Version 1.6.02026-10-17Design revision 6

Auto-scale

Runs: Run 8

Show

No replayThere is no recorded replay of this version; 2.7.0 is the one version with one. The results below are the study's own tables.

Five switchable rules: a late-request rebate, the orchestrator backing its fund, and stricter panel seats.

What went wrong

Too few inspection seats: checks waited, about 300 requests stood in line, and requests never expire.

Requests never expireabout 300 waited, while twelve inspectors on probation sat idle

Under version 1.5.0 the limit was not money but inspection seats. Thousands of checks waited with nobody to do them, and about 300 requests stood in line. Twelve inspectors on probation sat idle while checks waited. Requests never expire, so the line could keep growing. And bad work still slipped through, because an inspector's record took time to show what it was. Baruchi said a request left on hold hurts the network's name. He asked that it go back in line with priority, and that the orchestrator take some of the blame in money. He also said that if the project controls the network and the servers, the orchestrator should be allowed to scale up.

Run 7, under version 1.5.0, found that the reviewing seat, not the treasury, capped the market: reviews that found no reviewer ran to 7,300–11,400 case-ticks a run, the open-case queue stood around 300, and twelve probationers sat idle while reviews waited. It also found that the reserve counted a queue that never expires, that the one-period reset saw-toothed, and that the probation leak was the detection lag on the auditor. Baruchi's reading was that a request on hold is a problem for the network's rating: re-queue it with priority, have the orchestrator take some of the blame in money, and, if ELEF controls the network and the servers, allow auto-scale.

Run 7's finding was that the reviewing seat, not the treasury, caps the market — reviews that found no reviewer 7,300–11,400 case-ticks a run, the open-case queue ~300, vesting pinned near the posting-rate ceiling whoever delivered, twelve probationers idle while reviews waited — that the demand stock counted a never-expiring queue, that the one-period reset saw-toothed, and that the probation leak was the plant record's detection lag on the auditor (55–75% at fixed splits, 73–80% under drift). Baruchi's reading: a request on hold is a problem for the network's rating — re-queue it with priority and have the orchestrator take some of the blame in money; and if ELEF controls the network and the servers, allow auto-scale.

Source: HANDOFF.md l.1204–1210

What changed

Five rules, each switchable: a lateness rebate, orchestrator backing, idle seating, panel history, two-period reserve.

request waiting past 8 steps rebate each step, capped requester

orchestrator backs the fund, after a reserve for planted cases operating account orchestrator's inspector

A panel seat needs enough history on planted cases

Letting requests expire: it turns customers away

Version 1.6.0 added five rules, and each could be switched off, so the runs could tell which rule did what. First, a clock: once a request has waited more than 8 steps, the orchestrator owes the requester money for every extra step, up to a cap. Second, the orchestrator backs the fund for its own inspectors from its own money, after keeping aside what its test cases need. Third, idle inspectors on probation are seated while checks are waiting. Fourth, only an inspector with enough history on the test cases may sit on a panel. Fifth, the reserve is measured over two periods again. Letting requests expire was turned down, because it turns customers away.

Design revision 6 added five rules, each behind a switch. The service clock times every request from posting to its first verdict; past 8 ticks the orchestrator owes the requester a rebate each tick, capped, paid at once or carried as the orchestrator's liability. Auto-scale removes the old parameter A_T and has the orchestrator back the troops' fund from its own balance above a reserve for plants (cases whose correct answer the orchestrator knows by construction), so that the honesty clock is never starved. Seat the queue puts idle probationers on duty while reviews wait. A panel seat now requires a determined record: a minimum number of bad plants seen. And the reserve is measured over two periods again. Letting requests expire was rejected: it bounds the queue by turning customers away.

Mechanism. Five rules, each behind a flag. The service clock (§3.1, §6.1): every root request is timed from posting to its first verdict or its final, oldest first, and the orchestrator owes the requester $r_s$ a tick past $T_s$ , capped at $B$ , from $\Omega$ or as $\Omega$ 's liability — the queue is bounded in liability instead of unbounded in count, and the orchestrator has a cost for not scaling. Auto-scale (§10.4, §12.3): $A_T$ removed, $\Omega$ backs the fund above a plant reserve so the honesty clock is never starved, in a stated order with liabilities older than a period first; review seats are summoned before deliveries; a delivery is summoned only when its own capital and two review seats can be carried. Seat the queue (§12.1): the reserve target is at least the reviews waiting, by kind, so idle probationers are seated while reviews wait. A panel seat requires a determined record (§7.1), dormant until the world's first plant is adjudicated, so the lagging auditor is gone at a fixed split. The reserve over two periods (§12.3).

