Version 1.1.02026-10-17Design revision 1

The verification deposit is two pots

Runs: Runs 2 and 3

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.

The checking money splits into two pots, so careful inspectors stop paying for their care.

What went wrong

One pot paid back costs and the rest was the fee, so care cut a careful inspector's fee.

Care paid back from the fee's own pota careful inspector earned 7.9 per review, a stamper 9.5

Under version 1.0.1, each task came with one pot of money for checking it. The pot first paid back the inspector's costs. What was left over became the inspector's fee. So every credit a careful inspector spent on checking was a credit of fee it lost.

The first run showed what this did. With half the inspectors stamping, the stampers ended richer in four of five repeat runs. A careful inspector earned 7.9 per review and a stamper 9.5. The gap was the careful inspector's own costs.

The first draft of the rulebook had agreed on a fixed fee for checking work. A later draft made the fee whatever was left over, but the rulebook's own reasoning still assumed the fixed fee.

Under v1.0.1 each case carried one verification deposit doing two jobs: it reimbursed the inspectors' capped expenses, and paid what remained as their fee. Every credit reimbursed was a credit of fee lost; in the record's words, "the accounting closed; the incentive inverted."

Strategy run 1 measured it. With half the inspectors on duty (the "reserve") stamping, stampers ended richer in four of five seeds. A careful inspector earned 7.9 per review and a stamper 9.5, a gap equal to the careful inspector's own expenses. 47% of truly failed work was paid in full, and 88% with nine stampers in ten.

The record traces the rule's history. v0.1 had agreed a capped expense reimbursement plus a fixed fee, "for review work, not for the answer". v0.3, repairing a shortfall the second review found, made the fee the deposit's remainder. The spec's §7.3 kept reasoning with the fixed fee.

Finding [S1]. Strategy run 1 (October 16, 2026): at the defaults, with half the reserve stamping, stampers ended richer in four of five seeds; 47% of truly failed work was paid in full, 88% with nine stampers in ten. The channel was the fee: a careful reviewer earned 7.9 per review and a stamper 9.5 — a gap equal to the careful reviewer's own expenses.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1138

Mechanism. Under 1.0.1 §4.3 the deposit was one pot doing two jobs: reimbursing capped expenses, and paying what remained as the fee. Every credit reimbursed was a credit of fee lost. The rule charged for diligence at the margin, while 1.0.1 §7.3 still reasoned with the fixed fee of v0.1 R16–R17 ("capped expense reimbursement plus a fixed fee … for review work, not for the answer"), which v0.3 had replaced by the remainder when it repaired the second review's shortfall finding. The accounting closed; the incentive inverted.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1140

What changed

One pot pays back costs up to a cap; the other pays a fixed fee set by role.

expense pot costs, up to a cap inspector

fee pot fixed fee by role inspector

what is left in both when the task and its sub-tasks close requester

Care within the cap costs the inspector nothing

Version 1.1.0 splits the pot in two. One pot pays back the inspector's costs. The other pays a fixed fee set by the inspector's role. Costs are paid back up to a cap, so care within that cap now costs the inspector nothing. When the task and all its sub-tasks are closed, whatever is left in both pots goes back to the person who asked for the work.

The change was reviewed twice before it was accepted. The reviews found mistakes in the first draft. One was the claim that "nothing else moves": in fact, in the common case, the money inspectors earn from these pots falls by about three quarters.

Design revision 1 was the first change to the rulebook since the freeze and the first use of the §19 design-revision path. The deposit becomes two pots: an expense allowance and a fee pot. Expenses are paid from the allowance alone. Each inspector whose verdict stood gets a fixed fee set by role from the fee pot, scaled down only when the pot is short, plus its share of emission, a separate payment to inspectors whose verdicts stood. When the whole delegation tree closes, the fee pot's remainder pays what the orchestrator advanced, and then both pots' remainders return to the requester. Care within the expense cap costs the inspector nothing.

The record's arithmetic for a stamper's money term per review turns negative: about −0.3 at run 1's effective catch rate of 3%, −0.7 at the nominal audit odds. An appeals term is added once actors appeal.

The revision went through two implementation reviews before acceptance as rc.3. rc.1 had computed "unowed fee money" per case, which in a tree would have paid a root's expenses with money its nested case's inspectors were still owed. rc.1 also said "nothing else moves", when inspector income from deposits falls by about three quarters in the common case. The tie band and the audit floor, also implicated by run 1, were left as defaults to be swept separately.

