diff --git a/AGENTS.md b/AGENTS.md
index 56320dd6b..231b9295c 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -34,13 +34,15 @@ Make an efficient learning agent that can do anything.
 
 ## Docs
 
-- [`docs/architecture.md`](docs/architecture.md) — Package dependency graph, per-package details, architectural patterns
-- [`docs/request-flow.md`](docs/request-flow.md) — Full request lifecycle from CLI through server and back
-- [`docs/error-schema.md`](docs/error-schema.md) — Server error response formats and client-side handling
-- [`docs/development.md`](docs/development.md) — Dev setup, worktrees, logs, package management, DB migrations
-- [`docs/testing.md`](docs/testing.md) — DI over mocking, tmux CLI testing
-- [`docs/environment-variables.md`](docs/environment-variables.md) — Env var rules, DI helpers, loading order
-- [`docs/agents-and-tools.md`](docs/agents-and-tools.md) — Agent system, shell shims, tool definitions
-- [`docs/patterns/handle-steps-generators.md`](docs/patterns/handle-steps-generators.md) — handleSteps generator patterns and spawn_agents tool calls
-- [docs/evalbuff/interpreting-task-prompts.md](docs/evalbuff/interpreting-task-prompts.md)
-- [docs/patterns/discover-before-implement.md](docs/patterns/discover-before-implement.md)
+IMPORTANT: Prefer retrieval-led reasoning over pre-training-led reasoning. Always read the relevant docs below before implementing changes.
+
+- `docs/architecture.md` — Package dependency graph, per-package details, architectural patterns
+- `docs/request-flow.md` — Full request lifecycle from CLI through server and back
+- `docs/error-schema.md` — Server error response formats and client-side handling
+- `docs/development.md` — Dev setup, worktrees, logs, package management, DB migrations
+- `docs/testing.md` — DI over mocking, tmux CLI testing
+- `docs/environment-variables.md` — Env var rules, DI helpers, loading order
+- `docs/agents-and-tools.md` — Agent system, shell shims, tool definitions
+- `docs/patterns/handle-steps-generators.md` — handleSteps generator patterns and spawn_agents tool calls
+- `docs/evalbuff/interpreting-task-prompts.md`
+- `docs/patterns/discover-before-implement.md`
diff --git a/evalbuff/src/run-carve-eval.ts b/evalbuff/src/run-carve-eval.ts
index 1d627d87b..d53b6e54f 100644
--- a/evalbuff/src/run-carve-eval.ts
+++ b/evalbuff/src/run-carve-eval.ts
@@ -24,6 +24,40 @@ import type { CarvedFeature, CarveResult, FileOperation } from './carve-features
 import type { JudgingResult, ReviewerAgentType } from './judge'
 import type { RunnerResult } from './runners/runner'
 
+// --- Doc read stats ---
+
+/** Extract doc file reads from an agent trace (JSONL of PrintModeEvents). */
+function extractDocReads(agentTrace: string): Record<string, number> {
+  const counts: Record<string, number> = {}
+  for (const line of agentTrace.split('\n')) {
+    if (!line.trim()) continue
+    try {
+      const event = JSON.parse(line)
+      if (event.type !== 'tool_call' || event.toolName !== 'Read') continue
+      const filePath: string = event.input?.file_path ?? ''
+      // Normalize to repo-relative path
+      const match = filePath.match(/(?:^|\/)(?:docs\/.*|AGENTS\.md|CLAUDE\.md)$/)
+      if (!match) continue
+      const relPath = match[0].startsWith('/') ? match[0].slice(1) : match[0]
+      counts[relPath] = (counts[relPath] || 0) + 1
+    } catch {
+      // not JSON
+    }
+  }
+  return counts
+}
+
+/** Merge multiple doc-read count maps into one (summing counts). */
+function mergeDocReads(maps: Record<string, number>[]): Record<string, number> {
+  const merged: Record<string, number> = {}
+  for (const m of maps) {
+    for (const [k, v] of Object.entries(m)) {
+      merged[k] = (merged[k] || 0) + v
+    }
+  }
+  return merged
+}
+
 // --- Apply carve operations to a repo directory ---
 
