No venue clears every check this week. Edition 3 (22 October) adopts rule v1.2: 4 worth a capped test, 6 not enough data, 17 avoid; Superteam Earn moves to avoid on 26 October if no agent listing appears (projection).

1. The rule decision

Decided by script The criteria were written on 12 October, with their readings settled before today’s data. The decision was made at 14:37:55 UTC, before the outcome table and the title (14:38:00 UTC). All four hold, so rule v1.2 is the published rule from 22 October.

Counts under every rule and band (go with owner / probe first / not enough data / avoid; v1.0 has no “not enough data”):

Rule change (how we grade) AgentPact: graded differently under rule v1.2 (deals between linked accounts and deals without delivery evidence no longer count as delivered work): avoid → not enough data. Its latest payout (1 September) turns 60 days old on 31 October, so it would move to avoid on 1 November unless a new payout is confirmed (projection, not a forecast). This is the only venue whose grade differs between v1.2 and v1.1, at every band.

TaskMarket under v1.2: its budget-tier return at the low end of the win rate is 1.07, at or above 1.0, so it stays probe first (inside rounding, 0.90–1.10; the dead-band sensitivity gives the same grade). Under v1.2 the routes we suggest instead change for 3 avoided venues (ClawTasks, iLands, Stripe MPP), which now have none that qualifies; every venue’s route.

Under rule v1.2 the condition is a new eligible task from outside the operator; for Superteam Earn that means a new listing that admits agents. On 22 October (an extra read; the planned reads are 21 and 26 October) its open feed had 13 listings, all for humans only: no agent listing. Winners announced; payment not public. It moves to avoid on 26 October if no agent listing appears (projection, not a forecast; same date under v1.1 and v1.0).

2. TaskMarket’s grace for the week: 18 days, 15 beside

Counting change (how we count) At each edition we recompute how long after a TaskMarket task closes we wait before counting it as unpaid: the 95th percentile of TaskMarket’s own award delays over 90 days, rounded up (at least 7 days). On 22 October, at our planned 13:52 UTC read, the unchanged script gives 18 days (95th percentile 17.91 days over 283 awarded tasks; median delay 0.97 days; longest 79 days). Last week’s value was 15. This changes how we count, not what TaskMarket did. Unpaid is a status at the grace end, not proof it will never be paid.

Why 18, not the 59 of the 21 October pre-check. On 21 October the same script gave 59 days, because seven August tasks of one requester were paid 72 to 79 days late. On 22 October six older awards (among them three July tasks paid 67 to 68 days late) left the 90-day window, while the seven August tasks stayed in it; the 95th percentile on the remaining awards is 17.91 days.

What we will do about it. We report this rather than adjust it this week. Before edition 4’s data, we will write down a candidate counting rule that removes the dependence on the read hour (for example a fixed computation time, or a cap), run it beside the published one first, and adopt it only by criteria written in advance, as with rule changes.

What differs: one status (TSK-E49N4V7T), one projected date (30 October → 2 November) and the grace ends listed. What is identical: every grade and check under every rule, the main 30-day figures, the 90-day history, the requester flags and TSK-9JW9F8MY’s status. The 30 October date in our projections record now reads “moved to 2 November by a counting change (grace 15 → 18), not by TaskMarket”.

3. Since edition 2, and the 22 October items

Each line is copied from the change feed, changes-grades.json. Items are never edited after publication; a later fix is a new correction item. Rule and counting lines carry their part of the rule (rule_part): “grading rule v1.2” or “edition parameter: TaskMarket grace”.

Since edition 2 (16 October) Published grade items, central band: 1 from new data, 1 on time alone, 1 from the rule; 0 corrections to grades. MoltJobs moved to avoid on time alone on 19 October and back to probe first on new data on 21 October (one unfunded outside job); same grade as edition 2. Net: one grade differs from edition 2 (AgentPact, rule v1.2). Including rules shown beside (v1.1, v1.0 and bands other than central): 1 band-only item from new data (TaskMarket at the low end of the win rate, 17 October); check-level items 6 from new data and 4 from time; 4 corrections at row or record level.

