reviewed-domains.json is the canonical offline catalog. scripts/build-rules.ps1
validates it, builds browser rules, and exports the same catalog to
rules/reviewed-domains.json. scripts/export-cyberassist.ps1 -CyberAssistRoot <root>
validates the export and copies it into CyberAssist source with a backup. Add
-Check to verify equality without changing files. CyberAssist embeds that snapshot
in its core assembly; a build is required to use a new catalog.
The structural review contains 94 ad, 80 tracker and 73 unverified domain records.
There are zero independently evidenced malicious domains in this v1 catalog. The categories come
from inherited local lists and are not current reputation assertions. Former malware
entries are retained for observation only and generate no browser blocking rules.
review-decisions.json records category decisions and the two excluded miner filename
patterns. Six ad URL heuristics remain in browser-patterns.json, with their original
third-party and resource-type conditions; they are never exported as domain indicators.
requested-ad-hosts.json separately records explicit user ad-block requests, including
the request date and reason. The 12 November exact pxl.wheelroot.com request is
retained; the 14 November user report adds wheelroot.com and all its subdomains,
including page navigation. The report is not independent malware verification.
These entries are never exported to the
reviewed-domain contract or threat list. They use priority 4 in the ad ruleset, like
custom request filters: community exceptions cannot undo them, while explicit
Always Allowed and custom request exceptions still take precedence. Turning off
the ad layer or pausing protection disables them. The builder accepts canonical DNS
hosts with explicit exact or subdomains scope, rejects duplicate or malformed
entries, and performs no lookups.
The separate passive-domain-review-2026-10-18.json records a later passive search of
all 73 unverified entries. No entry met the user-selected recent dated-evidence
threshold, so it adds no blocks. Four explicit vendor threat descriptions were undated.
This research record is not a version-1 catalog or build input; it must not be imported
as one. See the complete findings for scope and limits.
Each record has a canonical ASCII hostname, exact or subdomains matching (the latter
includes the apex), its category, source ID and original rule ID, and original browser
priority, party scope and resource types. Parent and child entries are retained when
present in the source: matching reports every applicable entry and its source scope.
Exact duplicate domains or source rule IDs are refused, rather than silently merged.
Source records name the local pre-review rules file and its SHA-256. These identify
the bytes reviewed, not the domain's operator, current reputation, or an upstream
provider. The initial files are preserved in the review backup. reviewedAtUtc dates
only the structural review. lastVerifiedUtc and expiresUtc must be null in version 1:
there is no dated reputation evidence to supply. This version rejects malware categories,
confidence claims, fabricated verification dates and unknown fields. Future evidenced
malware uses the separate browser-only contract below; a download alone is insufficient.
Both the PowerShell importer and CyberAssist validate a closed schema, duplicate JSON
properties, UTF-8/JSON depth, a 1 MiB byte limit, at most 5,000 domains, provenance shape,
category/source relationships, bounds, label lengths, IDN encoding, special/reserved
names, malformed suffixes and browser scope. Production records must be lowercase
ASCII without a trailing dot or URL/filter syntax. Query matching normalizes case,
a single root dot and IDNs, then compares whole DNS labels. It never strips www or
looks up an address. Common public-suffix-only shapes such as co.uk are rejected;
there is no downloaded public suffix registry, so syntax checks cannot prove that a
name is registered, active, controlled by its previous owner, or currently malicious.
Old online/CSV build switches now fail explicitly. No URLhaus cache is activated by this workflow. Each build is serialized, stages every output before replacement and rolls back on publication failure. Rebuilds do not refresh the structural-review date. The digests detect changed snapshots; they are not signatures against a compromised local administrator or source tree.
Run pwsh -NoProfile -File tests/rules.test.ps1 for synthetic offline import, rejection,
scope, rollback, export and drift checks. These checks do not resolve or request any
catalog hostname and do not establish live browser enforcement or reputation.
community-lists.json (reviewKind: community-filter-snapshots) records each filter-list
snapshot stored under community/: id, title, file, HTTPS source, updates, license,
downloadedAtUtc and sha256. The builder refuses unknown or missing fields, duplicate ids,
files outside community/, non-HTTPS sources, an updates other than download or release,
future download times, files over 8 MiB, and any file whose SHA-256 differs from its record. It
never downloads anything: a new snapshot is placed there by a person, and its record is updated
deliberately. The records also reach rules/community-info.json. For a download list, the
extension downloads newer copies from source at runtime and converts them with
src/blocking/community-lists.js, a port of this converter that must give the same rules from
these snapshots (see Development).
