diff --git a/CONTEXT.md b/CONTEXT.md
index 78858dd..82d1f94 100644
--- a/CONTEXT.md
+++ b/CONTEXT.md
@@ -297,6 +297,15 @@ must all be present and pass aggregate, query-form, and repository guardrails. F
 diagnostic and inherit that decision. The last grouped fold is the explicit untouched final test and
 must pass independently. Artifacts retain exact blocker values, deterministic grouped bootstrap
 intervals, and fold-level selection/local-perturbation stability diagnostics.
+Schema 26 adds binary NDCG@5/10/20/50 without changing the corpus or Production retrieval. Unique chunks
+matching at least one resolved exact file-plus-symbol target receive gain one; duplicates and chunks
+matching multiple targets cannot add repeated gain, and zero ideal gain returns NDCG zero. The direct
+objective selects NDCG@5 first and is compared with a `direct-recall-first` ablation under identical
+folds, seeds, and guardrails. Recall@20, Recall@50, and ContextRecall@4096 remain hard direct-retrieval
+coverage guardrails; reranker objectives remain recall-first with the full Recall@5/10/20/50 and
+ContextRecall@4096 guardrail set. Empty authored ground truth is rejected by the manifest schema. The
+Recall-first direct ablation selects from the same archive without changing the three-scenario beam
+budget, and both proxy-promotion and successive-halving perform objective-specific archive selection.
 
 ### Scorer
 
diff --git a/benchmarks/BASELINE.md b/benchmarks/BASELINE.md
index 4eb97f6..666bbd3 100644
--- a/benchmarks/BASELINE.md
+++ b/benchmarks/BASELINE.md
@@ -1,7 +1,26 @@
 # Preliminary Retrieval Baseline
 
 The schema-17 entries below are historical artifacts from the benchmark-owned Sparse implementation.
-Current schema-24 runs use the production SparseEmbedder and IndexStore without benchmark vector caches.
+Current schema-26 runs use the production SparseEmbedder and IndexStore without benchmark vector caches.
+
+## Schema 26: Binary NDCG Direct-Objective Smoke
+
+The `fd` MiniLM smoke artifact `retrieval-2026-08-06T18-27-13.604Z.json` compares the new NDCG@5-first
+`direct` objective with the historical `direct-recall-first` priority under the same grouped folds,
+seeds, DBSF fusion, and hard Recall@20/50 plus Context@4k guardrails. The artifact is local and ignored.
+
+| Objective      | NDCG@5 | NDCG@10 | NDCG@20 | NDCG@50 |   R@5 |  R@20 | Context@4k |
+| -------------- | -----: | ------: | ------: | ------: | ----: | ----: | ---------: |
+| NDCG@5-first   |  57.0% |   61.5% |   62.5% |   64.2% | 73.3% | 91.3% |      87.3% |
+| Recall@5-first |  59.8% |   63.2% |   64.2% |   65.8% | 76.7% | 91.3% |      87.3% |
+
+This single-repository smoke run validates the complete measurement and ablation path, but does not
+support promoting the NDCG-first candidate: Recall-first is stronger on every reported holdout NDCG
+cutoff and R@5 while tying the coverage guardrails. A full multi-repository run is still required for
+any Production decision; issue #172 itself remains benchmark-only.
+The matched Successive-Halving smoke artifact `retrieval-2026-08-06T18-27-50.835Z.json` also completed
+with objective-specific archive selection, validating that the ablation is executed rather than emitted
+as repeated labels in both supported search strategies.
 
 ## Schema 24: Selectable Router Search
 
diff --git a/benchmarks/README.md b/benchmarks/README.md
index c863ac9..31b9d22 100644
--- a/benchmarks/README.md
+++ b/benchmarks/README.md
@@ -9,7 +9,7 @@ surface.
 The suite answers five questions independently:
 
 1. Do identity, CamelCase, BM25, dense, and sparse retrieval each contribute useful candidates?
-2. Does production-weighted RRF improve `Recall@K` over individual channels and combinations?
+2. Does the production DBSF configuration improve `Recall@K` and binary `NDCG@K` over individual channels and combinations?
 3. Does full RRF reduce the quality gap between a small general embedder and a code-specific one?
 4. How much authored ground truth fits into fixed context budgets?
 5. Can one evidence-based router outperform static weights without knowing the authored query form?
@@ -53,6 +53,12 @@ representations remain in the same validation fold. Gold targets are exact `file
 Matching uses extracted identifiers or exact declaration syntax inside the specified file; loose
 symbol-name segment matching is not accepted.
 
+The existing corpus needs no NDCG-specific labels. Every resolved chunk matching one or more exact
+gold targets has binary gain `1`; non-matching chunks have gain `0`. A chunk that resolves multiple
+targets and a duplicate occurrence of the same chunk each contribute at most once. NDCG is `0` when
+no gold target resolves. Manifest decoding rejects empty ground-truth arrays, and corpus preparation
+rejects authored targets that do not resolve to a chunk.
+
 | Corpus    | Revision                                   | Language   | Size band | Indexed scope                 |
 | --------- | ------------------------------------------ | ---------- | --------- | ----------------------------- |
 | FastAPI   | `95f8322ee1dcda7ceace7b1c4f6c9915b36d748f` | Python     | medium    | `fastapi/**/*.py`             |
@@ -95,7 +101,7 @@ vp run bench:retrieval:full
 
 `bench:retrieval` aliases `bench:retrieval:validate`. Every profile measures the same physical
 rankings and retrieval variants; profiles only control matrix size, holdout coverage, and expensive
-diagnostics. The selected profile is recorded in schema-24 artifacts without changing retrieval
+diagnostics. The selected profile is recorded in schema-26 artifacts without changing retrieval
 semantics. The full profile includes all three fusion methods; short profiles intentionally omit RRF
 to keep development runs fast.
 
@@ -158,8 +164,8 @@ defaulting to MiniLM. Select another with `PIX_BENCH_MODELS`. Supported values a
 The router search defaults to `proxy-promotion`. Set `PIX_BENCH_ROUTER_STRATEGY` to
 `successive-halving` to select the historical Successive-Halving variant. It uses the original
 lexicographic `R@20`, `R@10`, `Context@4k`, and MRR comparator plus its `halvingKeepFactor`.
-Both strategies use the same candidate evaluator and native worker queue, so their artifacts can be
-compared directly.
+Both strategies use the same candidate evaluator and native worker queue. Their final archive selection
+is objective-specific, so Direct and Reranker rows are real comparisons rather than repeated labels.
 
 The Jina code model cannot embed Effect's longest 7,103-token AST chunk on the tested DML GPU even as
 a single-item batch. Do not silently truncate, re-chunk only one model, or mix CPU and GPU vectors to
@@ -252,6 +258,14 @@ single-coordinate perturbations, epsilon-neighbor fraction, and median/worst hol
 currently deterministic with one seed and restart; those counts are explicit rather than implying
 unmeasured restart stability.
 
