The orchestrator's troops

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

agentic_sim/results/studies/1.4.0-troops/ (and 1.4.0-troops.zip): controls/ (160 runs: 1.2.0 and 1.3.0 with the roll held at 50–80%), tipping/ (80), adaptive/ (40), drift/ (40), windows/ (180), cost/ (120), float/ (40), redteam/ (120: camouflage 80, reformer 40), ruling/ (160, after the reviewer's reply) — 940 runs, 20 seeds per cell, 20 actors, every final audit passed, engine f0779ac. Appeals truthful, no disputes. The §6 reproduction is test_v14.

Source: HANDOFF.md l.945–948

The controls with the roll held barely moved. 1.2.0: false approvals 53 / 67 / 89 / 121 at 50–80% (run 4: 53 / 67 / 89 / 121); 1.3.0: 53 / 66 / 90 / 112 (run 5: 53 / 69 / 90 / 113), probationers' 7 / 12 / 14 / 15 (run 5: 7 / 15 / 14 / 16). The dropped roll was hiding almost nothing at m_max = 4; what it hid was the m_max = 2 drift of run 5 (65 of 97). Every comparison below is against these.

Source: HANDOFF.md l.950–953

stampers	1.4.0 vested / 1.3.0 / 1.2.0	troop deliveries (vested)	state's share of vested work	FA 1.4.0 / 1.3.0 (probationers')	troop FA	stamper − careful 1.4.0 / 1.3.0	members' awards 1.4.0 / 1.3.0	careful fees + uplift	award summonses refused for funds (case-ticks)
50%	2,136 / 1,923 / 2,399	349 (257)	12%	42 / 53 (7 / 7)	2	−2,761 / −2,130	164 / 152	1,921 (545)	7,243
60%	2,237 / 1,719 / 2,412	828 (613)	27%	49 / 66 (7 / 12)	4	−3,196 / −2,386	147 / 153	2,549 (922)	8,768
70%	2,361 / 1,443 / 2,396	1,390 (1,033)	44%	69 / 90 (12 / 14)	4	−3,965 / −2,489	118 / 145	3,691 (1,666)	4,888
80%	2,376 / 1,173 / 2,024	1,784 (1,338)	56%	98 / 112 (13 / 15)	5	−5,094 / −2,426	73 / 137	5,392 (2,709)	4,832

Source: HANDOFF.md l.957–962

Hypothesis 2 holds: vesting recovers to the 1.2.0 control at 50–70% and above it at 80%, and the state's share reads the capture — 12% of vested work at 50/50, 56% at 80/20. Hypothesis 1 holds in rate, not in count: false approvals fall below the controls (troop reviewers at ε_T = 0.05 and no dropped audit), but the probationers' count is unchanged at m_max = 4 because there was little to close there; in the m_max = 2 drift (below) the count stays at 62 while throughput doubles, so the rate per vested case halves (3.6% → 1.7%) — the residual is now the audit's 5% miss and the ruling's 20% error, not the missing auditor. Hypothesis 4 holds in sign and fails in its clause: the gap widens by the uplift as predicted (545–2,709 per careful actor), but members' awards are not unchanged — at 70–80% they fall by a fifth to a half (118 and 73 against 145 and 137), because every troop delivery needs an initial review and most need an audit, and that work keeps the members on duty; careful wealth still rises (fees 5,392 against 1,602 at 80/20). Hypothesis 5 fails: the fund refuses thousands of award summonses at every split (4,800–8,800 case-ticks a run) while paying 6,400–12,700 in uplift. The reserve R_T is the peak capital of troops that were fundable; the uplift drains the fund to it every batch; so the fund can never carry more troops than it once did — a self-limiting reserve. The treasury, not the market's need, sets the state's share. Hypothesis 6 holds: cases held for the market's audit live 6–16 ticks at most, and the held count at any tick is 0–1.

