finished !

Author: l0uisgrange

index.md

@@ -132,23 +132,23 @@ To answer this, we analyzed the "linguistic signature" of hostility. We looked a
 
 If political anger is ideological and complex, and sports anger is simple and tribal, their signatures should look different.
 
-**They don't. They are nearly identical.**
-
-The correlation between the hostility signature in politics and sports is **r = 0.937**.
-
-
-### The process
+**We found that they don't. They are nearly identical.**
 
 We treated hostility like a **forensic trace**: first we measured *how often* it appears, then *what it looks like in language*, then we asked whether that “signature” is **portable** across domains, and finally whether the same forces also show up at the level of **events** and **network structure**.
 
 Instead of betting everything on one metric, we built a chain of evidence that checks the same claim from multiple angles:
 
-#### 1) Proportion tests & effect sizes: “Is politics really bloodier?”
+### Proportion tests & effect sizes
+
+<p><span class="bg-zinc-100 rounded px-3 py-2 mt-2">"Is politics really bloodier?"</span></p>
+
 We started with the simplest question: **how often do hostile links occur**?
 We compared the proportion of hostile hyperlinks (`LINK_SENTIMENT = -1`) in Politics vs Sports using a **two-proportion test** (χ² / z-test), reported a **95% confidence interval** for the difference, and quantified how big the gap is with **Cohen’s h**, plus **risk ratio** and **odds ratio**.
 This is the “frequency layer”: it tells us *how much* conflict there is, not yet *how it sounds*.
 
-#### 2) LIWC hostility signatures: “What does hostility sound like?”
+### LIWC hostility signatures
+
+<p><span class="bg-zinc-100 rounded px-3 py-2 mt-2">"What does hostility sound like?"</span></p>
 Next, we built a linguistic fingerprint for each domain.
 
 For every LIWC feature, we computed a delta between hostile and non-hostile posts:
@@ -160,7 +160,11 @@ We then compared the two vectors with **Pearson** (shape similarity) and **Spear
 
 If hostility is domain-specific, these fingerprints should diverge. If it’s universal, they should line up.
 
-#### 3) Bootstrap & permutation tests: “Is the resemblance real, or just luck?”
+### Bootstrap & permutation tests
+
+<p><span class="bg-zinc-100 rounded px-3 py-2 mt-2">"Is the resemblance real, or just luck?"</span></p>
+
+: “Is the resemblance real, or just luck?”
 A strong correlation can still be a coincidence if the sample is weird.  
 So we stress-tested the signature similarity:
 
@@ -169,15 +173,18 @@ So we stress-tested the signature similarity:
 
 This is our “robustness layer”: we make sure the resemblance isn’t a fragile artifact.
 
-#### 4) Cross-domain classifier transfer: “Can a model trained on Politics recognize Sports?”
+### Cross-domain classifier transfer
+
+<p><span class="bg-zinc-100 rounded px-3 py-2 mt-2">"Can a model trained on Politics recognize Sports?"</span></p>
+
 Similarity on a plot is one thing; **portability** is stronger.
 
 We trained a **logistic regression** on LIWC features to predict hostility in one domain, then tested it **as-is** on the other:
 
 - train on Politics → test on Sports  
 - train on Sports → test on Politics
 
-We evaluated with **AUC**.  
+We evaluated with **AUC** (_Area Under the Curve_).  
 If hostility has a different meaning across domains, transfer should collapse toward **0.5**. If the signal is shared, AUC should stay high.
 
 <img
@@ -186,7 +193,10 @@ If hostility has a different meaning across domains, transfer should collapse to
   class="flourish-embed"
 />
 
-#### 5) Logistic regression coefficient comparison: “Do the same features do the work?”
+### Logistic regression coefficient comparison
+
+<p><span class="bg-zinc-100 rounded px-3 py-2 mt-2">"Do the same features do the work?"</span></p>
+
 Even when two models perform well, they might rely on different cues.  
 So we compared the *mechanisms* of the Politics and Sports classifiers:
 
@@ -196,25 +206,23 @@ So we compared the *mechanisms* of the Politics and Sports classifiers:
 
 This is the “interpretability layer”: it tells us whether the same linguistic levers are being pulled.
 
-#### 6) Difference-in-differences + network metrics: “Do events and structure tell the same story?”
+### Difference-in-differences + network metrics
+
+<p><span class="bg-zinc-100 rounded px-3 py-2 mt-2">"Do events and structure tell the same story?"</span></p>
+
 Finally, we zoomed out from language into dynamics and structure.
 
 - **Difference-in-differences (events):** for major shocks (elections, finals, referendums…), we compared the *change* in hostility for “involved” camps versus “less involved/observer” camps, before vs during the event window. This isolates event effects from normal baseline fluctuations.
 - **Network metrics (structure):** we model camps as a **directed network** where edges carry hostile-link flow. We then summarize behavior with metrics like **density** (how connected the conflict is), **reciprocity** (do attacks get returned?), and aggressor/target patterns via **outgoing vs incoming hostility**.
 
-Together, these checks cover the full stack:
-rates → language → robustness → portability → mechanism → system dynamics.
-
-When all six layers agree, we can say it with confidence:
-**the topic changes, the script of hostility doesn’t.**
+Together, these checks cover the full following stack. When all six layers agree, we can say it with confidence: **the topic changes, the script of hostility doesn’t.**
 
+---
 
-### What we found (and why it matters)
+## What we found (and why it matters)
 
 The resemblance is not subtle, it’s *statistically overwhelming*.
 
-#### The headline: r = 0.937
-
 A Pearson correlation of **r = 0.937** between the Politics and Sports signatures means that when a feature increases during political hostility, it almost always increases during sports hostility too (and vice versa). In other words: **the same “psychological knobs” get turned**.
 
 And the social-psychology “tell” is the same in both arenas: **THEY goes up**. Hostile posts are less about “us” and more about labeling and attacking *them*.
@@ -269,7 +277,7 @@ Even the *lower bound* of the confidence interval still implies a very strong re
 
 Our investigation reveals four undeniable truths about the nature of the conflict:
 
-1.  **Anger is universal:** politics is a bloodier battlefield (3.3x more hostile), but the soldiers use the exact same weapons. The linguistic signature of conflict is identical (r=0.937), proving that hostility is a fundamental human instinct, not a topic-specific reaction.
+1.  **Anger is universal:** politics is a bloodier battlefield (3.3x more hostile), but the soldiers use the exact same weapons. The linguistic signature of conflict is identical (r = 0.937), proving that hostility is a fundamental human instinct, not a topic-specific reaction.
 
 2.  **Hate is portable:** the patterns are so consistent that a machine trained to spot political vitriol can instantly recognize sports trash-talk. The DNA of abuse is shared, suggesting that moderation tools could work across any domain.