Results at study close — Day 43, 21 April 2026
The seven C.A.R.A. indicators were designed before the study began. For each one, a calculation formula and a significance threshold were established and publicly registered prior to data collection. This is a methodological requirement: if thresholds are defined after seeing the results, it is always possible to adjust them to confirm what one wanted to find. C.A.R.A. declares in advance what would constitute a significant finding, and then accepts what the data say.
| Account |
Type |
Avg amplification |
CIS |
Max level |
CIER |
| A |
New / Prosocial |
1.075 |
0.278 |
4 (Day 1) |
1.000 |
| B |
New / Neutral |
1.250 ⬆ highest |
0.273 |
3 (sustained) |
0.925 |
| C |
New / Traditional |
1.225 |
0.337 |
4 (Day 22) |
0.950 |
| D (@demachosahombres, >475K) |
Existing |
1.200 |
0.300 |
4 |
0.960 |
| E (Retiro p/Hombres, ~1.6K) |
Existing |
1.380 |
0.381 |
4 |
0.880 |
| F (MIRACLE, ~1.7K) |
Existing |
1.040 |
0.283 |
4 |
0.980 |
Gate A activated (A Day 1, C Day 22) · Gate F not activated (max new-account CIS: C=0.337; pre-registered threshold: 0.60) · κ=0.68 · 90.2% global agreement · 420 items classified
On calibration, what the CIS result means
The composite CIS threshold was not triggered. The study is honest about this: it documents significant, measurable exposure to Level 3-4 content without active search, and calibrates the regulatory claim accordingly. A study that claims more than the data supports does not serve regulatory incidence. The evidence is policy-relevant without being maximalist.
The audience-scale finding, a regulatory signal
Account D has more than 475K followers. Account E has approximately 1,600, a 294-fold difference. In the CIER indicator (Cross-Identity Exposure Ratio), Account E showed greater algorithmic enclosure than Account D. Smaller audience, more closed recommendation environment.
This has a direct regulatory implication: audience scale does not predict relational risk. Regulatory frameworks that use follower count as the primary criterion for oversight will systematically misidentify where the risk actually is. Relational positioning, how the system situates an account within a network of content and affinity, is the signal that matters.
What the community said
Alongside the field study, a structured survey was distributed to followers of @demachosahombres (N=147). Two findings are directly relevant:
| Finding |
Data |
| How did you first find IDMAH? |
50.6% arrived via direct algorithmic recommendation — not active search. |
| Have you received hostile or polarising content about masculinity via algorithm? |
28.6% (42/147) said yes. |
Note: the survey did not reach the threshold for full statistical robustness (N=200) and is used for qualitative pathway traceability, not as statistically representative inference. The data are consistent with the field study patterns.
What this study reframes
Radicalization is not the main phenomenon.
Radicalization is a by-product. C.A.R.A. does not study radicalization as an extreme event. It studies something prior: the progressive intensification of identity frames, what happens before the problem becomes visible.
The manosphere operates precisely in that prior stretch, and it does so from inside conventional platforms like Instagram, not from a dark corner of the internet. Red Pill, MG-TOW, incels, dominance narratives: none of these communities exist in a vacuum.
They exist because algorithmic recommendation systems build them, connect them to each other, and deliver them to users who never sought them out. What C.A.R.A. documents is that Instagram reproduces that same pattern autonomously — without any external actor intervening. Radicalization is not the starting point of the problem. It is the final result of trajectories that begin much earlier, with apparently innocuous content.
The algorithm compresses time.
What might previously have taken months or years to consolidate as part of an identity can now begin to take shape in weeks. Not because the content is more extreme, but because the system is more efficient at connecting. A new account with no prior signal reached Level 4 content on Day 1.
Another reached it by Day 22. The architecture does not wait.
The risk is not individual. It is populational.
What matters is not only what happens to one person. What matters is that the same pattern reproduces at scale. When millions of accounts receive similar intensification trajectories, what changes is not the individual. It is the social conversation. The manosphere is not a collection of extreme individuals, it is a structural output of a recommendation architecture operating at population scale.
That is what makes it a regulatory problem, not just a content moderation problem.
