In run 10, the market with 50 percent stampers recovered, from this correction alone.
The line of open requests fell from 609 to 38, requests were on time from the second period on, the debt was zero, and every test case was paid in every seed.
The orchestrator's inspectors were now sent steadily, not in bursts.
Approved bad work did not change beyond noise, and careful inspectors' wealth stayed within 2 percent.
Two things were flagged.
A request that free members do not take within two steps now goes to the orchestrator's inspectors, so members got fewer jobs and the orchestrator's inspectors did 46 to 67 percent of the work.
And no request was late, so the rules for late requests were never used in these runs: not tested, not wrong.
Run 11 tested sudden rushes of requests.
A rush of nine requests a step for 50 or 100 steps strained nothing: no request was late, and the orchestrator's inspectors grew in number and shrank back on their own.
Every seed was back to its quiet level 110 steps after the rush ended.
A harder rush, fifteen a step for 100 steps, did strain the system: requests waited, and the debt reached 16,607 and was still 7,351 at the end of the run.
What limits the system is money: about 2,200 carries nine deliveries a step, not fifteen.
Past that edge the rules worked as written and nothing broke, but requesters waited 20 to 40 steps with nothing said.
A longer waiting window, 3 or 5 steps instead of 2, moved almost nothing.
Run 10 (80 runs, 50–80% stampers, 20 seeds each) found that 50% stampers clears by the summons correction alone: the queue fell from 609 to 38, the on-time share from the second period on rose from 0.25 to 1.00, the liability went from 20,643 to zero, and plants were funded in every seed. Summonses were steady rather than bursty. False approvals were unchanged within noise and careful wealth stayed within 2% at every split. Two readings were flagged, with no rule moved: the award window now binds, so the members' awards fell and the orchestrator delivered 46–67% of what vested; and the lateness clock was dormant, never exercised in the study, untested rather than wrong.
Run 11 (460 runs) asked for riots. The asked spike, nine requests a tick for 50 or 100 ticks at 60% stampers, bound nothing: the orchestrator's cash, about 2,200, plus the troop account carried every summons, and the auto-scale rose and retracted on its own, every seed back at the quiet level 110 ticks after the spike, because the reserve's two-period memory is the recovery time. A hard spike, fifteen a tick for 100 ticks, added to find the edge, did bind: the liability accrued to 16,607 and stood at 7,351 at tick 1,000, with one seed in twenty cleared. The record's reading: what binds is capital; past the edge the clock and the orchestrator's order of payment work exactly as written and nothing breaks, but requesters waited 20–40 ticks with nothing said. Moving the award window to 3 or 5 ticks changed almost nothing; the members' share is their capacity.
80 runs — 50–80% stampers × 20 seeds, 20 actors, 1,000 ticks, logged every 20 ticks, every final audit passed, engine 8d9b866
Source: HANDOFF.md l.1545–1546
| stampers |
queue |
on-time (2+) |
lateness |
Ω's liability |
rebates paid |
plants unfunded |
troop deliveries |
refused for funds |
R_T end / mean / max |
Ω's income: surplus / fee |
Ω at the end |
FA (probationers') |
lag kind |
| 50% |
38 (609) |
1.00 (0.25) |
0 (90) |
0 (20,643) |
0 (1,782) |
0 (138) |
1,400 (500) |
0 (23,497) |
448 / 403 / 485 (223 / 400 / 1,129) |
4,639 / 2,996 (2,743 / 2,418) |
4,306 (59) |
11 (1.5) [11 (3.1)] |
0 |
| 60% |
42 (42) |
1.00 (0.89) |
0 (6.8) |
0 (317) |
0 (1,256) |
0 (0.7) |
1,753 (1,719) |
0 (607) |
568 / 503 / 599 (680 / 754 / 1,587) |
5,624 / 2,996 (6,029 / 2,993) |
5,406 (4,352) |
14 (3.1) [14 (3.1)] |
0 (0.05) |
| 70% |
41 (41) |
1.00 (0.97) |
0 (4.0) |
0 (0) |
0 (295) |
0 (0.1) |
1,831 (1,859) |
0 (95) |
607 / 559 / 638 (655 / 671 / 1,096) |
7,100 / 2,996 (7,191 / 2,996) |
7,856 (7,967) |
15 (4.3) [14 (3.5)] |
0 |
| 80% |
42 (40) |
1.00 (0.95) |
0 (5.0) |
0 (226) |
0 (387) |
0 (3.6) |
2,054 (2,132) |
0 (1,642) |
689 / 638 / 714 (733 / 740 / 1,060) |
8,231 / 2,997 (8,528 / 2,996) |
9,883 (9,215) |
18 (5.6) [16 (4.2)] |
0 |
Source: agentic_sim/results/studies/1.7.1-summons-notes-for-reviewer.md l.11–16
50/50 clears, by the summons alone (run 9 in brackets): queue 38 (609), on time 1.00 from the second period on (0.25), lateness 0 (90), Ω's liability 0 (20,643), rebates paid 0 (1,782), plants unfunded 0 in every seed (138), troop deliveries 1,400 (500), summonses refused for funds 0 (23,497), Ω at the end 4,306 (59), the surplus 4,639 beside the base fee 2,996 (2,743 / 2,418).
