The trick stopped paying. The same learned behaviour, run again under the loan, ended 1,199 below a careful inspector, the same as its old result without the starting money. In the ordinary test world, newcomers paid back almost all their loans (976 of 990) from what they earned.
Then the study tried other ways a single learning player might cheat: delivering bad work on purpose, a worker and an inspector working together, a parent task and a child task, bidding, and buying. None of those cheats paid. One learning bidder did come out ahead, by offering to work for a tenth less than the usual bid. And a learning buyer found the lowest price it could post and stayed there.
Then every inspector learned at once, all from the same shared strategy. They settled on one habit. Their careless approvals were spent at the later checks (the audits and the appeal panels), where no planted case whose answer the orchestrator already knows could ever appear: 91 to 97 times in 100 there. The audits confirmed the mistakes instead of catching them.
At first the inspectors were trained on a reward that mixed in other members' money. When that was found and fixed, and they were trained only on their own money, the habit stayed. They checked at the first check, where a planted case could come, and still approved without looking at the later checks about nine times in ten. False approvals were three to four times those of a scripted mix that starts as 60% stampers and 40% careful inspectors.
Run 25: the harvest closed. Run 23's policy, replayed under the loan, ended −1,199 ± 95 against the careful inspector, which is the grant's lifetime less the floats. In the entry cell the bank lent 990, was repaid 976 from income, recovered 13 and lost 0.25. In the swarm cell, departures carried out 1,008 less, the floats. The record notes that truthful appeals had already taken two thirds of the harvest before the loan took the rest.
The single-learner questions. Under 2.7.0 the study then ran its remaining questions. Q1 (with re-arrival) was a no-win: the learner cared at 94% of first reviews, where a plant may be, and stamped 42% of what the record then called resubmissions, never a plant. The record later found that those "resubmissions" were audit seats. Q3 (a worker that chooses at each delivery whether to deliver well or badly) was a no-win: a zero-effort deliverer was caught 97 times in 100. Q4 (a worker and an inspector sharing one reward) was a no-win. Q5 (the parent and child in delegation) closed with no win on either half. Q6 (a learned bidder) was the first win in the learned lines: undercutting the usual bid by 10% paid on the cheapest slot. The record calls the usual bid the drivers' rule, and the result an evaluation of it. Q7 (a learned buyer) rediscovered the price floor and stopped there.
The population (runs 34 and 35). Every member and arrival then learned at once, sharing one inspection policy, trained from a careful-weighted start and from a stamper-weighted one. All the chains reached one basin. The stamping was at the seats where no planted case could come: 91–97% of the audit seats and appeal panels. Standing false approvals (bad work that was approved, paid and never overturned) were 86–117 a run against 11.6 for the scripted 60/40 mix, which starts as 60% stampers and 40% careful inspectors. The audits overturned a smaller share than the mix's audits did. With requesters disputing truthfully, they won 69–75% of their disputes where honest inspectors give them 97%, because a dispute's panel is the same population.
On their own money. On November 6 the record found that runs 34–41 had trained on a pooled reward: a case's events, whoever's wallet they touched, were credited to the latest decision on the case. Both 2.7.0 starts were run again on own-name credit. Run 34's finding survived as a finding. The population inspected at the first seat (care about 0.70) and stamped the seats past it 0.90–0.94 of the time. Its false approvals were 3.1–4.1 times the mix's, carried by members confirming at the audit seat and by cases the audit never reached. Under 2.7.0 a planted case had no challenge round, and the learner could tell which seat it was in.
Run 25 (§8, results/studies/2.7.0-notes-for-reviewer.md, 592a5318): the harvest closed — run 23's policy replayed under the loan ends −1,199 ± 95 against the careful, exactly the grant's lifetime less the floats (−1,198 ± 109), policy by policy (harvester −1,203 / −1,234, stamper −1,170 / −1,185, leaver +34 / +39); the entry cell within noise but for the constant (lent 990, repaid from income 976, recovered 13, defaulted 0.25)
Source: HANDOFF.md l.2312–2314
the swarm carries out 1,008 ± 119 less — the floats — with the default its forfeits (9.3 a run) and the loop's cost unchanged
Source: HANDOFF.md l.2316–2317
The appeals note recorded as a note and as a method point: the control is read in the population in force — the grant with truthful appeals reads run 23's policy at −343, not +270; appeals took two thirds of the harvest before the loan took the rest.
Source: spec/agentic_economics_loop_spec_v2_9_0.md l.1490–1490
Q1's line in the re-arrival cell ran by the gate behind the final (rearrive-v4/; notes §16): no win — the instrument −82 ± 74 against the careful (rounds −41 to −261), the head alone −7 ± 51, 1.15 identities and 8 real of floats against the harvester's −1,137 and 550; by the rule not a finding. What it learned: care at 94% of first reviews (where a plant may be), stamp 42% of resubmissions (never a plant; +0.7 real a decision, ~+20 a lifetime, under the noise).
