constitution · epochs · watch · epoch 3

comparison

c_92734e554a25 (tommy-mor) vs c_48aeaf9b52c3 (tommy-mor)

download prompt · raw event · cmp_ca9942ca8beb91

council reasoning

~anthropic/claude-sonnet-latest · winner B · 9:1 · permalink

Side B fixes a real algorithmic bug (bipartite Markov chain causing incorrect ranking convergence) with a mathematically justified change grounded in cited literature, plus regression tests in both Rust and Clojure across multiple topologies. Side A merely deletes two existing tests and tweaks CI node version/npm install, providing no new functionality or bugfix and reducing test coverage.

~x-ai/grok-latest · winner B · 1:12 · permalink

B fixes a real correctness bug in core ranking (star topologies yielding uniform/oscillating scores) by switching to degree-based d_max per the Negahban–Oh–Shah definition, with Rust and Clojure regression coverage. A only bumps Node 22→24, drops a one-line npm upgrade step, and deletes two obsolete unit tests—useful chore work but far less lasting substance.

openai/gpt-chat-latest · winner B · 10:1 · permalink

Side B fixes a substantive algorithmic bug in Rank Centrality by changing the normalization from summed edge weights to degree-based d_max, preventing oscillating bipartite Markov chains and restoring correct rankings for star topologies. It also adds focused Rust and end-to-end regression tests with fixtures covering star, inverse star, chain, and cycle cases. Side A mainly updates the CI Node version, removes an npm upgrade step, and deletes obsolete tests, providing comparatively limited lasting value.

sides

A — c_92734e554a25 (tommy-mor)

message

[558dfca8] fix

diff preview

diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index 6f3e27e491b56aaaa1ef64c547cc68921e010b3f..5ed117dddd0bc36f89e3b3788ae55e7f871312a9 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -88,12 +88,9 @@ jobs:
 
       - uses: actions/setup-node@v4
         with:
-          node-version: 22
+          node-version: 24
           registry-url: "https://registry.npmjs.org"
 
-      - name: Upgrade npm for OIDC trusted publishing support
-        run: npm install -g npm@latest
-
       - name: Copy binaries into npm platform packages
         shell: bash
         run: |
diff --git a/server/src/api/mod.rs b/server/src/api/mod.rs
index f42e907775645166c60aeae2d98c5855223a71be..e3299ee353f8c62935fa82708762490a6d5c54e3 100644
--- a/server/src/api/mod.rs
+++ b/server/src/api/mod.rs
@@ -70,14 +70,6 @@ mod tests {
         }));
     }
 
-    #[test]
-    fn validate_ingest_document_requires_actor() {
-        let reduced = ReducerState::default();
-        let text = "~/t/a {a}\n~/t/b {b}\n";
-        let v = validate_ingest_document(&reduced, text, &crate::reducer::ScopeId::Public).unwrap();
-        assert_eq!(v.raw_text.trim(), text.trim());
-    }
-
     #[test]
     fn validate_ingest_document_parse_error() {
         let reduced = ReducerState::default();
@@ -108,14 +100,6 @@ mod tests {
         assert!(err.1.contains("undefined item"));
     }
 
-    #[test]
-    fn validate_ingest_document_requires_tag() {
-        let reduced = ReducerState::default();
-        // tags are no longer DSL routing metadata; validation no longer requires them.
-        let text = "~/t/a {a}\n~/t/b {b}\n";
-        validate_ingest_document(&reduced, text, &crate::reducer::ScopeId::Public).unwrap();
-    }
-
     #[test]
     fn validate_ingest_document_accepts_quoted_thread_title() {
         let reduced = ReducerState::default();

download full diff A

B — c_48aeaf9b52c3 (tommy-mor)

message

[595b3850] Fix star-topology ranking by using degree-based d_max (#146).

A pure forward star at the default `>` ratio (2:1) produced uniform 1/3
scores, and the alphabetical-fallback sort placed the unambiguous winner
last. Root cause: `compute_scores_from_edges` divided by the max sum of
pairwise-normalized weights, so every node ended up with P_ii = 0 — a
bipartite Markov chain whose power iteration oscillated and, after the
configured even iteration count, returned to the uniform initial state.

