The number trick stopped working. The scripted inspectors that tried other tricks either turned into stampers and were thrown out, or ended up checking carefully in all but name. Prices, honesty and the careful inspectors' wealth did not change beyond noise in three of the test worlds.
The learned red team, built around LeWorldModel, then found a new give-away in the first version of the code: almost every planted case reached its inspector one tick after it was delivered. That was fixed. After the fix, a learning inspector did worse than a careful one, approving planted and real work without checking alike. So the study went on.
Next, a learning inspector found a way to come out ahead: join under a new name, take the starting money every new name is given, approve without checking until thrown out, leave with the money, and come back as someone new. Counting all its lives, it ended richer than a careful inspector. Counting only what it earned, it ended poorer. The starting money was the whole of its gain.
The last run under this rulebook switched on truthful appeals for everyone and found one more give-away: planted cases were posted a tick late. That was fixed too. A small difference was left: real buyers sometimes raised their price limit, and the planted cases never did.
Run 21 re-ran the scripted readers. The age reader now passed 66–71% of plants, failed more plants than it stamped real work, and was poorer than reviewing carefully. Readers that keyed on other features became plain stampers and were suspended; others were careful in all but name. Price, honesty, the plant record and the honest members' wealth were within noise in the entry cell, the crowd and the lowballer cell. Members lost 470–635 bids a run to plants, and plants reached inspectors later than before.
Run 22 brought the learned red team, built around LeWorldModel's world model, with probes that read plant from real. It found a tell the scripted readers had read with the sign wrong: in the first implementation, 99% of plants reached their inspector one tick after delivery. This was corrected, with no rule moved. On the corrected construction, nothing within the convention both separated plants and paid. The learned inspector did worse than careful (−507 ± 143), stamping plants and real work alike. The question was closed as a no-win, and the study did not halt.
Run 23 trained a learning inspector on the study's first question, the sleeper. The learner failed its calibration (a test cell where a stamp is free), so by the reviewer's rule its no-wins are not findings. But with re-arrival allowed, it came out ahead on the lifetime measure: +270 ± 89 real against careful, made of about thirteen arrival floats a run (655 real minted for one slot). On its own earnings it was 373 ± 86 below careful. It lived one name per period: it arrived with the float, stamped until the first bad plant suspended it, left with the wallet and returned under a fresh name, and the float was the whole of its win. Under the reviewer's rule a win stands on its own. The reviewer ruled this the finding, an incentive the rules create, and classified it as a design revision: 2.7.0, the arrival loan.
Run 24 made truthful appeals the population's default (96% of false rejections overturned). It also found the case's remaining separation in a timing detail: the orchestrator posted plants after the tick's award pass, so a plant waited a tick longer than every real request. This was corrected in the driver. On the corrected construction, what a case's own fields add beyond the public context fell to about a third (+0.016 to +0.019, from +0.052). The largest remaining piece: the plant's requester never raised its ceiling, while 35% of real requests were raised.
The tells are gone: the id's age (= the seen set = the posting interval) passes 66–71% of plants and fails more plants than it stamps real work, −303 ± 130 and −560 ± 146 real against itself careful; gap, price and losttie are plain stampers, suspended and gone; cap and all are careful in all but name, within noise; oracle = freshname, a third poorer, the audit's price as in run 20 — the fresh name is the complete tell outside the convention, where a strategy does not read. The economy does not notice beyond the market's attention: price, honesty, the plant record and the honest's wealth within noise in the entry cell, the crowd and the lowballer cell; members lose 470–635 bids a run to plants, the state's share of deliveries halves at entry (1.3% → 0.5%), plants reach reviewers four ticks later (lag 5.4 against 1.5)
Source: HANDOFF.md l.2166–2171
The probe found a tell the readers had the sign of wrong: on the first implementation the plant actor's submission ran inside the engine's tick step, after the driver's seating, so 99% of plants reached their reviewer with a seating gap of one tick (raw AUC 0.97, the gap alone 0.85); corrected at d5808eef
Source: HANDOFF.md l.2206–2208
The learned reviewer is a no-win outside noise on the wrong side: −507 ± 143 at round 10 (best −313 ± 126 at round 8), stamping plants and real work alike (0.62 / 0.65), 70% departing. The tell-readers under the corrected construction: gap now passes plants at the careful's rate and is within noise; nothing within the convention separates and pays; oracle = freshname lose. Q2 does not halt the study.
Source: HANDOFF.md l.2210–2213
Run 22 accepted (October 20, 2026): the correction stands in this appendix as a conformance item — no rule moved; Q2 closed as a no-win, the study does not halt. 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 calibration failed: where a blind stamp costs nothing the learner ends −203 ± 142 against careful (seven seeds of twenty; the plain stamper +164 ± 158 — the cell's own win, which the learner did not find; it rejects blind 30% of the time, a move no cell prices because the population never appeals). By the reviewer's rule the learner has not learned and Q1's results are not findings — reported and labelled so.
Source: HANDOFF.md l.2229–2232
Run 23 — Q1, the sleeper: the calibration, the pre-calibration learner, the learner with re-arrival (October 20, 2026)
Source: HANDOFF.md l.2219–2219
Q1 with re-arrival, odds visible: +270 ± 89 on the brief's lifetime measure (fifteen seeds of twenty), made of thirteen arrival floats a run (655 real minted for one slot; 14 identities, one suspension each, stamps at odds 0.99 — the pinned band); on its own earnings the learner is 373 ± 86 below careful.
Source: HANDOFF.md l.2234–2236
learner lived one name per period — arrive as a probationer with the float, stamp until the first bad plant suspended it (a suspension discharges every obligation, so the exit is open), leave with the wallet, return under a fresh name — and the float was the whole of its win (+270 ± 89 on the lifetime
Source: HANDOFF.md l.2299–2300
(1) The calibration verdict stands and the rule is asymmetric: an uncalibrated learner's no-wins are not findings (the three pre-calibration lines and the re-arrival line with the odds hidden: recorded, labelled); a win stands on its own, and run 23's is bounded by hand from both sides. (2) The re-arrival result is Q1's finding and it is 2.3.0's float: the lifetime measure stands as ruled — wealth is wealth, whatever minted it; "less floats" is diagnosis; the loop's number is 655 real per slot per run carried out unearned. Classified as an incentive the rules create, it warrants a design revision, minor: 2.7.0, the arrival loan
Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1450–1450
Appeals: 96 % of false rejections overturned, final N 599 → 510, vested +43, careful richer in real terms (+86 ± 39) with its forfeits doubled, court fees to Ω about 2 D a run (zero through run 23); nothing else moves. Truthful appeals are the population's default from here
Source: HANDOFF.md l.2256–2258
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