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1250–1250

past T_s = 8 the orchestrator owes the requester r_s = ω₀B a tick, capped at B, from Ω when the clock stops or as Ω's liability; a request withdrawn before award is owed nothing; the clock runs only while β ≥ B̄.

Source: HANDOFF.md l.1220–1222

Rule. The orchestrator may introduce into the review stream a plant: a case whose submission's fulfillment is known to the orchestrator by construction. A bad plant is known not fulfilled; a good plant is known fulfilled.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.383–383

A panel seat requires a determined record (≥ n_min bad plants seen), dormant until the world's first plant is adjudicated — so in every run the first period's panels are the state's, and §8 reads from the second period on. R_T over two periods (the 1.4.0 reading restored).

Source: HANDOFF.md l.1227–1229

Recorded and rejected. Requests expire (the design reviewer's run-7 proposal) — bounded the queue by turning customers away; replaced by the clock and the rebate.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1254–1254

What the runs found

False approvals fell 72 to 82%, but with 50 to 60% stampers the orchestrator ran out of money.

The panel rule alonefalse approvals cut by 70%

All five at 50% to 60% stampersthe orchestrator owed 20,000 to 26,000 with about 90 left

Requesters withdraw requests after 12 stepsdebt fell from 19,800 to 830

False approvals, 1.6.0 against 1.5.0

false approvals

1.6.01.5.0

All five rules on; 1.5.0 from run 7. 20 seeds per cell. Source: HANDOFF.md l.1291–1296

Open cases, one rule at a time, 60/40

open cases

60% stampers, 20 seeds per cell. Source: HANDOFF.md l.1270–1279

The orchestrator's liability by duty cap

credits

The half cap is the 1.6.0 default; two thirds and all were swept. Source: HANDOFF.md l.1312–1319

The team tried each rule alone first. The panel rule was the honesty gain: on its own it cut approved bad work by about 70 percent. The orchestrator's backing was the capacity gain: the line of requests halved. The clock alone bankrupted the orchestrator. Seating the idle inspectors alone made things worse. All five together kept the honesty of the panel rule and the shorter line, with the clock's debt on top. With all five rules, approved bad work fell by 72 to 82 percent and stampers ended 45 percent poorer. But when 50 to 60 percent of the inspectors were stampers, many requests were late and the orchestrator ran out of money. It owed 20,000 to 26,000 and had about 90 left. Test cases went unpaid, 80 to 150 a run, where that number was meant to be zero. Letting two thirds of the inspectors be on duty, instead of half, fixed much of this at 60 percent stampers, with honesty unchanged. More starting money did not help, because the orchestrator's income was the limit. When requesters pulled back requests that had waited 12 steps, the orchestrator's debt fell from 19,800 to 830, at the cost of 11 percent of the work.

Run 8 was 840 runs, 20 seeds per cell, 20 actors, every final audit passed. One rule at a time at 60% stampers: the panel rule was the honesty gain, cutting false approvals by 70%; the orchestrator's backing was the capacity gain, halving the queue and cutting refusals by two thirds; the clock alone bankrupted the orchestrator (liabilities of 13,800 against a balance of 57); seat the queue alone hurt; the two-period reserve was a wash.

With all five rules, the lagging-auditor route was zero at every split, false approvals fell 72–82%, and the stampers ended 45% poorer. But Hypothesis 2 came out inverted: the on-time share was 0.9 at 80% stampers, while at 50–60% the on-time share was 0.4–0.6, mean lateness 38–48 ticks, and the orchestrator was insolvent, owing 20,000–26,000 against a balance near 90, with 80–150 plants unfunded a run where the reading must be zero. Raising the members' duty cap from half to two thirds at 60% stampers halved the queue, lifted the on-time share from 0.6 to 0.9 and cut the liability by four fifths, with honesty unchanged and careful wealth within 3%. A larger starting balance for the orchestrator (1,000 against 300) did not help: the orchestrator's income, about 2 a tick, was the limit. When requesters withdrew unawarded requests after 12 ticks, the liability fell from 19,800 to 830 at the cost of 11% of vesting.

