No venue clears every check this week, and the grades hold: no grade moved after 112 of 131 rule rows were re-read and token prices refreshed

1. The answer

• The grades held. At the central cost band, rule v1.0 gives 0 go with owner, 9 probe first and 18 avoid. That is the same after each of the three steps and the same as on 10 October. No grade moved at any band or at the win-rate lower bound. 3 of 27 records flip between cost bands (TaskMarket, Execution Market, Virtuals ACP), so no venue's grade is named in our headline.
• What moved, and at which step. Only gates moved, and only in step 1 (rules re-read). Rules allow (D5) went from weak back to pass at dealwork.ai, opentask.ai and toku.agency in every band. HackerOne's and x402's D5 would have gone stale on the date alone (step 0); the re-read kept them at pass. Step 2 (venue inputs) and step 3 (prices) moved no gate under either engine.
• Rules re-read. 112 of the 131 rule rows behind the grades were re-read: 0 changed, 17 were unfetchable and 19 were not re-read. 46 of the 50 rows that decide D5 freshness were re-read. One D5 gate still rests on a row older than 7 days: ugig.net's, read 2 October. Its pages now sit behind an HTTP 402 pay-per-crawl wall.
• Prices. On 11 October 2026, list prices were unchanged except the mid-tier output price: $7.25 → $6.75 per million tokens (−6.9%). Budget ($0.10 in / $0.50 out) and frontier ($10 / $50) are unchanged. Mid-tier returns rose by 1–3%. Budget-tier returns did not change.
• DeskCrew ticket 465 ended unpaid: still undecided 17.1 hours after the 7-day grace. Its unpaid share (D4) went 23.5% → 24.4% and stays weak.
• The 0x6349…4d24 trace was not possible: no keyless address-history route worked (Blockscout 403, routescan "chain not supported", Base eth_getLogs 429). Its 5 payouts ($18.50 net) stay unlinked: untraced, which is not proof of independence.

• 0 / 9 / 18go with owner / probe first / avoid at the central band, unchanged after every step
• 112 of 131rule rows behind the grades re-read 11 October; 0 changed, 17 unfetchable
• 3 gatesrules allow back to pass (dealwork.ai, opentask.ai, toku.agency); 0 grades moved
• −6.9%mid-tier output list price on 11 October ($7.25 → $6.75 per million); budget and frontier unchanged
• 0 / 17 / 6open gates one capped probe could settle: yes / partly / no (untested; we have not run a probe)

2. Rules behind the grades

The set that feeds a graded venue's rules gate (D5) is counted by script from the grading inputs and rules.json: 131 rows on 77 unique addresses (the plan said 78; the script's count is used). 124 rows were checked in one sub-pass and TaskMarket's 7 rows in another.

• Re-read OK means HTTP 200 and the quote test run. It includes 2 rows with no stored baseline hash and 1 row re-confirmed by another route (superteam-earn-account_identity-2). Rows not re-read: 131 − 112 = 19.
• Rows not re-read:
  • a Superteam Earn row whose Notion page has no recorded address and whose host was outside the request caps (no verdict effect: unclear, not a freshness row), and the Superteam census row;
  • 17 unfetchable rows: ugig.net ×3 (HTTP 402 x402 pay-per-crawl for a non-browser client; no payment was made, and browser and API access were not tested); Upwork ×4 (the live legal page answered 403; its 3 freshness rows were read on 10 October and stay fresh to 17 October, and need a working route by run 14); HackerOne programme pages ×5 and Kaggle ×3 (JavaScript shells; snippet rows, read as unclear); huntr guidelines (404); archestra's contributing file (now 404; not covering; a catalogue follow-up).
• Changed rows: none. No quote disappeared and no forbid row appeared or disappeared. The 28 hash changes have the quote present. They could not be fully diffed, because earlier raw bodies were not kept; this run's hashes are stored as the new baselines.
• Two venues: TaskMarket's terms still read "Draft for counsel review … not approved or active". DeskCrew's refund conflict still stands: its terms and API say entry fees are not refunded, while its bounties page says fees come back when no answer qualifies.
• Still stale: ugig.net's D5 rests on ugig-net-account_identity-1, read 2 October (9 days old). ugig.net is graded avoid on D1, so its grade does not depend on this row.
• Outside the check: 69 rows feed no gate. 54 are catalogue rows last read 2 October, next read at edition 2 (about 15 October). 15 are community rows read 10 October, next read in run 14 (about 13 October). The 7 triage rows are also next read in run 14. The list on the rules page.
• Standing scope (from today): every row listed under a graded venue's D5 is re-read in run 12, then every second run and every edition run, counted by script.

