# Redaction log — public copies of the five frozen captures (teardown #1)

**This log is published with the claim-verification record for teardown #1, which cites these copies.** Every removal and every transformation applied to a copy is recorded below. The sealed originals remain unaltered and unpublished.

These are **redacted derivatives**, not the originals. The sealed originals are retained unaltered in our own records, are **not published**, and each is identified here by its SHA-256 so it is clear which capture each copy was made from — not so that you can reproduce it: the originals are not released, so the derivation cannot be re-run from outside. They are named here by hash rather than by filename on purpose: one sealed filename carries the marketplace supplier's name, and this log is published. **The sealed originals are held internally and are not released to anyone.** A subject company that exercises the right of reply receives the redacted public copy of the capture it appears in, and the reply is published alongside it; it does not receive the sealed original, because the sealed captures carry third parties' personal data — reviewer names, their email addresses and their tracking keys — which is not ours to hand over. The originals stay in our control, unaltered and identified by hash.

**Method.** Each public copy is made by *deleting* content from the sealed original and, where an individual or a supplier is named, substituting a labelled placeholder. Nothing is rewritten, re-flowed or re-typed: no claim sentence, price, product description, specification or page structure is altered. Removal of executable content, redaction of identities, and neutralisation of remote loads (the load hints deleted from the markup and the content-security policy injected) are recorded as three separate kinds of change. Each public copy also carries a one-line provenance marker (HTML copies) or caption band (images) naming it as a derivative, and each HTML copy carries a visible note band saying that images, stylesheets and fonts are not loaded from the original sites.

| # | Public copy (file on this site) | What was captured | Removed — category, count | Other transformations | Sealed original SHA-256 | Public copy SHA-256 | Bytes (sealed → public) |
|---|---|---|---|---|---|---|---|
| 1 | `raoptics-homepage-2026-09-28.public.html` | raoptics.com homepage as served, 2026-09-28 08:20 UTC, HTTP 200 (carries the "100%" claim in the closed FAQ panel) | 114 script blocks; 3 inline event handlers; 0 `javascript:` URLs; 27 customer/reviewer names, each replaced by `[name redacted]` in both places it appears (the review credit line and the avatar `alt` text) — 54 placeholder substitutions. Inside the removed script blocks only: 5 email addresses, 2 error-tracking DSN keys, 2 agency staff names — counted here, removed with those blocks, none surviving (scan: 0) | provenance marker inserted; 11 remote load hints (`<link rel=dns-prefetch\|preconnect\|prefetch\|preload>`) deleted; content-security policy injected; visible "no remote loads" note band inserted | `9958aa33735281952ac655e0275d744b9df781f219f51b48f2fe8075668580af` | `558f00ba12fe6fc66a0c6da79a82e41c7931a52db872ac9fd8029bb01015db6e` | 706,649 → 406,257 |