A release list changes only with a new version of Module Engine; its source records where the
snapshot came from and is never requested.
ublock-filters.txt is uBlock filters, the main list of uBlock Origin, from the filters/
folder of uBlockOrigin/uAssets at commit
14c32aba13089811163690c95d66377c0424f013 (19 November 2026, 02:49 UTC). The archive was
downloaded that day, and the 25 files of its filters/ folder matched the archive's CRC-32
records. scripts/make-ublock-snapshot.cjs made the snapshot offline from that folder:
filters.txt, with each file its !#include lines name copied in after the line, between
! >>>>>>>> name and ! <<<<<<<< name comments. The %timestamp% in ! Last modified: is
replaced with the commit time, as the publisher does when it publishes. Nothing else changes,
and the same files give the same bytes. Its updates is release. It is packaged as text
(scripts/extension-files.cjs), because an update of the downloaded lists converts it again
beside them, after checking it against its recorded SHA-256. Converting the new copies without
it would switch its rules off.
The converters read uBlock Origin's own syntax as follows:
!#if,!#elseand!#endifblocks. A block is used only when its condition holds in every browser Module Engine supports, and its!#elsepart only when the condition fails in all of them.false,cap_html_filtering,cap_ipaddress,env_safari,env_mobileandext_ubolare false: Module Engine has no HTML filtering or IP-address matching, supports desktop Chrome, Edge and Firefox, and is not uBlock Origin Lite. Every other name, such asenv_firefoxorenv_chromium, is unknown. One bundle serves every browser, so a block that depends on an unknown name is left out, with its!#elsepart.!,&&,||and parentheses combine names with three-valued logic, soenv_firefox && falseis false. Mixing&&and||without parentheses is unknown. Nested blocks are supported; a block without!#endifends with its list. Lines left out are counted under twoconditional:reasons.- HTML filters (
##^…), which edit a page's source before it is parsed, are skipped and counted with snippet filters. $ghideand$ehideare read as$generichideand$elemhide.
uBlock filters also uses syntax the converters skip and count: $redirect, $removeparam,
$csp, $denyallow, $badfilter, name.* sites, regular expressions and trusted scriptlets,
among others. $badfilter cancels another filter in uBlock Origin. Here it is skipped, so the
filter it names stays. In this snapshot that keeps two exceptions uBlock Origin would cancel:
@@||succeedscene.com^$script from EasyList (in effect before this list was added) and
@@||zefoy.com/assets/$script,1p from uBlock filters itself. The other cancelled filters use
options the converters skip anyway.
scripts/community-list-tools.ps1 converts the snapshots into rules/community.json (the
switchable community_ruleset), rules/community-cosmetic.json (generic and site-specific hiding rules
and their exceptions), the registered generic stylesheets rules/community-generic.css and
rules/community-generic-excepted.css, and rules/community-info.json (versions, counts and skipped filters by
reason). Blocks get priority 1, exceptions 2 and $important blocks 3; the allow list is 4, reported threats 5, custom blocks 50 and site pauses 10000.
Exceptions for domains on Module Engine's own ad, tracker or site-check lists are dropped, and path
exceptions exclude those domains. The build fails if the snapshots would exceed 29,000
rules, which keeps room in Chrome's guaranteed 30,000-rule budget; the three lists give 11,910.
The lists are third-party data: EasyList and EasyList China under the EasyList licence, and
uBlock filters under the GNU GPL version 3. See the main README for attribution.
The cosmetic bundle now uses schemaVersion: 3: hide and allow retain the
domain-to-selector maps (allow["*"] holds global selector exceptions), generic
contains global selectors, and scoped contains {include, exclude, selector, exception} rules with excluded sites. disableGeneric and disableAll hold
host-wide generichide and elemhide exceptions. Path-specific and unsupported
option combinations are counted as skipped rather than expanded to a whole site.
A site the builder cannot represent, such as an amazon.* wildcard (which needs a
public-suffix lookup), is dropped from a rule's site list and counted; the rule is kept
for its other sites. A rule is skipped instead when every included site was dropped,
since its exclusions alone would make it generic, or when an excluded site is
unsupported, since dropping an exclusion would widen the rule.
procedural contains {include, exclude, selector, exception, procedure} records.
The procedure holds a CSS subject and ordered text/child/ancestor operations, with
nested child plans preserved. These are data, never JavaScript or regex source.