3. Prices and the tokens-covered input (D3)

List prices on 11 October 2026: the OpenRouter public list, cross-checked on 4 of 4 first-party pages. These are list prices, not costs. Mid is a component-wise median (estimated).

• Model set: run 6's 17 models. One delisted model was replaced by its successor under the plan's rule 3. One model was excluded again because its aggregator price is more than 5% off its first-party price.
• Sensitivities: with no exclusions, mid would be $1.00 / $6.00; with the delisted model dropped instead of replaced, $1.50 / $7.50; taking the higher frontier cached price on the tie, +300% cached.

• TaskMarket's budget move (1.045 → 1.079) comes from the new win rate (0.01363 → 0.01407), which is window roll-off, not from prices. At the win-rate lower bound it is 0.859 at budget tier, so avoid there, as before.
• MoltJobs' D3 date: its pass rests on a venue-side input from 5 October. It turns stale on 13 October unless run 13 re-reads it.
• Cost translations, presets and verdicts were all repriced, and a DeskCrew preset was added to the calculator on Getting started. No economics verdict changed.
• Membership costs were recomputed at the new input list prices, which are unchanged, so no value moved with prices. OpenClawCity's range moved from $2.67–$15.99 to $2.62–$15.74 a day at mid-tier input prices because its newly measured file sizes changed, not because of prices. The others range from about $0.12 to $1.84 a day. Communities pay nothing.

4. Shortlist

Revision entry (shortlist.json revisions[], 11 October). Gate moves against the 10 October publication: D5 weak → pass at dealwork.ai, opentask.ai and toku.agency (step 1, rules re-read). Grade moves: 0. The go guard did not fire (no record was graded go in any band). Transient moves, net none: HackerOne and x402 D5 pass → weak in step 0 (time) and weak → pass in step 1.

• Probe first (central, ranked): MoltJobs, TaskMarket, Execution Market, DeskCrew, Superteam, Bugcrowd, Virtuals ACP, HackerOne, opentask.ai. Operator caveat: the operator or its own accounts post 50% or more of the work at MoltJobs, Execution Market and DeskCrew.
• Flips: 3 of 27, under one third.
• v1.1 shadow (not the published grade): 0 go / 4 probe first / 5 not enough data / 18 avoid at the central band; 2 of 27 flip. 102 record-band differences from v1.0, 0 unexplained. Adoption criteria 1 to 3 are met so far (runs 10 to 12); the edition 2 check is pending.

5. What an owner's probe would settle (untested)

probe-record-template.json is untested: we have not run a probe, and one probe is one observation, not a rate. It maps every weak or unknown gate on the 9 probe-first venues: 23 open gates after this run (the template first mapped 24; opentask.ai's D5 has since been settled by the rules re-read).

• Partly: D1 and D2 (one owner's own view), D3 (the owner's own cost per attempt), and the verifiability of D4 at Superteam, Bugcrowd and HackerOne.
• No: a D4 unpaid share at MoltJobs, TaskMarket, Execution Market, DeskCrew, Virtuals ACP and opentask.ai. A share needs at least 5 to 10 final units.
• Record fields: per entry, the owner's own reference, tokens, estimated cost at the stated list price, owner minutes and outcome; also any payout in the owner's own wallet (with its evidence), fees, KYC requests and stop conditions. A count of task texts with instructions aimed at the agent is recorded, never their text. The gate map per venue.

6. Standing

TaskMarket (11 September to 11 October, without MolTrust)

• Supply: 6 open eligible tasks and 82 new in 30 days (2.73 a day).
• Win rate: p_win_final 0.01407 (90%: 0.0112 to 0.0180).
• Unpaid share: 14.8% of submissions and 14.7% of reward went to tasks that paid nobody. Run 11 had 13.6%; the move is window roll-off, and the 8 final-unpaid tasks are unchanged. If all 9 still-payable tasks end unpaid, the unpaid share would be 23.7%.
• At 11 October list prices: about $0.08 back per $1 at mid tier and $1.08 at budget tier, which is inside rounding; at $1.50 a day, about $0.07 expected per day and a first cash after about 19 days (median).
• Requesters: 0 flags changed. Who posts the work: 0 of 82 new tasks came from the operator or linked wallets; 16 are unknown, including 0x6349…4d24's.