| 2 | `swanwick-homepage-2026-09-28.public.html` | swanwicksleep.com homepage as served, 2026-09-28 08:55 UTC (carries the "99%" claim and the "Independently validated" badge) | 88 script blocks; 4 inline event handlers; 0 `javascript:` URLs. Inside the removed blocks only: 1 person-naming affiliate/referral code (2 occurrences, in two navigation links) — counted here, removed with those blocks, not surviving (scan: 0) | provenance marker inserted; 6 remote load hints (`<link rel=dns-prefetch\|preconnect\|prefetch\|preload>`) deleted; content-security policy injected; visible "no remote loads" note band inserted | `d0eb3bedb224dbc0139f06a192938617bb14d75ea0c0153275e83e574907ffef` | `4a64fc29bc85201868f8101c06a14517f81338f7852a1b3976b6ed37b90de703` | 602,585 → 417,127 |
| 3 | `wholesale-listing-2026-09-28.public.html` | wholesale-marketplace factory listing, desktop page as served, 2026-09-28 12:51 UTC (listing title spec, published price tiers, minimum order) | 34 script blocks; 1 inline event handler; 24 `javascript:` hrefs (removed — no script runs from this copy); **the supplier identity, masked as `[supplier]` in all 126 places it occurs** (marketplace subdomain in every href, company-name URL path, image `alt` text, registered company-name string, store/account username); **the supplier's marketplace account id, masked as `[supplier-id]` (38)**; **the listing's own id, masked as `[listing-id]` (43, plus 30 occurrences of the same listing's id on the supplier's international store)**. No email, phone number, token, affiliate code or personal name was present in this capture | provenance marker inserted; 8 remote load hints (`<link rel=dns-prefetch\|preconnect\|prefetch\|preload>`) deleted; content-security policy injected; visible "no remote loads" note band inserted | `7578b5028b718f4c818e2a29a79d95e751445034bc117084c5e89ccd1fa1cc48` | `e53dfdfed4b3c2b95cf54b0a2d2ab13d0680770d6e14bceb96ec2547eb19c019` | 257,771 → 205,732 |
| 4 | `meta-adlibrary-swanwick-us-2026-09-28.public.png` | Meta Ad Library result page, country United States, keyword "swanwick", filter active ads, "~48 results", three live creatives with start dates | 1 individual advertiser's account name + 1 profile avatar (a third-party account, boxed and labelled `[name withheld]`). No unrelated advertiser appears in this capture, so no crop was applied | provenance caption band added below the captured region; captured region otherwise untouched (a still image makes no network requests) | `2ef56faed51441d47d0443dd8f3f43f29f7a279bfa7ec32653b13f30cea5a5a9` | `8bbb6f407bf1860495c395a03c309258bc823a756f0212c88700b26e9e1d8e53` | 172,869 → 184,026 (1280×900 → 1280×962) |
| 5 | `tiktok-library-country-list-2026-09-28.public.png` | TikTok Commercial Content Library, "Ad target country" list open — the United States is not among the options | **Nothing.** No personal data is present in this capture | provenance caption band added; captured region untouched (a still image makes no network requests) | `e784eae6add308a915971441769fe03c98cc2f21df5db2d53a2597f5325b95b4` * | `b2ca44262e398b5cb435eda6073fda7da75b95938058906fb81764103f9363d5` | 103,836 → 118,205 (1280×900 → 1280×962) |