840 runs, 20 seeds per cell, 20 actors, every final audit passed, engine 195896a.

Source: HANDOFF.md l.1263–1264

cell	vested	open cases	reviews unassigned (case-ticks)	troop deliveries	award summonses refused	FA (probationers')	members' awards	Ω at the end	Ω's liability	plants unfunded
1.5.0 control	2,186	298	11,423	545	11,211	47 (8.2)	153	1,004	—	34
clock alone	2,151	320	6,777	514	8,184	46 (7.8)	151	57	13,810	136
Ω's backing alone	2,284	152	8,874	691	4,054	47 (7.3)	141	1,134	—	9
seat the queue alone	2,058	455	6,843	637	36,373	71 (6.4)	161	971	—	0
panel rule alone	2,016	433	17,251	631	19,642	14 (5.0)	176	1,150	—	80
two-period reserve alone	2,189	288	11,082	550	9,082	51 (8.2)	151	1,064	—	28
all five, literal horizon	2,259	92	7,760	1,010	12,330	12 (3.4)	141	754	24,839	22
all five (tipping 60/40)	2,269	94	7,746	999	13,049	13 (3.6)	151	94	19,778	81

Source: HANDOFF.md l.1270–1279

The panel rule is the honesty gain: alone it cuts false approvals by 70% and the probationers' by 40%, with the queue longer because undetermined members cannot fill panel seats and the fund-limited troops are slow to. Ω's backing alone is the capacity gain: the queue halves, refusals fall by two thirds, troop deliveries rise a quarter, Ω stays above 1,000. The clock alone bankrupts Ω: liabilities 13,800 against a balance of 57, plants unfunded quadrupled. Seat the queue alone hurts: more actors on duty, fewer free to deliver, three times the refusals, false approvals up by half. The two-period reserve alone is a wash. Together, the five give the honesty of the panel rule and the queue of the backing, and the clock's liability on top.

Source: HANDOFF.md l.1281–1287

stampers	vested	open cases	reviews unassigned	on-time share	mean lateness	Ω's liability	Ω at the end (min)	plants unfunded	troop deliveries	FA (probationers')	lagging-auditor kind	members' awards	stamper wealth
50%	2,178 (2,175)	208 (313)	6,238 (11,431)	0.40	48	26,100	87 (41)	151	671 (305)	12 (3.5) [42 (6.4)]	0 [3.6]	162 (163)	714 (1,417)
60%	2,269 (2,186)	94 (298)	7,746 (11,423)	0.60	38	19,778	94 (48)	81	999 (545)	13 (3.6) [47 (8.2)]	0 [4.3]	151 (153)	770 (1,389)
70%	2,293 (2,347)	44 (55)	5,976 (7,581)	0.80	21	5,587	413 (347)	14	1,302 (1,031)	16 (3.5) [66 (9.8)]	0 [6.7]	118 (121)	863 (1,347)
80%	2,282 (2,375)	40 (41)	5,875 (7,322)	0.90	21	5,422	590 (550)	13	1,590 (1,331)	17 (6.5) [96 (11.8)]	0 [7.7]	52 (74)	911 (1,394)

Source: HANDOFF.md l.1291–1296

Hypothesis 4 holds in full: the lagging-auditor kind is zero at every split (run 7: 3.6–9.1 of 6–12 upheld), and false approvals fall by 72–82%. The stampers end 45% poorer than under 1.5.0. Hypothesis 1 holds in part: the queue falls by a third to two thirds at 50–60% and reviews unassigned by 30–45% — not to a few hundred; vesting is at the ceiling as predicted. Hypothesis 2 is inverted: the on-time share is 0.9 at 80/20 and 0.4–0.6 at 50–60%, with mean lateness 38–48 ticks and Ω insolvent — liabilities 20,000–26,000 against a balance near 90, plants unfunded 80–150 a run (the must-be-zero reading fails).

