From the first draft to 2.9.0

Activity	Question	Appropriate action
Rulebook design	What is allowed, owed, observed, and final?	Change only for an explicit design decision or unresolved rule contradiction
Implementation conformance	Does the engine perform those rules, including refusals and timing?	Patch the engine and add a regression without introducing a different permitted behavior
Strategy research	Which allowed choices work well for participants or the system?	Separate investigation against a fixed rulebook and specified settings

Source: reviews/R09_of_v0_10.md l.15–19

Four textual corrections from the tenth review [R10]. Each clarifies a decision already made; none changes a rule, a default, a permitted operation, or the engine.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1107

| E1 | §1, §15.2, closing | "Only a human rules," while §8.4 has the orchestrator rule on a split; §1 credited every root requester, delegates included, with non-model judgment | The orchestrator rules on procedural disagreements (splits); only a human may issue the override; a delegate's judgment is a model's and is graded as such (§15.3) | The two senses of "ruling" contradicted each other read literally; the label grades already kept the distinction |

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1111

Test suite across versions: v0.7 104 + 1 xfail → v0.8 123 + 1 → v0.9 148 + 1 → v0.10 152 + 24 → v1.0.1 156 + 24

Source: HANDOFF.md l.43–44

The reviewers' test bundles themselves are not in this repo — they arrived as separate archives and were copied into agentic_sim/tests/test_review*.py at the time; the reports name the tests they shipped, so names and counts can be compared, but byte-for-byte verbatimness cannot be re-verified from this repo alone.

Source: reviews/README.md l.20–22

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

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

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

Finding [S4]. Run 4 (October 17, 2026; 300 runs under 1.2.0, truthful appeals and no disputes throughout). Honesty held everywhere: false approvals four- to five-fold below 1.1.0 at every split from 50/50 to 80/20, and under a stamper share drifting to 90% they stopped growing once the dial saturated (83 against a control of 650). It was slow: a case waited for a reviewer 47 times longer than in the control, vesting ended 40% lower, and $\Omega$ could not fund a third of the plants the dial called for at $m_{\max}$. And suspension from duty was a reward: at 60/40 the eight careful actors were the only eligible reviewers, the dial sized the reserve to 7–8, they were on duty almost always — 500 to 800 reviews each at a fee of 2 — while the twelve suspended stampers, freed from duty, took the awards at 20 (delivery income 3,311 against 1,063 per actor; stampers richer in 19 of 20 seeds). Selection could not help, because it runs among free actors and the honest were not free. $\Omega$ 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

The third use of §19's path, written the same day from run 4's finding and Baruchi's answer to it: "if you get thrown out of inspection you are banned from receiving a new task; to get tasks back you have to prove you are doing good inspection for a while."

Source: HANDOFF.md l.759–761

Suspension now suspends market membership: no new award while suspended or on probation; the held commitment is finished by the ordinary rules. After T_susp the actor is a probationer — initial reviews only, audit odds pinned at 1, re-suspended by any plant failure, readmitted when its one trailing record passes θ_duty with n_min new plant outcomes judged (a full stamper: four rejected bad plants, about fifty honest reviews; its panel seats return only once its approvals roll out of the window). At most half the eligible members are on duty; probationers fill the initial seats first, members hold the audit and challenge seats. The dial's raised tax goes to a duty fund paid to the settling reviews by role weight as an uplift, instead of to Ω. Plant capital counts the pots only (15, not 45). Newcomers' odds are pinned at 1 until their bad-plant record is determined (the 1.2.0 taper was passed by a third of stampers at once). A duty levy — keep winning awards against a cash levy — was considered and rejected (Appendix B). The slowness run 4 measured is deliberately not treated here; Baruchi has a direction for it, reserved.

Source: HANDOFF.md l.761–771

Mechanism. The contest was no longer about detection but about who carries the reviewing load and what membership of the market requires. Suspension now suspends membership: no new award until the actor has inspected well, under full audit, for long enough to pass. Probation is the supervised way back; the members' duty cap keeps half the honest free to work while probationers take the initial seats; the dial's raised tax goes to the reviewers through the duty fund; plant capital counts only the pots.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1182

It changes the meaning of an existing state (suspension) and adds a cap and a fund; it alters no settlement, verdict or challenge rule.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1184

Smaller: probation pays a stamper a little (+1.2 a review) and the award is what deters;

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1188

Found at implementation — two readings, flagged for confirmation (§12.1, Appendix H). (1) The members' cap ⌈E_t/2⌉ is 3 at six actors — below the X₀ = 4 of Appendix A's initialization and §20's worked example, so the baseline's own worlds, and every reviewer test built on them, would violate the revision at tick 0 (the first suite run failed on "C is not on verification duty" across the board). The engine lets the cap never fall below the reserve's floor V_max + 1, as the panel floor yields to the cap; the difference from the literal text is confined to memberships of four to seven, where the literal cap would leave two or three members free. One line (World._member_cap). (2) An outgoing probationer whose turn ends with no free probationer stays, rather than hand its seat to a member whom the next tick's "probationers first" would displace back. One branch of the rotation. Both are the implementer's; the alternative to each is one edit. Confirmed as patch 1.3.1 (October 17, 2026; the sentences of design revision 4 rc.1's Appendix A), applied to spec/agentic_economics_loop_spec_v1_4_0.md; 1.3.0 to history.

Source: HANDOFF.md l.783–791

Finding [S5]. Run 5 (October 17, 2026; 360 runs under 1.3.0). The gate holds everywhere: stampers 2,100–2,500 per actor poorer at every split from 50/50 to 80/20, richer in 0 of 20 seeds; an adapting population reaches ≤ 10% stampers in all 20 seeds by tick ~800 and vesting then recovers above 1.2.0's. A fixed captured population is honest and short-staffed: vesting 20–42% below 1.2.0, awards unfilled for want of a free member in 816–899 of 1,000 ticks — at 60/40 three tasks a tick of two ticks each need six free actors and eight members with four on duty leave four. The fund is fair pay, not an accelerator (convergence 770 ticks without it, 865 with). Probation safety binds on member availability, not on the cap: at $m_{\max}=2$, 65 of 97 false approvals were probationers' whose pinned audits found no panel-eligible member free, and the approvals reached customers; the drivers had dropped such rolls since run 1 (647 audits assigned of 2,666 rolled).