+Schema 26 reports binary NDCG at 5, 10, 20, and 50 for raw query-form/repository/model rows and for
+aggregate, fusion, and partitioned holdouts. The `direct` objective prioritizes NDCG@5, while the
+`direct-recall-first` ablation retains the historical Recall@5-first priority under the same folds,
+guardrails, and seeds. Both direct objectives retain Recall@20, Recall@50, and ContextRecall@4096 as
+hard coverage guardrails. Reranker-top20 and reranker-top50 remain recall-first and retain the complete
+Recall@5/10/20/50 plus ContextRecall@4096 guardrail set. The direct ablation selects from the same archive
+without consuming another beam-diversity slot, preserving the original search budget.
+
 Each router result also records a deterministic `random-scout` baseline using the same parameter grid and
 global-scout budget. `Random R@20` and `Random Ctx@4k` show whether the structured search beats that
 equal-budget random sample; the artifact records the effective seed and candidate count for reproduction.
diff --git a/benchmarks/retrieval/evaluation/collect.ts b/benchmarks/retrieval/evaluation/collect.ts
index 7090058..fbdae95 100644
--- a/benchmarks/retrieval/evaluation/collect.ts
+++ b/benchmarks/retrieval/evaluation/collect.ts
@@ -36,6 +36,7 @@ import { foldKey } from "./folds.js"
 import {
   contextRecallAtBudget,
   goldTargetRanks,
+  normalizedDiscountedCumulativeGain,
   recallAt,
   reciprocalRank,
   resolveGoldTargets,
@@ -257,6 +258,10 @@ const buildModelSamples = (
           recallAt10: recallAt(ranked, targets, 10),
           recallAt20: recallAt(ranked, targets, 20),
           recallAt50: recallAt(ranked, targets, 50),
+          ndcgAt5: normalizedDiscountedCumulativeGain(ranked, targets, 5),
+          ndcgAt10: normalizedDiscountedCumulativeGain(ranked, targets, 10),
+          ndcgAt20: normalizedDiscountedCumulativeGain(ranked, targets, 20),
+          ndcgAt50: normalizedDiscountedCumulativeGain(ranked, targets, 50),
           successAt10: successAt(ranked, targets, 10),
           successAt20: successAt(ranked, targets, 20),
           reciprocalRank: reciprocalRank(ranked, targets),
diff --git a/benchmarks/retrieval/evaluation/metrics-core.d.mts b/benchmarks/retrieval/evaluation/metrics-core.d.mts
new file mode 100644
index 0000000..e161c2c
--- /dev/null
+++ b/benchmarks/retrieval/evaluation/metrics-core.d.mts
@@ -0,0 +1,6 @@
+/** Binary NDCG at K for ranked chunk IDs and resolved exact-target chunk IDs. */
+export declare const binaryNdcgAt: (
+  ranked: readonly number[],
+  targets: readonly (readonly number[])[],
+  k: number,
+) => number
diff --git a/benchmarks/retrieval/evaluation/metrics-core.mjs b/benchmarks/retrieval/evaluation/metrics-core.mjs
new file mode 100644
index 0000000..7511ad5
--- /dev/null
+++ b/benchmarks/retrieval/evaluation/metrics-core.mjs
@@ -0,0 +1,19 @@
+/** Binary NDCG at K for ranked chunk IDs and resolved exact-target chunk IDs. */
+export const binaryNdcgAt = (ranked, targets, k) => {
+  const relevant = new Set(targets.flatMap((target) => target))
+  if (relevant.size === 0 || k <= 0) return 0
+
+  const seen = new Set()
+  let discountedGain = 0
+  for (let rank = 0; rank < Math.min(k, ranked.length); rank++) {
+    const chunkIndex = ranked[rank]
+    if (seen.has(chunkIndex)) continue
+    seen.add(chunkIndex)
+    if (relevant.has(chunkIndex)) discountedGain += 1 / Math.log2(rank + 2)
+  }
+
+  let idealDiscountedGain = 0
+  for (let rank = 0; rank < Math.min(k, relevant.size); rank++)
+    idealDiscountedGain += 1 / Math.log2(rank + 2)
+  return discountedGain / idealDiscountedGain
+}
diff --git a/benchmarks/retrieval/evaluation/metrics.ts b/benchmarks/retrieval/evaluation/metrics.ts
index 24f19aa..cdf2de2 100644
--- a/benchmarks/retrieval/evaluation/metrics.ts
+++ b/benchmarks/retrieval/evaluation/metrics.ts
@@ -1,5 +1,6 @@
 import type { Chunk } from "../../../src/domain/chunk.js"
 import type { RankedChunk } from "../../../src/domain/ports.js"
+import { binaryNdcgAt } from "./metrics-core.mjs"
 import type { GoldLocation } from "./types.js"
 
 /** Indexed identifiers retained per chunk so gold symbols can be matched exactly. */
@@ -63,6 +64,19 @@ export const reciprocalRank = (
   return rank < 0 ? 0 : 1 / (rank + 1)
 }
 
+/** Binary NDCG at K over unique chunks that resolve at least one exact gold target. */
+export const normalizedDiscountedCumulativeGain = (
+  ranked: readonly RankedChunk[],
+  targets: readonly ReadonlySet<number>[],
+  k: number,
+): number => {
+  return binaryNdcgAt(
+    ranked.map((entry) => entry.chunkIndex),
+    targets.map((target) => [...target]),
+    k,
+  )
+}
+
 /** One-based best rank for every authored target, or null when the channel did not retrieve it. */
 export const goldTargetRanks = (
   ranked: readonly RankedChunk[],
diff --git a/benchmarks/retrieval/evaluation/optimization-profiles.ts b/benchmarks/retrieval/evaluation/optimization-profiles.ts
index 3c46402..95df699 100644
--- a/benchmarks/retrieval/evaluation/optimization-profiles.ts
+++ b/benchmarks/retrieval/evaluation/optimization-profiles.ts
@@ -9,6 +9,10 @@ import {
 
 const NonNegativeNumber = Schema.Number.check(Schema.isGreaterThanOrEqualTo(0))
 const QualityMetrics = [
+  "ndcgAt5",
+  "ndcgAt10",
+  "ndcgAt20",
+  "ndcgAt50",
   "recallAt5",
   "recallAt10",
   "recallAt20",
@@ -73,7 +77,10 @@ const objective = (
 ): typeof MetricObjectiveSchema.Type => ({
   name,
   priority,
-  guardrailMetrics: ["recallAt5", "recallAt10", "recallAt20", "recallAt50", "contextRecallAt4096"],
+  guardrailMetrics:
+    name === "direct"
+      ? ["recallAt20", "recallAt50", "contextRecallAt4096"]
+      : ["recallAt5", "recallAt10", "recallAt20", "recallAt50", "contextRecallAt4096"],
   guardrailTolerance: 0.01,
   contextBudget: 4_096,
 })
@@ -147,6 +154,9 @@ export const SEARCH_PRIORITY_PROFILE = profile(
   COMPATIBILITY_PROFILE_CONFIG,
   { identifier: 1, agentTask: 2, naturalQuestion: 3, searchPhrase: 4 },
   objective("direct", [
+    "ndcgAt5",
+    "ndcgAt10",
+    "ndcgAt20",
     "recallAt5",
     "recallAt10",
     "contextRecallAt4096",
@@ -183,6 +193,9 @@ export const OPTIMIZATION_PROFILES = {
     BASIC_EXPLORATION_PROFILE_CONFIG,
     { identifier: 1, agentTask: 3, naturalQuestion: 4, searchPhrase: 2 },
     objective("direct", [
+      "ndcgAt5",
+      "ndcgAt10",
+      "ndcgAt20",
       "recallAt5",
       "recallAt10",
       "recallAt20",
diff --git a/benchmarks/retrieval/evaluation/prepared-fusion-core.mjs b/benchmarks/retrieval/evaluation/prepared-fusion-core.mjs
index d8a9f64..fdfd21c 100644
--- a/benchmarks/retrieval/evaluation/prepared-fusion-core.mjs
+++ b/benchmarks/retrieval/evaluation/prepared-fusion-core.mjs
@@ -1,3 +1,5 @@
+import { binaryNdcgAt } from "./metrics-core.mjs"
+
 export const evaluatePreparedContributions = (matrix, weights) => {
   const activeMask =
     Number(weights.identity > 0) |
@@ -64,6 +66,10 @@ export const evaluateCandidate = (snapshot, candidate) => {
   let recall10 = 0
   let recall20 = 0
   let recall50 = 0
+  let ndcg5 = 0
+  let ndcg10 = 0
+  let ndcg20 = 0
+  let ndcg50 = 0
   let contextRecall = 0
   let meanReciprocalRank = 0
   let totalWeight = 0
@@ -81,6 +87,10 @@ export const evaluateCandidate = (snapshot, candidate) => {
     recall10 += weight * recallAt(ranked, sample.targets, 10)
     recall20 += weight * recallAt(ranked, sample.targets, 20)
     recall50 += weight * recallAt(ranked, sample.targets, 50)
+    ndcg5 += weight * binaryNdcgAt(ranked, sample.targets, 5)
+    ndcg10 += weight * binaryNdcgAt(ranked, sample.targets, 10)
+    ndcg20 += weight * binaryNdcgAt(ranked, sample.targets, 20)
+    ndcg50 += weight * binaryNdcgAt(ranked, sample.targets, 50)
     contextRecall += weight * contextRecallAtBudget(ranked, sample, 4096)
     meanReciprocalRank += weight * reciprocalRank(ranked, sample.targets)
     totalWeight += weight
@@ -92,6 +102,10 @@ export const evaluateCandidate = (snapshot, candidate) => {
       recallAt10: 0,
       recallAt20: 0,
       recallAt50: 0,
+      ndcgAt5: 0,
+      ndcgAt10: 0,
+      ndcgAt20: 0,
+      ndcgAt50: 0,
       contextRecallAt4096: 0,
       meanReciprocalRank: 0,
     }
@@ -101,6 +115,10 @@ export const evaluateCandidate = (snapshot, candidate) => {
     recallAt10: recall10 / totalWeight,
     recallAt20: recall20 / totalWeight,
     recallAt50: recall50 / totalWeight,
+    ndcgAt5: ndcg5 / totalWeight,
+    ndcgAt10: ndcg10 / totalWeight,
+    ndcgAt20: ndcg20 / totalWeight,
+    ndcgAt50: ndcg50 / totalWeight,
     contextRecallAt4096: contextRecall / totalWeight,
     meanReciprocalRank: meanReciprocalRank / totalWeight,
   }
diff --git a/benchmarks/retrieval/evaluation/promotion-evidence.ts b/benchmarks/retrieval/evaluation/promotion-evidence.ts
index 7a892bf..a5c065c 100644
--- a/benchmarks/retrieval/evaluation/promotion-evidence.ts
+++ b/benchmarks/retrieval/evaluation/promotion-evidence.ts
@@ -14,6 +14,10 @@ import type {
 const BOOTSTRAP_SAMPLES = 1_000
 const STABILITY_EPSILON = 0.005
 const QUALITY_METRICS: readonly QualityMetric[] = [
+  "ndcgAt5",
+  "ndcgAt10",
+  "ndcgAt20",
+  "ndcgAt50",
   "recallAt5",
   "recallAt10",
   "recallAt20",
@@ -124,7 +128,8 @@ const buildUncertainty = (rows: readonly PromotionHoldoutRow[]): readonly Holdou
 }
 
