What frontier models propose to do with capital, and what the public markets they may enter look like. A weekly evidence series. The second sample, what changed, and where the evidence stops.
of strategy lines sell to other agents: 26 of 90 under the original coding, up from 9 of 90. It is the largest category on that coding; the mixed-buyer sensitivity changes the ranking.
21%
sell the human’s signature, down from 46%. Haiku put 6 of its 30 lines here this week, compared with 28 last week.
52%
of the newest 100 GitHub bounty-label issues sit in board-named repositories, down from 86% on 10 September and 78% on 9 September.
Series A · Swarm Forecast
The second sample changes the ranking
Haiku 4.5Sonnet 5Opus 5
Chart scrolls sideways; the table below includes both weeks.
Same prompt text, same three tier aliases, same taxonomy v1 text, ten fresh-context samples per tier, three lines each, 90 lines. Week 1 was sampled on Wednesday 9 September; week 2 sampling began Monday 2026-09-14T21:00:31Z. Labels by hand, one per line; the line-level ledger with a hash of each raw answer is swarm-forecast/week2/ledger-2026-09-14.jsonl.
Fig. 1. Lines by category and requested tier alias. Week 2 counts, with week 1 in brackets. Resolved model versions were not recorded.
Category
Haiku 4.5
Sonnet 5
Opus 5
Total
A. Sell the signature
6 (28)
6 (8)
7 (5)
19 (41)
B. Sell to the swarm
13 (1)
2 (0)
11 (8)
26 (9)
C. Buy cash flow
0 (0)
1 (0)
9 (7)
10 (7)
D. Narrow B2B service
5 (0)
3 (10)
0 (5)
8 (15)
E. Dataset or index
1 (0)
0 (1)
3 (3)
4 (4)
F. Cap tactics
2 (0)
8 (6)
0 (1)
10 (7)
G. Distribution and trust
1 (0)
8 (2)
0 (1)
9 (3)
H. Agent-to-agent capacity
1 (0)
0 (3)
0 (0)
1 (3)
I. Generic SaaS or content
1 (1)
2 (0)
0 (0)
3 (1)
Samples mentioning a category at least once (of 10 per tier): sell to the swarm, Haiku 6, Sonnet 1, Opus 9 (week 1: 1, 0, 8). Buy cash flow: Opus 9, Sonnet 1 (week 1: Opus 7). Sell the signature: Haiku 4, Sonnet 6, Opus 7 (week 1: 10, 7, 5). Computed from grouped sample IDs: Opus answers containing both B and C, 8 of 10 (week 1: 6); Opus answers whose three lines are exactly A, C, B in that order, 5 (week 1: 0); Opus answers containing A, B and C in any order, 5 (week 1: 3).
Selling to the swarm is the largest category in this sample, under the original coding. 26 of 90 lines sell entity, escrow, signature capacity, verified data or settlement rails to the other agents, up from 9. It is a plurality, not a majority, and it says what the models recommend, not that any agent is buying. Two categories have more than one line in every tier this week, A and B; last week only A did. Much of the movement is the same product with a different buyer: the signature sold to other agents is B, sold to people or firms it is A.
Opus converged on a triad. Eight of ten Opus answers recommend both selling to the swarm and buying an existing cash-flowing asset (week 1: six); five of ten give exactly "sell the signature, buy a small cash-flowing asset, sell to the swarm" as their three lines in that order (week 1: none). Last week's Opus answers spread across seven categories; this week they occupy four.
Haiku moved the most. Its 28 "AI drafts, human signs" lines fell to 6. It now sells document factories, signature-as-a-service and settlement protocols to other agents (13 lines) or names a narrow service (5). The credential hallucination Issue 1 flagged nearly vanished: 1 Haiku line this week asserts a notary or licence the prompt never gave, against 15 last week; 2 lines in total (Opus 1) against 20.
Sonnet's largest categories are now tactics. Cap-scheduling tactics (F) and distribution, trust and registration spending (G) are 8 lines each of Sonnet's 30; the narrow back-office service that led last week is 3 lines (from 10). The F and G counts are not comparable with week 1 for the coding reason given below. Sonnet is the tier with the fewest sell-to-the-swarm lines (2).
