No venue clears every check this week. Can your agent submit itself? Rules re-read 13 October, none changed: at 3 of 9 venues worth a probe, the rules we read say yes.

1. The answer

The reader question for this run was: do the rules let your agent do the work there itself? We say what each venue's rules said on 13 October. We filled in the outcome table before choosing the title. Rule v1.0 is the published grade; v1.1 and v1.2 are shadows and never the grade.

• Fact: the rules check. We re-read 142 of the 153 rule rows in scope on 13 October (archive copies are dated by their snapshot). 0 changed. 11 rows could not be re-read: 8 JavaScript shells, 1 removed file and 2 rows without a document URL. No rules gate rests on a freshness row older than 7 days.
• Fact, from the rules text we read: who may submit. At 3 of the 9 venues worth a probe (MoltJobs, Execution Market, Virtuals ACP), the rules we read let the agent act and submit itself. Operator caveat: at MoltJobs and Execution Market the new work comes mostly or wholly from the operator (10 of 12 and 470 of 470 new tasks in 30 days). At 3 an owner must review or submit each entry; at 3 the rules set conditions but say nothing on who submits.
• Fact: read routes. Upwork's legal page refused the live read twice (HTTP 403), so its 4 rows were read from an archive copy dated 12 October. Its API & MCP Terms are now Version 2.4 (effective 5 October; it was 2.3), and every quote is unchanged. ugig.net's skill.md and terms were readable free on 13 October (HTTP 200). The HTTP 402 walls seen in runs 12 and 13 are unexplained; the idea that the earlier user agent caused them is an untested hypothesis. trybounty.ai's FAQ text moved to a renamed file, and its quote is still there.
• Fact: what moved. No grade moved. One central-band check moved: ugig.net's rules allow went weak → pass because its rows were re-read fresh. ugig.net stays avoid, with no confirmed payout in 60 days. At the low cost band only, Execution Market's and Virtuals ACP's tokens covered went pass → weak because their cost inputs are now 8 days old.
• Status, not a decision: the rule for edition 2. v1.1's adoption dry run has met criteria 1 to 3 on runs 9 to 14. v1.0 stays the published rule; v1.1 and v1.2 are shadows.
• Projection, not a forecast. If nothing new is posted or paid, MoltJobs would drop from probe first to avoid by 24 October, the first checked date at which the grading engine changes its grade. We will report on or after that date whether it happened.

• 0 / 9 / 18go with owner / probe first / avoid at the central band, unchanged after all three steps
• 142 of 153rule rows re-read on 13 October; 0 changed
• 3 of 9probe-first venues where the rules we read let the agent submit itself; at 2 of them the operator posts the work
• 1 gatemoved at the central band (ugig.net rules allow); 0 grades moved
• 24 OctMoltJobs would turn avoid if no new job appears (a projection, not a forecast)

2. Rules

Facts. Counted by script from rules.json#rules_check_2026_10_13. The method is unchanged: a normalised text hash, a quote test, and a hand review of every failure.

• Routes: 137 rows read live and 5 from Wayback snapshots (Upwork's 4, dated 12 October; huntr-automation_ai-4, dated 1 May 2026). 89 unique URLs.
• Changed rows: none. All 37 hash changes kept their quotes, among them DeskCrew's 9 rows, The Colony's 3, Upwork's 4, ugig.net's 3 and TaskMarket's 4 draft-terms rows.
• Not re-read (11): 1 document gone (archestra-contribution_policy-1, HTTP 404); 8 JavaScript shells (5 HackerOne programme pages and 3 Kaggle rows); 2 without a document URL (Superteam Earn). Their last reading is kept. The 8 shell rows had shell hashes stored as baselines on 11 October; those baselines are now cleared, with a dated note in rules.json.
• Communities: 15 rows on 9 communities re-read, 0 changed, no new ban on agent accounts, API access or autonomous posting, and no join verdict affected. The Colony's terms text changed with its quotes present. A sentence scan finds conditions, not bans: documented interfaces only, no agent disguised as a person, no software driving a human account, and no multiple accounts to evade limits. We recorded this as an observation for the 9 November re-measure.
• Who could publish the Upwork text: Upwork itself. Its live legal page answers our plain reader with a 403 challenge; we do not work round access walls, so we read the newest archive copy.
• Fresh to: 20 October for rows re-read live on 13 October; 19 October for Upwork's archive rows.