 const objectiveMetric = (objective: RouterObjective): QualityMetric => {
-  if (objective === "direct") return "recallAt5"
+  if (objective === "direct") return "ndcgAt5"
+  if (objective === "direct-recall-first") return "recallAt5"
   return objective === "reranker-top20" ? "recallAt20" : "recallAt50"
 }
 
diff --git a/benchmarks/retrieval/evaluation/report.ts b/benchmarks/retrieval/evaluation/report.ts
index 3285f96..11c5161 100644
--- a/benchmarks/retrieval/evaluation/report.ts
+++ b/benchmarks/retrieval/evaluation/report.ts
@@ -137,6 +137,8 @@ export const renderMarkdownReport = (artifact: BenchmarkArtifact): string => {
       .map(([kind, weight]) => `${kind}=${weight}`)
       .join(", ")}.`,
     "",
+    "The `direct` objective selects NDCG@5 first; `direct-recall-first` is the matched historical-objective ablation. Reranker objectives remain recall-first.",
+    "",
     `Validation protocol: candidates use ${artifact.validationProtocol.selection}; holdouts are ${artifact.validationProtocol.holdouts.join(" and ")}; final promotion requires untouched ${artifact.validationProtocol.finalTest.strategy} fold ${artifact.validationProtocol.finalTest.fold}.`,
     "",
     `Search strategy: \`${artifact.searchStrategy.algorithm}\` (${artifact.searchStrategy.globalScouts} global scouts, beam ${artifact.searchStrategy.beamWidth}, ${artifact.searchStrategy.coordinatePasses} coordinate passes, ${artifact.searchStrategy.proxySampleFraction * 100}% proxy with minimum ${artifact.searchStrategy.proxyMinimumSamples}, ${strategyFactorLabel} factor ${strategyFactor}x).`,
@@ -162,14 +164,14 @@ export const renderMarkdownReport = (artifact: BenchmarkArtifact): string => {
         `| ${run.repository} | ${run.model} | ${run.tokenizerModel} | ${run.batchSize} | ${duration(run.chunkEmbeddingDurationMs)} | ${duration(run.queryTokenizationDurationMs)} |`,
     ),
     "",
-    "| Repository | Model | Query form | Variant | R@5 | R@10 | R@20 | R@50 | S@10 | S@20 | MRR | Ctx@2k | Ctx@4k |",
-    "| --- | --- | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |",
+    "| Repository | Model | Query form | Variant | R@5 | R@10 | R@20 | R@50 | NDCG@5 | NDCG@10 | NDCG@20 | NDCG@50 | S@10 | S@20 | MRR | Ctx@2k | Ctx@4k |",
+    "| --- | --- | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |",
   ]
 
   for (const [key, rows] of [...groups].sort(([left], [right]) => left.localeCompare(right))) {
     const [repository, model, queryKind, variant] = key.split("\0")
     lines.push(
-      `| ${repository} | ${model} | ${queryKind} | ${variant} | ${percent(average(rows, (row) => row.recallAt5))} | ${percent(average(rows, (row) => row.recallAt10))} | ${percent(average(rows, (row) => row.recallAt20))} | ${percent(average(rows, (row) => row.recallAt50))} | ${percent(average(rows, (row) => Number(row.successAt10)))} | ${percent(average(rows, (row) => Number(row.successAt20)))} | ${average(rows, (row) => row.reciprocalRank).toFixed(3)} | ${percent(average(rows, (row) => row.contextRecall["2048"] ?? 0))} | ${percent(average(rows, (row) => row.contextRecall["4096"] ?? 0))} |`,
+      `| ${repository} | ${model} | ${queryKind} | ${variant} | ${percent(average(rows, (row) => row.recallAt5))} | ${percent(average(rows, (row) => row.recallAt10))} | ${percent(average(rows, (row) => row.recallAt20))} | ${percent(average(rows, (row) => row.recallAt50))} | ${percent(average(rows, (row) => row.ndcgAt5))} | ${percent(average(rows, (row) => row.ndcgAt10))} | ${percent(average(rows, (row) => row.ndcgAt20))} | ${percent(average(rows, (row) => row.ndcgAt50))} | ${percent(average(rows, (row) => Number(row.successAt10)))} | ${percent(average(rows, (row) => Number(row.successAt20)))} | ${average(rows, (row) => row.reciprocalRank).toFixed(3)} | ${percent(average(rows, (row) => row.contextRecall["2048"] ?? 0))} | ${percent(average(rows, (row) => row.contextRecall["4096"] ?? 0))} |`,
     )
   }
 
@@ -297,7 +299,7 @@ export const renderMarkdownReport = (artifact: BenchmarkArtifact): string => {
     "",
     "## Evidence Router Holdouts",
     "",
-    "One shared search produces direct, reranker-top20, and reranker-top50 candidates. The current Production router is the guardrail baseline; static and dynamic validation columns use the same fusion method and excluded fold. A no eligible candidate result must not be promoted.",
+    "One shared search produces NDCG-first direct, Recall-first direct-ablation, reranker-top20, and reranker-top50 candidates. The current Production router is the guardrail baseline; static and dynamic validation columns use the same fusion method and excluded fold. A no eligible candidate result must not be promoted.",
     "",
     "| Model | Fusion | Objective | Strategy | Fold | Promotion | Params | Proxy evals | Full evals | Proxy agreement | Static I/C/B/D/S | Dynamic base I/C/B/D/S | Influence Score/Geometry/TermCoverage/PairwiseAgreement/DenseConfidence/Identifier/Length | Static R@5 | Dynamic R@5 | Static R@10 | Dynamic R@10 | Static R@20 | Dynamic R@20 | Dynamic R@50 | Static Ctx@4k | Dynamic Ctx@4k |",
     "| --- | --- | --- | --- | --- | --- | ---: | ---: | ---: | ---: | --- | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |",
@@ -330,6 +332,22 @@ export const renderMarkdownReport = (artifact: BenchmarkArtifact): string => {
     )
   }
 
