The loop's own clock

Finding [S2]. Runs 2 and 3 (October 17, 2026; 1,295 runs under 1.1.0). The actor polices rejections (appeals, $a\approx0.98$); the requester polices approvals (disputes). With appeals alone a careful reviewer's 5% false rejections are punished almost every time while a stamper's negligence is caught about 3% of the time, and stampers end richer (+200 per actor at 50/50, $X=8$); with the requester disputing one bad result in four, careful ends richer by 489; break-even is a dispute rate of about 0.07, and careful review pays only above an accuracy of about 92% at that rate. At 10/90 every panel is stamper-majority: disputes fail (23 of 135 won), only the careful minority's errors are adjudicated, their $\rho$ ends below the stampers', selection hands the stampers the awards (159 to 66), and no dispute rate recovers it. At $G^V=20$ about 7% of duty slot-ticks are held by actors who cannot fund the bond and the reserve starves.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1158

Mechanism. Every signal the loop produced was panel-produced, so under capture the loop was blind to its own state; the design assumption that humans do not intervene removes the one external oracle the baseline leaned on. The orchestrator therefore owns some truth itself: the truth of the work it writes. Planted work (§7.6) gives a measurement no panel can paint; the plant record (§11.2) turns it into eligibility a captured majority cannot protect; the dial (§12.2) turns it into effort that rises with the threat and eases without it, and into the reserve's size and the security tax.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1160

Run 3's reading was that three measured inputs decided the careful–careless contest under 1.1.0 — whether the party who knows the truth challenges, how often the careful reviewer errs, and who sits on the panel — and that the third was the rule's: a stamper-majority reserve produces every signal the loop has, reliability included, so no remedy that leans on ρ can cure it. The design discussion the same day added the assumption the baseline had left implicit — humans do not intervene — which removes the one external oracle the loop leaned on.

Source: HANDOFF.md l.603–608

The revision gives the orchestrator truth of its own: planted work whose fulfillment it knows by construction (§7.6), a plant record per actor adjudicated by that truth and not by the panel, in two eligibility tiers (§11.2), and an effort dial that reads the reserve's bad-plant approval rate once a period and moves the plant rate, the audit odds, the reserve's size and the service fee with it — "the block-difficulty adjustment with negligence in place of block time" (§12.2). A reserve that runs short of eligible actors runs short and says so; the loop does not seat the negligent to make a number.

Source: HANDOFF.md l.608–614

It adds a mechanism; it does not alter any settlement, verdict or challenge rule of 1.1.0. It changes what participants may do (the orchestrator creates plants; an actor can be suspended from duty; no challenge is admitted on a plant), what they may see (§14), and makes two fixed quantities endogenous ($X$, $\omega$).

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1162

Reviews. Three release candidates, each read on its own (spec/design_revisions/dr2_1.2.0-rc.*-review.md). rc.1: the plant actor was named in a view that carried the actor (§14) — the actor leaves the view;

Source: HANDOFF.md l.616–618

rc.2: clearing the record on return would have readmitted every stamper with a clean slate for ~50 reviews and kept panels captured at ~70% strength — the record persists;

Source: HANDOFF.md l.622–624

Found at implementation — flagged for confirmation. With the record persisting, "at the end of the suspension it is readmitted only if its record passes, and otherwise remains suspended for a further T_susp" makes every suspension permanent: the record is a trailing window of plant outcomes and a suspended actor judges no plants, so a record that failed cannot pass later. That leaves T_susp inert and contradicts the revision's §2.4 and its §8 run 2. The engine implements the smallest reading that gives T_susp its meaning: readmission on probation — to duty, not to panels — until the actor's next plant re-checks the record, with a still-failing record suspended again at once (§11.2 of the consolidated text, Appendix G, World._rotate).