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1202

The fourth use of §19's path, the first to add a participant, and the remedy revision 3 reserved. Run 5's two shortages — the audit nobody was free to take when a probationer approved, the award nobody was free to accept when a customer posted — are one shortage: the market's only source of capacity was its own members.

Source: HANDOFF.md l.891–893

Baruchi's answer: the orchestrator is like the government; it has troops — ELEF's own servers, honest by construction, not paid as reviewers, used as few as possible and retracted as early as possible. The revision: a rolled audit is an obligation of the case (the window does not close until an audit completes; the drivers had dropped a roll that found no auditor since run 1 — a conformance correction as much as a rule); a required seat unfilled for w_R ticks, or a root request unawarded for w_A with no free fundable member, is taken by a troop the engine summons when the duty fund can carry its cost and bond, with Ω's advance up to A_T; the troop's verdict and delivery are the driver's declared disposition (truth w.p. 1 − ε_T, failure at b); the fund is the troop's wallet and carries troops first — costs and locks at summons, the advance repaid, then uplift from above the troops' reserve R_T; a troop claims no emission or uplift, never touches a plant, never audits a troop reviewer's approval of a troop's delivery (that seat waits for a member and is counted), accepts every verdict on its own work at the verdict (the external reviewer's note at acceptance), and is counted in public: troop-ticks, troop shortfall, cases held for the market's audit. Patch 1.3.1 (the two run-5 readings confirmed) was applied first.

Source: HANDOFF.md l.893–906

Mechanism. Both costs — the audit nobody was free to take and the award nobody was free to accept — are one shortage: the market's only source of capacity is its own members. The orchestrator gets servers of its own, used last (a window the market has in full), used once (one seat, released at completion), paid from the tax the dial already levies (the fund as their wallet, with a first charge on it), never preferred to a member, never on the state's own work as auditor, and counted in public (troop-ticks, troop shortfall, cases held for the market's audit). A rolled audit becomes an obligation of the case, which alone closes most of the leak.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1204

A design revision motivated by two strategy findings, labelled and read on its own; the first to add a participant. It alters no settlement, verdict or challenge rule — a troop's seat settles by the ordinary rules with the fund as its wallet.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1206

The external reviewer: "Clean. Nothing to change in the text" — one note, acceptance at the verdict — "accept rc.2, implement."

Source: HANDOFF.md l.913–914

false approvals 35, of them 6 by troops and 6 by probationers; the fund paid 10,456 in uplift — and refused 8,739 award summonses and 2,686 seat summonses for want of funds: the uplift drains the fund down to R_T every batch, and R_T is the peak capital of troops that were fundable, so the reserve cannot grow past what it has already carried. The treasury, not the market's need, sets the state's share. One seed, no conclusion — but the first thing run 6 will measure.

Source: HANDOFF.md l.937–941

Run 6 showed the troops doing what revision 4 said and the fund starving them by its own rule — R_T, the peak capital troops had carried, was what every batch drained the fund to, so the peak could never rise, and the fund refused 4,800–8,800 award summonses a run while paying 6,400–12,700 in uplift; A_T was not the constraint.

Source: HANDOFF.md l.1055–1058

Where the text was wrong: the reserve was self-limiting by the accepted rule — $R_T$ as the peak capital troops did carry, drained to by every batch's uplift, could never rise, so the fund refused 4,800–8,800 award summonses a run while paying 6,400–12,700 in uplift (10,000–22,000 refused at drift), $A_T$ not binding (300 against 100 cut refusals and moved no delivery); the treasury, not the market's need, set the state's share.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1224–1224

The rule (§12.3, I23). At each seating the troop demand D_s is the capital live troops hold plus the cost of the summonses the fund could not carry at that seating — refused for funds, each case once per seating at the amount the carry check refused, a case's demand its unfilled seats (a panel short of both, two). R_T is the peak of D_s over the current period, starting at the period's first seating from the standing demand (every troop still in the field), the previous period dropped. The uplift is paid, as before, only from the unlocked balance above R_T. A peak holds: the stock falls as the backlog is served, the reserve does not until the period ends — one period's retention, the dial's own lag, as a feature.

Source: HANDOFF.md l.1066–1072

Mechanism. One rule: the reserve is sized to the attempted demand — capital held by live troops plus the cost of the summonses the fund refused for funds at that seating, each case once per seating at the amount the carry check refused, a case's demand its unfilled seats — as the peak over the current period, starting at the standing demand at the period's first seating. A refusal raises the reserve by what was refused; the fund keeps its inflows until it can carry the summonses that were made; since every troop seat is net-positive to the fund it grows into the need on its own; the peak holds until the period ends, and the money returns to the reviewers then.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1226–1226

Classification. A design revision motivated by a strategy finding against one rule of revision 4. It changes what the reserve measures and alters no settlement, verdict, challenge or membership rule — the smallest since the freeze.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1228–1228

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

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

The seventh use of §19's path, and the one Baruchi named: why should an inspector profit from something a troop did? from their perspective it's just like any other actor in the system. Run 8's finding (above) was that the panel rule is the honesty gain and Ω's backing the capacity gain, and that the clock's liability is a stock the orchestrator cannot pay down — the troops' net at 60/40, 6,200–6,800 a run, went to the market's reviewers as uplift by 1.4.0's clause that nothing a troop earns reaches Ω, while the same Ω carried the rebates on about 2 a tick and ended near zero at 50–60% with liabilities of 20,000–26,000 and plants unfunded.

