From 7eef5877f8caf47e7a73c4478106c480221f27e0 Mon Sep 17 00:00:00 2001
From: ParsaVictor <1.parsa.karkooti@gmail.com>
Date: Fri, 9 Oct 2026 15:20:57 +0330
Subject: [PATCH 1/2] WIP: classes stay out of the function-level list

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---
 crates/neuromesh-graph/src/chunk_rank.rs | 40 +++++++++++++++---------
 crates/neuromesh-graph/src/graph.rs      | 15 +++++++--
 2 files changed, 38 insertions(+), 17 deletions(-)

diff --git a/crates/neuromesh-graph/src/chunk_rank.rs b/crates/neuromesh-graph/src/chunk_rank.rs
index 1930533..08c20f6 100644
--- a/crates/neuromesh-graph/src/chunk_rank.rs
+++ b/crates/neuromesh-graph/src/chunk_rank.rs
@@ -36,6 +36,9 @@ pub(crate) struct ChunkSource {
     pub spans: Vec<std::ops::Range<usize>>,
     /// The definition each span is (`Class.method`), same order as `spans`.
     pub names: Vec<String>,
+    /// Whether each span holds other definitions (a class): it still votes
+    /// for its file, but a function-level list names what is inside it.
+    pub containers: Vec<bool>,
 }
 
 /// One definition ranked for a report: where it is and how well it matched.
@@ -48,11 +51,10 @@ pub struct RankedDefinition {
 }
 
 /// A file's chunks as term counts, before they are numbered.
-type FileChunks = (
-    NodeId,
-    PathBuf,
-    Vec<(HashMap<String, u16>, String, usize, usize)>,
-);
+type FileChunks = (NodeId, PathBuf, Vec<(HashMap<String, u16>, Label)>);
+
+/// A chunk's definition name, first and last line, and whether it is a container.
+type Label = (String, usize, usize, bool);
 
 #[derive(Default)]
 pub(crate) struct ChunkRankIndex {
@@ -62,7 +64,7 @@ pub(crate) struct ChunkRankIndex {
     avg_len: f32,
     post: HashMap<String, Vec<(u32, u16)>>,
     /// Per chunk: definition name (`<head>` for the file head) and lines.
-    label: Vec<(String, usize, usize)>,
+    label: Vec<Label>,
 }
 
 impl ChunkRankIndex {
@@ -79,10 +81,15 @@ impl ChunkRankIndex {
                 };
                 let path_terms = terms_of(&stem_path, &st, true);
                 let lines: Vec<&str> = f.source.lines().collect();
-                let head = (1..HEAD_LINES.min(lines.len()), "<head>".to_string());
+                let head = (1..HEAD_LINES.min(lines.len()), ("<head>".to_string(), true));
                 let mut chunks = Vec::new();
-                let named = f.spans.iter().cloned().zip(f.names.iter().cloned());
-                for (span, name) in std::iter::once(head).chain(named) {
+                let named = f.spans.iter().cloned().zip(
+                    f.names
+                        .iter()
+                        .cloned()
+                        .zip(f.containers.iter().copied().chain(std::iter::repeat(false))),
+                );
+                for (span, (name, container)) in std::iter::once(head).chain(named) {
                     let start = span.start.max(1) - 1;
                     let end = span.end.min(lines.len()).min(start + MAX_CHUNK_LINES);
                     if start >= end {
@@ -97,7 +104,7 @@ impl ChunkRankIndex {
                         *c = c.saturating_add(1);
                     }
                     if !counts.is_empty() {
-                        chunks.push((counts, name, start + 1, end));
+                        chunks.push((counts, (name, start + 1, end, container)));
                     }
                 }
                 (f.id, f.path, chunks)
@@ -107,10 +114,10 @@ impl ChunkRankIndex {
         for (id, path, chunks) in per_file {
             let file = idx.files.len() as u32;
             idx.files.push((id, path));
-            for (counts, name, start, end) in chunks {
+            for (counts, label) in chunks {
                 let chunk = idx.file_of.len() as u32;
                 idx.file_of.push(file);
-                idx.label.push((name, start, end));
+                idx.label.push(label);
                 idx.len.push(counts.values().map(|&c| c as u32).sum());
                 for (t, c) in counts {
                     idx.post.entry(t).or_default().push((chunk, c));
@@ -122,19 +129,21 @@ impl ChunkRankIndex {
         idx
     }
 