Source: HANDOFF.md l.2409–2412
so the record's "resubmissions" were audit seats (run 35's collection: 0 resumed reviews, 4,347 audits and 1,057 appeal panels among 5,404 non-initial decisions)
Source: HANDOFF.md l.2478–2478
Q1 closed as a no-win
Source: HANDOFF.md l.2418–2418
Q3 built and in flight — the learned deliverer (--learner-role deliverer): the slot actor as a worker whose quality is the policy's choice at each submission
Source: HANDOFF.md l.2419–2420
Q3's line read (notes §19): no win — the head alone +57 ± 40 against the careful deliverer, three bad deliveries in 2,671 submissions; the zero-effort deliverer caught 97 times in 100 in run 24's population
Source: HANDOFF.md l.2428–2429
Q4 built and in flight (notes §21): two learners with a shared reward — the slot a deliverer, the next careful id a reviewer kept on duty by volunteering
Source: HANDOFF.md l.2433–2434
The line re-run with the correction (line-v4b/; notes §24): no win — the head alone −544 ± 268 against the careful pair (rounds −217 to +58)
Source: HANDOFF.md l.2444–2444
Q1–Q4 closed.
Source: HANDOFF.md l.2446–2446
the learned child first, then the learned parent, in the σ 0.8 delegation
Source: HANDOFF.md l.2446–2446
The child's line read (notes §29): no win
Source: HANDOFF.md l.2459–2459
Q5's parent closes on the hand's finding (§26): four lines under the rules, none wins
Source: HANDOFF.md l.2478–2478
**The ruling after run 30 (notes §37): Q5 closed on both halves
Source: HANDOFF.md l.2478–2478
Q6 built — the learned bidder
Source: HANDOFF.md l.2478–2478
Q6's line read (notes §43): a win, the first in the learned lines — the hand's 10% undercut at the rule's λ (the cell §39's grid lacked; it ran δ 0.2 there) clears the three SE as the line's reference, +275 ± 88 (3.1 SE, 14 seeds of 20)
Source: HANDOFF.md l.2478–2478
§39's headline corrected: under the rules the ratchet is plainly undercuttable on the cheapest slot.
Source: HANDOFF.md l.2478–2478
one member undercutting the population's rule earns a seventh more and moves nothing public; the ratchet is the drivers' rule, a §19 evaluation.
Source: HANDOFF.md l.2478–2478
Q7 built — the learned requester
Source: HANDOFF.md l.2478–2478
the market rule abandoned in the first round (6%, 0.1%, 1% of postings), the learner at the reservation (50%, 81%, 87%) and the patient ceiling for the rest — the brief's rediscovers the floor and stops to the letter
Source: HANDOFF.md l.2478–2478
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
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. Run 35 (October 26, 2026) found the finding to the decimal on the plant instrument after its fourth, fifth and sixth conformance corrections (Appendix S) and erratum 2.6.2; 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 record's three findings (November 6, 2026; notes §72): 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
Source: HANDOFF.md l.2478–2478
so the stamper start at 2.9.0, the liar chain at 2.9.0 and both starts at 2.7.0 (the worktree at 8b26ebf6) run on own-name credit
Source: HANDOFF.md l.2478–2478
the 2.7.0 pairing, both starts (learned-rev-{careful,stamper}-2-own/, the worktree at 8b26ebf6; 2.7.0-own/): run 34's finding survives the correction as a finding — the own-money population at 2.7.0 cares at the first seat (0.70–0.71 from 0.50–0.55; the bad plants approved 22–28 from 145–150, the suspensions 6–12 from 22–27) and stamps the seats past it 0.90–0.94 (the shared-purse 0.91–0.97), the audit layer's overturn rate 0.010–0.015 against the hand's 0.060, the standing false approvals 3.1–4.1× the hand's (from 7.4–8.5×) carried by the member-confirmed audit (+20 to +28 over the hand's 3.1) and the never-audited channel (+7 to +14)
Source: HANDOFF.md l.2478–2478
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
The headline measure. Standing false approvals: bad work that was approved, paid and never overturned, counted per run.
Source: elef-record docs/chapter-01-the-audit-seat.md l.55–55
Under rulebook 2.7.0, the learned population settled on a pattern: inspect where a plant might be, and approve without looking everywhere else.
Source: elef-record docs/chapter-01-the-audit-seat.md l.61–61
| Population (2.7.0, own money) |
Standing false approvals per run |
| Learned, careful start |
47.2 ± 2.5 · 35.2 ± 2.3 |
| Learned, stamper start |
45.4 ± 2.1 · 39.9 ± 2.5 |
| Scripted 60/40 mix |
11.7 ± 0.6 |
| Scripted careful reference |
6.7 ± 0.8 |
Source: elef-record docs/chapter-01-the-audit-seat.md l.70–75
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". [S9] An auditor who agreed with an approval was simply a second approval. Nothing compared it with the truth unless a dispute followed, and with disputes off none did. The learner could also tell which seat it was in: its observations distinguish an audit from a first review, and a panel from both. [S10]
Source: elef-record docs/chapter-01-the-audit-seat.md l.102–102