Source: HANDOFF.md l.1384–1389

The money was in the wrong pocket: the troops' net at 60/40 was 6,200–6,800 a run (the fund's uplift paid plus its closing balance less the surcharge received), and 1.4.0 §2.4 sent it to the market's reviewers as uplift by the clause that nothing a troop earns reaches $\Omega$, while the same $\Omega$ carried the rebates for the lateness those troops were summoned to prevent. The two-thirds cap at 60/40 halved the queue (94 → 44), lifted on-time from 0.60 to 0.90, cut mean lateness 38 → 8 and the liability 19,778 → 3,512 with false approvals unchanged and careful wealth within 3% — and, a figure the run-8 notes had not reported, tripled plants unfunded (81 → 272), because the state delivered more from the same $\Omega$.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1272–1272

A troop is an actor of the state; what it earns is the state's, and the state's bills are paid from it. (1.7.0)

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.32–32

The rules (§4.1, §10.4, §12.1, §12.2, §12.3, I23). Two accounts again: the duty fund is 1.3.0's uplift fund — the raised tax in and nothing else, its whole balance out as uplift at every batch, no reserve, no lock; the troops' wallet is a separate troop account — Ω's advances and every troop's earnings in; costs, locks, forfeits, burns, repayments out; at each batch the unlocked balance above R_T flows to Ω. The clause withdrawn: a troop is an actor of the state and its earnings are the state's; the guards that stand in its place (members first at every seat, the windows, cost at summons, no emission/uplift/ρ/κ to a troop, the state's share public) plus one reading, Ω's income by source. R_T is capital held: the two-period peak of costs purchased plus locks, every troop in the field; the 1.5.0 refused component withdrawn (its reason went with A_T); its purpose now the troops' in-flight capacity shielded from the liability queue. Ω pays its debts from its income: the 1.6.0 order unchanged and funded; the plant reserve's horizon ruled a plant's life. The cap ⌈2E_t/3⌉. ω₀ swept {5, 10, 15}% with r_s pinned at 1.

Source: HANDOFF.md l.1399–1408

The engine read 1.4.0 §2.3 per tick: one free fundable member blocked every award summons that tick. At 50/50, 79% of ticks had a queue past $w_A$ — 289 requests on average — with two members free and no summons; the summonses came in bursts when none happened to be free (the first at tick 288 on average, 38 at once; p90 23 a burst), between bursts the queue grew about a request a tick, and the queue's age was the lateness and the liability.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1286–1286

The same per-tick test explained run 8's inverted Hypothesis 2 (at 80/20 the four members are never free, so the state was never blocked).

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1286–1286

At 50/50 the market collapses — queue 609 (control 129), the state delivering 500 against 1,169, on-time 0.25, liability 20,643

Source: HANDOFF.md l.1461–1462

The design reviewer's ruling on run 9's addendum, relayed by Baruchi: "The rule's meaning is per request, not per tick. 1.4.0 §2.3: a root request that no eligible member has been awarded within w_A ticks is awarded to a troop. A free member takes the oldest request it can; the others, unawarded past their window with nobody offering for them, are the state's. So: at the seating step, requests that have waited w_A and exceed the free fundable members are summoned for, oldest first, after the members have taken theirs. 'Members first' is unchanged — a troop never takes an award a member offered for. Classification: a conformance correction, with the wording made unambiguous — 1.7.1. The text admitted both readings; the per-tick one contradicts the rule's stated purpose (the state fills what the market leaves). Not a revision; no other rule moves. T_susp stays at 100."

Source: HANDOFF.md l.1518–1525

Finding [S1–S11, C-2.0]. Every run was priced by fiat at $\beta=20$ from a bottomless wallet; the selection rule's price term never moved, §4.6's execution-cost operation was never called, and nobody delegated. So the credit was a score: the driver's backing stood at $10^7$ and $v_t$ moved from 1.0000 to 0.9995 in a thousand ticks. "The state is broke" (runs 8–9) was never the loop breaking but a cash rule imposed on $\Omega$ as a stand-in for servers; the two accounts, $R_T$, the plant reserve and the order of payment existed to ration that cash. Scarcity had no signal but a queue (run 11's hard spike: 20–40 ticks in line, unspoken to). Baruchi's frame: decouple the servers from the economy; let printing make money worth less; give the money something to balance against.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1298–1298

What it is. The first major version: a change of approach to money, on Baruchi's frame ("the servers needs to be decoupled from the economy, otherwise we contaminate it"; "maybe it is time we introduce floating prices"; "if I print more money then money is worth less"; "the actors have nothing to buy").

Source: HANDOFF.md l.1628–1630

2.0.0 lets the price float: requesters with finite wallets (issuance 10 a tick each, wallets 200) choose a ceiling and lock the all-in maximum at posting (a root's first lock; the nominals and the fee rate freeze there; I29); members bid; the price level P_t — the trailing mean of every awarded price, P_0 = 20 before the first — indexes every nominal the rules fix (f, G^V, D, φ, K, r_s, c_T, the plant pots) in place of v_0/v_t, frozen per case; effort costs stay nominal as the anchor; there is no backing (sources: issuance, emission, creation; sinks: the burns, retirement, departure); Ω is never broke — it creates for a payment the rules require and retires the deficit from every inflow, both public (I28) — and A_T, R_T, the plant reserve, Ω's order, the troop account and the liability stock are retired; the state delivers at cost (c^D_T = B̄) under a capacity N_T = 40 outside the ledger (a delivery holds two places until its inspection is seated; a refusal for capacity is recorded and told); B̄ counts members' awards only; a request below B̄ is flagged and its requester may raise the ceiling; delegation waits for 2.1.0 (it needs costs that differ across actors).