+  lines.push(
+    "",
+    "## Evidence Router NDCG Holdouts",
+    "",
+    "These aggregate excluded-fold rows compare each fusion and objective with the current Production router under identical validation partitions.",
+    "",
+    "| Model | Fusion | Objective | Strategy | Production NDCG@5 | Dynamic NDCG@5 | Production NDCG@10 | Dynamic NDCG@10 | Production NDCG@20 | Dynamic NDCG@20 | Production NDCG@50 | Dynamic NDCG@50 |",
+    "| --- | --- | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |",
+  )
+  for (const [key, rows] of routerGroups) {
+    const [model, fusion, objective, strategy] = key.split("\0")
+    lines.push(
+      `| ${model} | ${fusion} | ${objective} | ${strategy} | ${percent(weightedAverage(rows, (row) => row.productionValidation.ndcgAt5))} | ${percent(weightedAverage(rows, (row) => row.validation.ndcgAt5))} | ${percent(weightedAverage(rows, (row) => row.productionValidation.ndcgAt10))} | ${percent(weightedAverage(rows, (row) => row.validation.ndcgAt10))} | ${percent(weightedAverage(rows, (row) => row.productionValidation.ndcgAt20))} | ${percent(weightedAverage(rows, (row) => row.validation.ndcgAt20))} | ${percent(weightedAverage(rows, (row) => row.productionValidation.ndcgAt50))} | ${percent(weightedAverage(rows, (row) => row.validation.ndcgAt50))} |`,
+    )
+  }
+
   const holdoutRows = [
     ...artifact.fusionSearch.flatMap((result) =>
       result.holdoutBreakdown.map((holdout) => ({
@@ -368,6 +386,20 @@ export const renderMarkdownReport = (artifact: BenchmarkArtifact): string => {
     )
   }
 
+  lines.push(
+    "",
+    "### Holdout NDCG Breakdown",
+    "",
+    "| Model | Fusion | Objective | Strategy | Fold | Partition | Candidate NDCG@5 | Baseline NDCG@5 | Candidate NDCG@10 | Baseline NDCG@10 | Candidate NDCG@20 | Baseline NDCG@20 | Candidate NDCG@50 | Baseline NDCG@50 |",
+    "| --- | --- | --- | --- | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |",
+  )
+  for (const row of holdoutRows) {
+    const holdout: HoldoutQuality = row.holdout
+    lines.push(
+      `| ${row.model} | ${row.fusion} | ${row.objective} | ${row.strategy} | ${row.fold} | ${holdout.dimension}:${holdout.name} | ${percent(holdout.candidate.ndcgAt5)} | ${percent(holdout.baseline.ndcgAt5)} | ${percent(holdout.candidate.ndcgAt10)} | ${percent(holdout.baseline.ndcgAt10)} | ${percent(holdout.candidate.ndcgAt20)} | ${percent(holdout.baseline.ndcgAt20)} | ${percent(holdout.candidate.ndcgAt50)} | ${percent(holdout.baseline.ndcgAt50)} |`,
+    )
+  }
+
   lines.push(...renderPromotionEvidence(artifact))
 
   lines.push(
diff --git a/benchmarks/retrieval/evaluation/types.ts b/benchmarks/retrieval/evaluation/types.ts
index 513671e..d00c170 100644
--- a/benchmarks/retrieval/evaluation/types.ts
+++ b/benchmarks/retrieval/evaluation/types.ts
@@ -29,7 +29,7 @@ const BenchmarkQuestionSchema = Schema.Struct({
   queries: QueryFormsSchema,
   category: Schema.String,
   difficulty: Schema.Literals(["easy", "medium", "hard"]),
-  groundTruth: Schema.Array(GoldLocationSchema),
+  groundTruth: Schema.NonEmptyArray(GoldLocationSchema),
 })
 
 /** Versioned manifest for one pinned real-world benchmark repository. */
@@ -59,7 +59,12 @@ export type QueryKind = keyof typeof QueryFormsSchema.Type
 type FusionMethod = ProductionFusionMethod
 
 /** Product retrieval scenario used to choose a router candidate. */
-export const ROUTER_OBJECTIVES = ["direct", "reranker-top20", "reranker-top50"] as const
+export const ROUTER_OBJECTIVES = [
+  "direct",
+  "direct-recall-first",
+  "reranker-top20",
+  "reranker-top50",
+] as const
 export type RouterObjective = (typeof ROUTER_OBJECTIVES)[number]
 
 /** Versioned evidence-router search strategies recorded in benchmark artifacts. */
@@ -145,6 +150,10 @@ export interface QueryMeasurement {
   readonly recallAt10: number
   readonly recallAt20: number
   readonly recallAt50: number
+  readonly ndcgAt5: number
+  readonly ndcgAt10: number
+  readonly ndcgAt20: number
+  readonly ndcgAt50: number
   readonly successAt10: boolean
   readonly successAt20: boolean
   readonly reciprocalRank: number
@@ -168,6 +177,10 @@ export interface QualitySummary {
   readonly recallAt10: number
   readonly recallAt20: number
   readonly recallAt50: number
+  readonly ndcgAt5: number
+  readonly ndcgAt10: number
+  readonly ndcgAt20: number
+  readonly ndcgAt50: number
   readonly contextRecallAt4096: number
   readonly meanReciprocalRank: number
 }
@@ -424,7 +437,7 @@ export interface BenchmarkTimings {
 
