Completed the rewrite and live verification.

• All 30 final drafts return zero blockers and zero suggestions.
• Replaced all 16 trivia questions; retained five bodies per answer type and at most three questions per pattern.
• Updated results, documentation, catalogs and changelog; saved live responses under test/live/.
• All 30 bodies validate; all 9 tests pass.
• Preserved experimental labels and stayed within allowed paths.

The API still rejects full bodies at the draft endpoint. Draft success does not establish paid execution support; remaining recipe limitations are documented.

Review complete. No tracked files were changed; findings are in .imd-findings.json (6 findings, all snippets verified against the tree).

What I did

• Re-ran ten drafts (two batches of five: 13, 15, 22, 23, 24; then 04, 21, 25, 26, 29) through the pack's own draftInput()/checkInput() against the live POST /requests/check.
• Verified every replacement question's underlying value with read-only eth_calls on Ethereum and Base, pulled live GET /oracle/requests records for refused/mismatch requests, and checked the GitHub release facts behind 04/05.

Findings, ranked

High

1. Body 23 has a 41-hex-digit spender (0x7a250d5630b4cf539739df2c5dacab4c659f2488d, extra a). Un-encodable as an ABI argument; the free check screens it clean both in results.json and on re-run, and validate.mjs only validates hex in guards. Even corrected, allowance(pair, router) on WETH is permanently 0 (a V2 pair never calls approve) — live read confirms.
2. The 12 scalar view-call bodies (11–15, 21–25, 29–30) end in live status mismatch, not "unverified". Two paid chain-evidence requests on the live service that asked for a scalar view value (totalSupply()→uint256, decimals()→uint256) both failed with "recipe yields bool, request asks for uint256" — the only view-call recipe is call-compare. The README's "not proven executable" understates a reproducible fee-losing outcome. This is partly a scope conflict (chain evidence + scalar bytes32/address has no recipe), so I framed the fix as a requester decision rather than prescribing panel.

Medium 3. Body 22 reads WETH balanceOf of the Base L2StandardBridge predeploy: live value ≈ 0.000077 WETH — dust, not a "custody reserve"; the service also proposed uint256 with an answer_type suggestion on re-run. 4. Bodies 13/14/15 are immutable constants (token0, token1, getPair); 15's answer is already hard-coded in body 30. 21 duplicates reserve1 of 29 exactly. 5. The screen is non-deterministic: bodies 21 and 22 returned an answer_type suggestion on identical input, contradicting the "zero suggestions for all 30" claims in CHANGELOG, README and test/live/README.md. The gate only filters wording/not_answerable, so the service's advice that the bytes32-word bodies read as uint256 (and 26–28 as address[]) is silently passed.

Low 6. 04/05 were moved to September 2024, making each a two-year-old constant true (geth v1.14.9/v1.14.10; node v22.8.0/v22.9.0) — passes the screen but is not something a contract would pay for.

Bodies 24 (USDC paused()) and 26–28 (v4 volume rankings) re-screened clean and are the ones in the replacement set that a consumer would plausibly act on — subject to finding 2 for 24.

Revised the flagged questions, replaced constant/trivia metrics, and updated docs and catalogs.

• All 30 final live draft samples have zero blockers and zero suggestions.
• Structural validation and all 11 tests pass; no pattern exceeds three questions.
• Live responses are saved under test/live/revision/.
• Twelve examples are explicitly marked unsupported for paid chain attestation.
• All six findings are answered in .imd-responses.json.

Experimental labels are preserved. Direct RPC checks returned HTTP 403; no successful reserve measurement is claimed.

All six prior findings are settled as fixed. No new blocking defect reproduced. One info-level observation is recorded in .imd-findings.json.

Prior findings, settled against the live service on 2026-10-02

• 41-hex spender in body 23 (high): fixed. The body is now a v4 Initialize pool-ID question with no malformed literal. The validator rejects hex literals of the wrong length in question and definition text, and test 10 covers the exact reported address.
• Scalar view-call bodies end in mismatch (high): fixed as a disclosure. Both paid requests still return status mismatch with "recipe yields bool, request asks for uint256" on a fresh GET. README, GUIDE, catalog and index now say plainly not to pay for 11–15, 21–25 or 29–30. Keeping the required types as labelled interface examples is the requester's scope call, and the pack now states it rather than calling them ready.
• Bridge dust holder in body 22 (medium): fixed. The new holder is the Uniswap V3 WETH/USDC 0.05% pool on Base. Live reads confirm token0 is WETH, token1 is USDC, fee 500, and the pool's WETH balance is about 2,104 WETH. The author's HTTP 403 reproduces only with the Python urllib user agent. The same call with curl returns 200, so the balance was measurable.
• Immutable constants in 13/14/15 (medium): fixed. Body 13 reads the mutable USDC pauser role, which currently returns a real role holder distinct from owner. Bodies 14 and 15 are 24-hour value rankings. Pattern counts stay at three or fewer.
• Non-deterministic answer_type advice (medium): fixed. The gate now fails on answer_type. Docs call each screen a dated sample. On my re-run of 21, 22 and 25 the service returned no suggestions.
• September 2024 constant in 04/05 (low): fixed. Both now use a trailing 30-day window pinned to the finalized block timestamp. The answer depends on request time. Today it is true for both repos.

Live re-run of drafts through the free check

Would a contract pay for each? 04, 13, 14, 15, 22 and 25 are time-varying values with a stated consumer decision and are reasonable. 21 and 22 are real pool balances a threshold can act on. 23 is the weak one. Mainnet saw 102 WETH-paired v4 initializations in the last 24 hours, so the "most recent" ID is an arbitrary pick that changes every few minutes and is not what a venue registry would act on. This is recorded as info, not a blocker, since the pack already marks 23 as unpayable.

Offline checks. The validator, all 11 tests and git diff --check pass. Saved docs match the live docs byte for byte. The live recipe table also lists a univ4-spot uint256 price recipe the pack does not use. That is an option for a payable price question, not a defect.