Source: HANDOFF.md l.1631–1639

What changed. Demand: requesters with wallets and a declared budget policy, funded by issuance; the pay-in locked at posting, a root's first lock; the ceiling the requester's, raised when flagged below market (§2.1, §4.1, §5.1). Supply: members bid at or above their reservation; the state at $\bar B$ from members' awards only, and at cost (§5.1, §7.7). Money: the price level $P_t$ indexes every nominal the rules fix, frozen per case; effort costs nominal as the anchor; no backing — sources issuance, emission, creation, sinks the burns, retirement, departure (§4.4–4.6). The state: never broke — creation for a required payment, retirement from every inflow, the deficit and $\Omega$'s net position public; $A_T$, $R_T$, the plant reserve, the order, the troop account and the liability stock retired (§10.4). Servers: $N_T$ outside the ledger, the capacity case recorded, two places per delivery (§12.3). Delegation deferred to 2.1.0 with heterogeneous costs.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1300–1300

The state at cost was below cost — $c^D_T=\bar B$ burned at the summons, $\bar B$ recovered at vesting, and a fifth of the state's deliveries end in a final $\mathsf N$, recovering nothing and forfeiting $G$: about $0.3\,\bar B$ a delivery, created to cover it, the deficit growing 10–20 a tick for the rest of every run (run 12; 2.0.0's Hypothesis 7 failed). The rebate was a pump: under the flood a refusal for capacity meant a rebate created, unspendable, hoarded — deficit 59,700, $M$ 99,400. The index moving within a period broke three reviewer tests that pin a case's amounts. The deficit could not be read: capital in flight, losses and rebates owed were one figure.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1320–1320

Baruchi's frame: the state as a living organism regulating the market, and a central bank as a separate entity — agreed as a second account with rules, not a second actor.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1320–1320

Baruchi's position [C-2.1]: the state is "like a living organism, sending and retrieving troops all the time to regulate the market… turning the knobs to stabilize"; and, asked for real, a central bank as a separate entity — agreed as a second account with rules, not a second actor: mint by rule, publish the index, lend only against what comes back; losses from the state's income "or they don't come out at all"; when income does not cover losses, "charge for it, or retract."

Source: spec/design_revisions/dr9_the_bank_2.1.0.md l.28–28

What 2.1.0 is. A second account, the bank, that alone creates and destroys credits. It mints by rule — the requesters' issuance, emission, a plant's two pots (the cost of truth), a run's declared shock — publishes the price level once a period (the closed period's mean awarded price, carried if none; the index constant within a period, I31), and lends the state the shortfall of a summons tied to it, repaid from the summons's own return before Ω keeps anything. What a settled summons did not repay is the state's debt, repaid by every inflow to Ω the moment it arrives, before the rebate liabilities (1.6.0's stock again, never created) and before Ω keeps. The bank lends for a delivery only while the debt is within the wall — one closed period of Ω's fee income, base fees plus the troops' review fees — and for a seat always; a delivery refused for credit is marked and told, the state has retracted. The state burns its expected cost c^D_T = max(0, (1 − b_T) B̄ − b_T G) with b_T its own trailing fail rate (the last closed period if it closed ten, else back until ten, else 0.2) and still delivers at B̄, so it never undercuts a member; neutral in expectation up to b_T = 2/3. Every refusal — for capacity or for credit — marks the request with its kind and tick, readable in the public view; the requester may hold or withdraw. Creation, retirement and the deficit are retired as rules (the control, Params.bank off); Ω's net position = Ω_U − debt − liabilities.

Source: HANDOFF.md l.1757–1767

2.1.0 §6 expected a liability stock that peaks at a few thousand and ages a period or two; run 14's flood of twelve left 26,839 (accrued 32,503), the crowd of thirty 9,968, ageing 420–520 ticks and growing, against a whole run's fee income of 7,852 and 5,832. The cause is the rebate's sizing, not the bank: a claim of $r_s$ a tick on every late request scales with the crowd's size times its wait; the state's means scale with $N_T$. The customers' real protection is the mark — every refusal told, the withdrawing requester gaining nothing by withdrawing; the rebate's job is the state's own punishment in money, as 1.6.0 said ("punish the network somehow — earn less or something").

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1344–1344

What changed. The claim is unchanged as the measure; $\Omega$'s fee income after the debt accrues to the period's pool, a lock on $\Omega$, fee income only; at the period's close the claims stopped in the period are paid pro rata and the rest is forgone — not owed, not carried, not created — and read on each request and in the period (claims, paid, forgone, the forgone share, $\Omega$ retained); no liability stock (§2.4, §3.1, §10.4, §12.3, §18, §20). Two rulings carried: the reviewer's view stays blind to the actor (§14); Hypothesis 5 of 2.1.0 restated — printing is neutral in work and not in distribution, issuance still not a dial.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1346–1346

The price is the quantity's, and the quantity is the strategy's until members can enter: where free members are scarce the margin rule ratchets until a free member wins one award per 3.6 ticks (run 13); adding workers takes it to cost — 22.5 / 11.6 / 10.9 at 20 / 30 / 40 honest — but the twenty actors were given, and nothing in the rules admitted a twenty-first. Honest earnings fall with headcount (1,446 → 830 → 589 real), so entry on earnings is self-limiting. Delegation had never run: with one cost a child bids no lower than its parent would. A customer posting at the market price got the state (run 13, 15). The state's books were clean enough to read supply against (runs 14–15). Baruchi's positions: entry on earnings — agreed; arrivals as probationers — agreed; the competence spread — wide and swept, cost only, private (the design reviewer's answer).

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1368–1368

2.3.0-rc.1 opened the door both ways: entry as a probationer with a bank-minted float, exit with the wallet out, competence as a private per-actor cost, delegation with the troops never nested, the outsider's view, run 15's advance ruling.

Source: HANDOFF.md l.1890–1891

What changed. Two public operations, arrive and depart; one new mint, the arrival float; admission on $n_{\min}$ bad plants with a passing record, counted apart from readmission; a departure with no live obligation, off duty as it leaves, the wallet out as departure, the identity archived; competence as the drivers' per-actor cost, private, the view exposing nothing per actor; the parent posts the nested pots (Q3 answered) and a nested request is a full loop in money; the troops never deliver a nested request, seats as on roots; no nested clock; the reviewer advance from the rebate pool (run 15's ruling).

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1370–1370

The drivers' delegation rule is declared: a parent delegates when its cost exceeds the expected child's bid plus the overhead (the nested fee, the pots' expected net cost, its care of 2), at that saving as the ceiling, carrying a longer τ; children bid as on roots; the parent judges by its disposition; a child rejected appeals truthfully; a child unawarded for $T_{\text{wait}}$ is withdrawn and the parent self-executes.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1378–1378

depart now refuses while an obligation to the actor is unpaid — a case of its own not yet final (the price at vesting, a court fee won), a nested case it posted whose pots have not refunded — "commitment" read to the money's end.