Source: HANDOFF.md l.1298–1303

cell	open cases	reviews unassigned	on-time	lateness	Ω's liability	Ω at the end	troop deliveries	members' awards / delivery income / fees	careful wealth	FA
60/40 half	94	7,746	0.60	38	19,778	94	999	151 / 1,755 / 2,919	4,956	13
60/40 two thirds	44	3,424	0.90	8	3,512	47	1,282	111 / 1,755 / 3,336	4,808	13
60/40 all	50	3,335	0.80	13	6,809	30	1,382	94 / 1,498 / 3,356	4,614	13
80/20 half	40	5,875	0.90	21	5,422	590	1,590	52 / 683 / 6,323	6,945	17
80/20 two thirds	40	5,360	0.90	7	1,409	1,135	1,551	43 / 684 / 6,392	6,881	17
80/20 all	39	5,520	0.90	8	1,335	957	1,768	40 / 640 / 6,675	7,128	17

Source: HANDOFF.md l.1312–1319

Hypothesis 6 holds, and the two-thirds cap is the answer to run 7's seat: at 60/40 it halves the queue and reviews unassigned, lifts the on-time share from 0.6 to 0.9, cuts mean lateness from 38 to 8 ticks and Ω's liability by four fifths, with honesty unchanged and the careful members' wealth within 3% — their income moves from awards to fees, as 1.5.0's restated hypothesis said.

Source: HANDOFF.md l.1321–1324

Float and target. At 60/40, Ω₀ = 1,000 does not help: liabilities 18,000–19,400 against 19,800–22,400 at 300, Ω ends near 100–190 at both, plants unfunded 72–97 at both; T_s moves nothing.

Source: HANDOFF.md l.1327–1328

The float is not the limit at 50–60%: Ω's income (ω₀B a member award, about 2 a tick) against plant costs and rebates is, and no float changes a flow.

Source: HANDOFF.md l.1329–1331

The late withdrawer (H withdraws unawarded requests after 12 ticks): gains nothing, as expected — and Ω's liability falls from 19,800 to 830, mean lateness from 38 to 3.5 ticks, the queue from 94 to 43, at the cost of 11% of vesting (362 requests withdrawn).

Source: HANDOFF.md l.1356–1358

Conclusion

The rebate debt sits in old untaken requests; the orchestrator's income rule comes before every 1.6.0 rule.

Candidate: re-queue or withdraw a request nobody took after some time

Planted cases go unpaid because the orchestrator is poor

The clock's debt is a pile the orchestrator cannot pay down, and it lives in old requests nobody took. The record's candidate is a rule that re-queues or withdraws such a request after some time. Letting two thirds of the inspectors be on duty is a default for Baruchi to decide. Test cases go unpaid because the orchestrator is poor, and the rule for its income comes before every rule of 1.6.0.

The record leaves four items in order. The clock's liability is a stock the orchestrator cannot pay down, and it lives in the stale unawarded request; a rule that re-queues or withdraws such a request after some time is the candidate. The two-thirds duty cap fixes the seat at 60% stampers, a default and Baruchi's product decision. The plant reserve's horizon is for the reviewer to rule on. And plants go unfunded for want of money in the orchestrator's hands: its income rule is upstream of every 1.6.0 rule.

The clock's liability is a stock the orchestrator cannot pay down, and it lives in the stale unawarded request. Ω's income is about 2 a tick; a late request accrues 1 a tick up to 20; at 50–60% the liability reaches 20,000–26,000 and no float, target or window changes it, while the withdrawer's cell shows what bounds it: a request that waits unawarded past 12 ticks is the lateness. The rc rejected expiry for turning customers away; the data say the queue that never expires is also what the clock prices. A rule that re-queues or withdraws an unawarded request after some T, with the clock's stop at that point, is the candidate — the design reviewer's and Baruchi's "re-queue with priority" read literally.

The members' duty cap at two thirds fixes the seat at 60/40 with honesty unchanged and careful wealth within 3%; a default, and Baruchi's product decision with the numbers beside it. The reserve-sizing rule alone made things worse.

The plant reserve's horizon — the implementer's reading (a plant's life) is in force; the text's literal horizon leaves Ω nothing to advance at Ω₀ = 300 (the ofat/ literal cell shows the same honesty, a shorter queue, and the fund carrying the troops on its own earnings). The reviewer should rule.

Plants unfunded for want of Ω are not zero at 50–60% (80–150 a run) and under drift (20–84): the honesty clock is being starved by Ω's poverty, not by the advances — the reserve cannot hold back money Ω does not have. Ω's income rule (ω₀B per award) is upstream of every 1.6.0 rule.

Source: HANDOFF.md l.1363–1377