 /** Reproducible machine-readable output of one complete benchmark run. */
 export interface BenchmarkArtifact {
-  readonly schemaVersion: 25
+  readonly schemaVersion: 26
   /** Profile controlling benchmark coverage without changing retrieval behavior. */
   readonly benchmarkProfile: BenchmarkProfile
   /** Versioned objective profile used for candidate selection and aggregate metrics. */
diff --git a/benchmarks/retrieval/evaluation/weight-search.ts b/benchmarks/retrieval/evaluation/weight-search.ts
index d6e0125..05595a8 100644
--- a/benchmarks/retrieval/evaluation/weight-search.ts
+++ b/benchmarks/retrieval/evaluation/weight-search.ts
@@ -28,7 +28,12 @@ import {
   type CandidateEvaluationQueue,
   type EvaluationCandidate,
 } from "../execution/candidate-evaluation-pool.js"
-import { contextRecallAtBudget, recallAt, reciprocalRank } from "./metrics.js"
+import {
+  contextRecallAtBudget,
+  normalizedDiscountedCumulativeGain,
+  recallAt,
+  reciprocalRank,
+} from "./metrics.js"
 import { SEARCH_PRIORITY_PROFILE, type OptimizationProfile } from "./optimization-profiles.js"
 import { prepareFusion, type PreparedFusionEvaluator } from "./prepared-fusion.js"
 import { buildGuardrailBlockers } from "./promotion-evidence.js"
@@ -177,6 +182,10 @@ const summarizeRanked = <T>(
       recallAt10: 0,
       recallAt20: 0,
       recallAt50: 0,
+      ndcgAt5: 0,
+      ndcgAt10: 0,
+      ndcgAt20: 0,
+      ndcgAt50: 0,
       contextRecallAt4096: 0,
       meanReciprocalRank: 0,
     }
@@ -185,6 +194,10 @@ const summarizeRanked = <T>(
   let recall10 = 0
   let recall20 = 0
   let recall50 = 0
+  let ndcg5 = 0
+  let ndcg10 = 0
+  let ndcg20 = 0
+  let ndcg50 = 0
   let contextRecall = 0
   let mrr = 0
   let totalWeight = 0
@@ -197,6 +210,10 @@ const summarizeRanked = <T>(
     recall10 += weight * recallAt(ranked, sampleTargets, 10)
     recall20 += weight * recallAt(ranked, sampleTargets, 20)
     recall50 += weight * recallAt(ranked, sampleTargets, 50)
+    ndcg5 += weight * normalizedDiscountedCumulativeGain(ranked, sampleTargets, 5)
+    ndcg10 += weight * normalizedDiscountedCumulativeGain(ranked, sampleTargets, 10)
+    ndcg20 += weight * normalizedDiscountedCumulativeGain(ranked, sampleTargets, 20)
+    ndcg50 += weight * normalizedDiscountedCumulativeGain(ranked, sampleTargets, 50)
     contextRecall += weight * contextRecallAtBudget(ranked, sampleTargets, chunks(sample), 4_096)
     mrr += weight * reciprocalRank(ranked, sampleTargets)
     totalWeight += weight
@@ -207,6 +224,10 @@ const summarizeRanked = <T>(
       recallAt10: 0,
       recallAt20: 0,
       recallAt50: 0,
+      ndcgAt5: 0,
+      ndcgAt10: 0,
+      ndcgAt20: 0,
+      ndcgAt50: 0,
       contextRecallAt4096: 0,
       meanReciprocalRank: 0,
     }
@@ -216,6 +237,10 @@ const summarizeRanked = <T>(
     recallAt10: recall10 / totalWeight,
     recallAt20: recall20 / totalWeight,
     recallAt50: recall50 / totalWeight,
+    ndcgAt5: ndcg5 / totalWeight,
+    ndcgAt10: ndcg10 / totalWeight,
+    ndcgAt20: ndcg20 / totalWeight,
+    ndcgAt50: ndcg50 / totalWeight,
     contextRecallAt4096: contextRecall / totalWeight,
     meanReciprocalRank: mrr / totalWeight,
   }
@@ -297,6 +322,7 @@ const evidenceRouterGuardrailsMet = (
   fusion: FusionMethod,
   baselines: GuardrailBaselines,
   profile: OptimizationProfile,
+  objective: RouterObjective,
 ): boolean => {
   const holdoutProfile = unweightedProfile(profile)
   const evaluate = (partition: GuardrailBaselinePartition): boolean => {
@@ -306,13 +332,14 @@ const evidenceRouterGuardrailsMet = (
       fusion,
       holdoutProfile,
     )
-    return isWithinGuardrails(candidate, partition.baseline, holdoutProfile)
+    return isWithinGuardrails(candidate, partition.baseline, holdoutProfile, objective)
   }
   return (
     isWithinGuardrails(
       summarizeEvidenceRouter(prepareEvidenceSamples(samples), config, fusion, profile),
       baselines.overall,
       profile,
+      objective,
     ) && baselines.partitions.every(evaluate)
   )
 }
@@ -366,6 +393,9 @@ const buildProxySamples = (
 
 const OBJECTIVE_PRIORITIES: Readonly<Record<RouterObjective, readonly QualityMetric[]>> = {
   direct: [
+    "ndcgAt5",
+    "ndcgAt10",
+    "ndcgAt20",
     "recallAt5",
     "recallAt10",
     "contextRecallAt4096",
@@ -373,6 +403,17 @@ const OBJECTIVE_PRIORITIES: Readonly<Record<RouterObjective, readonly QualityMet
     "recallAt50",
     "meanReciprocalRank",
   ],
+  "direct-recall-first": [
+    "recallAt5",
+    "recallAt10",
+    "contextRecallAt4096",
+    "recallAt20",
+    "recallAt50",
+    "meanReciprocalRank",
+    "ndcgAt5",
+    "ndcgAt10",
+    "ndcgAt20",
+  ],
   "reranker-top20": [
     "recallAt20",
     "recallAt10",
@@ -391,13 +432,31 @@ const OBJECTIVE_PRIORITIES: Readonly<Record<RouterObjective, readonly QualityMet
   ],
 }
 
+const OBJECTIVE_GUARDRAILS: Readonly<Record<RouterObjective, readonly QualityMetric[]>> = {
+  direct: ["recallAt20", "recallAt50", "contextRecallAt4096"],
+  "direct-recall-first": ["recallAt20", "recallAt50", "contextRecallAt4096"],
+  "reranker-top20": ["recallAt5", "recallAt10", "recallAt20", "recallAt50", "contextRecallAt4096"],
+  "reranker-top50": ["recallAt5", "recallAt10", "recallAt20", "recallAt50", "contextRecallAt4096"],
+}
+
+// The recall-first direct objective is an ablation over the same searched archive, not another
+// source of beam candidates. This preserves the pre-ablation budget of two candidates per scenario.
+const SEARCH_OBJECTIVES = ROUTER_OBJECTIVES.filter(
+  (objective): objective is Exclude<RouterObjective, "direct-recall-first"> =>
+    objective !== "direct-recall-first",
+)
+
 const isWithinGuardrails = (
   quality: QualitySummary,
   baseline: QualitySummary | undefined,
   profile: OptimizationProfile = SEARCH_PRIORITY_PROFILE,
+  objective?: RouterObjective,
 ): boolean =>
   baseline === undefined ||
-  profile.metricObjective.guardrailMetrics.every(
+  (objective === undefined
+    ? profile.metricObjective.guardrailMetrics
+    : OBJECTIVE_GUARDRAILS[objective]
+  ).every(
     (metric) => quality[metric] >= baseline[metric] - profile.metricObjective.guardrailTolerance,
   )
 
@@ -458,6 +517,7 @@ const buildHoldoutBreakdown = (
     evaluationProfile: OptimizationProfile,
   ) => QualitySummary,
   profile: OptimizationProfile,
+  objective?: RouterObjective,
 ): readonly HoldoutQuality[] => {
   const holdoutProfile = unweightedProfile(profile)
   const baselines = buildGuardrailBaselines(samples, profile)
@@ -473,7 +533,9 @@ const buildHoldoutBreakdown = (
       partition.name,
       candidateQuality,
       partition.baseline,
-      evaluationProfile.metricObjective.guardrailMetrics,
+      objective === undefined
+        ? evaluationProfile.metricObjective.guardrailMetrics
+        : OBJECTIVE_GUARDRAILS[objective],
       evaluationProfile.metricObjective.guardrailTolerance,
     )
     return {
@@ -488,15 +550,16 @@ const buildHoldoutBreakdown = (
   })
 }
 
-const compareObjectiveQuality = (
+/** Compare two quality summaries using one benchmark objective's deterministic priority. */
+export const compareObjectiveQuality = (
   left: QualitySummary,
   right: QualitySummary,
   objective: RouterObjective,
   baseline?: QualitySummary,
   profile: OptimizationProfile = SEARCH_PRIORITY_PROFILE,
 ): number => {
-  const leftGuardrails = isWithinGuardrails(left, baseline, profile)
-  const rightGuardrails = isWithinGuardrails(right, baseline, profile)
+  const leftGuardrails = isWithinGuardrails(left, baseline, profile, objective)
+  const rightGuardrails = isWithinGuardrails(right, baseline, profile, objective)
   if (leftGuardrails !== rightGuardrails) return leftGuardrails ? -1 : 1
   const priorities =
     profile.metricObjective.name === objective
@@ -1140,8 +1203,8 @@ const selectObjectiveCandidates = (
   profile: OptimizationProfile,
 ): readonly RouterCandidate[] => {
   const selected = new Map<string, RouterCandidate>()
-  const perObjective = Math.max(1, Math.floor(limit / ROUTER_OBJECTIVES.length))
-  for (const objective of ROUTER_OBJECTIVES) {
+  const perObjective = Math.max(1, Math.floor(limit / SEARCH_OBJECTIVES.length))
+  for (const objective of SEARCH_OBJECTIVES) {
     const ranked = [...candidates]
       .sort((left, right) => compareRouterCandidates(left, right, objective, baseline, profile))
       .slice(0, perObjective)
@@ -1337,6 +1400,32 @@ export const selectEligibleCandidate = <T>(
   }
 }
 