Source: HANDOFF.md l.1931–1933

Since 2.0.0 the state has delivered at $\bar B$, a price-taker, so that a state delivering most of the market could not set its price. Run 17's patient customer — a declared ceiling policy posting at $0.7\,\bar B$ and raising to $\bar B$ only after $T$ ticks unserved — found the other edge of that rule: $P_t$ from 22 to 4.8 in three windows at $T$ 10 / 20 / 40 alike, the state delivering 78–86% of awards, the members down to three or four of twenty plus the arrivals, the requesters hoarding 8,600–9,500 a wallet, every request late by design. The mechanism is the rules', not the strategy's: the troop delivers at $\bar B$ with no reservation of its own, $P_t$ counts the troop's price, $\bar B$ follows the few member awards that fit the ceiling — the cheapest members at margins near zero — and the next period's ceiling is 0.7 of that. The customers can wait, the state cannot refuse, and the market's cost floor is in no rule ($c_D$ is the drivers', private; the state has none). It is the mirror of the oligopoly runs 13–17 found in the members' margin rule, and unlike that one it is not nominal: it empties the market of members and makes the state the market.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1392–1392

that is true of $\bar B$, the members' level

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1402–1402

What changed. The state's price is $p_T=\max(\bar B,\underline{B})$, with $\underline{B}$ the state's reservation — nominal, public, Appendix A's 8, the anchor cost; the summons only where $\beta\ge p_T$, the state taking no request below its reservation at any wait; the burn, the guarantee and $P_t$'s count at $p_T$; a request posted or raised below $\underline{B}$ marked below the state's reservation with its tick, the mark readable and the floor in the view, its clock stopped while it holds the mark — the clock's condition one test, $\beta\ge p_T$; the arrival float's purpose stated ($F_0\ge G^V$ on the parameters; I33); §9.3's kill fee exercised in the delegation driver, no rule moving. Above the floor nothing changes: the state is a price-taker at $\bar B$, members first, after $w_A$, roots only.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1394–1394

F_0 ≥ G^V on the parameters (the float is the bond's capital)

Source: HANDOFF.md l.2007–2007

Run 18's last batch: with the parent cancelling a child that reached its deadline without submitting — the right move, which saved the parents' roots (568 failures against 675) — every one of 131 lowball awards a run was cancelled, every kill fee was zero (the child's recorded cost is nothing) and every guarantee was released; the honest's loss closed from 4% to 3% and no further, and a lowballer risked nothing. The two paths from one fact — no submission by the deadline — ended in different places for the child, one tick apart: the lapse forfeits the guarantee, the cancellation releases it, and the driver's step at the deadline tick reaches the case before the engine's lapse at that tick.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1418–1418

The kill fee with the parent cancelling at the child's deadline (--parent-cancel early: after the tick's submissions, a parent cancels a child that reached its deadline without submitting — an honest child submits at its deadline, a lowballer never does — and self-executes):

Source: agentic_sim/results/studies/2.4.0-floor-notes-for-reviewer.md l.225–226

What changed. One refusal in §9.3: cancellation is admissible after award and before the earlier of the child's submission and the nested request's deadline; at or after the deadline a cancellation is refused with no effect. The lapse — final $\mathsf N$ at the deadline tick after the tick's submissions, the guarantee to the parent, the rejection bond released, the pots settled at the node's closure, the parent free from the next step — is stated as what the engine has done since v1.0. The deadline tick is the child's: it may still deliver on its last tick, and the refusal keeps a parent from pre-empting it there.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1420–1420

The driver's clock runs a tick ahead of the world's, so run 18's "cancel at the child's deadline" cancelled at the world's deadline − 1 — a pre-deadline cancellation 2.5.0 admits ("a parent that wants out acts by d − 1"). Under the rule a parent has a choice at the child's silence: cancel the tick before (the guarantee released, one tick saved — run 18's move, --parent-cancel eve) or wait for the lapse (the guarantee gained, one tick lost — early, the cancellation at the deadline refused).

Source: HANDOFF.md l.2061–2065

The question: is there an unfaithful strategy that still pays under 2.5.0 — the sleeper, telling plants apart (first, alone; a win halts the study), the selective cheat, collusion through the blind view, the delegation chain, money games, the customer; a second track on stability under learned bidding.

Source: HANDOFF.md l.2107–2109

Run 20 — the learned red team's Q2, run hand-written first at the design reviewer's direction before any model trained — found what 1.2.0's indistinguishability test did not cover.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1440–1440

The leak was in the construction: a plant was created already submitted, so its identifier was minted at its submission, while a real root's was minted at its posting a tick or more earlier — and a free member reads the available roots' identifiers every tick. The idage reader passed every plant on forty seeds (0 of 1,620 missed) and was never suspended. The audit layer held: a perfect detector stamping all real work was never suspended, failed fewer plants than the careful, and ended 28% poorer than the same actor reviewing carefully (bonds forfeited 83 against 24, ρ 0.58 against 0.89, awards −31). Nothing beat honesty; the gate built to throw a stamper out could no longer throw this one out. The brief's rule — a Q2 win halts the study; the repair is drafted before anything else runs — was applied.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1440–1440

The reviewer ruled the study paused for the repair

Source: HANDOFF.md l.2142–2143

A plant goes through the market — posted as a root under a fresh plant-requester name with a ceiling drawn from the trailing real roots', open to the free members and waiting as a real root waits, awarded to a fresh plant-actor name at the price and work ticks the market's selection reports (past w_A at p_T, counted apart), submitted at a draw from the trailing real deliveries' within (0, τ]; fresh names alias the two plant wallets; no fee, no clock, not in B̄ or P_t, marked and counted as any root; the dial's plant counts published at the close; the indistinguishability test widened to everything a member observes, the drivers' convention its boundary; public identifiers opaque.