Switch the divisor to the unweighted max neighbor degree, matching the
canonical Rank Centrality definition in Negahban–Oh–Shah 2012 §3.1
(arXiv:1209.1688, eq. defP and the d_max definition in §6). This gives
every non-saturated node a positive self-loop, makes the chain aperiodic,
and converges the star to π_zebra = 1/2, π_alpha = π_beta = 1/4.

Add Rust regression test and a Clojure test that drives the sorterc
binary against four .sorter fixtures (star, inverse star, chain, cycle).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

diff preview

diff --git a/server/src/ranking.rs b/server/src/ranking.rs
index 3710c9f64437f5bef3b2121905b6f3bcb7611047..38e6d09b4370e5f8cbae09c0e5760b4e7f1ef7db 100644
--- a/server/src/ranking.rs
+++ b/server/src/ranking.rs
@@ -1,4 +1,4 @@
-use std::collections::HashMap;
+use std::collections::{HashMap, HashSet};
 
 use crate::path_types::ItemId;
 use crate::reducer::GroupState;
@@ -143,23 +143,35 @@ fn compute_scores_from_edges(n: usize, edges: impl Iterator<Item = ((usize, usiz
         }
     }
 
+    // Rank Centrality (Negahban, Oh, Shah 2012, §3.1):
+    //   P_ij = (1/d_max) * A_ij           for i ≠ j compared
+    //   P_ii = 1 - (1/d_max) * Σ_k A_ik
+    // where d_i is the *degree* (number of distinct neighbors compared) and
+    // d_max = max_i d_i. Using the unweighted degree — not the sum of
+    // pairwise-normalized weights — is what guarantees aperiodicity: it
+    // forces P_ii > 0 for every non-maximum-degree node, and for max-degree
+    // nodes whenever any neighbor weight is below 1 (i.e. not a unanimous
+    // loss). Without this, regular comparison graphs (e.g. a pure star at
+    // ratio 2:1) produce a bipartite chain that oscillates instead of
+    // converging — see issue #146.
     let mut out_edges: Vec<Vec<(usize, f64)>> = vec![Vec::new(); n];
-    let mut out_deg: Vec<f64> = vec![0.0; n];
+    let mut neighbors: Vec<HashSet<usize>> = vec![HashSet::new(); n];
 
     for ((src, dst), w) in &normalized {
         out_edges[*src].push((*dst, *w));
-        out_deg[*src] += w;
+        neighbors[*src].insert(*dst);
+        neighbors[*dst].insert(*src);
     }
 
-    let mut max_out = 0.0f64;
-    for &d in &out_deg {
-        if d > max_out {
-            max_out = d;
-        }
-    }
-    if max_out <= 1e-12 {
+    let weight_sum: Vec<f64> = out_edges
+        .iter()
+        .map(|es| es.iter().map(|(_, w)| *w).sum())
+        .collect();
+    let d_max = neighbors.iter().map(|s| s.len()).max().unwrap_or(0);
+    if d_max == 0 {
         return vec![1.0 / n as f64; n];
     }
+    let d_max_f = d_max as f64;
 
     let mut scores = vec![1.0 / n as f64; n];
     let mut next = vec![0.0f64; n];
@@ -167,14 +179,14 @@ fn compute_scores_from_edges(n: usize, edges: impl Iterator<Item = ((usize, usiz
     for _ in 0..max_iters {
         next.fill(0.0);
         for i in 0..n {
-            let stay_prob = (max_out - out_deg[i]) / max_out;
+            let stay_prob = (d_max_f - weight_sum[i]) / d_max_f;
             next[i] += scores[i] * stay_prob;
 
             if out_edges[i].is_empty() {
                 continue;
             }
             for &(dst, w) in &out_edges[i] {
-                next[dst] += scores[i] * (w / max_out);
+                next[dst] += scores[i] * (w / d_max_f);
             }
         }
 