-    /// Definitions by score for `prompt`, best first (file heads left out).
+    /// Functions and methods by score for `prompt`, best first. File heads and
+    /// containers (a class) are left out: on SWE-bench dev a class body outranked
+    /// the method the patch edits (func Acc@5 0.268 -> 0.311 without them).
     pub(crate) fn rank_definitions(&self, prompt: &str, limit: usize) -> Vec<RankedDefinition> {
         let mut scored: Vec<(u32, f32)> = self
             .scores(prompt)
             .into_iter()
-            .filter(|(c, _)| self.label[*c as usize].0 != "<head>")
+            .filter(|(c, _)| !self.label[*c as usize].3)
             .collect();
         scored.sort_by(|a, b| b.1.total_cmp(&a.1).then_with(|| a.0.cmp(&b.0)));
         scored
             .into_iter()
             .take(limit)
             .map(|(c, score)| {
-                let (name, start, end) = &self.label[c as usize];
+                let (name, start, end, _) = &self.label[c as usize];
                 RankedDefinition {
                     path: self.files[self.file_of[c as usize] as usize]
                         .1
@@ -231,6 +240,7 @@ mod tests {
             path: PathBuf::from(path),
             source: source.to_string(),
             names: spans.iter().map(|s| format!("def_{}", s.start)).collect(),
+            containers: vec![false; spans.len()],
             spans,
         }
     }
diff --git a/crates/neuromesh-graph/src/graph.rs b/crates/neuromesh-graph/src/graph.rs
index 1a841d6..bb7b0c8 100644
--- a/crates/neuromesh-graph/src/graph.rs
+++ b/crates/neuromesh-graph/src/graph.rs
@@ -825,8 +825,18 @@ impl NeuralProjectGraph {
                         }
                     }
                 }
+                // A definition is a container when another one in the file
+                // names it as its parent (a class with methods): it votes for
+                // its file but is left out of the function-level list.
+                let parents: HashSet<&str> = ids
+                    .iter()
+                    .filter_map(|id| data.mesh.node(id))
+                    .filter_map(|n| n.parent.as_deref())
+                    .collect();
+                let containers: Vec<bool> =
+                    names.iter().map(|n| parents.contains(n.as_str())).collect();
                 if let Some(id) = file_id {
-                    files.push((id, path.clone(), spans, names));
+                    files.push((id, path.clone(), spans, names, containers));
                 }
             }
             (key, files)
@@ -834,7 +844,7 @@ impl NeuralProjectGraph {
         let t0 = std::time::Instant::now();
         let sources: Vec<crate::chunk_rank::ChunkSource> = files
             .into_par_iter()
-            .filter_map(|(id, path, spans, names)| {
+            .filter_map(|(id, path, spans, names, containers)| {
                 let source = self.read_source(&path)?;
                 Some(crate::chunk_rank::ChunkSource {
                     id,
@@ -842,6 +852,7 @@ impl NeuralProjectGraph {
                     source,
                     spans,
                     names,
+                    containers,
                 })
             })
             .collect();

From 56853d5475cd7c3cf0ebdb76ce15ebdb57a8b35e Mon Sep 17 00:00:00 2001
From: ParsaVictor <1.parsa.karkooti@gmail.com>
Date: Fri, 9 Oct 2026 16:36:00 +0330
Subject: [PATCH 2/2] function level: classes out, file prior in (dev func
 Acc@5 0.268 -> 0.323); functions_to_look over MCP; Verified + ablation
 measured

- A definition that contains others (a class) still votes for its file but
  leaves the function list; `definition_order` fuses the function ranking with
  each definition's file rank (Agentless's function step, without the LLM).
  SWE-bench dev, 189 issues: func Acc@1/5/10 0.137/0.268/0.337 -> 0.175/0.323/0.402.
- MCP: a pasted report also gets `functions_to_look` (5, `path:Class.method Lx-Ly`).
- Measured (v1.2.0 engine, run once): SWE-bench Verified 496/500 Acc@1/3/5/10
  0.405/0.669/0.732/0.804 vs BM25 0.216/0.391/0.490/0.641; ablation on the Lite
  holdout: without the definition ranking 0.388/0.569/0.652, without report
  hygiene 0.467/0.707/0.743, packet order only 0.366/0.536/0.558.