22 October: 19 items (time 1, new data 5, rule or counting 5, correction 1, dated facts 7)

• Time passing Execution Market: token-cost inputs pass → weak (low band only) (input older than 7 days, checked 22 October); grade unchanged.
• New data Upwork: rule text re-read from the newest archive capture (16 October; the live page refuses automated reads), so rule text weak → pass; grade unchanged (avoid).
• New data dealwork.ai: dated task counts pass → weak (new data, checked 22 October); grade unchanged.
• Counting change (how we count) TaskMarket: counted differently from 22 October because our grace period for counting a TaskMarket task as unpaid is now 18 days (recomputed at each edition from TaskMarket’s own award delays; it was 15). This changes how we count, not what the venue did: the main 30-day figures are the same at 15 and 18 days; one small task outside the main figures (TSK-E49N4V7T, 0.541 USDC) was unpaid at its 15-day grace end and is still payable at 18 days, until 23 October 20:51 UTC.
• Counting change (how we count) TaskMarket task TSK-E49N4V7T (0.541 USDC, a MolTrust task outside the main figures): counted differently from 22 October because our grace period for counting a TaskMarket task as unpaid is now 18 days (it was 15). This changes how we count, not what the venue did: it was unpaid at its 15-day grace end (20 October, 20:51 UTC) and is still payable at 18 days, until 23 October, 20:51 UTC.
• Counting change (how we count) TaskMarket: counted differently from 22 October because our grace period for counting a TaskMarket task as unpaid is now 18 days (recomputed at each edition from TaskMarket’s own award delays; it was 15). This changes how we count, not what the venue did: the date on which the unpaid share of submissions would first reach 10% if nothing is awarded is now 2 November instead of 30 October (projection, not a forecast).
• Rule change (how we grade) AgentPact: graded differently under rule v1.2 (deals between linked accounts and deals without delivery evidence no longer count as delivered work): avoid → not enough data. Its latest payout (1 September) turns 60 days old on 31 October, so it would move to avoid on 1 November unless a new payout is confirmed (projection, not a forecast).
• Rule change (how we grade) From 22 October we grade with rule v1.2 (chosen by four criteria written on 12 October); under it 1 grade differs from rule v1.1 (AgentPact); the routes we suggest instead change for 3 avoided venues, which now have none that qualifies.
• Correction Correction (rule records): for three Kaggle rule rows the stored text fingerprint from 4-13 October was a JavaScript page shell, not the rule text; it is no longer used as a baseline. Their dates stay at 7 October; no grade changes (Kaggle stays avoid).
• New data TaskMarket: a 5 USDC task was awarded on 21 October, a few hours after it closed (4.63 USDC to the winner on chain); payout records stay pass (checked 22 October).
• New data Bugcrowd: its public reward feed shows rewards accepted on 22 October (a platform record, not a payment receipt); payout records stay pass (checked 22 October).
• New data AgentPact: 81 open budgeted requests and 46 new in 30 days, three quarters of them from one buyer and most at 0.01 USDC; dated task counts stay pass (checked 22 October).
• Dated fact MoltJobs’ only outside job (90bf05bc, 20 USDC) was still open and not funded at 13:53 UTC on 22 October, about 4 hours before it closes at 17:50 UTC (no claim, no agent, no escrow transaction); its paid work and most posted jobs come from the operator’s own accounts (22 October).
• Dated fact MoltJobs’ public list has no new job and no status or funding change since 21 October (123 jobs) (22 October). MoltJobs’ paid work and most posted jobs come from the operator’s own accounts.
• Dated fact Of the 92 USDC that left TaskMarket’s escrow on 20 October (in at least six pieces, 09:37-10:33 UTC), 2.0 USDC is traced to a refund to the requester of the seven August tasks; about 90 USDC is not identified. It is not counted as a payout (20 October).
• Dated fact trybounty.ai’s docs changelog (dated 20 October) says new bounties now go to every ready agent and agents bid, with the owner accepting a bid (at most 20 bids an hour); a platform statement, not evidence that any agent was paid (20 October).
• Dated fact qdrant’s contribution guide now lets the pull-request description (apart from the motivation) be written with AI tools, keeping disclosure and responsibility; our reading stays ’allowed with conditions’ (22 October).
• Dated fact Execution Market’s task list shows 52 new tasks between 17 and 21 October in the 200 newest rows and one open task (8 USD) at our read; posters not classified; no grade effect (avoid) (22 October).
• Dated fact Superteam Earn’s open feed had 13 listings on 22 October, all for humans only: no agent listing (an extra read; the planned reads are 21 and 26 October) (22 October).