The first change to the rulebook since the freeze, and the first use of §19's design-revision path. The stamping result above was read against the text and the history: v0.1 R16–R17 had agreed "capped expense reimbursement plus a fixed fee … for review work, not for the answer"; v0.3, repairing the second review's shortfall finding, made the fee the deposit's remainder after reimbursements, and from then on every credit of care reimbursed was a credit of fee lost — while §7.3 still reasoned with the fixed fee. The revision was written as a release candidate (spec/design_revisions/), read on its own through two implementation reviews — rc.1 computed "unowed fee money" per case, which in a tree would have paid a root's expenses with money its nested case's reviewers were still owed (reviewer test r4_07 is the shape); rc.1 said "nothing else moves" when reviewer income from deposits falls by about three quarters in the common case; the hypothesis needed its appeals term; I10 and record mode gain the tree consequences; the rerun needs twenty seeds or more — and accepted as rc.3.

Source: HANDOFF.md l.506–516

One payout rule (1.1.0): expense debts for every review from the allowance alone → each survivor's fixed fee w_r·f (2 / 3 / 3) from the fee pot, scaled only when the pot is short, plus its emission share in full → at tree closure the pot's remainder pays the tree's Ω claims and advances, then both pots' remainders return to the requester. Care within K costs the reviewer nothing.

Source: HANDOFF.md l.61–64

Consequence. Within the cap, care costs the reviewer nothing: a careful reviewer that spends $c\le K$ is made whole and earns the same fee as one that spent nothing.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.357

Emission E_t = E_0·2^(−t/H) pays reviewers whose verdicts stood, by role weight; nothing carried forward.

Source: HANDOFF.md l.59–60

Under 1.1.0 the fee is fixed and care within K is made whole, so the money term per review is −p·b·(1−ε)·(f̄ + G^V): about −0.3 at run 1's effective catch rate (3%), −0.7 at the nominal p = 0.5, b = 0.2, with an appeals term a·ε·(1−b)·(f̄ + G^V) once actors appeal.

Source: HANDOFF.md l.99–101

The selection tie band and the audit floor, which the run also implicated, were left as defaults, to be swept separately.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1142

What the runs found

Careful inspectors now ended richer; appeals then favoured stampers, and disputes favoured careful inspectors.

The new rule in run 1's settingcareful inspectors richer by 173 per actor

Truthful appeals, eight on dutystampers richer by 136 per actor

Every false approval disputed, eight on dutycareful inspectors richer by 1,331 per actor

Careful against stamper wealth, rerun of run 1 under 1.1.0

credits

carefulstamper

Per actor, means over 20 seeds; 20 actors, 1,000 ticks. X is the number on duty, f the fee. Source: HANDOFF.md l.263–274

Appeals expose careful inspectors, disputes expose stampers

credits

Per actor, means over 20 seeds; truthful policies. Positive means stampers richer. ± is the standard error over the seeds. Source: HANDOFF.md l.463–470

Stamper minus careful wealth by dispute rate (appeals on, band 0)

credits

50/50, X 450/50, X 810/90, X 8

Per actor; positive means stampers richer. The 10/90, X 4 column is in the source table. ± is the standard error over the seeds. Source: HANDOFF.md l.555–562

The first run was repeated under the new rule, with twenty repeat runs for each setting. The result flipped. In the same setting as before, careful inspectors now ended richer, by 173 per actor, and stampers were richer in only 6 of 20 runs. With eight inspectors on duty instead of four, so that no task waited for one, careful inspectors were ahead by 290 and stampers were richer in 1 of 20.

Fees per review were now about equal for both kinds. The difference came from the careful ones winning more tasks to do, and from stampers losing more of their bonds.

When nine of ten inspectors were stampers, nothing separated the two kinds. The rule change stopped charging careful inspectors, but it did not rescue an economy where almost every inspector is careless.

Then the runs added challenges. A worker whose work was rejected could appeal. When workers appealed only work that was truly good, 98% of wrongly rejected work was appealed and overturned, so a careful inspector's mistakes were almost always punished. A stamper's false approvals were caught 3% of the time. With eight on duty, appeals alone made stamping win: stampers ended richer by 136 per actor, in 15 of 20 runs. Where most inspectors were stampers and every rejection was appealed, the appeal became a weapon against the careful inspectors.

