I found three problems and nothing high: two medium and one low. Items A, B and D are met. Items C and E each have a gap the builder can close in the repository. The findings are in .imd-findings.json. No other files were changed: the probe screenshots went to the ignored test/scratch/.

Area	Count
Brief requirements	8 met, 2 missing, 0 faked, 2 after review
Broken features	0 (no wallet, so nothing was signed)
Product on the first screen at 390	yes
taste@1	15 of 17 passed (the two failures are the site's own choices)
Copy findings	0
Share card	yes (title, og:title, og:image)

Requirements

• A, block bad shared top lists — met. checkSharedSpec in web/src/shared-safety.mjs reads the spec's pinned log window and builds the head list. It blocks payment when any address is below 2**80 or the list is shorter than the head. It runs before a registration or shared request is prepared and again before every step. This matches how Registry._list and the spec rules decide.
• B, Agents menu entry, skill file and llms.txt — met.
  • The menu entry is in web/src/App.tsx:183 and links straight to agents/sworn/SKILL.md.
  • The skill has every section the brief lists: deployment and how to verify it, reading facts, DON'T, the four examples, wallet lists (including the 19,522 wallets and the sybil caveat), costs and the Missed reasons table.
  • The root llms.txt has the links, Claude Code and Codex install steps, and a four-criteria Hackathon section.
  • The copies in dist/ are identical to the sources.
  • I could not check the quoted WhitelistCurator.sol lines 32 to 53 word for word, because that file is not in the repository. It is quoted with a source SHA-256.
• C, reads that survive a failing RPC — met. web/src/rpc.mjs falls back across four RPCs, checks the chain, and requires two endpoints to agree on logs.
• C, a card holder never sees a mint offer — missing. While the ownership read is pending or has failed, the page still shows the mint pitch and the sample card (claim.tsx:768).
• D, empty community spec form — met. defaultForm is all empty strings, and the template examples are placeholders only.
• E, one dialog with Prepare buttons, balance, card preview, checkbox, approve only when short, step locks, simulation and step states — met (transactions.tsx).
• E, "the live price, where the IMD goes … and the result" — missing. Only a spec registration shows a result. The price and destination text is the same for every action.
• Contract addresses and chain — met. Chain 1, and the three addresses match deployment.json.
• Keyless and static, no request to api.imdnotdead.fun — met. The only api.imdnotdead.fun mention is a link in llms.txt; the site never fetches it.
• Hosting under the same site name and the IPFS publish — after review.

First screen and probe. At 390 wide the first screen shows the logo on the left, Connect on the right, the headline, the wallet input and the green "Explore wallet" button. The probe fails two checks:

• casing: the uppercase headings "WHY SWORN" and "BUILT BY THE SWARM".
• type-faces: monospace prose.

Both are recorded in DESIGN.md and docs/DESIGN.md as choices that must stay, so they are not findings.

Findings

1. Medium — fidelity: a card holder can see the mint offer (web/src/claim.tsx:768). The page shows "Mint your card now", a disabled "Prepare mint" button and the sample card whenever ownership is unknown. That covers every page load until the card read returns, and stays that way if it fails. The fix is to show the mint copy, button and sample card only when the wallet is known to hold no card.
2. Medium — fidelity: the dialog shows no result (web/src/transactions.tsx:696). After a mint, a fact request or a shared request it only says "All transaction steps confirmed. Refresh wallet facts to see the result." The fix is to read the minted card id or the

The review is complete. Nothing critical or high was found, so the site can be hosted. Three advisory findings are in .imd-findings.json, and the working tree is otherwise unchanged.

What I checked

• Export integrity. Rebuilt the site from the source in web/ into a temp directory. The output matches the committed dist/ byte for byte, so the hosted scripts are what the source says. The index page loads one module script and one stylesheet, both local. No analytics, no request to api.imdnotdead.fun, no seed-phrase or key prompts, no remote code loading. External hosts in the bundle are wallet SDK deep links, the four read RPCs, the explorer, GitHub and X.
• A, kind-3 guard. The deployed Registry returns Missed reason 9 for any list word below 2**80 or a wrong head length (src/Registry.sol lines 449 to 456). The site mirrors this in checkSharedSpec and runs it before every registerSpec and requestSharedFact, again before each approval and the paid call. Unit tests for address(0), 0x…dEaD, short lists and incomplete logs pass.
• B, Agents skill. The nav has an Agents entry linking to agents/sworn/SKILL.md. The quoted WhitelistCurator lines 32 to 53 match the pinned source lines exactly. Addresses, commit, launch number, ABI hashes, the 25 ids, the DON'T section, the four read-only examples, list construction with the 19,522 wallets and sybil caveat, costs and the Missed table are all present. llms.txt has links, install steps for Claude Code and Codex, and the hackathon section. The explorer job link resolves to a real Sworn job record.
• C, resilient reads. Failover across four endpoints with chain-id verification; all actions disable when health is null. A wallet with a nonzero tokenOf sees its card number and no mint button; prepare and the pre-signature check both re-read tokenOf.
• D, spec form. The form starts from an all-empty default and examples appear only as placeholder attributes.
• E, one dialog. A single dialog serves mint, fact request, shared fact request and spec registration. It shows total, balance, live price, the BurnExecutor and Intake destinations, the card preview, the disclaimer, Approve only when allowance is short, and gates each step on ack, previous confirmation and a passed simulation.
• Tests in a real browser. All 50 unit tests pass. The Playwright Chromium run against the export completed with no errors across 40 pages and the fake wallet flows. WebKit was not installed, so only Chromium was exercised.

Findings (all low)

1. The Claim page's "Read the contract README" link goes to a third-party specification branch rather than the pinned source.
2. The read RPC list drops eth.drpc.org from network.json and adds three undisclosed third-party endpoints.
3. The payment dialog restores a ticked disclaimer box from localStorage after a reload.