 function applyCarveOperations(repoDir: string, operations: FileOperation[]): void {
@@ -274,6 +308,8 @@ interface CarveEvalResult {
   docsKept: Array<{ path: string; reasoning: string; scoreBefore: number; scoreAfter: number }>
   docsRejected: Array<{ path: string; reasoning: string; scoreBefore: number; scoreAfter: number }>
   totalCost: number
+  /** Which doc files agents read and how many times (summed across all parallel runs). */
+  docsRead: Record<string, number>
 }
 
 async function runCarveEval(options: CarveEvalOptions): Promise<void> {
@@ -357,6 +393,7 @@ async function runCarveEval(options: CarveEvalOptions): Promise<void> {
         docsKept: [],
         docsRejected: [],
         totalCost,
+        docsRead: {},
       })
       continue
     }
@@ -368,6 +405,15 @@ async function runCarveEval(options: CarveEvalOptions): Promise<void> {
       `  Baseline: ${currentScore.toFixed(1)}/10 (${baselineScores.map((s) => s.toFixed(1)).join(', ')})`,
     )
 
+    // Track which docs agents read across all runs for this feature
+    const baselineDocReads = mergeDocReads(validBaseline.map((r) => extractDocReads(r.agentTrace)))
+    const docReadEntries = Object.entries(baselineDocReads).sort((a, b) => b[1] - a[1])
+    if (docReadEntries.length > 0) {
+      console.log(`  Docs read (baseline): ${docReadEntries.map(([p, n]) => `${p} (${n}x)`).join(', ')}`)
+    } else {
+      console.log(`  Docs read (baseline): none`)
+    }
+
     const docsKept: Array<{ path: string; reasoning: string; scoreBefore: number; scoreAfter: number }> = []
     const docsRejected: Array<{ path: string; reasoning: string; scoreBefore: number; scoreAfter: number }> = []
 
@@ -510,6 +556,7 @@ async function runCarveEval(options: CarveEvalOptions): Promise<void> {
       docsKept,
       docsRejected,
       totalCost,
+      docsRead: baselineDocReads,
     })
   }
 
@@ -525,6 +572,12 @@ async function runCarveEval(options: CarveEvalOptions): Promise<void> {
     console.log(`    Baseline: ${r.baselineScore.toFixed(1)}/10`)
     console.log(`    Final:    ${r.finalScore.toFixed(1)}/10`)
     console.log(`    Docs kept: ${r.docsKept.length}, rejected: ${r.docsRejected.length}`)
+    const readEntries = Object.entries(r.docsRead).sort((a, b) => b[1] - a[1])
+    if (readEntries.length > 0) {
+      console.log(`    Docs read: ${readEntries.map(([p, n]) => `${p} (${n}x)`).join(', ')}`)
+    } else {
+      console.log(`    Docs read: none`)
+    }
     console.log(`    Cost: $${r.totalCost.toFixed(2)}`)
     totalCostAll += r.totalCost
   }
@@ -538,6 +591,18 @@ async function runCarveEval(options: CarveEvalOptions): Promise<void> {
   console.log(`  Average final:    ${avgFinal.toFixed(1)}/10`)
   console.log(`  Total cost: $${totalCostAll.toFixed(2)}`)
 
+  // Aggregate doc read stats across all features
+  const allDocReads = mergeDocReads(results.map((r) => r.docsRead))
+  const allReadEntries = Object.entries(allDocReads).sort((a, b) => b[1] - a[1])
+  if (allReadEntries.length > 0) {
+    console.log(`\n  Doc read stats (all features):`)
+    for (const [docPath, count] of allReadEntries) {
+      console.log(`    ${docPath}: ${count} reads`)
+    }
+  } else {
+    console.log(`\n  No docs were read by any agent.`)
+  }
+
   // Save results
   const outputPath = path.join(
     repoPath,