Signal 01 — View-to-Engagement Mismatch
// the core proof
0: 1
median views per like
10:1100:11,000:113,000:1 worst
4,096,118 views produced 6,706 likes and
2 comments across 73 posts. Median engagement rate
0.19% (authentic reels run 1–3%). This is
26× outside the organic band — views are
inflated relative to any genuine human reaction. Worst
single post:
51,837 views / 4 likes = 12,959:1.
Engagement Funnel
A funnel this collapsed — 4.1M views, a trickle of
likes, essentially zero comments — is the fingerprint
of purchased reach with no human audience behind it.
Like/comment ratio 3,353:1 (organic 50–200:1).
Signal 02 — Coordinated Cluster
0accounts
58.9% post coverage · density
1.0 · co-liking 3.0× above chance
for their activity level. 34.5% of all pairs pass a
Bonferroni-corrected significance test — an
independent-accounts control scored 0%. Not username-based;
pure behaviour.
p < 1e-634.5% sig pairs
Example of Bot-Pattern accs
| Account |
Follow |
Ratio |
|
| @alvess2003_ |
1,144 / 5 |
228.8× |
BOT |
| @zeekids.pk |
1,313 / 9 |
145.9× |
BOT |
| @harkeysireyy |
2,066 / 25 |
82.6× |
BOT |
| @petercervantes86 |
1,528 / 19 |
80.4× |
BOT |
| @houfani_rouaim |
8,434 / 316 |
26.7× |
BOT |
| @shobe_jaidee |
2,068 / 111 |
18.6× |
BORDERLINE |
The follow-for-follow signature of rented or farmed
accounts, not an organic audience. Account gained
253 followers/day during the sampled window (typical
organic range: 10–50/day).
Co-Like Network · 955 likers · live
simulation, coordinated core highlighted
Cluster Breakdown
| Size |
Cov |
Coord |
Sig |
Verdict |
| 93 |
37.0% |
3.5× |
0% |
likely real |
| 91 |
38.4% |
3.6× |
0% |
likely real |
| 41 |
58.9% |
3.0× |
34% |
COORDINATED |
Two dense groups (93, 91) test clean — high density
fully explained by member activity. Only the 41-account
group beats chance. Honest result: one confirmed coordinated
cluster, not everything flagged.
Like-Count Distribution & View/Like Ratio Curve
· generated from raw data
Deduplicated Scorecard
✓ View/engagement mismatch — 26×
outside band · HIGH
✓ Coordinated 41-account cluster ·
HIGH
✓ Like/comment decoupling 3,353:1 ·
MED
✓ Non-uniform like digits (χ² 18.64)
· MED
✓ Named bot-pattern followers (228× follow
ratio) · MED
— No username-pattern reliance (farms disguise
these)
Weighted, evidence-linked findings — not a single
0–100 score. The v1 report's “100/100
DEFINITIVE” implied a precision the data can't
support.
Methodology — how every number above was produced
01 / COLLECT
Authenticated CDP session
Attached to an AdsPower browser profile over the
Chrome DevTools Protocol. All requests ride the real
session's cookies, TLS fingerprint and proxy —
no synthetic client, so liker caps and bot-detection
are avoided. Post IDs derived locally via
shortcode→pk base64 decode,
independent of any intercepted response.
02 / EXTRACT
GraphQL + private API
Likers pulled from
/media/{pk}/likers/ with cursor
pagination; counts from
/media/{pk}/info/ (like, comment,
view/play, share fields probed across six locations
incl. clips_metadata). Recency-ordered
sample of ~40–90 likers per post — the
early window where injected engagement concentrates.
03 / ANALYSE
Four independent modules
static_signals: view/like ratio vs
10–30:1 band, round-number & last-digit
χ², like/comment decoupling.
liker_graph: bipartite→co-like
projection, edge-weight threshold, Louvain, then a
Poisson test of co-occurrence vs each pair's own
activity. deep_signals: cadence
z-score, bootstrap CIs.
04 / VALIDATE
Control-calibrated
The coordination test is calibrated against a
synthetic control of independent accounts at matched
activity levels, which produced
0% significant pairs — so any non-zero
fraction is a real deviation, Bonferroni-corrected
for the number of pairs tested. Small-sample
percentages carry 95% bootstrap CIs; anything wider
than 25pp is flagged, not scored.
Supporting detection layer.
Isolation Forest / LOF — flags likers whose
activity pattern sits statistically outside the rest
of the crowd
Network cluster analysis — groups likers by
shared behaviour to surface coordinated pockets
before testing them