All 12 unpayable bodies are replaced or retyped, and all 30 drafts pass the live check as bare questions. The last run of scripts/check.mjs gave 30 of 30 judged, no blockers, and no wording, not_answerable or answer_type suggestion. Each body's type equals the type the check proposed in that run. npm run validate and the 12 offline tests pass. No paid request was made.

What changed

• Retyped bodies: There is no recipe for a scalar address or bytes32, so the pack has no body of those types any more.
  • 11/12 became USDC paused() bools.
  • 13 counts Blacklisted logs over 720 hours.
  • 14/15 sum WETH Deposit and Withdrawal amounts.
  • 21 is a v2 pair WETH balance floor.
  • 22 counts Base WETH deposits.
  • 23 counts mainnet v4 native-ETH pool Initialize logs.
  • 24, 29 and 30 are address[] rankings.
  • 25 sums Base USDC Transfer values.
• Near-constants: 01 and 02 (totalSupply() > 0) became Uniswap pool WETH balance floors.
• Further replacements:
  • 09 and 10 (Robinhood block and transaction counts) drew wording on every attempt and have no recipe. They are now Base v4 native-pool and USDC transfer counts.
  • 26 and 28 now use mainnet WBTC and Base USDC, because their old tokens drew wording or an address[] proposal.
  • Several questions (03, 04, 05, 07, 13, 18, 19 and others) were reworded to pass bare. The old 05 drew not_answerable.
• Check script: scripts/check.mjs now sends the bare question instead of pasting definitions into it.
• Validator: scripts/validate.mjs requires the four recipe-backed types and rejects scalar address/bytes32, instead of requiring five of each type.
• Docs and tests: README, GUIDE, catalog and index no longer list anything as unpayable, and the pattern audit has at most 3 per pattern. I added one test that drafts are bare and each type matches the recorded proposal, and fixed one existing assertion that expected definitions in the draft. There is a CHANGELOG entry, and the experimental labels are untouched.
• Live evidence: results.json holds the bare-question checks. A copy, the checks of the previous bodies' bare questions, and fresh docs and health GETs are in test/live/bare/.

Things to know

• The service is noisy. The same text sometimes drew wording or flipped the proposed type, for example 27 once came back as address[]. I chose wordings from repeated samples and re-ran the full check until one run was clean. That run is a sample, not a guarantee.
• I did not fix 03, 18 and 19. Their block and transaction scans still have no listed recipe, and they are documented as a limitation.
• Several thresholds are guesses I did not check on chain: 10,000 WETH in the mainnet v3 pool, 500 WETH in the Base pool, 1,000 WETH in the v2 pair.
• The renamed files are in the working tree but nothing is committed.

The review is complete. The findings are in .imd-findings.json; no other file in the tree was changed. Summary:

Assignment result: all re-typed drafts match the live proposal. I sent the bare question of every re-typed body to the free check at api.imd.fun on 2026-10-03, using the pack's own draftInput and checkInput. The eight required (11, 13, 14, 21, 22, 24, 25, 29) and the other four (12, 15, 23, 30) all returned no blockers, judged true, no suggestions, and a proposed type equal to the body's type at confidence 1.0. I also sampled the three bytes32[] bodies twice each and 03, 18, 19 once: every proposal matched, with no wording or not_answerable suggestion anywhere. The recorded results.json is byte-identical to the saved live copy, every body hash matches, and npm run validate plus all 12 offline tests pass.

Findings, ranked:

• Medium, bodies 18 and 19. Both rank senders by transaction count "including system transactions". A live block read today showed every chain 4663 block carries one ArbOS internal transaction from the address ending in a4b05, and every Base block carries one deposit from the dead...0001 address. That sender therefore has count equal to the block count and is a guaranteed top entry, and effectively always first on chain 4663. This is the near-constant answer the job's item 3 asked to remove, and body 03 already excludes the same transaction type. Fix is to exclude the system type or deny-list the system sender.
• Medium, bodies 03, 18, 19. They keep chain evidence with a block-enumeration metric, but the saved live docs list no recipe that enumerates blocks, and say a recipe that cannot yield the type is refused before a vote. The free check cannot see this since it rejects window, definitions and guards. README line 43 and GUIDE line 47 still say these three have no listed recipe, so item 4 is met only in wording. This was disclosed in the CHANGELOG, so it is a scope call for the requester: re-type the three onto a listed recipe or accept them as not payable.
• Info, bodies 26 to 28. The bytes32[] proposals matched in all eight samples but at confidence 0.38 to 0.58, against 0.98 to 1.0 for every other body. The quality gate compares the value only, so an unchanged pack could fail a future re-run. No failing input exists today.

The draft screen's default 24-hour window differing from the bodies' 1-hour and 720-hour windows is already documented in README and GUIDE, so I did not report it.