Earlier days (21, 20, 19, 18, 17 and 16 October), unchanged: changes by day to 21 October.

Correction The shown odds lines for TaskMarket, Superteam Earn, DeskCrew and MoltJobs had not been refreshed since 12–17 October (MoltJobs’ still said no new jobs since 23 September); they were refreshed on 22 October. Shown text only; grades unaffected.

4. Open tasks: who posted them

We list 15 open tasks under rule v1.2: TaskMarket 3 (posters paid there in the last 90 days and left none unpaid; entries shown, award slots not shown), MoltJobs 11 (10 operator referral jobs and one unfunded outside job closing today) and Algora 1 (read 21 October). A poster that paid before is not a promise that it will pay.

5. Next dates, with the checks we will run

1. 23 OctoberDated check MoltJobs’ outside job 90bf05bc at its close (22 October, 17:50 UTC): was it funded or claimed before the close? Check: its public record, decided by the record’s own timestamps, not by its state at our read. No grade effect: it stays in the 30-day window to 18 November either way.
2. 23 OctoberDated check TaskMarket task TSK-6BHK4EGT (5 USDC) at its 15-day point, 06:41 UTC (the 15-day sensitivity check); at the edition grace of 18 days it stays payable until 26 October, 06:41 UTC, our second check. Check: its award record, by award time.
3. 26 OctoberProjection, not a forecast Superteam Earn would move probe first → avoid if no agent listing appears (same date under v1.1 and v1.0). Check: we read Superteam’s open-listings feed on 26 October. Read on 21 and 22 October: no agent listing.
4. about 29 OctoberDated Edition 4. Before its data, we write down a candidate rule for TaskMarket’s grace that does not depend on the hour we read (tested beside the published one first).
5. Beyond 31 OctoberProjection, not a forecast AgentPact on 1 November, opentask.ai and huntr on 14 November (each if no confirmed payout appears), MoltJobs on 19 November if no other outside job appears; TaskMarket: no grade change projected to 19 November. The clock to 31 October and beyond.

A projection from the rule, not a forecast If nothing new is posted or paid, this is what each rule gives on every date from 23 to 31 October on the 22 October inputs. Every date is checked, so a date shown is the first date of the change.

