rule with the grant withheld whenever the two craft’s true height gap was already inside the class standard, the refused pair falling back on ordinary handling. The oracle reads true heights that no craft or authority possesses; it locates the mechanism and is not a design. Every contact was traced to the code path of both craft 0.05 s before it. 72 blocks of 48 runs, 3,456 runs, no error rows. The oracle switch is off by default and reproduces every earlier result bit for bit when off; an identity check showed the flight state unchanged step by step with every grant sent down the new path.
What happened
The contact excess replicated: calm L1 153 contacts against 5 with ordinary handling (+148, t 9.5), calm L2 479 against 105 (+374, t 14.5); CITY 34 on other layouts had +182 and +452. 152 of 153 and 479 of 479 of those contacts were between two craft both holding a grant a twentieth of a second before contact, in grant runs lasting about 12 s with a heard gap near 19 m against a true gap near 7 m; in 152 of 152 and 470 of 479 cases both grants began with the true gap already inside the standard.
With those grants withheld, calm L1 fell to 5 contacts, equal to ordinary handling, and calm L2 to 121 (ordinary 105; the remaining +16, t 2.3, mostly air taxis). The descending case read the same way (L1 4 against 1 and 43; L2 22 against 16 and 84). The oracle still lost fewer craft than ordinary handling on every tier and condition (calm L1 119 fewer of 6,006; calm L2 543 fewer of 2,666; descending L1 277 fewer, L2 162 fewer), but less than the unguarded 8 s rule, whose calm-L2 reduction of wall losses from 2,456 to 3 was paid in 479 contacts. Six of seven predictions held; the miss was a loss saving on calm L1 that the plan expected to vanish and that stayed.
Mid-air contacts · calm · 4 m own-height error · six class-and-spacing cells × 48 layouts
Descriptive contact counts, not risk estimates. The oracle reads true heights no craft has. CITY cycle 35; proposed claim awaiting the cycle-36 review.
What we learned later
Amendment (2026-10-10, CITY cycle 36, provisional; two reviewer-contested candidates run under one contract). A guard that reads only the two craft’s own height estimates and an assumed declared error bound (4 m per craft, the error’s own standard deviation, not a worst case) was built into the simulator, off by default, and scored on 48 new layouts against ordinary handling, the unguarded 8 s rule and the cycle-35 oracle. On the calm crossing with the weakest link (L1) a 6 m allowance brought contacts from 188 to 10 (ordinary handling 6, oracle 5; the +4 is within noise) while still losing 107 fewer craft than ordinary handling, at the price of withholding 7.2 per cent of grant halves, of which about one in four was truly too close at that moment. On the better link (L2) and when a craft is pushed down, no allowance removed the contact excess while keeping the loss saving (L2 at 6 m: 185 contacts against 121, the oracle 146; 1,065 of the 8 s rule’s 1,549 fewer losses kept). The remaining contacts are pairs whose estimates lean apart while the craft are truly close, which no estimate-only rule can see. Nothing is adopted; the guard stays off. The proposed map sentence awaits the cycle-37 review. Alongside the guard, switching off speed-matching with the other height level (the alignment term LINK 29 identified, an existing switch) cut ordinary handling’s lost craft from 6,027 to 3,852 on L1 and from 2,761 to 270 on L2 (wall strikes 2,519 to 30) with contacts unchanged, and the 8 s rule’s from 5,820 to 3,626 on L1: LINK 29’s mechanism seen again, now with own-height error. On L2 the registered comparison (the 8 s rule with and without the switch) is empty by construction, because under that rule every cross-level pair is granted before the alignment term acts; the ordinary-handling comparison carries the result there. LINK 29’s caveats carry: one crossing geometry, band exits rose there, throughput was not measured, and nothing is adopted.
research/city/design/B-CITY-ASG-GUARD-record.md
What this does not establish
One crossing, one assumed error process (4 m standard deviation, independent between craft), calm air, one class per run, and an assumed contact geometry: a contact is an overlap of two padded safety spheres (0.6 to 10.4 m apart by class), not a demonstrated airframe impact. The oracle uses information a real craft does not have; a guard built on estimates will both miss truly close pairs and refuse safe ones, and the refused halves fall to handling that LINK 27 found can steer craft into walls (64 per cent of the withheld L1 halves reached the plain avoidance rule). The proposed map sentence awaits the cycle-36 review; CITY 34’s own sentence was disputed by all three cycle-35 reviewers and rewritten for that same review. The step’s contract close failed on a bookkeeping item (five harness scripts declared as deliverables had been written before the contract opened); the record discloses it, and the git order of the files holds.
The next question
Nothing is adopted; the oracle stays off. The next step (CITY 36) scored a guard that uses only the craft’s own estimates against both the unguarded rule and the oracle, together with an alignment question handed over by LINK 29 (see the later note).
Source references for reproduction
Optional repository references, pinned to this snapshot. Repository access may be required.