Fourteen sets unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---
 crates/neuromesh-cli/src/commands/packet.rs | 20 +++++-------
 crates/neuromesh-context/src/gold.rs        | 34 +++++++++++++++++++++
 crates/neuromesh-mcp/src/tools.rs           | 28 ++++++++++++-----
 docs/measured.md                            | 11 ++++++-
 docs/research/contributions-log.md          | 34 +++++++++++++++++++++
 docs/research/paper-draft.md                | 10 +++---
 6 files changed, 111 insertions(+), 26 deletions(-)

diff --git a/crates/neuromesh-cli/src/commands/packet.rs b/crates/neuromesh-cli/src/commands/packet.rs
index f793557..f7698e5 100644
--- a/crates/neuromesh-cli/src/commands/packet.rs
+++ b/crates/neuromesh-cli/src/commands/packet.rs
@@ -104,6 +104,12 @@ pub fn execute(args: &[String]) -> Result<()> {
 
     let selected_paths = neuromesh_context::gold::packet_file_order(&graph, &prompt, &view);
     let localization = neuromesh_context::gold::localization_order(&graph, &prompt, &view, 20);
+    let definitions =
+        if prompt.split_whitespace().count() >= neuromesh_parser::text_normalize::REPORT_WORDS {
+            neuromesh_context::gold::definition_order(&graph, &prompt, &localization, def_depth())
+        } else {
+            Vec::new()
+        };
     let mut files: Vec<String> = packet_file_names(&view).into_iter().collect();
     files.sort();
     let reduction = if workspace_tokens > 0 {
@@ -134,19 +140,7 @@ pub fn execute(args: &[String]) -> Result<()> {
         selected_files: files.clone(),
         selected_paths,
         localization,
-        definitions: if prompt.split_whitespace().count()
-            >= neuromesh_parser::text_normalize::REPORT_WORDS
-        {
-            graph
-                .rank_definitions(&prompt, def_depth())
-                .into_iter()
-                .filter(|d| {
-                    !neuromesh_context::selector::is_noise_path(std::path::Path::new(&d.path))
-                })
-                .collect()
-        } else {
-            Vec::new()
-        },
+        definitions,
         ranked_paths: graph
             .file_rank(&prompt, 10)
             .into_iter()
diff --git a/crates/neuromesh-context/src/gold.rs b/crates/neuromesh-context/src/gold.rs
index 9730ee6..a584ca5 100644
--- a/crates/neuromesh-context/src/gold.rs
+++ b/crates/neuromesh-context/src/gold.rs
@@ -1135,6 +1135,40 @@ pub fn localization_order(
     out
 }
 