Source: HANDOFF.md l.628–634

Confirmed as the rule in force by patch 1.2.1 (October 17, 2026; the sentence of design revision 3 rc.1's Appendix A), applied to spec/agentic_economics_loop_spec_v1_2_1.md; 1.2.0 to history.

Source: HANDOFF.md l.638–639

agentic_sim/results/studies/1.2.0-clock/ (and 1.2.0-clock.zip beside it): drift/ (60 runs), adaptive/ (40), tipping/ (120), redteam/ (80) — 300 runs, 20 seeds per cell, 20 actors, every final audit passed; engine 544ff19 (2708f64 adds only test records). Appeals truthful and no disputes throughout — the design assumption's regime: the actor polices rejections, nobody polices approvals but the loop itself. The §6 reproduction is test_v12. "FA" is false approvals that vested; 1.1.0 cells are the control (--rules 1.1.0).

Source: HANDOFF.md l.665–669

Drift (run 2 of the revision's §8): the honest regime holds, at a capacity price. Stampers rise from 50% to 90% in steps of 5% every 100 ticks over 1,500 ticks, 4,500 tasks posted.

cell	FA	vested	stamper − careful (se)	stampers richer	m at end	alarms	eligibility shortfall (actor-ticks)	reviews waiting (review-ticks)	plants unfunded
1.2.0, dial on	83	2,518	−2,213 (49)	0 of 20	4.0	10.6	2,515	119,620	351 of 918
1.2.0, dial off (plants at π₀, m = 1)	257	3,213	−827 (69)	0 of 20	1.0	0	33	52,528	0
1.1.0 control	650	4,209	+233 (186)	12 of 20	—	—	—	2,283	—

Source: HANDOFF.md l.671–678

Under 1.1.0 false approvals grow with the stamper share (80 at tick 300, 650 at 1,500) and the stampers end richer in 12 of 20. With plants alone (dial off) they fall to 257; with the dial they fall to 83, and they stop growing once the dial saturates: 66 at tick 600, 83 at 1,500.

Source: HANDOFF.md l.680–682

The price is capacity: on duty falls from 6 at tick 300 to about 3 from tick 900, against an X_t of 5–8; the eligibility shortfall accumulates 2,515 actor-ticks; a submitted case waits for a reviewer 119,620 review-ticks in all (47× the control); vesting ends at 2,518 cases against 4,209; Ω cannot fund 351 of the 918 plants the rate called for at m = 4 (Ω net still +1,603, the raised ω). Hypothesis 2 holds as stated — "honest and slow rather than captured and fast" — and the slow is measured.

Source: HANDOFF.md l.685–689

cell	stamper share at end	stampers richer	FA	m	alarms
1.2.0	0.29	6 of 20	59	3.2	4.2
1.1.0	0.43	9 of 20	249	—	—

Both populations move toward careful under truthful appeals; 1.2.0 moves them further and faster. One adaptation rule, one seed of starting shares: a direction, not an equilibrium.

Source: HANDOFF.md l.694–700

Adaptive dispositions (run 3). Actors start 50/50; every 100 ticks each actor takes the richer disposition by mean public wealth with probability 0.2.

Source: HANDOFF.md l.691–692

split (stampers)	1.2.0: stamper − careful (se), richer	1.2.0 FA	1.1.0: stamper − careful (se), richer	1.1.0 FA	1.2.0 reviews careful / stamper	1.2.0 awards careful / stamper	1.2.0 delivery income careful / stamper	1.2.0 fees careful / stamper
50% (red team, no camouflage)	−2 (51), 11 of 20	53	—	—	502 / 78	99 / 200	—	—
60%	+215 (34), 19 of 20	67	+22 (108), 9 of 20	306	636 / 82	64 / 207	1,063 / 3,311	2,106 / 277
70%	−10 (25), 8 of 20	89	+43 (77), 9 of 20	379	787 / 91	37 / 195	654 / 3,142	2,612 / 309
80%	−902 (25), 0 of 20	121	+223 (138), 13 of 20	442	799 / 100	33 / 143	626 / 2,373	2,915 / 354