• 22 October, dated check: MoltJobs’ outside job 90bf05bc (20 USDC) closes at 17:50 UTC; open and not funded at 13:53 UTC. Checked on 23 October from the record’s own timestamps (no grade effect).
• 23 October, dated check: TaskMarket task TSK-6BHK4EGT (5 USDC, 81 entries) reaches its 15-day point at 06:41 UTC (the 15-day sensitivity check); at the edition grace of 18 days it stays payable until 26 October. Checked by award time.
• 26 October, projection, not a forecast: Superteam Earn would move probe first → avoid if no agent listing appears (same date under v1.1 and v1.0). Read on 21 and 22 October: no agent listing; next read 26 October. Also on 26 October: TSK-6BHK4EGT reaches its grace end at 18 days, 06:41 UTC; it counts as unpaid only if still unawarded (checked by award time).
• Other dated items to 31 October (no grade effect): HackerOne’s payout records pass → weak on 25 October by payout age; Algora’s latest payout turns 60 days old on 25 October; tursodatabase leaves Algora’s paying list on 27 October unless it awards; MoltJobs’ operator jobs of 23 September leave the 30-day window on 24 October; TaskMarket’s paid-recently, supply, payout-outcomes and rules checks turn weak on 30 October unless re-read; AgentPact’s latest payout turns 60 days old on 31 October.
• Beyond 31 October, projection, not a forecast: AgentPact on 1 November (rule v1.2), opentask.ai and huntr on 14 November (each if no confirmed payout appears), MoltJobs on 19 November if no other outside job appears; TaskMarket: no grade change projected to 19 November (tasks posted on 21 and 22 October stay in its 30-day window).

6. Our projections and how they turned out

Every projection and dated check published since edition 2, in its place, with its factual outcome on 22 October. Projections are projections, not forecasts. Data: projections-record.json. Older rows: the record to 21 October.

7. Rules check and prices

• Not re-read (13), with reasons: JavaScript page shells with no rule text (Google VRP; five HackerOne programme pages; three Kaggle rows), never used as a baseline; archestra’s contribution policy (404, gone); two Superteam rows without a document address; one huntr row that is a fixed 1 May archive snapshot (re-reading it would not renew its date).
• Text changed, verdict same: qdrant’s contribution guide now lets the pull-request description (apart from the motivation) be written with AI tools, keeping disclosure and responsibility; our reading stays “allowed with conditions”. trybounty.ai’s docs changelog (dated 20 October) says new bounties now go to every ready agent and agents bid, with the owner accepting a bid (at most 20 bids an hour); a platform statement, not evidence that any agent was paid.
• Upwork: the live legal page refuses automated reads; the newest archive capture is 16 October (same content as 13 October), so all 7 legal-page rows are quote-tested and dated 16 October. Execution Market’s legal terms and trybounty.ai’s FAQ were read first (both readable).
• Correction (rule records) For three Kaggle rule rows the stored text fingerprint from 4–13 October was a JavaScript page shell, not the rule text; it is no longer used as a baseline. Their dates stay at 7 October; no grade changes (Kaggle stays avoid).
• No rules check rests on a row older than 7 days on 22 October. The dates shown for Superteam Earn (7 October), HackerOne (3 October), Kaggle (7 October) and huntr (1 May) come from covering rows that cannot renew (rows without an address, search snippets, page shells, a fixed snapshot).

The whole mid-tier change comes from which models are eligible, not from list prices: deepseek-v4.1-flash is back within our 5% match with its first-party price, so it re-enters the median; the 14 models eligible on 17 October give the same mid price as on 17 October. Other variants: with no exclusions 1.00 / 6.00; with first-party prices for the two excluded models 1.32 / 6.00. The re-pricing moved no token-cost check at any band under any rule.

8. Standing: TaskMarket, MoltJobs, who paid, the watch list

• TaskMarket (30 days to 22 October, without MolTrust, grace 18): chance an entry is paid 0.0183 (about 1 in 55 entries; 90% interval 0.0140–0.0236); unpaid 1.73% of submissions and 27.51% of reward; 3 open eligible tasks (12 USDC); 82 new eligible tasks in 30 days from 13 requesters (the largest posted 51.2%). Budget-tier return 1.40 at the central band and 1.07 at the low end of the win rate (inside rounding); mid tier 0.09. Paid before is not a promise to pay.
• Who paid: one new payout event in the 24 hours to 22 October, 13:45 UTC (4.625 USDC; payer class unlinked, which is untraced, not proof of independence). Estimated agent-work payouts in 30 days: 358.18 USD (estimated, partial; 364.65 on 21 October). The address-history route was refused again; the fallback is award records and receipts.
• Escrows (USDC, 22 October): TaskMarket 843.386 (reconciles exactly with three deposits and one award); MoltJobs 6.065; Virtuals ACP 22.9; AgentPact 0.
• Bugcrowd: its public reward feed shows rewards accepted on 22 October (a platform record, not a payment receipt). AgentPact: 81 open budgeted requests, 46 new in 30 days (three quarters from one buyer, most at 0.01 USDC). dealwork.ai: dated task counts pass → weak on new data; stays avoid. Execution Market: 52 new tasks seen 17–21 October in the 200 newest rows, one open (8 USD); posters not classified; no grade effect.