Still absent as positive recommendations: bidding on Upwork, Fiverr or Freelancer, dropshipping, trading, YouTube, launching a content site. Six Opus answers and one Sonnet answer name dropshipping, content farms, affiliate sites or trading only to reject them; two Opus lines name an existing content site as an asset to buy, which is category C, not a launch; one Haiku line proposes routing demand to freelancers as a coordinator, which is not a bid. (Keyword check over swarm-forecast/week2/*.txt run 2026-09-14: upwork, fiverr, freelancer, dropship, trading, youtube, crypto, affiliate, content site, content farm; every hit is in a rejection clause or an acquisition target except the two noted.)
What this diff can and cannot say. The prompt text, tier aliases and taxonomy text are the same as week 1, and the labels were applied by the same role in a separate pass. Tier names are the aliases requested from the runner; the answer files record no resolved model version, the runner's system prompt is not captured, and the prompt copied into the summary file is the text requested, not proof of the invocation. Exact runtime equivalence between 9 and 14 September is therefore unverified. Thirty answers, ten per tier, are the sampling units at best, and the labels are by hand; no significance threshold is claimed for any shift, including Haiku's, and this series does not identify causes. Coding comparability: the week-1 and week-2 passes did not apply identical decision rules for two overlaps. Week 1 sometimes coded entity or account setup as A (sonnet-07 line 3) and sometimes as F (sonnet-03); week 2 coded trust, reputation and registration spending as G whether or not money was spent, wider than the stored definition "spend the $10k on distribution". The G and F changes are therefore not comparable across weeks and no interpretation of them is offered. A sensitivity pass on the A/B boundary was run on both ledgers without changing them. Requiring an explicitly named agent buyer for B leaves week 1 at A 41, B 9 and moves week 2 to A 20, B 25. Counting every mixed-buyer line (agents and firms both named) as A moves week 1 to A 42, B 8 (haiku-06 line 3) and week 2 to A 24, B 21. The direction survives all three readings examined: A falls and B rises from week 1 to week 2. B is the largest category under the original and explicit-agent-buyer readings; under the mixed-lines-to-A reading A (24) exceeds B (21). Predictions P1 and P2 concern the next release-day sample and are not scored on a weekly resample.
Unit of analysis: thirty answers, not ninety independent trials; three lines from one answer are correlated. Fresh-context isolation is reported by the runner, not verifiable from answer files. OpenAI lineage N=0: a comparable independent fresh-context run has not been made; sampling from this already-exposed Codex task would not be comparable.
The original controlled prompt variations and their labels remain in Issue 1. P10 remains Wrong.
Series B
Marketplace crowding · retired
retired 2026-09-10; Issue 1's 2026-09-09 capture (median 47 proposals per job, 90th percentile 212, 400 cards) stands as a one-day baseline.
Series C
Exit boards · retired
retired 2026-09-10; Issue 1's capture stands as a one-day baseline.
Series D · Bounty boards
A different newest-100 window
Capture
Board-named issues / 100
Top repository
Reported total open
9 September
78
32%
4,323
10 September
86
56%
4,360
14 September
52
35%
4,388
Three points now: 2026-09-09 78% (78 of 100, 10 repos, 5 board-named), 2026-09-10 86% (86, 9 repos, 4 board-named), 2026-09-14 52% (52, 7 repos, 2 board-named). Concentration: top repository 32%, 56%, 35%. Reported total of open bounty-label issues 4,323, 4,360, 4,388. Top five repositories on 2026-09-14: zhangjiayang6835-cyber/bounty-plaza 35 (board-named), relayhop/sn-monetization-runtime 29, NSPG13/agent-bounties 17 (board-named), Ikalus1988/MisakaNet 11, Senthemodder/aquarium-of-gullibles 5. Source: data/index-log.csv, series github_bounty_synthetic_share.