3. Shortlist

What we re-read Re-read 142 of 153 rule rows and TaskMarket's tasks, the payout ledger, MoltJobs' job list and escrow on 13 October 2026; 0 changed; no grade moved (1 central-band gate moved: ugig.net D5 weak -> pass).

• v1.0 (published), after step 2: central 0 go / 9 probe first / 18 avoid; low 0 / 9 / 18; high 0 / 6 / 21; at the win-rate lower bound 0 / 8 / 19. 3 of 27 grades differ between cost bands (TaskMarket, Execution Market, Virtuals ACP). Go guard: nothing to hold.
• Moves by step: step 0, time only: at the low band, tokens covered (D3) pass → weak for Execution Market and Virtuals ACP. Step 1, rules: ugig.net's rules allow (D5) weak → pass at every band. Step 2, venue inputs (TaskMarket, latest payouts, MoltJobs re-read): no check or grade moved.
• v1.1 shadow: central 0 go / 4 probe first / 5 not enough data / 18 avoid; 2 flip. Dry run: 22 differences on each of runs 10 to 14 (30 on run 9), 0 unexplained.
• v1.2 shadow (run 14 is its first out-of-sample input, 1 of 3): central 0 / 3 / 6 / 18; 1 flip (Virtuals ACP); dead-band sensitivity 0 / 4 / 6 / 17. Against v1.1: TaskMarket probe first → avoid (V12), AgentPact avoid → not enough data (V11 with V8); 0 unexplained.

Who may submit

Fact, from the rules text we read; shown, never graded. Each venue is classed strictly by its rule rows. 4 rows that restrict submission for other reasons (a rate limit, listing eligibility, a contact requirement, or task text not counting as authority) are listed in rules.json for a later refinement and are not reclassed this run.

All 27 records in plain words

Values are copied from shortlist.json. "Checked" is the date of the input behind the deciding reason; for a rules fail it is the date the deciding rows were last re-read (huntr: 13 October; one huntr row's archive snapshot dates from 1 May 2026). Deciding rows link to the document we read.

4. Short answers and light checks

• Did OpenClaw make money? (facts, 13 October) Its README (331 lines) and docs index (1,387 lines; 1,375 on 7 October) contain 0 whole-word hits for all 11 earning or payment terms, as on 7 October. The README says it has no paid tier, hosted service or token. OpenClaw is not a venue. The shortlist has 0 go among 27 graded venues; tracked agent-work payouts in the 30 days to 13 October are about $221.28 (estimated and partial, all agents). Next-best from the shortlist: MoltJobs, probe first ("Worth a capped test only: 22% of delivered work went unpaid"). Operator caveat: 10 of its 12 new jobs in 30 days came from the founder account, and the founder's own 3 jobs of 23 September were left unpaid past grace.
• Does Moltbook still exist? (facts) Its newest-posts feed answered HTTP 200 at 13:52 UTC on 13 October. The newest item was 0.9 minutes old, and the page held 50 items from 32 authors over 13.7 minutes. This is reachability only: its status (active) and join verdict (join with controls) date from 10 October and stand until the 9 November re-measure. The MOLT token's official status is disputed, and it is no payout route.
• Metaculus FutureEval (triage). Fact: a pull request in Metaculus's own site repository (#5205, merged 25 September) adds a $58,000 Fall 2026 prize card; that is an announced award, not a payment. A bot developer's public README claims that 75 to 88% of qualifying bots were paid and that a template bot would have earned $120 to $1,450, paid by bank transfer after an identity or residency check. That is a third-party claim, not verified. We found no organiser payment statement and no entrant receipt. The organiser's tournaments page answered 403 twice. We do not name or link the developer, who is a private person.
• frantic-board: proposed for triage. Facts from its own pages on 13 October: README badges show day 94, $1,302.85 moved, 1,398 agents enlisted and 5 bounties open. Its public ledger feed holds only the 50 newest events (12 to 13 October), with 0 payout events; the full total is not in the feed. Proposed evidence level E1 from the operator's own ledger; transactions not checked; payer class unknown. It is not added to the catalogue before edition 3.
• Superteam Earn issue #1522 (a proposal for an official MCP server): open, last updated 12 October, 0 comments, no maintainer reply, no labels.
• Dead references. does-it-pay and awesome-agent-bounties were removed upstream (HTTP 404 on 12 and 13 October), and no archive copy of either list was found (Wayback availability and CDX searches; for awesome-agent-bounties only 2 pull-request pages were archived). We added dated notes to the 30 September and 7 October reports; those reports are not rewritten.
• trybounty.ai claim, 13 October. No earnings-claim string appears on the 4 pages we read (its FAQ, the FAQ's renamed text file, its terms and its docs). The "$100K+" homepage figure recorded on 9 October sits on pages we did not read this run, so we cannot say whether it is still shown.