+/// Functions and methods to look at for a report, best first, `k` deep: the
+/// definition ranking fused (RRF, k = 60) with the rank of each definition's
+/// file in [`localization_order`] — the function step Agentless takes inside
+/// the files it chose. SWE-bench dev (190 issues with function gold), func
+/// Acc@5/@10: 0.311/0.374 without the file prior → 0.321/0.400 with it.
+pub fn definition_order(
+    graph: &neuromesh_graph::NeuralProjectGraph,
+    prompt: &str,
+    files: &[String],
+    k: usize,
+) -> Vec<neuromesh_graph::RankedDefinition> {
+    const K: f32 = 60.0;
+    let file_rank: std::collections::HashMap<&str, usize> = files
+        .iter()
+        .enumerate()
+        .map(|(i, f)| (f.as_str(), i))
+        .collect();
+    let mut scored: Vec<(f32, usize, neuromesh_graph::RankedDefinition)> = graph
+        .rank_definitions(prompt, 300)
+        .into_iter()
+        .filter(|d| !crate::selector::is_noise_path(std::path::Path::new(&d.path)))
+        .enumerate()
+        .map(|(i, d)| {
+            let mut s = 1.0 / (K + i as f32 + 1.0);
+            if let Some(r) = file_rank.get(d.path.as_str()) {
+                s += 1.0 / (K + *r as f32 + 1.0);
+            }
+            (s, i, d)
+        })
+        .collect();
+    scored.sort_by(|a, b| b.0.total_cmp(&a.0).then_with(|| a.1.cmp(&b.1)));
+    scored.into_iter().take(k).map(|(_, _, d)| d).collect()
+}
+
 /// Indexed files a report names, best pointer first: traceback frames
 /// (`File "…/pkg/mod.py", line 12`) from the deepest up, then any other path
 /// or file name in order of mention. A path resolves by its longest suffix
diff --git a/crates/neuromesh-mcp/src/tools.rs b/crates/neuromesh-mcp/src/tools.rs
index 331b80b..2500746 100644
--- a/crates/neuromesh-mcp/src/tools.rs
+++ b/crates/neuromesh-mcp/src/tools.rs
@@ -514,15 +514,27 @@ impl McpToolHandler {
                         >= neuromesh_parser::text_normalize::REPORT_WORDS
                     {
                         if let Some(obj) = value.as_object_mut() {
-                            obj.insert(
-                                "where_to_look".into(),
-                                json!(neuromesh_context::gold::localization_order(
-                                    &self.graph,
-                                    &task_desc,
-                                    &view,
-                                    10
-                                )),
+                            let files = neuromesh_context::gold::localization_order(
+                                &self.graph,
+                                &task_desc,
+                                &view,
+                                10,
                             );
+                            // The functions inside those files to read first
+                            // (`path:Class.method Lx-Ly`), five deep.
+                            let functions: Vec<String> = neuromesh_context::gold::definition_order(
+                                &self.graph,
+                                &task_desc,
+                                &files,
+                                5,
+                            )
+                            .into_iter()
+                            .map(|d| format!("{}:{} L{}-L{}", d.path, d.name, d.lines.0, d.lines.1))
+                            .collect();
+                            obj.insert("where_to_look".into(), json!(files));
+                            if !functions.is_empty() {
+                                obj.insert("functions_to_look".into(), json!(functions));
+                            }
                         }
                     }
                     neuromesh_graph::timing("mcp: cache_and_build", t_s);
diff --git a/docs/measured.md b/docs/measured.md
index ce3248d..d073844 100644
--- a/docs/measured.md
+++ b/docs/measured.md
@@ -74,7 +74,16 @@ with Claude-3.5 0.726 / 0.792 / 0.796; LocAgent with Claude-3.5 0.777 / 0.920 /
 remain ahead at Acc@1; the engine is an LLM-free first stage they can start from.
 
 SWE-bench dev (tuning set, 225): plain BM25 0.160 / 0.347 / 0.427 / 0.538 → 0.249 / 0.484 / 0.600 /
-0.680. Raw results: `swebench/results-final-test.jsonl` (outside the repository); harness
+0.680.
+
+**SWE-bench Verified** (500 human-validated issues, v1.2.0 engine, run once; 496 scored, 4 lost to
+git checkout errors): plain BM25 0.216 / 0.391 / 0.490 / 0.641 → **0.405 / 0.669 / 0.732 / 0.804**
+(Acc@1/3/5/10). On the 403 Verified issues that are not in Lite: BM25 0.194 / 0.372 / 0.476 / 0.620
+→ 0.392 / 0.655 / 0.727 / 0.809.
+
+**Which part does the work** (Lite strict holdout, one run per removed component, Acc@1/3/5):
+definition-level ranking removed 0.388 / 0.569 / 0.652; report hygiene removed 0.467 / 0.707 /
+0.743; packet order only 0.366 / 0.536 / 0.558; full 0.486 / 0.710 / 0.750. Raw results: `swebench/results-final-test.jsonl` (outside the repository); harness
 `scripts/swebench_localize.py`, scoring `scripts/research/eval_loc.py`.
 