The person who asked for the work could also dispute an approval they knew was wrong. This exposed the stampers instead. When that person disputed every false approval and eight inspectors were on duty to hear the disputes, careful inspectors ended richer by 1,331 per actor, with no run going the other way. Each dispute takes two reviews where an audit took one. With only four on duty, disputes made tasks wait, and both kinds ended poorer than with no disputes at all. Where most inspectors were stampers, the disputes were heard by stampers and mostly failed.

A later check found the program had chosen inspectors by a random draw, where the rules say to take turns. With turns, every main result moved by less than its margin of error, and none changed direction.

With appeals and disputes together, the outcome turned on how often the person who asked for the work disputed. The break-even rate was about 0.07. It also turned on how often careful inspectors made mistakes: with one false approval in four disputed, a careful error rate of 10% left the stampers richer in 18 of 20 runs.

The rerun repeated run 1's declaration, changing only the rule: 160 runs, 20 actors, 1,000 ticks, 20 seeds per cell, with four or eight inspectors on duty, two fee levels, and 50/50 and 10/90 splits of careful inspectors and stampers. In run 1's cell the sign flipped. Careful inspectors ended richer by 173 per actor (standard error 73), and stampers were richer in 6 of 20 seeds. With eight on duty, so no case waited for an inspector, the gap was 290 (se 39), 1 of 20. Doubling the fee moved both groups up by about 400 and left the gap where it was. Fees per review were equal, 4.04 against 4.02. The gap came from delivery income and bonds: careful actors won more awards, and stampers forfeited three times the bonds. At 10/90 the difference was within one standard error of zero in every cell.

A sweep of defaults found the tie band to be "the knob that carries the reliability channel". Selection favours reliability, and duty credit only breaks ties; the band sets how close two reliability scores count as tied. Careful and stamper reliability differed by only 0.02, so a band of 0.05 already let most stampers through. With no band, careful actors won 165 awards to the stampers' 135 at 50/50, a gap of 530 per actor. At a band of 0.2 the gap vanished. No default rescued the 10/90 economy.

Appeals run against careful inspectors. When actors appealed truthfully, 98% of false rejections were appealed and overturned, so a careful inspector's 5% error was paid for in full, while a stamper's false approvals were caught 3% of the time. With four on duty the 50/50 gap shrank to −140 (se 120). A control with eight on duty showed it was a reversal: +136 (se 39), stampers richer in 15 of 20. With nine stampers in ten and every rejection appealed, the appeal panel was drawn from stampers. 147 of 155 true rejections were overturned, and stampers ended richer by 717 in 19 of 20 seeds.

Requester disputes run the other way. A requester who disputed every false approval at 50/50 won 72% of them and cut false approvals from 284 to 83. With eight on duty, careful inspectors led by 1,331 per actor (se 35), with none of twenty seeds the other way. With nine stampers in ten, the requester won 78 of 458 disputes, and the accuracy of procedurally established labels fell to 0.59. A dispute uses two reviews where an audit used one. With four on duty and truthful or half disputes, cases waited 7,100 to 8,300 review-ticks, and both populations ended poorer than with no disputes. The record's summary: adjudication "finds the error of whoever the truth-holder chooses to challenge, at the rate the truth-holder chooses to challenge."

Run 3 first corrected a conformance gap. The earlier sweeps assigned inspectors by a seeded random choice, where §7.5 says round-robin. Re-run under round-robin, every headline cell moved by less than its standard error, and no direction changed. With appeals and disputes combined, at no tie band and eight on duty, appeals alone gave stampers +200. Disputing one false approval in four turned it to −489, and the break-even dispute rate at 50/50 was about 0.07. The careful inspector's own error rate decided too: at a dispute rate of 0.25, a careful error rate of 10% put the stampers at +312, richer in 18 of 20.

drivers/sweep.py: 8 cells × 20 seeds = 160 runs of drivers.stamping, 20 actors × 1,000 ticks, 3 tasks per tick at β = 20, b = 0.2, careful accuracy 0.95 and cost 2–3, p_start 0.5, G^V 5, ruling accuracy 0.8 — run 1's declaration unchanged — at X ∈ {4, 8} on duty, f ∈ {2, 4} with φ^F = 2.5 f, splits 50/50 and 10/90.