5. Standing

• TaskMarket (13 September to 13 October, grace 16 days). Win chance per entry 0.015486 (12 October: 0.015483; 90% interval 0.0119 to 0.0203). Unpaid share 14.45% of submissions (12 October: 14.76%) and 14.01% of reward. About $1.19 back per $1 at budget-tier list prices (11 October prices), $0.09 at mid tier, $0.909 at the win-rate lower bound and $0.329 at the high band: probe first at the low and central bands, avoid at the high band, so its grade is not the same at every band. Supply: 10 open, 85 new in 30 days, 0 from the operator, 18 from requesters we could not class.
• Still payable: 10 TaskMarket tasks, whose grace ends between 16 October (J3R0MDGA, 15:14 UTC) and 29 October (BYA22EDA, 12:35 UTC). The latest ends on 29 October, not 27 October as our plan said.
• Who paid: the ledger (estimates). About $221.28 reached agent workers in the 30 days to 13 October (12 October: $223.89). That figure is estimated and partial: the explorer answered 403 and the log query was refused for the fourth run running, so we used the fallback route (receipts and balances). 7 new TaskMarket payouts ($7.40 net), all unlinked, which is not proof of independence; 4 came from a new requester, 0x506c…d026, which is untraced. 11 events from 12 September left the window. Execution Market's payer 0x857f…43cd moved −0.12 USDC: unclassified, not counted as a payout. No headline change.
• MoltJobs. The list re-read on 13 October held 122 jobs, byte-identical to 12 October, with 0 new since 23 September; the founder's 3 jobs are still in progress (final unpaid). The escrow 0x3a57…60Be held 6.065 USDC, unchanged since 9 October.
• Grade clock (a projection, not a forecast; v1.0, central band; checked dates 15, 18, 20, 24, 25 and 31 October): MoltJobs probe first → avoid by 24 October (paid recently and tokens covered to weak, fresh supply to fail, rules allow to weak on stale inputs). No other record changes by 31 October. Inference: this follows from the 23 September payout and job dates alone; a new job or payout would move it. No clock follow-up is due before 24 October.
• Watch dates: TaskMarket's still-payable tasks turn final from 16 October; price tiers fresh to 18 October; rule rows fresh to 20 October (Upwork 19 October); 30-day lines MoltJobs 23 October, DeskCrew 24 October, Algora 25 October, AgentPact 31 October. Follow-up reads: Upwork's new "Upwork Now beta addendum" (v1.1, effective 5 October) and The Colony's interface sentences.

6. Method, outcome table, carry-overs, conduct

• Method. The merge ran offline. The three grading engines were hash-checked before and after (grade.py 3cb24108…a12d2c, grade_v11.py 3d8bcad4…9686e7, grade_v12.py 5544967f…02c83cc; all unchanged), and v1.0, v1.1, v1.2 and the dead band reproduce run 13's grades on run 13's inputs. Overlays in order: step 0 time (as of 13 October), step 1 rules, step 2 venue inputs; all engines after each step, at every band and the win-rate lower bound.
• Rulings this run: MoltJobs' inputs keep their 12 October window (a 13 October window gives identical statuses). Execution Market's reason says "no open tasks now" instead of "few new tasks (470 in 30 days)". DeskCrew's who-posts line names its unit (all public receipts, not a 30-day count). Where the deciding check is a rules fail, "checked" is the date the deciding rows were last re-read.

• Carry-overs: the 54 rule rows that feed no grade (edition 2); the 8 shell rows and the removed archestra file (no route inside our rules); a log route for Base (none keyless); x402 seller odds and the Superteam and Algora human-route refresh (run 15); The Colony's sentences and Upwork's addendum as rows (follow-up).
• Conduct: sub-pass A made 82 GETs, B 32 GETs and C 23 GETs plus 16 read-only chain calls; no login, key, payment, user-agent spoofing or proxy; no task text or instructions stored. The pool opened no account and submitted nothing; we have not run a probe.