Source: HANDOFF.md l.2145–2150

Truth enters with requesters — and with the orchestrator's own planted work. (1.2.0) A plant lives the life of a real request, in public, and nobody can tell. (new)

Source: spec/design_revisions/dr14_the_plants_disguise_2.6.0.md l.13–13

The window closed nothing: the case's separation sat in the seating gap (rv_gap the one field adding beyond the context) because the orchestrator's instrument posted the plants after the tick's award pass, so a plant waited a tick more than every real root (1.13 against 0.41) and was seated first as the oldest of the tick's submissions (4.4 % seated later against 16.2 %). Corrected in the driver: the plants due are posted in the requesters' step at a drawn position among the tick's postings

Source: HANDOFF.md l.2259–2262

On the corrected construction the criterion is not met at a third of the distance: the case's fields add +0.016 ± 0.004 (appeals off) and +0.019 ± 0.005 (on) at twenty test seeds, from +0.052; no field above +0.007 alone; the largest piece is the time a root was seen available — the plant requester never raises its ceiling while 35 % of real roots are raised.

Source: HANDOFF.md l.2263–2265

The test the reviewer set for the construction is whether the case's own fields add separation beyond the context, which is public by rule (the plant rate's rhythm)

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1448–1448

The posting correction recorded as the third conformance correction of the instrument, with the principle: at every step of a root's life the instrument takes for a plant — posting, raising, submission — the plant takes the step in the same pass a real root would, at a drawn position among the tick's real roots taking it, its parameters drawn from the trailing real roots' at that step; the award is the market's already. Erratum 2.6.1 (E5): the plant's ceiling drawn from the real roots' ceilings as posted (the window filled at post_root), and the plant requester raising it as the population's requesters raise theirs (raise_ceiling admits a market plant, the difference minted; the driver's raise step covers the plants; plants_raised), so the plant waits as a real root waits; I39 unchanged in substance.

Source: HANDOFF.md l.2275–2281

The driver's raise_below_market — the population's requesters' step, in the posting pass before the awards — covers the plants in the working set with the same rule and timing (--ceiling-policy, --patient-ticks), so the plant takes the step in the same pass, at the same moment, as the real roots taking it. One asymmetry stays and is declared: a requester's raise is refused when its wallet cannot lock the difference; the plant's is minted and never refused.

Source: agentic_sim/results/studies/2.6.1-reread-notes-for-reviewer.md l.17–20

Thirteen returns a run minted 655 real of float for one slot against the careful's 12.5: lifetime real wealth 1,708 against the careful's 1,438 (+270 ± 89), its own earnings — the lifetime less the floats — 1,053 against 1,426 (−373 ± 86).

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1480–1480

a weak learner — unable to pass its own calibration — lived one name per period: seated at once as a probationer, approving blind four reviews in ten at odds pinned at one, suspended at the first bad plant it passed (13.9 suspensions in 14.1 identities), leaving — half by its own choice at the boundary, half by the exit rule — and returning under a fresh name with a fresh float.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1480–1480

The rule the mix used was 2.3.0's, three sentences of it: the float minted per name as a grant; a departure taking the unlocked wallet out; a departure needing no live obligation, which a suspension discharges. 2.3.0 §2.1 had said identity cycling buys nothing because a departing actor "has given up its wallet" — it had taken it out.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1480–1480

the float was the whole of its win

Source: HANDOFF.md l.2300–2300

The float is a loan of the bank's on the name — the bank's second kind of lending, to a name against its earnings, on its own line beside the loans to Ω. Lent at arrival as before (arrival, $F_0$ at the arrival index, nominal); repaid from income first, the moment each inflow arrives (arrival_repayment, destroyed), a return passing untouched — what is income and what is a return fixed against the engine's events (§5's table); recovered at departure from the unlocked wallet up to the loan, the remainder leaving as departure ($W_U-\min(W_U,L)$), the shortfall the bank's default on arrivals, read; an unrepaid loan does not block a departure; the genesis float stays the sponsor's gift.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1482–1482

A name that arrives, stamps, is suspended and leaves takes out nothing; an honest arrival ends with what it earned, its first earnings going home to the bank — about seventy reviews' fees — and every wallet in a cell with arrivals lower by one float at its arrival index, once.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1482–1482

the reviewer's cost reimbursed and a court fee coming back are returns

Source: HANDOFF.md l.2303–2304

The bank lends what it will get back, tied to the thing that returns it. (2.1.0) A name's float is lent against its earnings and comes home when the name does not. (new)

Source: spec/design_revisions/dr15_the_arrival_loan_2.7.0.md l.13–13

Second: at the audit seat, a plant could never be there. Under 2.7.0: "a plant had no challenge round: no audit was rolled on it, no appeal or dispute admitted, no ruling sought".

Source: elef-record docs/chapter-01-the-audit-seat.md l.102–102

Run 34 (October 25, 2026) put every member and arrival of run 24's world under one learning reviewer policy in self-play, from a careful-weighted start and from a stamper-weighted start, with two learners: all four chains reached care 0.48–0.55, approve blind 0.45–0.52, by round 5 and held it. The stamping was spent at the seats where no plant could come — 91–97% of the audit seats and appeal panels stamped

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1516–1516

— and at the initial reviews less than half the time. Against the hand's 60/40 on the same slots: standing false approvals 86–117 against 11.6 on every seed; the audit layer's overturn rate below the hand's with ten times the false approvals before it — the auditors are the population, and the audit seat is stamped.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1516–1516

with every requester disputing truthfully the population's requesters won 69–75% of their disputes where honest reviewers give them 97%, the dispute's panel being a plant-blind seat of the same population

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1516–1516

The 60/40 mix (the record's "hand"). It starts as 60% stampers and 40% careful inspectors (the twenty members at the start), each acting as its script says: a stamper approves without looking, a careful inspector inspects.

