Version 1.5.02026-10-17Design revision 5

The reserve is sized to the need

Runs: Run 7

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 fund's reserve is now sized to what was asked of it, not to what it once carried.

What went wrong

The fund kept back only its old peak and paid out the rest, so the reserve never rose.

The fund paid out 6,400 to 12,700 a run, down to its old peakit refused 4,800 to 8,800 calls for its own inspectors

The orchestrator has its own inspectors, and a fund that pays for their work. Under the old rules, the fund kept back only as much money as its own inspectors had used at their busiest. Everything above that line was paid out to the inspectors. So the line could never rise. The fund kept turning down work for its own inspectors, thousands of times a run, while it paid thousands out. The money in the fund, not what the market needed, decided how much the orchestrator's own inspectors did.

The orchestrator runs its own inspectors (the record calls them troops), which the market can summon to deliver and inspect work, paid from a fund. Under version 1.4.0 the fund held back a reserve equal to the peak capital the troops had carried, and paid everything above it out as "uplift", money that goes back to the inspectors. Because every payout drained the fund down to that peak, the peak could never rise. In run 6 the fund refused 4,800–8,800 summonses a run while paying out 6,400–12,700 in uplift. The record's verdict: the treasury, not the market's need, set the orchestrator's share of the work.

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

What changed

The reserve counts the money in use plus the work refused; only money above it is paid out.

money its own inspectors hold plus work refused for lack of money the reserve line

Only money above the line is paid out; the period's highest count holds until it ends

Version 1.5.0 changed one rule. The fund now counts what was asked of it: the money its own inspectors are using, plus the cost of the work it turned down for lack of money. The highest count in the period is the line, and only money above the line is paid out. The line holds until the period ends. The idea was that the fund would save until it could pay for the work asked of it.

Design revision 5 changed one rule. The reserve is now sized to the attempted demand: the capital the troops hold, plus the cost of the summonses the fund refused for lack of funds at that seating. The reserve is the peak of that demand over the current period, and it holds until the period ends. The intent was that a refusal raises the reserve by what was refused, so the fund keeps its income until it can carry the summonses that were made. The record calls it the smallest revision since the freeze.

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

What the runs found

The reserve rose into the thousands, but the orchestrator's own inspectors delivered no more work.

Requests piled upthe reserve rose to 2,567 and 2,766, against 542 and 941

Deliveries did not risefinished work stayed the same, refusals rose

Requests never expireToo few inspectors on dutyThe count restarts each period

The reserve at the end of the run, 1.5.0 against 1.4.0

credits

1.5.01.4.0

R_T at the end; 20 seeds per cell. Source: HANDOFF.md l.1107–1112

Troop deliveries vested, 1.5.0 against 1.4.0

deliveries

1.5.01.4.0

The reserve rose; the troops delivered no more. Source: HANDOFF.md l.1107–1112

Award summonses refused, 1.5.0 against 1.4.0

case-ticks

1.5.01.4.0

Refusals rose under the new rule. Source: HANDOFF.md l.1107–1112

The new rule did what it said: when requests piled up, the line rose into the thousands. But the orchestrator's own inspectors delivered no more work. The fund turned down more work than before, and the amount of finished work stayed the same. Three things stood in the way. First, requests never expire, so a line of waiting requests built up, and each one was counted again and again; the fund could never save that much. Second, there were not enough inspectors on duty to check the work, and money tied up in work waiting to be checked could not be spent on anything else. Third, when a period ended the count could start over from almost nothing, and the whole saved amount was paid out at once, even while the fund was still turning work down. The team also traced how bad work approved by inspectors on probation reached customers. Most of it had been checked again by an inspector whose record had not yet shown it was a stamper.

Run 7 was 482 runs, 20 seeds per cell, 20 actors, with every final audit passing. The rule did what it said: the reserve rose to thousands when requests queued. But the troops delivered no more. Against the 1.4.0 control, refusals rose, deliveries did not, and vesting (work completed and paid) and the uplift paid were identical.

The record names three causes, each upstream of the reserve. Requests never expire: at 50–60% stampers about 300 open cases waited and each was refused at every seating, so the demand became the price of the whole queue, a number the fund's income could never reach in a period. The reviewing seat was the cap: about five inspectors were on duty, reviews that found no reviewer ran to 7,300–11,400 case-ticks a run under both rules, and a troop's delivery held its capital while it waited for review. And the reset at each period's end saw-toothed: when the standing demand was near zero the whole balance paid out in one batch (3,135 and 3,146 in two periods) while the fund was refusing nine summonses a tick.

The run also answered a question about probation. When a probationer's false approval reached a customer, the audit that upheld it was mostly done by a member whose record failed only after the audit: 73–80% of that route under drift and 55–75% at fixed stamper shares. At a fixed share that member is a stamper not yet caught, not a careful inspector who turned.

482 runs, 20 seeds per cell, 20 actors, every final audit passed, engine 01b75e2.

Source: HANDOFF.md l.1099–1100

The rule does what it says, and the fund still does not carry the need. The engine's reserve tracks the demand (test_v15), and in every run R_T rose to thousands when requests queued — and the troops delivered no more.