Source: HANDOFF.md l.254–257

Per actor, means ± sd over 20 seeds; "richer" counts seeds:

cell careful wealth stamper wealth stamper − careful (se) stampers richer false approvals / 3,000 reviews unassignable ticks

50/50, f 2, X 4 — run 1's cell 3,275 ± 156 3,102 ± 180 −173 (73) 6 of 20 284 1,046

50/50, f 2, X 8 3,320 ± 87 3,031 ± 94 −290 (39) 1 of 20 284 0

50/50, f 4, X 4 3,670 ± 176 3,491 ± 205 −179 (83) 9 of 20 284 1,057

50/50, f 4, X 8 3,700 ± 87 3,427 ± 94 −273 (39) 1 of 20 283 0

10/90, f 2, X 4 3,687 ± 339 3,626 ± 62 −61 (89) 8 of 20 522 946

10/90, f 2, X 8 3,578 ± 254 3,644 ± 30 +66 (62) 11 of 20 533 0

10/90, f 4, X 4 4,112 ± 396 4,020 ± 68 −92 (103) 9 of 20 522 946

10/90, f 4, X 8 3,952 ± 263 4,038 ± 31 +85 (65) 11 of 20 534 0

cell	careful wealth	stamper wealth	stamper − careful (se)	stampers richer	false approvals / 3,000	reviews unassignable ticks
50/50, f 2, X 4 — run 1's cell	3,275 ± 156	3,102 ± 180	−173 (73)	6 of 20	284	1,046
50/50, f 2, X 8	3,320 ± 87	3,031 ± 94	−290 (39)	1 of 20	284	0
50/50, f 4, X 4	3,670 ± 176	3,491 ± 205	−179 (83)	9 of 20	284	1,057
50/50, f 4, X 8	3,700 ± 87	3,427 ± 94	−273 (39)	1 of 20	283	0
10/90, f 2, X 4	3,687 ± 339	3,626 ± 62	−61 (89)	8 of 20	522	946
10/90, f 2, X 8	3,578 ± 254	3,644 ± 30	+66 (62)	11 of 20	533	0
10/90, f 4, X 4	4,112 ± 396	4,020 ± 68	−92 (103)	9 of 20	522	946
10/90, f 4, X 8	3,952 ± 263	4,038 ± 31	+85 (65)	11 of 20	534	0

Source: HANDOFF.md l.263–274

The sign flipped in run 1's cell. Under v1.0.1 stampers ended richer by 249 per actor (4 of 5 seeds); under 1.1.0, same declaration, same seeds among the twenty, careful reviewers end richer by 173 (se 73), stampers richer in 6 of 20. With the capacity ceiling removed (X = 8: no tick in which a submitted case waited for a reviewer) the gap is 290 (se 39), stampers richer in 1 of 20. The fee level does not change the gap (f = 4 moves both populations up by about 400 and leaves the difference where it was), which is what a fee that is fixed by role should do.

Source: HANDOFF.md l.276–281

The channel moved as the revision said it would. Fees per review are now equal: 4.04 careful against 4.02 stamper at f = 2 (run 1: 7.9 against 9.5). The whole gap is delivery income and bonds — careful actors win more awards (154.7 against 145.3 per actor; +193 in delivery) and stampers forfeit three times the bonds (28.2 against 9.9) with 3.0% of their reviews overturned against 1.1%.

Source: HANDOFF.md l.283–286

At 10/90 nothing separates the two. With eighteen stampers the reserve is stamper-on-stamper, audits agree, 522–534 of about 600 truly failed tasks are paid in full, and label accuracy is 0.82–0.83 under both grades; the two careful actors' wealth varies by ±340 across seeds and the difference is within one standard error of zero in every 10/90 cell. The rule change does not rescue an economy whose reviewers are almost all careless; it stops charging the ones who are not.

Source: HANDOFF.md l.289–293

The tie band is the knob that carries the reliability channel. With no band, selection is argmax ρ and careful actors win 165 awards to the stampers' 135 at 50/50: a gap of 530 per actor, three times the default's. At band 0.2 the κ tiebreak hands most awards to whoever has duty credit and the gap vanishes.

Source: HANDOFF.md l.331–333

The ρ separation itself is only 0.02 (0.93 against 0.91), so a band of 0.05 already lets most stampers through; the deterrent the rulebook relies on (§7.3, §11) works exactly as far as selection is allowed to see ρ.

Source: HANDOFF.md l.334–336

ρ reliability (trailing 100 vested outcomes, flat prior) feeds selection; κ duty credit is a tiebreak only.

Source: HANDOFF.md l.78

What this says for the next runs. Within the declared regime no default rescues the 10/90 economy: the reserve is stamper-on-stamper and no knob changes who audits whom.