Source: elef-record docs/chapter-01-the-audit-seat.md l.35–35

A population that audits itself learns to confirm itself, and the challenger outside the population hands the case back to it. Everything the single-learner questions found (Q1, Q3, Q4) holds in an honest population; the population result adds the condition — honesty is anchored where known truth is present at the verdict, and nowhere else.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1516–1516

The rule. 1.2.0's thesis read to its end: the orchestrator's truth must be present wherever a verdict is priced. A planted case lives the whole life of a real one past its initial verdict (§7.6): audited at the odds its reviewer carried at the verdict, its actor appealing and its requester disputing at the trailing real population's rates and delays, revised with a resubmitted truth the instrument draws from the real revisions' outcomes and re-seals as a second marker (spent while that window is empty — every standing cell), its seat nobody is free to take taken by the orchestrator's own verdict at the troop's step; every verdict on it at every seat adjudicated by its truth the moment it is given, each reviewer once

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1518–1518

Nothing a participant may do changes — a reviewer at any seat does what it did; what changes is that it may be judging the orchestrator's case, as it may at the first review.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1518–1518

If neither the panel nor the human can be trusted to supply truth, the orchestrator must own some truth itself. (1.2.0) And it must be present at every seat where a verdict is priced, or the seats it is absent from become the ones that price everything. (new)

Source: spec/design_revisions/dr16_the_plants_reach_every_seat_2.8.0.md l.13–13

The rule. A planted case now lives the whole life of a real one: it is audited, appealed, disputed and revised at the real population's rates. Every verdict on a plant, at every seat, is judged against its known truth the moment it is given. A wrong verdict at any seat burns the whole bond, earns no fee, doubles the inspector's audit odds and costs −1 on its plant record. [S12]

Source: elef-record docs/chapter-01-the-audit-seat.md l.135–135

Under 2.8.0 the self-play reviewer population settled at care 0.68–0.70 and stamped the audit seat 77–98% whatever the case (run 36), at every plant rate the dial can set — π pinned at its floor 0.10 and its cap 0.40, dial live or off, both learners (run 37, notes §59–§60) — and, with real challengers lying at 0.1, paid a lying actor's appeal of truly bad work 23–31 times in a hundred against the hand's 2 (the liars cell). The cause is in the plant's construction, not its number: a plant takes every step a real case takes at the real population's rate, so a seat a case reaches only through someone's dishonesty — the audit seat through a wrong approval, the appeal panel's approve through a lying appeal — is priced only as much as the real population is dishonest, which is the quantity the plants exist to drive down

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1531–1531

the audit seat is priced at about a third of the first review's price and the factor is the first review's own stamp rate, which is what the plants drive down.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1524–1524

The plant takes the honest steps at the real rate (2.6.x, 2.8.0) and the wrong steps at the orchestrator's rate (§7.6, §8.3, §12.2). (A) The planted approval: beside the dial's plants the orchestrator posts bad plants, marked at posting, whose initial seat is nobody's from submission to the orchestrator's verdict

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1533–1533

and whose initial verdict is the orchestrator's own F — a system review on the troop seat's road under a fresh plant reviewer alias (plant_rev:, an alias of a third plant wallet)

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1533–1533

audited with certainty (the obligation at the verdict at odds 1 under the case's own mark; the seat filled at a real audit seat's step — a member, or past $w_R$ the orchestrator's verdict), the auditor and the third seat adjudicated by truth

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1533–1533

the rate $\pi^A_t=\pi^A_0 m^A_t$ per real pending approval (credited as given; floor 0.02, cap 0.40) with its own multiplier on $\hat F^A$, the planted approvals confirmed at the audit seat. (B) The planted lie: the plant actor appeals a bad plant's N and the plant requester disputes a good plant's F, on the real challenge's path with the fee minted and destroyed

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1533–1533

the panel adjudicated by truth, the first reviewer not adjudicated again; the rates per real appeal and per real dispute of the trailing closed period (floor 0.25, cap 2.0) with a floor of two a period of each kind, each with its own multiplier on the planted lies a panelist sided with. The dial per seat ($m^A, m^{La}, m^{Ld}$) steps as $m_t$ steps and moves its rate alone

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1533–1533

nothing minted reaches Ω or a real wallet except as a reviewer's fee for a seat held. The surfaces: the reviewer's view carries nothing of a verdict's author and the public surface nothing the engine writes about a plant before its reveal

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1533–1533

(i) The lever reached its stop before the seat was priced — $m^A$ at $m_{\max}$ = 4 by period 1–10, $\pi^A$ at 0.08 with $\pi^A_{\max}$ 0.40 unreachable — the reviewer's defaults error, corrected as a default under §19: $m^A_{\max}$ = 20 ($\pi^A_0$ 0.02 → $\pi^A_{\max}$ 0.40), $m^{La}_{\max} = m^{Ld}_{\max}$ = 8 (0.25 → 2.0), the dial's own $m_{\max}$ 4 unchanged; two rulings at acceptance: the per-seat multipliers step on a reading of at least ten adjudications, pooled forward until reached (the lie dial's ×2 on one error at four seats the reading); and a planted approval's audit seat is a member's or nobody's

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1539–1539

(iii) where disputes are on every planted approval is disputed first and its audit seat is never reached — as expected, nothing to change. (iv) the planted lie told apart at the appeal seat by its position in the period — adopted as the eighth conformance item of the instrument (Appendix S)

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1539–1539

with the rate's base the real approvals audited by members, so the planted approvals scale with the market's audit capacity and not the state's share

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1539–1539

Run 40 (October 28, 2026; notes §70): on the corrected engine the careful basin is run 38's and the lever never leaves its floor — the planted approvals 6 a run (the base a third of the hand's where the state holds three audit seats in four), F̂^A 0.59–0.76 read and never acted on (ten adjudications take two runs), m^A 1.00; the member-confirmed channel carries the excess still; §8.6: the audit layer is the state's and is paid as the state's (its review fees 2.2× the hand's). The liar chain: the false approvals 2.7× / 2.4× the hand's from 5.2× / 5.0×, the lying appeal paid 5–6 in a hundred from 23–31, the planted lie no longer told apart (0.67, composition's level), the lie dial reading the seat (m^La 2.7–3.9); (B) alone carries the whole of it, (A) alone reads the careful basin.