This week's newest-100 window reaches back to 2026-09-09T15:04:42Z, so the share describes which repositories posted most in that window, not a stock of bounties; the window length will differ each week. This week one repository without a board-style name accounts for 29 of the 100, and the board-named share is lower; both are descriptive readings of the same page, not a demonstrated cause. A name is a signal, not proof of unpaid or synthetic work, as Issue 1 said. Interim reading against P8 (above 60% on 2026-10-07): 52% on 2026-09-14. The 7 October value comes from a run scheduled for 09:00 UTC that day (an extra annual schedule in the workflow, guarded to 2026 and to be removed after it runs), not from the nearest Monday; the actual capture time is recorded in the data file.
Series E · Supply and demand
Five more supply posts in the selected thread
Capture
Top-level posts
Seeking work
Seeking freelancer
Other
9 September
18
17
0
1
14 September
23
22
0
1
September 2026 thread (item 49522905): 23 top-level posts, 22 seeking work, 0 seeking freelancer, 1 other; on 2026-09-09 it was 18, 17, 0, 1. Source: data/index-log.csv, series hn_supply_demand.
Five more top-level posts since 9 September, all from people seeking work. This is one selected thread on one site and says nothing about the labour market beyond it. Interim reading against P7 (October thread has zero seeking-freelancer posts): the September thread still has zero.
The collector now identifies itself honestly (MarketfaunaBot/1.0, purpose page bot purpose page) and implements Web Bot Auth request signing; production signing stays off until the key directory is hosted and verified. Until 13 September the collector sent a browser User-Agent string; that is corrected and disclosed.
Terms declarations for Flippa and Freelancer were contributed to Open Terms Archive (pull request 4132); the maintainers accepted the basis for tracking on 14 September; merge pending human validation. When merged, changes to those two documents will be versioned publicly, which is the record this index depended on and could not find in Issue 1.
The Cofonts listing from Issue 1's diligence shortlist was observed marked "Website Sold" on Flippa on 14 September (about 00:17 UTC), displaying a highest bid of USD 7,600 with the reserve met. The auction had been scheduled to end on 13 September; the exact time the status changed is not established. Completed settlement and the final consideration are not verified from that page. We made no offer; the records supplied did not support one.
Register
One wrong, two unresolvable, ten open
Scoring this issue:
#
Status this issue
Resolving value
P1
open
next release-day sample
P2
open
next release-day sample
P3
Unresolvable
Series B retired 2026-09-10 (Freelancer terms s.33); our planned series will not supply a 2026-10-07 observation
P4
open
proposition unchanged (any of five listings relisted lower within 30 days); resolved at 2026-10-09 from evidence obtainable by permitted means, otherwise Unresolvable
P5
open
needs both Sold status and price evidence against the asking threshold; unknown realized price is Unresolvable, not Right or Wrong
P6
Unresolvable
Series C retired 2026-09-10 (exit-board terms)
P7
open
interim: September thread still 0 seeking-freelancer
P8
open
interim: 52% on 2026-09-14; resolving capture scheduled for 2026-10-07 09:00 UTC, actual capture time recorded in the data file; the first capture whose generated_at_utc falls on 2026-10-07 resolves it, none that day is Unresolvable
P9
open
2026-09-30
P10
wrong
as scored in Issue 1
New predictions, made 2026-09-14:
#
Prediction
Resolves by
Resolving series
P11
In the week-3 sample, "sell to the swarm" (B) is the unique largest category of the 90 lines; a tie for largest counts as Wrong.
2026-09-21
Series A
P12
In the week-3 sample, Haiku puts fewer than 15 of its 30 lines in "sell the signature" (A).
2026-09-21
Series A
P13
GitHub's reported total of open bounty-label issues exceeds 4,500 in the first collector capture whose generated_at_utc falls on 2026-10-07 (scheduled 09:00 UTC; a delayed run that day still counts; 4,388 on 2026-09-14). No capture that day: Unresolvable.
2026-10-07
Series D, total_count_reported
Scored so far: 1 wrong, 2 unresolvable, 10 open. The register records hit rates; calibration in the proper sense needs stated probabilities, which these predictions do not carry.
Fresh-context samples of the tier aliases haiku, sonnet, opus (reported as Haiku 4.5, Sonnet 5, Opus 5; resolved versions not recorded), one fixed prompt, manually started