\* Hash computed at redaction time. The two image captures were stored without a recorded `.sha256`; when and how each was captured is recorded in our own records. The three HTML captures each still match their recorded `.sha256` (re-checked immediately before these copies were built, and again after).

**The public copy of capture 3 was renamed.** Its previous filename carried the supplier's name and the listing's id; it is now `wholesale-listing-2026-09-28.public.html`, a label carrying neither. The old filename is not printed here for the same reason it is no longer used, and it appears nowhere on this site; the old file is deleted rather than published beside the neutral one. The **sealed** original keeps its existing filename and hash and is untouched.

**No claim text was altered — per file.** Capture 1: the quoted "100% of blue light" sentence, the "$174.00" price and the "Clyde Sunset" product name read identically in the sealed original and the public copy. Capture 2: the quoted "99% of the blue light" sentence, "Independently validated" and "$129.99" read identically. Capture 3: "99% Blue Light Blocking", the three published price tiers, the minimum order and the supplied specification read identically. See *Proof* below for the method and result.

## What stays, and why — three decisions of record (2026-09-28)

- **Companies, brands, shops, product names, prices, product copy, claim text and page structure stay.** That is the evidence. For the two brand captures that includes the brands' own marketing content.
- **Named people appearing in a brand's own published marketing copy stay — our decision, applied without re-deriving those captures.** A brand's own advertising content is what this piece examines: 6 public figures in the celebrity showcase and 1 named expert endorsement in capture 1; 4 named professional athletes in capture 2. This was raised while the copies were being prepared, and the decision of record is that they stay as they are. Nothing in those two captures was re-derived for that decision.
- **The marketplace supplier is masked and the file renamed — our decision.** Our published promise (the site's claim piece) is that the factory behind the wholesale listing is *named nowhere and never linked*. Publishing capture 3 must not undo that, so every form of the identity is masked: the name in the marketplace subdomain of every href, the company-name URL path, the image `alt` text, the registered company-name string, the store/account username, the supplier's marketplace account id, and the listing's own id. The marketplace platform's own name is untouched (it is the platform, not the factory), as are the supplier-attribute lines that are evidence and not identification — the staff count, the export-experience and import/export-rights badges, and the customs HS code.
- **Remote loads are neutralised — later extended.** Each HTML copy carries a strict content-security policy — `default-src 'none'; img-src 'none'; style-src 'unsafe-inline'; font-src 'none'; form-action 'none'; base-uri 'none'` — injected as the **first** element of its `<head>`, so the copy loads nothing from any captured site, and none of the page's own forms can submit to one. `form-action` and `base-uri` are named separately because CSP inherits **neither** from `default-src`: without them a reader who submits one of the page's own forms (4 in capture 3, 6 in capture 2, 2 in capture 1) could still be navigated to the captured site, and a `<base>` element in the markup could still redirect every relative URL on the page. CSP does not govern load hints either, so every `<link>` element pointing at a captured host was **deleted from the copies** — `dns-prefetch`, `preconnect`, `prefetch` and `preload`, including `rel` values that combine several of those tokens: 11 in capture 1, 6 in capture 2, 8 in capture 3. Left in place, a hint would still make the reader's resolver look up a captured hostname (`dns-prefetch`) or open a connection to one (`preconnect`), however complete the policy was. A load hint carries no visible text and none was evidence, so this is recorded as a transformation, not a redaction, and no claim text, price, specification or page structure changed. The captured markup and the inline `<style>` blocks are unchanged, and a visible note band tells the reader why the page looks plain. A still image makes no requests, so no policy was added to the two image copies.
- **Honest limit, recorded rather than hidden.** Masking removes the name; it does not make the listing un-findable to someone determined to search the marketplace. Our promise is that we do not name or link the factory — not that a reader cannot find it. For the same reason this log identifies the sealed originals by hash and not by filename, and does not print the platform product/SKU identifier for capture 3 that remains in the markup (see below).
- **Retained, and why.** The marketplace's own product/SKU identifier for capture 3 remains in the markup, because the claim review quotes and checks it, and masking it would break that check. It is the platform's product number, not the supplier's name or account, so it does not identify the factory directly; masking it is available and is a one-line change; we have not taken that decision. The digits are not printed here, because this log is published and printing them would restore the findability the rename and the masking remove.

## Proof — the checks, and their result

Every check below was run on the rebuilt files, against the sealed originals, after this log was written. All checks pass: the checks were run against the published copies and the captures they came from, and their result is recorded below.