+/** Select one eligible archive candidate per objective using that objective's own comparator. */
+export const selectObjectiveArchiveCandidates = <T>(
+  candidates: readonly T[],
+  quality: (candidate: T) => QualitySummary,
+  isEligible: (candidate: T, objective: RouterObjective) => boolean,
+  baseline: QualitySummary,
+  profile: OptimizationProfile = SEARCH_PRIORITY_PROFILE,
+): ReadonlyArray<{
+  readonly objective: RouterObjective
+  readonly candidate: T | undefined
+  readonly promotionStatus: PromotionStatus
+}> =>
+  ROUTER_OBJECTIVES.map((objective) => {
+    const ranked = [...candidates].sort((left, right) =>
+      compareObjectiveQuality(quality(left), quality(right), objective, baseline, profile),
+    )
+    const selected = selectEligibleCandidate(ranked, (candidate) =>
+      isEligible(candidate, objective),
+    )
+    return {
+      objective,
+      candidate: selected.candidate ?? ranked[0],
+      promotionStatus: selected.promotionStatus,
+    }
+  })
+
 interface RouterSearchPreparation {
   readonly samples: readonly WeightSearchSample[]
   readonly evidenceSamples: readonly EvidenceSearchSample[]
@@ -1504,43 +1593,30 @@ const selectBestEvidenceRouter = async (
   const fallback = beam[0]
   if (fallback === undefined) throw new Error("Evidence router search produced no candidate")
   const candidates = [...searchContext.archive.values()]
-  const selections = useSuccessiveHalving
-    ? (() => {
-        const guardrailsMet = evidenceRouterGuardrailsMet(
-          samples,
-          fallback.config,
-          fusion,
-          guardrailBaselines,
-          profile,
-        )
-        return ROUTER_OBJECTIVES.map((objective) => ({
-          objective,
-          config: fallback.config,
-          quality: fallback.quality,
-          guardrailsMet,
-          promotionStatus: (guardrailsMet
-            ? "eligible"
-            : "no-eligible-candidate") as PromotionStatus,
-        }))
-      })()
-    : ROUTER_OBJECTIVES.map((objective) => {
-        const rankedCandidates = [...candidates].sort((left, right) =>
-          compareRouterCandidates(left, right, objective, productionQuality, profile),
-        )
-        const { candidate: eligibleCandidate, promotionStatus } = selectEligibleCandidate(
-          rankedCandidates,
-          (entry) =>
-            evidenceRouterGuardrailsMet(samples, entry.config, fusion, guardrailBaselines, profile),
-        )
-        const candidate = eligibleCandidate ?? rankedCandidates[0] ?? fallback
-        return {
-          objective,
-          config: candidate.config,
-          quality: candidate.quality,
-          guardrailsMet: eligibleCandidate !== undefined,
-          promotionStatus,
-        }
-      })
+  const selections = selectObjectiveArchiveCandidates(
+    candidates,
+    ({ quality }) => quality,
+    (entry, objective) =>
+      evidenceRouterGuardrailsMet(
+        samples,
+        entry.config,
+        fusion,
+        guardrailBaselines,
+        profile,
+        objective,
+      ),
+    productionQuality,
+    profile,
+  ).map(({ objective, candidate: selectedCandidate, promotionStatus }) => {
+    const candidate = selectedCandidate ?? fallback
+    return {
+      objective,
+      config: candidate.config,
+      quality: candidate.quality,
+      guardrailsMet: promotionStatus === "eligible",
+      promotionStatus,
+    }
+  })
   const randomCandidate = randomSearch?.candidate ?? fallback
   return {
     selections,
@@ -1918,6 +1994,7 @@ export const optimizeEvidenceRouter = async (
               evaluationProfile,
             ),
           profile,
+          selection.objective,
         )
         const guardrailsMet = holdoutBreakdown.every((holdout) => holdout.guardrailsMet)
         return {
diff --git a/benchmarks/retrieval/runner.ts b/benchmarks/retrieval/runner.ts
index 2b68326..c647108 100644
--- a/benchmarks/retrieval/runner.ts
+++ b/benchmarks/retrieval/runner.ts
@@ -210,7 +210,7 @@ export const runRetrievalBenchmark = (
     )
 
     const artifact: BenchmarkArtifact = {
-      schemaVersion: 25,
+      schemaVersion: 26,
       benchmarkProfile: profile,
       optimizationProfile,
       validationProtocol: {
diff --git a/benchmarks/tests/channels.test.ts b/benchmarks/tests/channels.test.ts
index 9173c44..9553043 100644
--- a/benchmarks/tests/channels.test.ts
+++ b/benchmarks/tests/channels.test.ts
@@ -12,6 +12,7 @@ import { buildIdentifierIndex } from "../../src/lib/retrieval/identifier-index.j
 import type { PreparedCorpus } from "../retrieval/corpus/prepare.js"
 import {
   contextRecallAtBudget,
+  normalizedDiscountedCumulativeGain,
   recallAt,
   reciprocalRank,
   resolveGoldTargets,
@@ -22,11 +23,13 @@ import {
   rankLexicalChannels,
   RETRIEVAL_VARIANTS,
 } from "../retrieval/evaluation/ranking.js"
-import { ROUTER_OBJECTIVES } from "../retrieval/evaluation/types.js"
+import { ROUTER_OBJECTIVES, type QualitySummary } from "../retrieval/evaluation/types.js"
 import {
+  compareObjectiveQuality,
   optimizeEvidenceRouter,
   optimizeWeights,
   selectEligibleCandidate,
+  selectObjectiveArchiveCandidates,
 } from "../retrieval/evaluation/weight-search.js"
 
 const texts = [
@@ -366,6 +369,22 @@ describe("retrieval benchmark fixture", () => {
     expect(recallAt(ranked, targets, 1)).toBe(1)
   })
 
+  it("measures binary NDCG without inflating duplicate or overlapping gold chunks", () => {
+    const ranked = [
+      { chunkIndex: 2, score: 4 },
+      { chunkIndex: 0, score: 3 },
+      { chunkIndex: 0, score: 2 },
+      { chunkIndex: 1, score: 1 },
+    ]
+    const targets = [new Set([0]), new Set([0, 1])]
+
+    expect(normalizedDiscountedCumulativeGain(ranked, targets, 4)).toBeCloseTo(
+      (1 / Math.log2(3) + 1 / Math.log2(5)) / (1 + 1 / Math.log2(3)),
+    )
+    expect(normalizedDiscountedCumulativeGain(ranked, [], 5)).toBe(0)
+    expect(normalizedDiscountedCumulativeGain(ranked, [new Set()], 5)).toBe(0)
+  })
+
   it("keeps equal-weight RRF separate from production routing", () => {
     const rankings = {
       identity: [
@@ -397,6 +416,10 @@ describe("retrieval benchmark fixture", () => {
       recallAt10: recallAt(ranked, sample.targets, 10),
       recallAt20: recallAt(ranked, sample.targets, 20),
       recallAt50: recallAt(ranked, sample.targets, 50),
+      ndcgAt5: normalizedDiscountedCumulativeGain(ranked, sample.targets, 5),
+      ndcgAt10: normalizedDiscountedCumulativeGain(ranked, sample.targets, 10),
+      ndcgAt20: normalizedDiscountedCumulativeGain(ranked, sample.targets, 20),
+      ndcgAt50: normalizedDiscountedCumulativeGain(ranked, sample.targets, 50),
       contextRecallAt4096: contextRecallAtBudget(ranked, sample.targets, sample.chunks, 4_096),
       meanReciprocalRank: reciprocalRank(ranked, sample.targets),
     })
@@ -542,6 +565,45 @@ describe("retrieval benchmark fixture", () => {
     expect(result.shapleyRecallAt20.identity).toBe(1)
   })
 