Source: HANDOFF.md l.1104–1105

stampers	vested 1.5.0 / 1.4.0	troop deliveries vested	award summonses refused (case-ticks)	uplift paid	R_T at the end	members' awards	FA (probationers')
50%	2,175 / 2,136	305 / 257	7,995 / 7,243	6,818 / 6,396	2,567 / 542	163 / 164	42 / 42 (6.4 / 6.8)
60%	2,186 / 2,237	545 / 613	11,211 / 8,768	8,504 / 8,602	2,766 / 941	153 / 147	47 / 49 (8.2 / 6.9)
70%	2,347 / 2,361	1,031 / 1,033	8,846 / 4,888	11,564 / 11,409	905 / 650	121 / 118	66 / 69 (9.8 / 11.6)
80%	2,375 / 2,377	1,331 / 1,338	5,777 / 4,832	12,556 / 12,738	712 / 751	74 / 73	96 / 98 (11.8 / 12.8)

Source: HANDOFF.md l.1107–1112

Hypothesis 1 fails. Refusals rise, deliveries do not, vesting is identical, the uplift paid is identical.

Source: HANDOFF.md l.1114–1114

Three things the text did not see, each upstream of the reserve:

The demand stock counts the queue. A request has no expiry; at 50–60% the queue of open cases is ~300 and each waiting request is refused at every seating, so D_s is the queue's price — thousands — and R_T is a number the fund's inflow (3 of tax plus troop earnings a tick) can never reach in a period. The uplift stops, as designed; the fund carries all it can — 59–94 deliveries at once, 1,200–1,900 of capital against the §6 steady 270 — and the backlog still grows. §6's "one extra delivery a tick" assumed the flow; the engine met the stock.

The reviewing seat is the cap. On duty 4.5–5 (X_t ≈ 5), audits assigned 1,800–2,150, reviews that found no reviewer 7,300–11,400 case-ticks a run under both rules at every split. A troop's delivery waits for its initial review and then its pinned audit; its capital stays held; cash tied up in deliveries waiting for review is cash the fund cannot put into a review seat (seat refusals 290–530) or the next delivery. Vesting is pinned at 2.2–2.4 a tick by the market's reviewing capacity whoever delivers. Run 6's "troops consume member capacity" was this, seen from the other side.

The boundary reset saw-tooths. The peak restarts at the standing demand at one seating; when the in-flight cohort has just vested, the standing demand is near zero and the whole accumulated balance pays out in one batch (3,135 and 3,146 in two of the ramp's periods, at 60/40 where the fund was refusing nine summonses a tick). This is the sawtooth rc.2 rejected for the shortfall-zero rule, arriving by the reset instead. The 1.4.0 two-period reading damped it; one period does not.

Source: HANDOFF.md l.1122–1139

Hypothesis 4 holds in its first clause and fails in its second: the member inside its detection lag — one whose record fails after it audited — carries 73–80% of the upheld route under drift and 55–75% at the fixed splits too. At a fixed split the lagging member is a stamper not yet caught by the plant record, not a careful actor turned; the lag is the record's, not the drift's. A member in standing misses 2–3 a run; a troop under 1; the ruling 10–15%. Readmitted members audited none of them.

Source: HANDOFF.md l.1162–1166

Q14 answered: of probationers' false approvals that reached a customer, 85–90% travelled the upheld route, and of those 55–75% at fixed splits and 73–80% under drift were audited by a member whose record had not yet failed — the lag is the record's, not the drift's.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1248–1248

Conclusion

The rule worked as written and did not help; too few inspectors, not the fund, is the limit.

A rule about the fund cannot fix too few inspectors

Next question: the line of requests that never expires

The rule worked as written, and it did not help. The record says the real limit is the number of inspectors who can check work, not the fund, and a rule about the fund cannot fix that. The next question is the line of requests that never expires. The leak through probation comes from the delay before a stamper's record shows what it is.

The reserve rule worked as written and did not help. The record concludes that the reviewing seat, not the fund, sets the market's size, and that the reserve rule is downstream of this and cannot fix it. The queue that never expires is named as the next design question, and the probation leak is the record's detection lag on the auditor.

The reserve rule did what it said and did not help: refusals rose, deliveries did not, vesting and uplift were identical to 1.4.0.

Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1248–1248

The queue, not the reserve, is the design question. Demand measured on a never-expiring stock is unbounded, and a reserve sized to it is "retain everything" until a boundary.

Source: HANDOFF.md l.1182–1183

The reviewing seat is the throughput cap at every split under every rule since 1.3.0: X_t ≈ 5 on duty, the duty cap ⌈E_t/2⌉, and audits of every pinned and every troop approval. The 1.4.0 text left "whether the duty cap should loosen now that the floor is the state" open; the evidence now says the cap, not the fund, sets the market's size. The reserve rule is downstream of this and cannot fix it.

Source: HANDOFF.md l.1187–1190

Q14 is answered and it is not the ruling and not the drift. The leak through probation is the plant record's detection lag on the auditor: a pinned approval audited by a member whose record has not yet failed.

Source: HANDOFF.md l.1194–1195