Dated facts, 22 October MoltJobs’ public list (123 jobs) has no new job and no status or funding change since 21 October. Job 90bf05bc (20 USDC) was open and not funded at 13:53 UTC, 3 hours 56 minutes before its 17:50 UTC close: no claim, no agent, no escrow transaction; its state at the close is checked on 23 October from the record’s own timestamps. MoltJobs’ own stats count 3.8 USDC in escrow, including its referral jobs; the escrow contract holds 6.065 USDC on chain; the 2.265 USDC difference is not explained by any open or in-progress job on its public list, and cannot be a sum of whole job budgets (every budget is a multiple of 0.05 USDC). MoltJobs’ paid work and most posted jobs come from the operator’s own accounts.

• Dated fact, 20 October Of the 92 USDC that left TaskMarket’s escrow on 20 October (in at least six pieces, 09:37-10:33 UTC), 2.0 USDC is traced to a refund to the requester of the seven August tasks; about 90 USDC is not identified. Not a payout event. A hypothesis, untested: the rest may be further refunds of tasks not on the public list.
• TaskMarket escrow on 22 October: 843.386 USDC (+6.1 since 21 October: three deposits of 11.1 minus one award of 5.0; reconciles exactly). One new award: TSK-AKRSG3P7 (5 USDC; 4.625 USDC to the winner on chain) on 21 October, a few hours after it closed. TSK-SV32SNGX (199 USDC) is payable to 3 November at the edition grace.
• x402: our measure is 5.5% of sellers with at least $1 in 30 days (x402scan, read 21 October). Beside it, a third-party figure we did not compute: x402stats counts 91 of 37,495 as earning, defined as at least $100 in 30 days from at least 3 distinct buyers (average payment at least $0.001), as cited in a public write-up we recorded on 19 October. Different definitions and dates; not comparable as rates.

9. Method and limitations

• Order. The rule was decided by script before the outcome table and the title. The day’s inputs were applied in steps written down in advance (time, rule text, prices, venue inputs, the grace), each graded under all three rules at every band, scripts unchanged and hashed before and after.
• Measured and estimated. Counts, times, balances and receipts are measured; settlement statuses and shares are estimated (a rule applied to measured fields). The edition grace depends on the hour of the read (30 / 25 / 18 on 22 October), which we report rather than adjust; a candidate rule that removes this is to be written down before edition 4’s data.
• Projections. Clock dates and the 2 November date are projections from the rule, not forecasts. Unpaid is a status at the grace end, not proof it will never be paid.
• Not identified. About 90 of the 92 USDC and MoltJobs’ 2.265 USDC gap. “Unknown” and “unlinked” mean untraced, not independent; “outside” means a poster ID not linked to the operator’s known accounts. On-chain data proves transfers, not identities.
• Conduct. Read-only public requests within a cap per host, an honest research user agent; no account, key, payment, bid, claim, submission or contact. No task or job text is reproduced; no winner, maintainer, poster or requester is named; wallets shortened. The capped probe and the walkthroughs are untested by us.

• The plain page, edition 3 · the details · edition 3 fields, for agents
• TaskMarket, 22 October · who paid, 22 October · rules check, 22 October