+  it("compares NDCG-first direct selection with the matched Recall-first ablation", () => {
+    const quality = (ndcgAt5: number, recallAt5: number): QualitySummary => ({
+      ndcgAt5,
+      ndcgAt10: ndcgAt5,
+      ndcgAt20: ndcgAt5,
+      ndcgAt50: ndcgAt5,
+      recallAt5,
+      recallAt10: recallAt5,
+      recallAt20: 1,
+      recallAt50: 1,
+      contextRecallAt4096: 1,
+      meanReciprocalRank: recallAt5,
+    })
+    const ndcgFirst = quality(0.9, 0.7)
+    const recallFirst = quality(0.8, 0.8)
+
+    expect(compareObjectiveQuality(ndcgFirst, recallFirst, "direct", recallFirst)).toBeLessThan(0)
+    expect(compareObjectiveQuality(ndcgFirst, recallFirst, "direct-recall-first")).toBeGreaterThan(
+      0,
+    )
+    expect(
+      compareObjectiveQuality(ndcgFirst, recallFirst, "reranker-top20", recallFirst),
+    ).toBeGreaterThan(0)
+    expect(compareObjectiveQuality(quality(0.9, 0.8), quality(0.1, 0.8), "reranker-top20")).toBe(0)
+    const selections = selectObjectiveArchiveCandidates(
+      [
+        { id: "ndcg", quality: ndcgFirst },
+        { id: "recall", quality: recallFirst },
+      ],
+      ({ quality }) => quality,
+      () => true,
+      recallFirst,
+    )
+    expect(selections.find(({ objective }) => objective === "direct")?.candidate?.id).toBe("ndcg")
+    expect(
+      selections.find(({ objective }) => objective === "direct-recall-first")?.candidate?.id,
+    ).toBe("recall")
+  })
+
   it("selects one evidence router across queries with different reliable channels", async () => {
     const samples = makeEvidenceRouterSamples()
 
diff --git a/benchmarks/tests/corpus.test.ts b/benchmarks/tests/corpus.test.ts
index c164b7e..ba77cce 100644
--- a/benchmarks/tests/corpus.test.ts
+++ b/benchmarks/tests/corpus.test.ts
@@ -1,10 +1,11 @@
 import { expect, it } from "@effect/vitest"
-import { Effect } from "effect"
+import { Effect, Schema } from "effect"
 
 import type { ChunkingOptions } from "../../src/domain/ports.js"
 import { prepareCorpus } from "../retrieval/corpus/prepare.js"
 import { loadCorpusManifests, prepareRepository } from "../retrieval/corpus/repository.js"
 import { resolveGoldTargets } from "../retrieval/evaluation/metrics.js"
+import { CorpusManifestSchema } from "../retrieval/evaluation/types.js"
 
 const validationChunkingOptions: ChunkingOptions = {
   maxTokens: Number.MAX_SAFE_INTEGER,
@@ -13,6 +14,36 @@ const validationChunkingOptions: ChunkingOptions = {
   onDiagnostic: () => Effect.void,
 }
 
+it("rejects benchmark questions without exact ground truth", () => {
+  expect(() =>
+    Schema.decodeUnknownSync(CorpusManifestSchema)({
+      schemaVersion: 2,
+      id: "fixture",
+      repository: "owner/repository",
+      revision: "abc123",
+      language: "TypeScript",
+      size: "small",
+      includeRoots: ["src"],
+      excludePaths: [],
+      extensions: [".ts"],
+      questions: [
+        {
+          id: "missing-ground-truth",
+          queries: {
+            identifier: "target",
+            searchPhrase: "target",
+            naturalQuestion: "Where is target?",
+            agentTask: "Find target",
+          },
+          category: "navigation",
+          difficulty: "easy",
+          groundTruth: [],
+        },
+      ],
+    }),
+  ).toThrow()
+})
+
 // This validation intentionally reads real pinned checkouts; memfs is used by other adapter tests.
 it.effect("resolves every authored gold symbol in each pinned corpus", () =>
   Effect.gen(function* () {
diff --git a/benchmarks/tests/optimization-profiles.test.ts b/benchmarks/tests/optimization-profiles.test.ts
index 34d5cb5..086360e 100644
--- a/benchmarks/tests/optimization-profiles.test.ts
+++ b/benchmarks/tests/optimization-profiles.test.ts
@@ -15,6 +15,12 @@ it("records the authored weighted search-priority objective without a profile sc
   })
   expect("schemaVersion" in SEARCH_PRIORITY_PROFILE).toBe(false)
   expect(SEARCH_PRIORITY_PROFILE.provenance).toBe("authored-seed")
+  expect(SEARCH_PRIORITY_PROFILE.metricObjective.priority[0]).toBe("ndcgAt5")
+  expect(SEARCH_PRIORITY_PROFILE.metricObjective.guardrailMetrics).toEqual([
+    "recallAt20",
+    "recallAt50",
+    "contextRecallAt4096",
+  ])
   expect(SEARCH_PRIORITY_PROFILE.fusionConfig.baseWeights).toEqual({
     identity: 3,
     camelcase: 1.5,
@@ -35,6 +41,12 @@ it("records the authored weighted search-priority objective without a profile sc
   expect(
     OPTIMIZATION_PROFILES["code-navigation"].fusionConfig.baseWeights.identity,
   ).toBeGreaterThan(OPTIMIZATION_PROFILES["natural-language"].fusionConfig.baseWeights.identity)
+  expect(OPTIMIZATION_PROFILES["code-navigation"].metricObjective.priority).not.toContain(
+    "ndcgAt20",
+  )
+  expect(OPTIMIZATION_PROFILES["natural-language"].metricObjective.priority).not.toContain(
+    "ndcgAt50",
+  )
 })
 
 it("rejects semantically empty benchmark profiles", () => {
diff --git a/benchmarks/tests/promotion-evidence.test.ts b/benchmarks/tests/promotion-evidence.test.ts
index d496739..1714492 100644
--- a/benchmarks/tests/promotion-evidence.test.ts
+++ b/benchmarks/tests/promotion-evidence.test.ts
@@ -8,13 +8,18 @@ import {
 import type { PromotionHoldoutRow } from "../retrieval/evaluation/promotion-evidence.js"
 import type { HoldoutQuality, QualitySummary } from "../retrieval/evaluation/types.js"
 
-const quality = (recallAt20: number): QualitySummary => ({
+const quality = (recallAt20: number, overrides: Partial<QualitySummary> = {}): QualitySummary => ({
+  ndcgAt5: recallAt20,
+  ndcgAt10: recallAt20,
+  ndcgAt20: recallAt20,
+  ndcgAt50: recallAt20,
   recallAt5: recallAt20,
   recallAt10: recallAt20,
   recallAt20,
   recallAt50: recallAt20,
   contextRecallAt4096: recallAt20,
   meanReciprocalRank: recallAt20,
+  ...overrides,
 })
 