Source: HANDOFF.md l.704–709

Honesty holds at every split: false approvals fall four- to five-fold against 1.1.0, the stampers' â is 0.76–0.90 and they are suspended within their first handful of plants (run 1's "seat" channel works). But the money contest at 50–70% is not what Hypothesis 1 predicted. The careful actors are the only eligible reviewers, the dial sizes the reserve up (X_t 7–8), and they are therefore on duty almost all the time — 500 to 800 reviews each against 80–100 for a stamper — while the suspended stampers, free of duty, take the awards (200 to 64 at 60/40) and the delivery income (3,311 to 1,063). The fee f = 2 per review does not replace the award B = 20 per delivery: at 60/40 the stampers end richer in 19 of 20 seeds by 215 per actor, under the rule whose purpose was to make stamping unprofitable. The ρ channel does not help, because selection is argmax ρ among the free actors, and the careful are not free. What decides the contest under 1.2.0 is therefore not the plant penalty (about 0.6 per review, as computed) but who carries the reviewing load, and the revision loads it onto the honest minority and pays them the ordinary fee while Ω collects the raised security tax (Ω net +5,898 at 60/40 against +3,106 under 1.1.0). Only at 80/20, where the stampers are too many for the market to absorb and the reserve is short most of the time, do they lose again (−902).

Source: HANDOFF.md l.711–723

Red team (run 5). A stamper that rejects 10% of its reviews at random: at 50/50 it buys time — plant failures per stamper 5.9 against 7.2, suspensions 45 against 54, alarms 2.1 against 4.3 — and turns the wash (−2) into +112 (se 74, 13 of 20), with false approvals still low (42) and false rejections up (26 against 18; its random rejections are appealed and 133 of 138 appeals won). At 90/10 it changes nothing that matters (−1,733 against −1,961; FA 154 against 165). Camouflage is a small edge, not a safety; the detection half of the red team has nothing to detect by construction (the view test).

Source: HANDOFF.md l.725–730

Cost (run 6). Ω's plant spend is 2,000–3,500 per 1,000 ticks at m ≈ 4 (the careful inspections' fees and expenses; a stamper's failed plant costs Ω nothing), against ω income that rises with m to about four times 1.1.0's; Ω net is positive in every cell (+900 to +5,900 per run) and lowest where the reserve is shortest (80/20, drift), because fewer awards mean less ω. The working capital binds before the income does: at m = 4 and π = 0.4 the rate calls for more plants than Ω's unlocked balance can carry at once — 17 to 663 unfunded per run, worst where vesting is slow and plant pools stay open longest.

Source: HANDOFF.md l.732–737

1. The probation reading (above) needs a decision. It is what makes F̂ stay high and the alarm fire every period in a captured market, and what gives the returning stamper its one plant before it is out again. The literal reading (permanent suspension) would make the clock quieter and the shortage larger. Neither is obviously right; the engine holds the first, and the alternative is one branch.
2. The reviewing load is the contest. Under 1.2.0 the honest minority is conscripted into the sized-up reserve and paid f = 2 a review while the stampers deliver at B = 20. Candidates, all design questions: the fee scaled with the dial as the tax is (f_t = f·m_t, funded from ω_t, which Ω now keeps); a reserve sized to the eligible pool rather than to demand; or the stampers' exclusion from duty extended to awards (a ρ or â floor for selection — which run 3 warned against for ρ, but â cannot be captured).
3. Ω's working capital, not its income, caps the clock at m_max; the float (300) carries one period at m = 1, and the rate at m = 4 is four times that with slower closures.

Source: HANDOFF.md l.742–752

$\Omega$ banked the raised tax (+5,898 at 60/40) levied for the effort the honest were supplying; and a plant tied up 45 credits of which 30 was $\Omega$ paying itself.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1180