Source: HANDOFF.md l.1549–1552
The summons is steady: 34 / 43 / 45 / 50 deliveries summoned per 20-tick window at 50–80%, max within 20% of the mean, windows with none 3.8% / 1.5% / 0.9% / 0.2% (run 9's 50/50: 62% of windows, bursts to 60); requests past w_A at the seating never above 3.
Source: HANDOFF.md l.1554–1557
False approvals unchanged within noise, 11 / 14 / 15 / 18 (11 / 14 / 14 / 16), the lagging auditor zero everywhere; careful wealth within 2% at every split (4,111 / 4,302 / 4,638 / 5,226 against 4,056 / 4,250 / 4,659 / 5,190), stampers five to seven times poorer.
Source: HANDOFF.md l.1558–1561
Two readings flagged, no rule moved. w_A now binds in fact: a request the free members have not taken within two ticks is the state's that tick, so the members' awards fall (162 → 129 at 50/50, 43 → 40 at 80/20, 8.5 from the second period on there) and the state delivers 46% / 57% / 60% / 67% of what vests — the market's share is now the default w_A's choice, at the cost of the lateness a longer window admits; the state's share is visible in "Ω's income by source" (the surplus 1.5 to 2.7 times the base fee). The clock is dormant: lateness zero at every split (awarded or summoned by w_A = 2, the fastest troop path 5 against T_s = 8), so the rebate, the liability and its order were exercised in no run of this study; untested, not wrong.
Source: HANDOFF.md l.1563–1569
riots/ (160), wa/ (240), adaptive/ (20), drift/ (40) — 460 runs, every final audit passed.
Source: HANDOFF.md l.1582–1583
Riots. The asked spike (nine a tick for 50 or 100 ticks from tick 400, 60/40) binds nothing: Ω's cash (about 2,200) plus the troop account carry every summons, advances of 9,400–10,300 in the spike window all repaid inside it, no request late, no rebate, no liability, no plant unfunded; the auto-scale alone rises and retracts — troops in the field 22 → 75–98 → 22, R_T 529 → 2,152–2,613 → 540 two periods after the spike — and every seed is back at the quiet level 110 ticks after the spike ends, 1.8–1.9 periods: R_T's two-period memory is the recovery time.
Source: HANDOFF.md l.1585–1589
The hard spike (mine: fifteen a tick for 100 ticks, two and a half times the asked, added to find the edge) is where it binds in a quiet market: the summons refused for funds from the first spike window (38,834 request-ticks in all), queue to 659, on-time 0.00 in the spike, lateness to 37, the liability accrued to 16,607 by tick 560, Ω at 77, plants unfunded 11, R_T 4,905 and back to 554 two periods after; the flows are back 160 ticks after the spike ends (2.4 periods); the stock is not — paid at Ω's whole income above the plant reserve, 220–330 a window, it stands at 7,351 at tick 1,000 and would clear about 25 periods after the spike, Ω holding only the plant reserve the whole time; one seed in twenty cleared. What binds is capital: about 2,200 at 60/40 carries nine deliveries a tick in flight and not fifteen; past the edge the clock and Ω's order work exactly as written and nothing breaks, but a 100-tick spike at 2.5× the quiet capacity leaves 25 periods of debt and requesters who waited 20–40 ticks with nothing said.
Source: HANDOFF.md l.1596–1604
| cell (60/40 unless stated) |
binds? |
on-time (2+) |
lateness |
liability peak → end |
rebates / liabilities paid |
refusals for funds (awards / seats) |
plants unfunded |
R_T quiet → peak → end |
troop deliveries a window quiet → peak |
Ω min → end |
FA |
| spike 9 × 50 |
no |
1.00 |
0 |
0 |
0 / 0 |
0 / 0 |
0 |
529 → 2,152 → 548 |
108 → 358 → 111 |
1,482 → 7,425 |
13 |
| spike 9 × 100 |
no |
1.00 |
0 |
0 |
0 / 0 |
0 / 0 |
0 |
529 → 2,613 → 544 |
108 → 418 → 110 |
1,482 → 10,004 |
15 |
| flood 6 a tick sustained |
ramp |
0.96 (0.03 in the first 100 ticks) |
14 (17 in the ramp) |
7,415 → 0 at tick ~750 |
211 / 9,275 |
13,496 / 986 (1,339 / 50 after the first period) |
0.45 |
1,607 → 2,385 → 1,343 |
— → 522 steady |
81 → 14,927 |
16 |
| hard spike 15 × 100 (mine) |
yes |
0.63 |
24 |
16,607 → 7,351 at tick 1,000 |
386 / 13,920 |
38,834 / 26 |
11 |
529 → 4,905 → 560 |
108 → 467 → 113 |
77 → 272 |
14 |
Source: agentic_sim/results/studies/1.7.1-riots-notes-for-reviewer.md l.7–16
w_A ∈ {2, 3, 5} at 50–80%. The window moves almost nothing: members' awards within 1–6% (129 / 130 / 130 at 50/50; 40 / 40 / 41 at 80/20), the state's share within 2% (46% → 45%; 67% → 69% → 68%), on-time 1.00 everywhere (w_A = 5 plus three troop ticks is T_s = 8), Ω's surplus against its fee unchanged, the queue up by the waiting ticks (38 → 47). The members' share is their capacity, not the window's artefact — a longer window leaves a request waiting for a member who is not coming.
Source: HANDOFF.md l.1606–1609