Source: HANDOFF.md l.350–351

with the actor's response to an N now declared by --appeal-policy: truthful — appeal an N on work that is truly fulfilled, accept otherwise; always — appeal every N.

Source: HANDOFF.md l.357–359

a is about 1, and the §7.3 condition then fails at the effective catch rate. With truthful appeals, 98% of false rejections are appealed and overturned. The careful reviewer's own 5% error on the 80% of work that is good is therefore paid in full — its bond, its fee, and the court fee — while a stamper's false approvals are caught 3% of the time.

Source: HANDOFF.md l.374–377

The money gap at 50/50 shrinks from −173 to −140 and is no longer distinguishable from zero (se 120); at 10/90 it turns against the careful reviewers (+179, 13 of 20), whose ρ falls below the stampers' (0.96 against 0.97) because their errors are the ones that get adjudicated.

Source: HANDOFF.md l.378–381

Frivolous appeals turn the appeal into a weapon in a stamper-majority reserve. With every N appealed at 10/90, the appeal panel is drawn from stampers and answers F: 147 of 155 true rejections are overturned, each careful actor is overturned 96 times and forfeits 483 in bonds, ends with ρ 0.89 against 0.98, and the stampers end richer by 717 in 19 of 20 seeds.

Source: HANDOFF.md l.383–386

with the human's response to a pending F declared by --dispute-policy: truthful — H disputes every pending F on work that truly failed; half — the same with probability 0.5; always — H disputes every pending F.

Source: HANDOFF.md l.402–404

The dispute does to stampers what the appeal did to careful reviewers. A requester who knows the truth and disputes every false approval at 50/50 wins 72% of them, overturns 17% of the stampers' reviews (33 per actor, bonds forfeited 167 against 33), cuts false approvals from 284 to 83, and widens the careful lead to 304 per actor (2 of 20 seeds the other way).

Source: HANDOFF.md l.421–424

Together the two runs say the rulebook's adjudication finds the error of whoever the truth-holder chooses to challenge, at the rate the truth-holder chooses to challenge.

Source: HANDOFF.md l.427–428

At 10/90 the dispute panel is the stampers, and the requester loses. H files 458 disputes and wins 78: the two dispute reviewers are drawn from a reserve that is nine-tenths stampers and say F, so 82% of true disputes fail, false approvals stay at 376, and the procedural grade's accuracy falls to 0.59 — the challenge path is now the least reliable label in the economy.

Source: HANDOFF.md l.430–433

Every dispute costs the reserve more than the audit it replaces. A dispute is two reviews where an audit was one, and with four on duty the economy pays for it in capacity: a submitted case waits for a reviewer in 7,100–8,300 review-ticks under truthful or half disputes, awards go unfilled in about 970 ticks, vesting falls 6–17%, and both populations end poorer than with no disputes at all — the careful reviewers by 500 per actor at 50/50.

Source: HANDOFF.md l.436–440

cell careful stamper stamper − careful (se) richer stamper overturned % false approvals reviews unassignable ticks

50/50, appeals, X 4 (above) 3,269 3,129 −140 (120) 8 of 20 3.4 248 2,981

50/50, appeals, X 8 3,194 3,330 +136 (39) 15 of 20 2.7 276 0

10/90, appeals, X 8 3,291 3,693 +403 (85) 16 of 20 0.6 529 0

50/50, disputes, X 4 (above) 2,748 2,444 −304 (50) 2 of 20 16.9 83 7,794

50/50, disputes, X 8 3,595 2,264 −1,331 (35) 0 of 20 14.2 131 0

10/90, disputes, X 8 3,661 3,545 −116 (80) 9 of 20 3.8 455 1

cell	careful	stamper	stamper − careful (se)	richer	stamper overturned %	false approvals	reviews unassignable ticks
50/50, appeals, X 4 (above)	3,269	3,129	−140 (120)	8 of 20	3.4	248	2,981
50/50, appeals, X 8	3,194	3,330	+136 (39)	15 of 20	2.7	276	0
10/90, appeals, X 8	3,291	3,693	+403 (85)	16 of 20	0.6	529	0
50/50, disputes, X 4 (above)	2,748	2,444	−304 (50)	2 of 20	16.9	83	7,794
50/50, disputes, X 8	3,595	2,264	−1,331 (35)	0 of 20	14.2	131	0
10/90, disputes, X 8	3,661	3,545	−116 (80)	9 of 20	3.8	455	1

Source: HANDOFF.md l.463–470

Capacity was hiding the size of both challenge effects, not their direction — and in the appeals case it was hiding a reversal. With eight on duty no case waits for a reviewer, and the appeals cell that read as "indistinguishable from zero" at X = 4 reads +136 (se 39), stampers richer in 15 of 20: truthful appeals alone, with the fee penalty gone and the reserve unconstrained, hand the contest to the stampers at 50/50.