The installed Node 22+ runtime runs scripts/compile-procedural.cjs offline so the
builder and extension use the same parser in cosmetic.js. Unsupported syntax and
unscoped advanced blocking are counted and skipped. Browser-native CSS validation
runs again before a procedure is activated. The runtime still accepts schema 2
bundles, which contain no advanced rules.
scriptlets contains {include, exclude, exception, name, args} records for
site##+js(...) and site#@#+js(...) filters; an exception with an empty name
turns every scriptlet off on its sites. scripts/compile-scriptlets.cjs checks them
offline with the extension's own parser in src/blocking/scriptlets.js, so a
filter whose scriptlet or arguments the extension does not support is counted and
skipped, as are scriptlet filters without a site. Names and arguments are data:
the extension runs only its own packaged scriptlets with them. EasyList and EasyList
China contain no scriptlet filters. uBlock filters gives 7,178 rules (scriptletRules
in community-info.json); 1,052 of its scriptlet filters name a trusted scriptlet, one
the extension does not have, or arguments it does not support, and are skipped. Sites
written as name.* are dropped as for site hiding. A bundle without the key has no
scriptlets. The converters accept at most 20,000 scriptlet candidates; this bundle has
about 8,200.
Generic selectors are also compiled into two stylesheets that the extension
registers for http(s) frames, instead of sending the list to every frame.
scripts/compile-generic-css.cjs, run offline by the builder like the procedural
compiler, writes every generic selector that no exception touches to
community-generic.css, and those that some site exempts to
community-generic-excepted.css, which is not injected on those sites. The
optional genericStatic record lists the excepted selectors, their
exceptedSites, and perFrame selectors that stay in each frame's state: those
with an exception that applies everywhere except listed sites, and any the
compiler cannot rewrite safely. allow["*"] selectors are in neither sheet.
Without genericStatic, the runtime sends generic selectors in frame state as
before.
Every selector in the sheets carries :not(:where([data-me-generic-off] *)), so
a frame can revoke both sheets by setting that attribute on its root element,
without changing any selector's specificity. A rule holds at most 256 selectors
and 8,000 selector characters, because Blink ignores selectors after the first
8,192 simple selectors of one rule. Escaped or uncommon selectors get a rule of
their own, so one that a browser rejects cannot void its neighbours.
tests/generic-css.test.cjs checks that the committed stylesheets match a
rebuild from the committed bundle.
This browser-only format does not change the shared reviewed-domain v1 contract.
site-ad-hosts.json (reviewKind: observed-site-ad-hosts) records ad hosts seen on a
named site by the opt-in tests/site-ad-check.cjs. Each entry has the canonical
domain, the UTC date observedOnUtc, the site it was seen on, its observed role, and
the committed evidence file under lists/evidence/ that records the request. Raw browser
output stays in the ignored verification/ folder; copy reviewed build inputs into
lists/evidence/ before adding them to the catalog. The builder turns each
entry into an ad rule (ids from 1001, ||domain^, page loads included so a click pop-up to
the host is blocked too). Like the URL patterns, the list is browser-only: it is not part of
the v1 catalog and is never exported to CyberAssist.
The builder refuses unknown or missing fields, non-canonical or duplicate hosts, future or
invalid dates, evidence paths outside lists/evidence/ or to missing files, and any host
already in the v1 catalog, so an unverified catalog entry cannot be activated through
this list. A recorded ad role is not a malware verdict, and generated-looking ad domains
may be replaced by their network; recheck the site and add the new host when that happens.
reported-threats.json is a distinct input with reviewKind: passive-dated-threat-reports.
It does not change or extend the v1 catalog and is not consumed by the CyberAssist exporter.
The first review contains 24 exact hosts from Proofpoint and Volexity HTML reports dated
2026-09-09. Each record preserves its source, report section, observed date precision,
and malicious role. No inferred subdomains, speculative shared-IP neighbors, undated
warnings, or former unverified entries are promoted.
scripts/threat-report-tools.ps1 validates the closed schema, strict UTF-8, 128 KiB / 100
host bounds, unique canonical names, exact-only scope, publisher/report URLs, supported
roles, real calendar dates, and observations recent at review. URLs are references only;
the builder never requests them. Only approved publisher HTTPS blog pages are accepted;
web-hosted text files and binary download URLs are rejected. Human source review remains
necessary: a well-formed record alone does not prove that its claims are true.
The builder emits rules/threats.json and rules/threat-policy.json together. The whole
batch expires 180 days after its earliest observation. Month-only evidence uses the first
day of that month conservatively, without inventing a precise last-seen date. The review
date and later rebuilds do not refresh evidence. Current expiry: 2027-02-28 00:00 UTC.
Old bundles remain reproducible; the runtime disables expired rules rather than silently
extending them. A manual new evidence review is required to renew coverage.
The window was 90 days until 18 November 2026. That day's review found no qualifying newer report from either publisher, and the user chose to keep these 24 hosts for 180 days rather than let the layer lapse on 30 November. No observation date or source changed; only the allowed age did. The builder and the evidence panel accept only this window, and the runtime refuses an expiry more than 180 days after the evidence date.
Run pwsh -NoProfile -File tests/threat-rules.test.ps1 for offline schema, expiry,
exact-boundary, legacy-exclusion and failed-publication checks. See the
dated-threat review for sources and browser validation.