Perceptual matching of video posters across 59
platforms · generated 27 Sep 2026 04:22 UTC from snifferX
· every figure below is computed from the live index by
scripts/measure.py, none are hand-entered
Headline
59/106
Domains indexed
55.7% of collectable
26/26
Match precision
pHash matches confirmed byte-identical
4
Mirror networks found
across 59 domains
71.4%
Detection rate
16 misses published, not hidden
1.855 MB
Index footprint
0 bytes of media stored
Coverage
Domain reach
Indexed, media retrieved
59
Declared collectable
106
Collection mode
homepage, 15 media/domain
of the domains marked collectable in data/mirror-groups.json, not of the 123 in the source report and not of the wider ecosystem.
Why the other 47 yielded nothing
unreachable
28
no-media
17
robots-blocked
2
Measured ceiling, not an engineering gap. Nothing was
worked around: no Referer spoofing past CDN 403s, no Cloudflare challenge
solving, no robots bypass.
Method validation
Match precision
pHash matches re-fetched and re-compared
26
Byte-identical by sha256
26
pHash false positives
0
Same URL refetched 3x
14/14 identical bytes across 3 fetches
Every perceptual match was verified against a content
hash, so no claim rests on a similarity judgement alone.
Detection ceiling
Alterations tested
56
Survived re-encode / resize
40
Missed
16
survived by sha256
0
Known blind spot. text/branding overlay onto a poster breaks both pHash and sha256.
These are false negatives measured on synthetic variants, not on real reposts. Real-world recall is unknown and probably lower.
Confirmed networks
Platform A
Platform B
Shared media
n
kamababa.desi
thekamababa.com
14
xvideos-india.com
xvideos.tube
10
xhaccess.com
xhamster19.com
1
indianptv.com
indianpworld.com
1
found in ~15-image homepage samples of 59 domains; this is a lower bound, not a census.
26 assets were seen on more than one domain.
Cost
Storage
MongoDB data
916,754 B
MongoDB indexes
1,028,096 B
Media bytes stored
0
Media fields in database
0
Monthly, at configured rate
$0.000038
Collection
Poster files fetched
1,345
Assets indexed
600
LSH prefilter bands
4
Query cost vs corpus size
constant
~1345 poster fetches totalled under 50KB each, so egress is well under 1GB and rounds to $0.00 at the configured rate.
Corpus
Assets with 16:9 poster geometry
354/600
Median aspect ratio
1.77
Median / max / min assets per domain
29 / 32 / 2
geometry consistent with video posters, which is what makes poster matching valid here.
Limitations
What this does not claim
4 networks is a lower bound. Discovery used
~15-image homepage samples of 59 of
106 collectable domains. A domain that served 2 images cannot be
expected to overlap with anything.
71.4% is a synthetic figure. The
56 alterations were generated from fetched
posters, not harvested from real reposts. Real-world recall is unknown
and is probably lower.
Branding breaks detection. A title or watermark overlay onto a
poster defeats both pHash and sha256, so heavily branded sites
under-report rather than fail loudly.
Recompressed reposts are invisible to sha256. It confirms
verbatim copies only, which is why it is a confirmation tier and not
the detection tier.
Groups are hypotheses. Only
1
group(s) in the manifest are supported by fingerprints; the rest rest
on contact, CDN or analytics heuristics from a source report, not on
measured media.