Ledger

• About $213.79 in 30 days (estimated, partial). No headline change.
• 2 new TaskMarket payouts ($5.55, both from 0x6349…4d24). 7 in-window flags were corrected.
• Unclassified flows: Virtuals ACP escrow outflows of −3.0 and −1.0 USDC on 11 October, at the same second of day as earlier refunds; small Execution Market Base outflows. Not counted as payouts.

Log route. Blockscout 403, routescan "chain not supported", and Base eth_getLogs refused on the first call, for the second run running. This bears on the open proposal for indexer access. Options for the pool's manager, as proposals only: approve block-by-block reads, test another keyless explorer, or decide the indexer proposal.

0x6349…4d24 (not traced: route refused). It funded 7 tasks since 6 October (31.0 USDC): 5 awarded to 4 workers, 2 open. Offline checks: it matches no TaskMarket operator address and not MolTrust's; no payee funded it. It stays unlinked, which is not proof of independence.

OpenClawCity (measured 11 October)

Measured on the read-only route its public feed page uses; the site's public key was found on the second script file and never stored. Status: active, with 182 authors and about 188 items a day in the 500-item sample.

• Templated posts from many accounts make up 41% of items under every reading.
• Verdict: not this week, robust under every reading. J3 fails (no legal entity named) and J6 fails (a server-sent setup script and a detached background process). No change from 10 or 4 October. skill.md and heartbeat.md are unchanged since 4 October.

The Colony: its agent files are unchanged since 10 October. No snapshot from on or before 4 October exists, so a diff was not possible (hash only). A person should still review the files.

Watch dates

• MoltJobs: 3 jobs leave grace on 12 October; its D3 input turns stale on 13 October.
• x402 seller odds: stale 16 October.
• TaskMarket still-payable tasks: turn final 16 to 27 October.
• D5 rows and price tiers: fresh to 18 October. Upwork rows: fresh to 17 October.
• D1 lines: MoltJobs 23 October, DeskCrew 24 October, Algora 25 October, AgentPact 31 October.

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

• Engines: the v1.0 grading script (SHA-256 3cb24108…a12d2c) and the v1.1 shadow script (3d8bcad4…9686e7) were hash-checked before and after use and run unchanged. On run 11's inputs they reproduce run 11's grades and gates in every band.
• Overlays: applied in the pre-registered order on run 11's inputs at as_of 2026-10-11, with step 0 (time only) added for attribution. Both engines ran after every step at the low, central and high bands and at the win-rate lower bound.
• D3 inputs: recomputed from the token bands × tiers. At the old prices the merge reproduces every run 11 D3 input and the TaskMarket ratios.
• Rulings and readings: earlier input rulings carried; eight new rulings applied; nine merge readings recorded, none of which changes a gate.
• Raw bodies: the sub-passes deleted their raw page bodies. Rule results, the DeskCrew read and the OpenClawCity feed are validated from the logged hashes, not by re-normalising bodies.

Carry-overs: a working read route for Upwork's legal page and ugig.net's rule pages (by run 14); the archestra contributing file (404); MoltJobs' venue-side D3 input (run 13); the ACP and Execution Market outflows; the 0x6349…4d24 trace (needs a log route); the 54 catalogue rows that feed no gate (edition 2); the v1.2 text (run 13).

Conduct. Read-only public GETs and read-only JSON-RPC POSTs only. No account, key, login, payment, signing or model call. The merge made 0 requests. Rule pages: 75 GETs (74 rule addresses of an 80 cap, 0 retries, no address fetched twice); prices, OpenClawCity and The Colony: 20 GETs; TaskMarket and chain: 41 requests, 1 refused (eth_getLogs 429), no automatic retries. No host outside the cap tables was contacted. The OpenClawCity public key was held in memory only; a scan finds no key-shaped string in any output.