@@ -270,6 +282,38 @@ mod tests {
         }
     }
 
+    /// Regression for issue #146: pure forward star at default `>` ratio (2:1).
+    /// Under the old (sum-of-weights) divisor every node had P_ii = 0 and the
+    /// chain was bipartite; power iteration oscillated and returned the
+    /// uniform initial distribution after an even number of steps. Using the
+    /// paper's degree-based d_max gives every node a positive self-loop and
+    /// the chain converges to the correct stationary distribution.
+    #[test]
+    fn star_topology_winner_at_top_via_subset() {
+        let mut g = mk_group();
+        g.apply_vote(vote(1, "zebra", "alpha", 2, 1));
+        g.apply_vote(vote(2, "zebra", "beta", 2, 1));
+
+        let mut items: Vec<(usize, String)> = g
+            .idx_to_item
+            .iter()
+            .enumerate()
+            .map(|(i, it)| (i, it.as_str().to_string()))
+            .collect();
+        items.sort_by(|a, b| a.1.cmp(&b.1));
+        let idxs: Vec<usize> = items.iter().map(|(i, _)| *i).collect();
+
+        let ranked = ranked_items_subset(&g, &idxs, 10000, 1e-8);
+        for r in &ranked {
+            eprintln!("{}: {}", r.item.as_str(), r.score);
+        }
+        assert_eq!(
+            ranked[0].item.as_str(),
+            "https://slug.social/zebra",
+            "zebra won both votes and should rank #1"
+        );
+    }
+
     #[test]
     fn group_ranking_cache_dirty_flow() {
         let mut g = mk_group();
diff --git a/test/fixtures/ranking/chain.sorter b/test/fixtures/ranking/chain.sorter
new file mode 100644
index 0000000000000000000000000000000000000000..912a96bdc08f631be27f3c9afc7e05004a2457c0
--- /dev/null
+++ b/test/fixtures/ranking/chain.sorter
@@ -0,0 +1,10 @@
+#t3
+
+~/t3/a { head of chain }
+~/t3/b { middle }
+~/t3/c { tail }
+
+{ a > b }
+~/t3/a > ~/t3/b
+{ b > c }
+~/t3/b > ~/t3/c
diff --git a/test/fixtures/ranking/cycle.sorter b/test/fixtures/ranking/cycle.sorter
new file mode 100644
index 0000000000000000000000000000000000000000..771731ae176e5d77e00ab89767ff2d7465bb7c6b
--- /dev/null
+++ b/test/fixtures/ranking/cycle.sorter
@@ -0,0 +1,12 @@
+#t4
+
+~/t4/a { node a }
+~/t4/b { node b }
+~/t4/c { node c }
+
+{ a > b }
+~/t4/a > ~/t4/b
+{ b > c }
+~/t4/b > ~/t4/c
+{ c > a }
+~/t4/c > ~/t4/a
diff --git a/test/fixtures/ranking/star.sorter b/test/fixtures/ranking/star.sorter
new file mode 100644
index 0000000000000000000000000000000000000000..135c9f9097d57b73c7ab18fac737ed3598d76fbe
--- /dev/null
+++ b/test/fixtures/ranking/star.sorter
@@ -0,0 +1,11 @@
+#repro
+
+~/repro/zebra { winner — beats both others }
+~/repro/alpha { loser — alphabetically first }
+~/repro/beta { loser — alphabetically middle }
+
+{ zebra beats alpha }
+~/repro/zebra > ~/repro/alpha
+
+{ zebra beats beta }
+~/repro/zebra > ~/repro/beta
diff --git a/test/fixtures/ranking/star_inverse.sorter b/test/fixtures/ranking/star_inverse.sorter
new file mode 100644
index 0000000000000000000000000000000000000000..dab842fa924e5124865ee216b9565f6c7471c812
--- /dev/null
+++ b/test/fixtures/ranking/star_inverse.sorter
@@ -0,0 +1,11 @@
+#t2
+
+~/t2/win { source of incoming edges (loses both) }
+~/t2/loss-a { winner }
+~/t2/loss-b { winner }
+
+{ loss-a beats win }
+~/t2/loss-a > ~/t2/win
+
+{ loss-b beats win }
+~/t2/loss-b > ~/t2/win
diff --git a/test/ranking.clj b/test/ranking.clj
new file mode 100644
index 0000000000000000000000000000000000000000..e1358783a84a264ee633bd79151ba29750b717cf
--- /dev/null
+++ b/test/ranking.clj
@@ -0,0 +1,74 @@
+(ns test.ranking
+  "Drives sorterc on .sorter fixtures and asserts ranking properties.