Source: HANDOFF.md l.964–978

Adaptive dispositions (3,000 ticks): under 1.4.0 the stamper share reaches ≤ 10% at a mean tick of 900 in 20 of 20 (1.3.0 with the roll held: 870); troops are active about 0.9 at a time at ticks 500 and 1,000 and zero from tick 1,500 in every seed; vesting 7,203 against 7,138. Hypothesis 3 holds: the state retracts on its own, within a few hundred ticks of the market turning honest.

Source: HANDOFF.md l.980–983

m_max	FA (probationers', troops')	vested 1.4.0 / 1.3.0 / 1.2.0 / 1.1.0	troop deliveries (vested)	troop audits	award summonses refused	fund paid out	stamper − careful
2	157 (62, 8)	3,621 / 1,856 / — / —	2,527 (1,950)	407	21,883	15,065	−6,763
4	78 (22, 7)	3,494 / 1,506 / 2,518 / 4,209	2,942 (2,206)	407	10,197	22,334	−10,595

Vesting doubles against 1.3.0 and reaches 85% of the 1.1.0 control's at m_max 4 with false approvals at an eighth of its; the state delivers 55–60% of what vests at 90% stampers and says so.

Source: HANDOFF.md l.987–993

Cost and competence (60/40): c^D_T = 20 makes a delivery a loss in expectation (0.8 × 0 − 0.2 × 30) and the fund stops carrying — troop deliveries 209 against 828 at c^D_T = 10 and 950 at 5, award summonses refused 23,900, vesting back to 1,894 (the 1.3.0 level). ε_T = 0.10 doubles troops' false approvals (3–4 → 6–8) and adds 20–30 false rejections; everything else within noise. A_T 300 against 100 at 80/20: refused award summonses fall from 4,832 to 1,640 but troop deliveries do not move (1,805 against 1,784) — the advance is not the binding constraint; the fund's inflow against its uplift is.

Source: HANDOFF.md l.1002–1007

Red team. Camouflage under troops buys nothing (−2,882 against −2,761 at 50/50; −7,221 against −7,559 at 90/10).

Source: HANDOFF.md l.1009–1010

The leak does not close. Wrong rulings fall from 34 to 9 a run at drift, but probationers' false approvals fall by ten and six, and at the tipping cells not at all; false rejections fall by a similar count, since a wrong ruling cuts both ways. So the ruling is a sixth of the probationers' leak where it is largest, and the rest is the audit itself: the auditor of a pinned approval is whichever member or troop was free, and a member whose record has not yet failed approves what the probationer approved.

Source: HANDOFF.md l.1042–1046

What run 6 leaves — three findings for the next decision, in order.

1. The fund's reserve is self-limiting. R_T is the peak capital of fundable troops and the uplift drains the fund to it every batch, so the treasury can never carry more troops than it once did, however many the market needs; A_T does not help, and c^D_T = 20 collapses it. Candidates, all design questions: a reserve sized to the unmet demand (the refused summonses) rather than the met one; uplift paid only when troop shortfall was zero in the period; or the raised tax split between the fund's reserve and the uplift.
2. Troops consume member capacity. Every troop delivery is reviewed, and every troop-reviewed approval of a member's work is audited at the floor: at 70–80% the members review more and deliver less. The "no crowding out" clause of Hypothesis 4 should be restated — the honest keep their income, but it moves from awards to fees — or the troops' reviews should carry less audit load (a default: the odds a troop reviewer's approval carries).
3. The roll held changed little at m_max 4; the residual probation leak is mostly the auditor's, not the ruling's. The reviewer asked for the sweep before any default moved; it is below. Ruling accuracy 0.80 → 0.95 removes three quarters of the wrong rulings and a sixth of the probationers' false approvals at drift, none at the tipping cells. What reaches a customer through a probationer is the auditor's own miss — a member whose record has not yet failed, or a troop at ε_T — more than the ruling's.

Source: HANDOFF.md l.1014–1030