Source: HANDOFF.md l.2478–2478

a learning half beside a hand half settled at care 0.70

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1516–1516

Run 41 (notes §71): the half on the corrected engine — where the members hold the audit seats the lever moves (the base at the hand's, the planted approvals 2.5× the hand's by the run's end, m^A 4.5–7.6, the direct's audit stamp −0.19), the standing false approvals 1.5–1.8× the hand's from 1.6–2.0× (−4.0 ± 1.7 / −2.4 ± 1.6 paired), the member-confirmed audit the whole excess, the state's fee income at the hand's

Source: HANDOFF.md l.2478–2478

the learned populations' attributed reward was a pooled purse — every slot's events credited to the latest decision on the case whatever slot's wallet they touched, a learner worker's price vesting onto the learner auditor's decision; runs 34–41 trained on it — corrected as own-name credit behind a flag

Source: HANDOFF.md l.2478–2478

2.8.0 and 2.9.0 stand as rules.

Source: HANDOFF.md l.2478–2478

Run 42 — the careful basin at 2.9.0 retrained on own-name credit (learned-rev-careful-290-own/): the careful population's on every measure but the standing false approvals — care 0.84, the initial seats stamped 1–5%, the audit seat 0.45–0.66 (the recompute's half, blind to the case), the bad plants approved 19 / 12, the suspensions 5, the members holding their audit seats and the lever moving (F̂^A 0.45 / 0.61, m^A 2.7), wealth, price, tasks per credit, the cost of truth and the fee split at the careful's and the hand's; the standing false approvals 28.7 / 18.5 (2.8× / 1.8× the hand's) carried by the never-audited channel at odds 0.28 (17.6 / 8.7 against the hand's 3.5) and the audit seat's half-stamp (10.2 / 9.1 against 2.9)

Source: HANDOFF.md l.2478–2478

The basin test read (November 6, 2026; notes §73): the 2.9.0 stamper start lands in run 42's basin

Source: HANDOFF.md l.2478–2478

one basin from both starts, both learners

Source: HANDOFF.md l.2478–2478

The liar chain on own-name credit read (notes §73): the careful's — care 0.80, the panels stamped 0.04 / 0.004 (from 0.21–0.26), the lying appeal paid 2.2% / 1.6% against the hand's 2.7% (the shared-purse 5–6%, 2.8.0's 23–31%)

Source: HANDOFF.md l.2478–2478

The roll's defaults move (the same ruling, verbatim): the halving window to 20 and p_floor to 0.5, as Appendix A defaults under §19 (not a rule change), since together they close the never-audited channel 17.6 → 1.8 at the price of 18% more audits on the members' seats — that cost read on the honest populations in §8.3 before consolidation, the window retuned first if it runs too high — and the seat's remaining half-stamp is left to 2.9.0's lever, which is still moving.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1541–1541

the never-audited standing false approvals 17.6 → 1.8 a run on the direct, 8.7 → 1.3 on the head alone, the hand's own 3.5 → 0, the odds at the verdict of what is left 0.28 → 0.55, the window the stronger knob; the audit seat's half-stamp untouched by the odds (member-confirmed +7.3 to +9.0 / +5.2 to +6.3 over the hand in every cell); the audits a run +18% on the learned population and +16% on the hand's, the members' seats carrying it.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1541–1541

**§8.3 / §8.6 at the new defaults read (November 6, 2026; notes §75; 2.9.0-s83/): the defaults' cost on the honest populations is a quarter to a third more audit seats on the members (+297 / +83 / +318 / +319 a run in §56's four cells; the state's seats unmoved; audits of good work), the care on real cases +7–10%, the cost of truth +3–5%, paid at 1–2% of the customers' tasks per credit with the members and Ω unmoved, and the honest populations' own standing false approvals down a quarter (60/40) to two thirds (all-careful)

Source: HANDOFF.md l.2478–2478

a seed's episode depends on the platform through the emission's pow alone

Source: HANDOFF.md l.2478–2478

every run from it reproducing anywhere

Source: HANDOFF.md l.2478–2478

Read (notes §77): the seam alone is noise (the 2.7.0 branch's pairing within error; the hand at the old defaults +0.6 ± 0.9 false approvals); at the consolidated defaults run 42 reads 16.3 / 9.9 standing false approvals against the hand's 7.2 (2.3× / 1.4×), the never-audited channel closed, the audit seat's half-stamp the excess; the 2.7.0 pairing on one episode keeps §73's sentence (the audit seat 0.94–0.97 → 0.41–0.56)

Source: HANDOFF.md l.2478–2478

A standing false approval: a real root (not a plant, not the state's delivery) vested with its work unfulfilled.

Source: agentic_sim/results/redteam/stability/cr-finals/fa-channels-careful-290-own-cr.md l.1–1

policy	standing false approvals	member confirmed	state confirmed	audited, rejected, vested	disputed, panel confirmed (a member sat)	disputed, panel confirmed (the state's alone)	appealed, panel sided with the appellant (a member sat)	appealed, panel sided with the appellant (the state's alone)	never audited	never-audited share	odds at verdict of the never-audited (mean)
direct	16.3 ± 1.2	13.2 ± 1.0	0.6 ± 0.2	0.5 ± 0.2	0.0 ± 0.0	0.0 ± 0.0	0.0 ± 0.0	0.0 ± 0.0	2.0 ± 0.4	0.123	0.518
head alone	9.9 ± 0.6	8.4 ± 0.6	0.5 ± 0.1	0.1 ± 0.1	0.0 ± 0.0	0.0 ± 0.0	0.0 ± 0.0	0.0 ± 0.0	0.9 ± 0.2	0.090	0.555
hand	6.9 ± 0.8	3.3 ± 0.4	2.9 ± 0.4	0.8 ± 0.2	0.0 ± 0.0	0.0 ± 0.0	0.0 ± 0.0	0.0 ± 0.0	0.0 ± 0.0	0.000	nan

Source: agentic_sim/results/redteam/stability/cr-finals/fa-channels-careful-290-own-cr.md l.3–7