+
+   Regression coverage for issue #146 — pure forward star at default ratio
+   (2:1 for `>`) used to produce tied uniform scores because the random walk
+   on the normalized edge weights was bipartite. Fixed by switching to the
+   degree-based d_max from Negahban–Oh–Shah rank centrality (§3.1)."
+  (:require [clojure.test :refer [deftest is testing]]
+            [babashka.process :as p]
+            [cheshire.core :as json]
+            [clojure.java.io :as io]))
+
+(def sorterc-bin
+  "Path to the locally-built sorterc binary. Builds on demand if missing."
+  (let [dbg     "target/debug/sorterc"
+        release "target/release/sorterc"]
+    (cond
+      (.exists (io/file release)) release
+      (.exists (io/file dbg))     dbg
+      :else
+      (do (println "building sorterc…")
+          (let [r (p/shell {:out :string :err :string :continue true}
+                           "cargo build -p sorterc")]
+            (when-not (zero? (:exit r))
+              (throw (ex-info "cargo build -p sorterc failed"
+                              {:stderr (:err r)}))))
+          dbg))))
+
+(defn compile-sorter [fixture-path]
+  (let [{:keys [out exit]} (p/shell {:out :string :err :string :continue true}
+                                    sorterc-bin "compile" fixture-path)]
+    (when-not (zero? exit)
+      (throw (ex-info "sorterc exit nonzero" {:fixture fixture-path :out out})))
+    (json/parse-string out true)))
+
+(defn first-component-ranking [result]
+  (-> result :rankings first :components first :ranking))
+
+(defn item-leaf [item]
+  (last (clojure.string/split item #"/")))
+
+(deftest chain-ranks-head-first
+  (let [ranking (first-component-ranking (compile-sorter "test/fixtures/ranking/chain.sorter"))
+        names   (mapv (comp item-leaf :item) ranking)]
+    (is (= ["a" "b" "c"] names)
+        "chain a>b>c should rank a, b, c in order")
+    (is (apply > (map :score ranking))
+        "scores should attenuate strictly down the chain")))
+
+(deftest inverse-star-puts-winners-on-top
+  (let [ranking (first-component-ranking (compile-sorter "test/fixtures/ranking/star_inverse.sorter"))
+        names   (mapv (comp item-leaf :item) ranking)]
+    (is (= "win" (last names))
+        "the item that lost to both others should be ranked last")))
+
+(deftest cycle-produces-uniform-scores
+  (let [ranking (first-component-ranking (compile-sorter "test/fixtures/ranking/cycle.sorter"))
+        scores  (map :score ranking)]
+    (is (every? #(< (Math/abs (- % 1/3)) 1e-3) scores)
+        "a perfectly symmetric 3-cycle should give every node ~1/3")))
+
+(deftest star-topology-winner-at-top
+  ;; Issue #146 regression: source-only star at default `>` ratio (2:1).
+  ;; Pre-fix produced uniform 1/3 scores; alphabetical fallback put the
+  ;; unambiguous winner at the bottom. Post-fix the chain is aperiodic and
+  ;; converges to π_zebra = 1/2, π_alpha = π_beta = 1/4.
+  (let [ranking (first-component-ranking (compile-sorter "test/fixtures/ranking/star.sorter"))
+        names   (mapv (comp item-leaf :item) ranking)
+        by-name (into {} (map (juxt (comp item-leaf :item) :score) ranking))]
+    (is (= "zebra" (first names))
+        "zebra won both votes and should rank #1")
+    (is (< (Math/abs (- (by-name "zebra") 0.5)) 1e-3))
+    (is (< (Math/abs (- (by-name "alpha") 0.25)) 1e-3))
+    (is (< (Math/abs (- (by-name "beta") 0.25)) 1e-3))))

download full diff B

Hardlinks — judgments / attempts / prompt

prompt download

judgments

attempts

Prompt text is loaded only by the download route.