 ## Speed (same machine, same hour, release binary without embeddings)
diff --git a/docs/research/contributions-log.md b/docs/research/contributions-log.md
index 59f96c7..c61fafe 100644
--- a/docs/research/contributions-log.md
+++ b/docs/research/contributions-log.md
@@ -361,3 +361,37 @@ functions in the top k). dev-fast (48 issues with function gold): func Acc@1/5/1
 0.062/0.229/0.292/0.396 (raw definition ranking, no file prior). Reference (LocAgent Table 4, Lite,
 function level Acc@5/@10): BM25 0.318/0.369, CodeRankEmbed 0.518/0.588, Agentless+Claude 0.588,
 LocAgent+Claude 0.734/0.774. File-prior orderings (`func_hier.py`) pending the full-dev run.
+
+### 8.11 SWE-bench Verified (second standard benchmark) and the ablation table (2026-10-09)
+
+**Verified, v1.2.0 engine, run once** (496 of 500; 4 lost to git checkout errors):
+
+| | Acc@1 | Acc@3 | Acc@5 | Acc@10 |
+|---|---|---|---|---|
+| engine, 496 | **0.405** [0.36,0.45] | **0.669** [0.63,0.71] | **0.732** [0.69,0.77] | **0.804** |
+| BM25, 496 | 0.216 | 0.391 | 0.490 | 0.641 |
+| engine, 403 never seen before (not in Lite) | 0.392 | 0.655 | 0.727 | 0.809 |
+| BM25, same 403 | 0.194 | 0.372 | 0.476 | 0.620 |
+
+**Ablation on the Lite strict holdout (276)**, one run per removed component (`NM_ABLATE` switch
+build, not shipped):
+
+| removed | Acc@1 | Acc@3 | Acc@5 | Acc@10 |
+|---|---|---|---|---|
+| nothing (v1.2.0) | 0.486 | 0.710 | 0.750 | 0.804 |
+| report hygiene + code tokens (C11) | 0.467 | 0.707 | 0.743 | 0.801 |
+| definition-level ranking (C12) | **0.388** | **0.569** | **0.652** | 0.764 |
+| localisation list — packet order only | 0.366 | 0.536 | 0.558 | – |
+
+Reading: the definition-level ranking (LocAgent's entity-content BM25 idea, made LLM-free) carries
+most of the gain over the packet (+0.10 @1, +0.14 @3). Hygiene is small on Lite (−0.019 @1 when
+removed) — its dev gains came mostly from sqlfluff's template-heavy issues; reported as is.
+
+### 8.12 Function level: classes out, file prior in (engine, dev 225)
+
+Classes (a definition another one names as its parent) stay in the file ranking but leave the
+function list; the list is fused (RRF) with each definition's file rank in `where_to_look`.
+Engine on SWE-bench dev (189 issues with function gold), func Acc@1/5/10/20: 0.137/0.268/0.337/–
+→ **0.175/0.323/0.402/0.460** (prototype predicted 0.174/0.321/0.400). Same run, file level with
+everything since v1.2.0 (named files): 0.249/0.484/0.600/0.680 → **0.308/0.527/0.621/0.692**.
+MCP: reports now also get `functions_to_look` (five, `path:Class.method Lx-Ly`).
diff --git a/docs/research/paper-draft.md b/docs/research/paper-draft.md
index 74255ed..a84d3cc 100644
--- a/docs/research/paper-draft.md
+++ b/docs/research/paper-draft.md
@@ -70,17 +70,19 @@ tokens, definition-level BM25 with title ×3, RRF with the packet order → `whe
 
 † LocAgent Table 4, their 274-instance subset (head-to-head on the same subset: TODO).
 
-### 5.2 SWE-bench dev (225) and Verified (500, TODO)
+### 5.2 SWE-bench dev (225) and Verified (496 of 500)
 
 dev: BM25 0.160/0.347/0.427/0.538 → engine 0.249/0.484/0.600/0.680 (Acc@1/3/5/10).
+Verified (v1.2.0, run once): BM25 0.216/0.391/0.490/0.641 → engine 0.405/0.669/0.732/0.804; on the
+403 Verified issues not in Lite: 0.194/0.372/0.476/0.620 → 0.392/0.655/0.727/0.809.
 
-### 5.3 Ablation (Lite holdout) — TODO (runs in progress)
+### 5.3 Ablation (Lite holdout, 276)
 
 | removed | Acc@1 | Acc@3 | Acc@5 |
 |---|---|---|---|
 | nothing | 0.486 | 0.710 | 0.750 |
-| report hygiene + code tokens | TODO | | |
-| definition-level ranking | TODO | | |
+| report hygiene + code tokens | 0.467 | 0.707 | 0.743 |
+| definition-level ranking | 0.388 | 0.569 | 0.652 |
 | localisation list (packet order only) | 0.366 | 0.536 | 0.558 |
 
 ### 5.4 Plain-language holdouts (R@3)