 const holdout = (candidate: number, baseline: number): HoldoutQuality => ({
@@ -131,4 +136,31 @@ describe("promotion evidence", () => {
       restarts: 1,
     })
   })
+
+  it("uses each direct objective's primary metric for stability diagnostics", () => {
+    const row = (objective: "direct" | "direct-recall-first"): PromotionHoldoutRow => ({
+      model: "fixture",
+      fusion: "dbsf",
+      objective,
+      strategy: "grouped-5-fold",
+      fold: "1",
+      validation: quality(1, { ndcgAt5: 0.9, recallAt5: 0.1 }),
+      productionValidation: quality(1, { ndcgAt5: 0.8, recallAt5: 0.9 }),
+      holdoutBreakdown: [holdout(1, 1)],
+      config: PRODUCTION_COMPATIBILITY_CONFIG,
+    })
+    const evidence = derivePromotionEvidence(
+      [row("direct"), row("direct-recall-first")],
+      ["grouped-5-fold"],
+      { strategy: "grouped-5-fold", fold: "1" },
+    )
+
+    expect(evidence.find(({ objective }) => objective === "direct")?.stability).toMatchObject({
+      medianHoldoutDrop: -0.1,
+      worstCaseHoldoutDrop: 0,
+    })
+    expect(
+      evidence.find(({ objective }) => objective === "direct-recall-first")?.stability,
+    ).toMatchObject({ medianHoldoutDrop: 0.8, worstCaseHoldoutDrop: 0.8 })
+  })
 })
diff --git a/benchmarks/tests/report.test.ts b/benchmarks/tests/report.test.ts
new file mode 100644
index 0000000..6f89143
--- /dev/null
+++ b/benchmarks/tests/report.test.ts
@@ -0,0 +1,86 @@
+import { expect, it } from "@effect/vitest"
+
+import { SEARCH_PRIORITY_PROFILE } from "../retrieval/evaluation/optimization-profiles.js"
+import { renderMarkdownReport } from "../retrieval/evaluation/report.js"
+import { ROUTER_SEARCH_STRATEGIES, type BenchmarkArtifact } from "../retrieval/evaluation/types.js"
+
+const artifact = {
+  schemaVersion: 26,
+  benchmarkProfile: "smoke",
+  optimizationProfile: SEARCH_PRIORITY_PROFILE,
+  validationProtocol: {
+    selection: "development-only",
+    holdouts: ["grouped-5-fold"],
+    finalTest: { kind: "untouched-grouped-fold", strategy: "grouped-5-fold", fold: "5" },
+  },
+  generatedAt: "2026-08-06T00:00:00.000Z",
+  searchStrategy: ROUTER_SEARCH_STRATEGIES["proxy-promotion"],
+  timings: {
+    totalDurationMs: 0,
+    corpusPreparationDurationMs: 0,
+    embeddingDurationMs: 0,
+    retrievalDurationMs: 0,
+    weightSearchDurationMs: 0,
+    fusionSearchDurationMs: 0,
+    evidenceRouterSearchDurationMs: 0,
+    candidateQueueStartupDurationMs: 0,
+    candidateQueueShutdownDurationMs: 0,
+  },
+  chunkConfig: { chunkTokens: 512, overlapLines: 0 },
+  contextTokenEstimator: "utf8-bytes-divided-by-four",
+  contextBudgets: [2_048, 4_096],
+  models: ["fixture-model"],
+  repositories: [],
+  evaluationCases: [],
+  embeddingRuns: [],
+  sparseEmbeddingRuns: [],
+  measurements: [
+    {
+      repository: "fixture",
+      language: "TypeScript",
+      size: "small",
+      revision: "abc123",
+      model: "fixture-model",
+      variant: "rrf",
+      questionId: "fixture-1",
+      queryKind: "agentTask",
+      query: "find target",
+      category: "navigation",
+      difficulty: "easy",
+      groupedFold: 1,
+      recallAt5: 1,
+      recallAt10: 1,
+      recallAt20: 1,
+      recallAt50: 1,
+      ndcgAt5: 0.5,
+      ndcgAt10: 0.6,
+      ndcgAt20: 0.7,
+      ndcgAt50: 0.8,
+      successAt10: true,
+      successAt20: true,
+      reciprocalRank: 0.5,
+      goldRanks: [2],
+      contextRecall: { "2048": 1, "4096": 1 },
+      queryDurationMs: 0,
+    },
+  ],
+  weightSearch: [],
+  recommendedWeights: [],
+  productionRouterSearch: [],
+  fusionSearch: [],
+  recommendedFusionWeights: [],
+  evidenceRouterSearch: [],
+  recommendedEvidenceRouters: [],
+  promotionEvidence: [],
+} satisfies BenchmarkArtifact
+
+it("renders the complete NDCG and direct-objective contract", () => {
+  const report = renderMarkdownReport(artifact)
+
+  expect(report).toContain("NDCG@5 | NDCG@10 | NDCG@20 | NDCG@50")
+  expect(report).toContain("| 50.0% | 60.0% | 70.0% | 80.0% |")
+  expect(report).toContain("`direct` objective selects NDCG@5 first")
+  expect(report).toContain("Recall-first direct-ablation")
+  expect(report).toContain("reranker-top20")
+  expect(report).toContain("reranker-top50")
+})
diff --git a/benchmarks/tests/retrieval.test.ts b/benchmarks/tests/retrieval.test.ts
index 44e8433..841f618 100644
--- a/benchmarks/tests/retrieval.test.ts
+++ b/benchmarks/tests/retrieval.test.ts
@@ -20,7 +20,7 @@ const foldQuestions = (prefix: string) =>
     },
     category: ["architecture", "routing", "symbol-lookup"][Math.floor(index / 4)],
     difficulty: ["easy", "medium", "hard"][index % 3] as "easy" | "medium" | "hard",
-    groundTruth: [],
+    groundTruth: [{ file: "src/fixture.ts", symbol: `target${index + 1}` }] as const,
   }))
 
 type TestQuestion = ReturnType<typeof foldQuestions>[number]
@@ -103,7 +103,7 @@ const runProfile = (profile: BenchmarkProfile, groupedFolds: number, fusionMetho
     expect(artifact.evaluationCases.every(({ groundTruth }) => groundTruth.length > 0)).toBe(true)
     expect(artifact.models.length).toBeGreaterThan(0)
     expect(artifact.measurements.length).toBeGreaterThan(0)
-    expect(artifact.schemaVersion).toBe(25)
+    expect(artifact.schemaVersion).toBe(26)
     expect(artifact.searchStrategy).toEqual(
       ROUTER_SEARCH_STRATEGIES[resolveRouterSearchStrategy(process.env.PIX_BENCH_ROUTER_STRATEGY)],
     )
@@ -127,6 +127,9 @@ const runProfile = (profile: BenchmarkProfile, groupedFolds: number, fusionMetho
     expect(artifact.evidenceRouterSearch.length).toBe(
       artifact.models.length * routerFusionMethods * ROUTER_OBJECTIVES.length * holdoutsPerModel,
     )
+    expect(new Set(artifact.evidenceRouterSearch.map(({ objective }) => objective))).toEqual(
+      new Set(ROUTER_OBJECTIVES),
+    )
     expect(new Set(artifact.evidenceRouterSearch.map((row) => row.fusion)).size).toBe(
       routerFusionMethods,
     )
@@ -145,6 +148,19 @@ const runProfile = (profile: BenchmarkProfile, groupedFolds: number, fusionMetho
     expect(artifact.measurements.every((row) => row.recallAt20 >= row.recallAt10)).toBe(true)
     expect(artifact.measurements.every((row) => row.recallAt10 >= row.recallAt5)).toBe(true)
     expect(artifact.measurements.every((row) => row.recallAt50 >= row.recallAt20)).toBe(true)
+    expect(
+      artifact.measurements.every(
+        (row) =>
+          row.ndcgAt5 >= 0 &&
+          row.ndcgAt5 <= 1 &&
+          row.ndcgAt10 >= 0 &&
+          row.ndcgAt10 <= 1 &&
+          row.ndcgAt20 >= 0 &&
+          row.ndcgAt20 <= 1 &&
+          row.ndcgAt50 >= 0 &&
+          row.ndcgAt50 <= 1,
+      ),
+    ).toBe(true)
     expect(outputPath).toMatch(/benchmarks[\\/]results[\\/]retrieval-.*\.json$/)
   })
 
diff --git a/benchmarks/tests/worker-pool.test.ts b/benchmarks/tests/worker-pool.test.ts
index 9bfbd7a..552e6b0 100644
--- a/benchmarks/tests/worker-pool.test.ts
+++ b/benchmarks/tests/worker-pool.test.ts
@@ -5,6 +5,7 @@ import { describe, expect, it } from "@effect/vitest"
 import type { Chunk } from "../../src/domain/chunk.js"
 import { SEARCH_PRIORITY_PROFILE } from "../retrieval/evaluation/optimization-profiles.js"
 import { prepareFusion } from "../retrieval/evaluation/prepared-fusion.js"
+import { ROUTER_OBJECTIVES } from "../retrieval/evaluation/types.js"
 import {
   fitRecommendedEvidenceRouter,
   fitRecommendedFusionWeights,
@@ -342,7 +343,7 @@ describe("benchmark candidate evaluation pool", () => {
   it("retains the serial evidence-router fit path", async () => {
     const result = await fitRecommendedEvidenceRouter("fixture", "dbsf", [searchSample])
 
-    expect(result).toHaveLength(3)
+    expect(result).toHaveLength(ROUTER_OBJECTIVES.length)
     expect(result.every((candidate) => candidate.fitQuality.recallAt20 >= 0)).toBe(true)
   })