1. **Independent re-derivation.** Re-running the recorded transformation rules from scratch on each sealed original reproduces a byte-identical public copy (SHA-256 compared), and a second derivation into a fresh directory reproduces all three byte-for-byte again. The public copies contain nothing that is not the original minus the recorded deletions and placeholders.
2. **Whole-document visible-text equality.** For each HTML capture: strip the executable content from the sealed original, apply the recorded name placeholders and the capture-3 identifier masks, strip all tags and comments, collapse whitespace — and compare with the same extraction from the public copy (with the added note band set aside, since it is an addition). **The two are identical character-for-character for all three files** (174,834 / 34,328 / 11,136 characters of visible text). **How that count is taken, so it can be reproduced as defined:** it is the length of the text remaining after comments and tags are removed and every run of whitespace is collapsed to a single space, with the ends stripped — a count of the *stripped document*, not of what a reader sees rendered. It therefore includes the text inside inline `<style>` blocks, because removing tags leaves that text behind; it excludes everything removed with the executable content (scripts), because that is stripped before the comparison. Any other extractor will return a different figure unless it strips the same things in the same order. This is the general proof: every character of visible text is preserved, and the only substitutions are the recorded placeholders.
3. **Quoted-string equality.** Each string the published claim piece leans on was located in both and compared: identical bytes, identical count and identical surrounding text. Result: **all 26 anchors pass, 0 fail** (7 in capture 1, 6 in capture 2, 13 in capture 3).
4. **Name-leakage check.** Every reviewer name was extracted from the sealed original and searched for in the public copy: 9 distinct names, 27 occurrences, **0 remaining in the public copy**.
5. **Supplier-leakage check (capture 3).** The supplier token before masking: **126 occurrences in the copy before masking** (146 in the sealed original, the difference being occurrences that sat inside removed script blocks). After masking: **0**. Listing id before: 43 → **0**; supplier account id: 38 → **0**; the same listing's id on the international store: 30 → **0**. Nothing carrying the supplier's name or account id remains anywhere in the copy, including its filename, its provenance marker and its note band.
6. **Personal-data scan of the public copies.** 0 email addresses, 0 error-tracking keys, 0 phone numbers, 0 analytics or pixel IDs, 0 affiliate/referral codes, 0 visitor/session IDs, 0 `<script>` tags, 0 inline event handlers, 0 `javascript:` URLs.
7. **A reader's browser contacts nothing — browser check, not an inspection of the markup.** Each copy was served exactly where the site serves it (`/evidence/…`) and loaded in a real Chromium with the network recorded. Every off-origin entry in the recording is a blocked attempt — status 0, no body, no HTTP exchange — and **0 of the page's images decoded**: capture 1, 149 off-origin entries, 0 of 83 images decoded; capture 2, 390 entries, 0 of 191 images; capture 3, 115 entries, 0 of 86 images. Each copy was then asked to submit a form, which is what `form-action 'none'` is for. On the wholesale copy, clicking the Send control of the marketplace inquiry form produced a **`securitypolicyviolation` with `violatedDirective=form-action` and `blockedURI=https://www.made-in-china.com/sendInquiry/…`**; the copy's own document did not navigate (its tab still shows the same URL and the same four forms), and the recording contains one document entry — the copy itself — and no new navigation. That form carries the capture's own `target="_blank"`, so the click does leave one empty popup tab behind: that is the browser allocating a window for a navigation it then refused, not a page that loaded. The same form, submitted with the popup taken out of the picture (its target set to `_self` in the live DOM only — nothing on disk was changed), blocked identically, and the copy's four forms cannot submit anywhere, including the two whose action is on the copy's own origin. **Control, so the harness is not simply blind:** a page carrying the same policy with a form aimed at a second local origin that logs every request it receives drew **zero** requests, while the identical page without the policy drew the POST. (The companion subresource control — six subresources aimed at that logging origin, none fetched from the policy page — was run in the same session.) Screenshots of all three copies are held in the lab's own records; the note band is present and visible at the top of each page.
8. **Sealed originals untouched.** Each sealed file was re-hashed after the copies were built: 13 of 13 recorded `.sha256` files still match, no sealed file or `.sha256` file was modified, and no sealed original exists byte-for-byte anywhere on this site.
9. **Published-directory scan.** No file published on this site — copies, this log, the checksum file or the README — carries the supplier token or the listing id.
10. **Load hints gone, policy first in `<head>`.** In each copy **0** `<link>` elements carry `rel` `dns-prefetch`, `preconnect`, `prefetch` or `preload` (11 were removed from capture 1, 6 from capture 2, 8 from capture 3), and the content-security policy meta is the **first** element inside `<head>`, so nothing is parsed before it. Nothing is published that fails either test, and both are re-checked independently against the copies as published. The checks need no third-party packages: only rebuilding the two image copies needs one, which the build asks for by name instead of failing part-way.
