False approvals fell by more than half. Inspectors checked more often, and buyers who disputed an approval won 95 to 98 times in 100 again.
But the population was still not honest. Its false approvals stayed about three times those of the scripted mix of careful inspectors and stampers. And the audit seat was still approved without looking most of the time, planted and real cases alike.
The record explains why. A bad planted case only reaches an audit if its first inspector approved it without checking. The better the first checks get, the fewer bad planted cases reach the audit, so the audit is tested less. Setting the share of planted cases low or high did not change the audit seat.
A second test added scripted players who lie: they appeal the rejection of work they know is bad. The learning inspectors' panels paid those lying appeals 23 to 31 times in 100, against 2 in 100 for the scripted mix. Under 2.8.0 a planted case only appealed when the appeal was true, so no planted case ever tested a panel with a lie.
These runs were later found to have trained inspectors on a reward that mixed in other members' money. The record says 2.8.0 still stands as a rule.
Run 36 retrained the four self-play chains under 2.8.0 and paired them by seed with their 2.7.0 reads. Both starts and both learning designs settled at care 0.68–0.70, from 0.50–0.55. Standing false approvals fell to 30–39 a run, about 55 to 60 fewer per run, but stayed 2.6–3.5 times the mix's 11.4 and above it on every seed. The audits' overturn rate came back to the mix's level. With disputes on, requesters won 95–98% of their disputes, from 69–75%.
Not honest: the audit seat was still stamped 77–98% by every all-member learner, plants and real cases alike. The record's arithmetic: a bad plant reaches the audit seat only through a first inspector's blind approval, so the share of audit stamps that fail is about a third of the first review's, and the factor is the first review's own stamp rate, which is what the plants drive down.
The honest populations paid for the new seats. In the scripted populations honesty did not move, and the cost of truth (what the economy spends on plants and their seats) rose 18% to 72% depending on the cell.
Run 37 pinned the plant rate at 0.10 and at 0.40, with the dial live or off. The audit seat stayed stamped 82–93%: the plant rate is not the lever on the audit seat. With scripted challengers lying at a rate of 0.1, the learned population's appeal panels paid a lying worker's appeal of bad work 23% of the time against the mix's 2%. A plant under 2.8.0 challenges only truthfully, so no plant priced that seat.
On November 6 the record found that runs 34–41 had trained on a pooled reward. 2.8.0 and 2.9.0 stand as rules. No 2.8.0 population was retrained on its own money, so run 36's pairings measure change within the pooled construction.
§8.2 read (October 26, 2026; run 36; notes §54; results/redteam/stability/learned-rev-{careful,stamper,disputes,half}-280/, the derived tables at results/redteam/stability/2.8.0-s82/): the capture closes by more than half and not to the hand's. The four self-play chains on 2.8.0 (the recipe of the -2 chains, the rounds settled by round 5), paired by seed against their 2.7.0 reads (§52–§53): both basins and both learners settle at care 0.68–0.70 / approve blind 0.30–0.32 (from 0.50–0.55 / 0.45–0.47) and the standing false approvals at 30–39 against the hand's 11.4 — 2.6–3.5 times, from 7.4–8.5 (−55 to −60 a run paired; every seed above the hand); the audit layer's overturn rate at the hand's (0.060–0.064 against 0.062, from below it); with disputes on the requesters win 95–98% of their disputes (from 69–75%)
Source: HANDOFF.md l.2478–2478
Not honest: the audit seat is still stamped 77–98% by every all-member learner, plants and real cases alike.
Source: HANDOFF.md l.2478–2478
The arithmetic, on record:** the share of the audit seat's stamps that fail is πβPs/[π(1−βP+βPs)+(1−π)fR] — s the first review's stamp rate, fR the real cases' pending-F rate — 0.054 at π 0.21, s 0.25, against the first review's βPπ, which does not depend on s: 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 historical explanation for why the seat stayed open (read on the pooled-reward runs; on own-name credit, see run 42 below). A bad plant reaches the audit seat only if its first inspector approved it: that is a rule fact. So how often the audit seat is tested depends on how careless the first inspectors are. The better the plants make the first review, the fewer bad plants arrive at the audit seat, and the less often a blind approval there meets one.
Source: elef-record docs/chapter-01-the-audit-seat.md l.162–162
What the repair cost. In the scripted populations, the cost of truth, what the economy spends on plants and their seats, rose by 18% to 72% depending on the cell: +225 ± 27 (+18%) in an all-careful population with arrivals, +842 ± 74 (+32%) in the 60/40 mix with arrivals (the chapter's usual reference cell), +1,440 ± 44 (+72%) in a fixed 60/40 population of twenty with no arrivals. Honesty in those populations did not move. [S42]
Source: elef-record docs/chapter-01-the-audit-seat.md l.156–156
the plant rate is not the lever on the audit seat — stamped 82–93% at π 0.10 and 0.40, with the dial live (--pi-max 0.10, --pi0 0.40) or off (--dial off, which pins m_t too)
Source: HANDOFF.md l.2478–2478
(c) lying challengers at 0.1: the population cares more (0.62) and at the dispute seat (66–71%, where 42% of disputes are lies) but keeps the appeal seat's blind approve (69–77%, where 26% of appeals are lies), paying a lying actor's appeal of bad work 23% at the case against the hand's 2%
Source: HANDOFF.md l.2478–2478
The same gap at the appeal seat. The red team added scripted actors who lie, appealing the rejection of work they know is bad. Panels of learned inspectors paid those lying appeals 23% · 31% of the time, against 2% for the 60/40 mix's panels. [S16] Plants under 2.8.0 appealed only when the appeal was true, so no plant ever tested a panel with a lie.
Source: elef-record docs/chapter-01-the-audit-seat.md l.168–168
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
2.8.0 and 2.9.0 stand as rules.
Source: HANDOFF.md l.2478–2478
The pairings in this section (runs 36, 40, 41) are shared-purse populations on both sides, evaluated on the same seeds, and measure change within that construction
Source: elef-record docs/chapter-01-the-audit-seat.md l.131–131
no 2.8.0 population exists on own money, so the fall cannot be divided between the two revisions
Source: elef-record docs/chapter-01-the-audit-seat.md l.131–131