Source: HANDOFF.md l.475–479

The dispute cell moves the other way by the same release: −1,331 (se 35), none of twenty seeds the other way — a requester who disputes every false approval, with a reserve that can hear the disputes, takes 17% of the stampers' reviews and most of their awards (112 against 188).

Source: HANDOFF.md l.481–484

Conformance first: assignment is round-robin (§7.5), and the earlier sweeps used a seeded choice.

Source: HANDOFF.md l.536

Every cell moved by less than its standard error and no direction changed, so the five earlier sweeps stand as declared, with this note.

Source: HANDOFF.md l.550–551

The combined regime (combined/): truthful appeals and truthful disputes at rate ∈ {0, 0.25, 0.5}, band 0, X ∈ {4, 8}, both splits. Stamper − careful wealth per actor (se), stampers richer of 20:

dispute rate 50/50, X 4 50/50, X 8 10/90, X 4 10/90, X 8

0 (appeals only) +2 (87), 10 +200 (52), 16 +1,336 (176), 19 +1,794 (155), 20

0.25 −77 (56), 8 −489 (70), 1 +145 (168), 7 +1,274 (171), 19

0.5 −144 (36), 3 −919 (77), 0 −141 (61), 7 +1,270 (201), 18

dispute rate	50/50, X 4	50/50, X 8	10/90, X 4	10/90, X 8
0 (appeals only)	+2 (87), 10	+200 (52), 16	+1,336 (176), 19	+1,794 (155), 20
0.25	−77 (56), 8	−489 (70), 1	+145 (168), 7	+1,274 (171), 19
0.5	−144 (36), 3	−919 (77), 0	−141 (61), 7	+1,270 (201), 18

Source: HANDOFF.md l.555–562

The requester's attention is the swing variable. With the reserve unconstrained and selection pure argmax ρ, appeals alone give the stampers +200; a requester who disputes one false approval in four turns it to −489 and one in two to −919 — the break-even dispute rate at 50/50 is about 0.07.

Source: HANDOFF.md l.564–566

The ε sweep (epsilon/): the careful reviewer's error rate at dispute rate 0.25, band 0. 50/50, X 8: ε 0.02 → −930 (58); 0.05 → −489 (70); 0.10 → +312 (63), stampers richer in 18 of 20.

Source: HANDOFF.md l.573–574

Conclusion

Who does better now turns on who challenges, how often careful inspectors err, and who hears challenges.

Whether the one who knows the truth challengesHow often careful inspectors errWho sits on the panel

No dispute rate in the runs recovered a pool of mostly stamper inspectors

Version 1.1.0 removed the fee rule that charged careful inspectors for their care. After that, three measured things decided who did better: whether the party who knows the truth challenges, how often careful inspectors make mistakes, and who sits on the panel that hears challenges. The first two come from the people using the system. The third comes from the rules, and when most inspectors are stampers, no rate of disputes in the runs recovered it.

With the fee penalty gone, the record names three measured things that decide the contest between careful and careless review: whether the party who knows the truth challenges (break-even dispute rate about 0.07 at 50/50), how often the careful inspector errs (break-even error rate about 0.08 at a dispute rate of 0.25), and who sits on the panel. A stamper-majority reserve was not recovered by any dispute rate. The first two are inputs from the product; the third is the rule. Under 1.1.0 at the defaults, when actors appeal truthfully and the reserve is not the bottleneck, stamping wins.

What this leaves for the revision 2 proposal. With the fee penalty gone (1.1.0), the contest between careful and careless review is decided by three things the runs have now measured: whether the party who knows the truth challenges (break-even dispute rate ≈ 0.07 at 50/50), how often the careful reviewer errs (break-even ε ≈ 0.08 at dispute rate 0.25), and who sits on the panel (a stamper-majority reserve is not recovered by any dispute rate). The first two are the product's inputs; the third is the rule.

Source: HANDOFF.md l.591–595

What changes in the reading. The appeals finding is stronger than the X = 4 run showed: under 1.1.0 at the defaults, when actors appeal truthfully and the reserve is not the bottleneck, stamping wins.

Source: HANDOFF.md l.492–494