diff --git a/adapters/github.py b/adapters/github.py
index a3cda52..dda90a5 100644
--- a/adapters/github.py
+++ b/adapters/github.py
@@ -107,14 +107,15 @@ class GitHubAdapter(SourceAdapter):
         return ""
 
     def _score(self, repo: dict, age_days: float) -> float:
-        """Score: stars weighted by recency. Higher = better signal."""
+        """Score: growth velocity (stars/day) on log scale. Higher = better signal."""
         stars = repo.get("stargazers_count", 0)
-        # Normalize: log scale on stars, decay by age
         import math
-        star_score = min(math.log1p(stars) / 2.0, 10.0)  # log(1000) ≈ 6.9 → ~3.5
-        # Recency bonus: newer repos get a slight boost
-        recency_bonus = max(0, 1.0 - age_days / 30.0) * 1.5  # up to +1.5 for very recent
-        return min(star_score + recency_bonus, 10.0)
+        # Velocity: stars per day — the real signal
+        velocity = stars / max(age_days, 1)
+        vel_score = min(math.log1p(velocity) / 1.8, 10.0)  # log(1000/d) ≈ 6.9 → ~3.8
+        # Absolute stars still matter (a 100K star repo is legit even if slow)
+        star_score = min(math.log1p(stars) / 3.0, 5.0)  # max contribution: 5.0
+        return min(vel_score * 0.7 + star_score * 0.3, 10.0)
 
     def _tags(self, repo: dict) -> list:
         """Generate category tags from repo metadata."""
@@ -148,12 +149,21 @@ class GitHubAdapter(SourceAdapter):
             print(f"  Searching: '{query}'")
             repos = self._search_repos(query, sort="stars", per_page=min(limit * 2, 100))
         else:
-            # "Trending" = AI repos created in last 14 days, sorted by stars
-            cutoff = (now - timedelta(days=14)).strftime("%Y-%m-%d")
-            # Two queries to get breadth
+            # "Trending" = repos created in last 30 days, sorted by GROWTH VELOCITY
+            # (stars per day), NOT absolute stars. A 5-day-old repo with 3K stars
+            # (600/d) is more interesting than a 25-day-old repo with 77K stars (3K/d).
+            #
+            # GitHub API sort=stars ranks total stars. We fetch by stars to get
+            # broad coverage, then re-sort by velocity locally.
+            cutoff = (now - timedelta(days=30)).strftime("%Y-%m-%d")
+            # Three queries for breadth: agents, LLM/infra, and security/tools
             repos = []
-            for q in ["ai agents created:>=2026-06-01", "llm inference created:>=2026-06-01"]:
-                batch = self._search_repos(q, sort="stars", per_page=50)
+            for q in [
+                f"ai agent created:>{cutoff}",
+                f"llm OR inference OR rag created:>{cutoff}",
+                f"autonomous agent OR AI tool created:>{cutoff}",
+            ]:
+                batch = self._search_repos(q, sort="stars", per_page=30)
                 repos.extend(batch)
                 time.sleep(1)  # polite spacing
 
@@ -167,9 +177,23 @@ class GitHubAdapter(SourceAdapter):
                     unique.append(r)
             repos = unique
 
-        # Sort by stars descending, take top limit
-        repos.sort(key=lambda r: r.get("stargazers_count", 0), reverse=True)
-        repos = repos[:limit]
+            # Compute velocity (stars/day) and sort by that, not total stars
+            def _velocity(repo):
+                created = repo.get("created_at", "")
+                if created:
+                    try:
+                        created_dt = datetime.fromisoformat(
+                            created.replace("Z", "+00:00")
+                        )
+                        age_days = max((now - created_dt).days, 1)
+                    except (ValueError, TypeError):
+                        age_days = 1
+                else:
+                    age_days = 1
+                return repo.get("stargazers_count", 0) / age_days
+
+            repos.sort(key=_velocity, reverse=True)
+            repos = repos[:limit]
 
         entries = []
         readme_budget = min(10, limit)  # Only fetch README for top 10 to stay under rate limit
@@ -212,6 +236,8 @@ class GitHubAdapter(SourceAdapter):
                 time.sleep(0.5)  # polite spacing between README fetches
 
             # Structured metadata
+            stars = repo.get("stargazers_count", 0)
+            velocity = stars / max(age_days, 1)
             raw_meta = {
                 "full_name": repo.get("full_name", ""),
                 "owner": repo.get("owner", {}).get("login", ""),
@@ -224,6 +250,7 @@ class GitHubAdapter(SourceAdapter):
                 "updated_at": repo.get("updated_at", ""),
                 "pushed_at": repo.get("pushed_at", ""),
                 "age_days": age_days,
+                "stars_per_day": round(velocity, 1),
                 "readme_length": len(readme_text),
                 "score_type": "actual",  # based on real star counts
             }