v1/mode: override knobs + staffing native runner + pass 2/3/4 harnesses

Setup for the corpus-tightening experiment sweep (J 2026-04-26 — "now
is the only cheap window before the corpus gets large and refactoring
costs go up").

Override params on /v1/mode/execute (additive — old callers unaffected):
  force_matrix_corpus      — Pass 2: try alternate corpora per call
  force_relevance_threshold — Pass 2: sweep filter strictness
  force_temperature         — Pass 3: variance test

New native mode `staffing_inference_lakehouse` (Pass 4):
  - Same composer architecture as codereview_lakehouse
  - Staffing framing: coordinator producing fillable|contingent|
    unfillable verdict + ranked candidate list with playbook citations
  - matrix_corpus = workers_500k_v8
  - Validates that modes-as-prompt-molders generalizes beyond code
  - Framing explicitly says "do NOT fabricate workers" — the staffing
    analog of the lakehouse mode's symbol-grounding requirement

Three sweep harnesses:
  scripts/mode_pass2_corpus_sweep.ts — 4 corpora × 4 thresholds × 5 files
  scripts/mode_pass3_variance.ts     — 3 files × 3 temps × 5 reps
  scripts/mode_pass4_staffing.ts     — 5 fill requests through staffing mode

Each appends per-call rows to data/_kb/mode_experiments.jsonl which
mode_compare.ts already aggregates with grounding column.

Pass 1 (10 files × 5 modes broad sweep) currently running via the
existing scripts/mode_experiment.ts — gateway restart deferred until
it completes so the new override knobs aren't enabled mid-experiment.

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

config/modes.toml

@@ -31,9 +31,13 @@ matrix_corpus = "chicago_permits_v1"
 
 [[task_class]]
 name = "staffing_inference"
-preferred_mode = "ladder"
+# Staffing-domain native enrichment runner — Pass 4 (2026-04-26).
-fallback_modes = ["consensus", "pipeline"]
+# Same composer architecture as codereview_lakehouse but with staffing
-default_model = "gpt-oss:120b"
+# framing + workers corpus. Validates that the modes-as-prompt-molders
+# pattern generalizes beyond code review.
+preferred_mode = "staffing_inference_lakehouse"
+fallback_modes = ["ladder", "consensus", "pipeline"]
+default_model = "openai/gpt-oss-120b:free"
 matrix_corpus = "workers_500k_v8"
 
 [[task_class]]

crates/gateway/src/v1/mode.rs

@@ -51,6 +51,7 @@ const VALID_MODES: &[&str] = &[
     "codereview_isolation",     // file + pathway only (no matrix)
     "codereview_matrix_only",   // file + matrix only (no pathway)
     "codereview_playbook_only", // pathway only, NO file content (lossy ceiling)
+    "staffing_inference_lakehouse", // staffing-domain composer (Pass 4)
 ];
 
 /// Whether a mode is handled natively in this gateway vs proxied to
@@ -63,6 +64,7 @@ fn is_native_mode(mode: &str) -> bool {
             | "codereview_isolation"
             | "codereview_matrix_only"
             | "codereview_playbook_only"
+            | "staffing_inference_lakehouse"
     )
 }
 
@@ -82,6 +84,7 @@ pub struct EnrichmentFlags {
 pub enum ReviewerFraming {
     Adversarial, // forensic, ranked findings + verdict (lakehouse default)
     Generic,     // "review this" — no codebase priors (null baseline)
+    Staffing,    // staffing-domain coordinator framing (Pass 4)
 }
 
 fn flags_for_mode(mode: &str) -> EnrichmentFlags {
@@ -114,6 +117,18 @@ fn flags_for_mode(mode: &str) -> EnrichmentFlags {
             use_relevance_filter: false,
             framing: ReviewerFraming::Adversarial,
         },
+        "staffing_inference_lakehouse" => EnrichmentFlags {
+            // Staffing reuses the same composer architecture but with
+            // domain-specific framing. file_content here = the request
+            // payload (e.g. "fill 2 welders in Toledo OH"), bug_fingerprints
+            // surface prior playbook patterns from this geo+role, matrix
+            // pulls candidate workers + city/state demand chunks.
+            include_file_content: true,
+            include_bug_fingerprints: true,
+            include_matrix_chunks: true,
+            use_relevance_filter: true,
+            framing: ReviewerFraming::Staffing,
+        },
         // Default (codereview_lakehouse): everything on.
         _ => EnrichmentFlags {
             include_file_content: true,
@@ -404,6 +419,20 @@ pub struct ExecuteRequest {
     /// runner uses its built-in forensic-review framing.
     #[serde(default)]
     pub user_question: Option<String>,
+    /// Override the matrix corpus the runner queries. Defaults to the
+    /// task_class's matrix_corpus from modes.toml. Use for the corpus-
+    /// tightening experiment (Pass 2 of the 2026-04-26 mode sweep).
+    #[serde(default)]
+    pub force_matrix_corpus: Option<String>,
+    /// Override the relevance filter threshold (default 0.3). Setting
+    /// to 0 keeps every chunk; raising rejects more aggressively. Used
+    /// to find the threshold sweet spot per task class.
+    #[serde(default)]
+    pub force_relevance_threshold: Option<f64>,
+    /// Override the LLM temperature (default 0.1). Used by Pass 3
+    /// variance testing to measure run-to-run stability.
+    #[serde(default)]
+    pub force_temperature: Option<f64>,
 }
 
 #[derive(Serialize, Debug, Default)]
@@ -442,10 +471,20 @@ proven. Do NOT hedge.";
 const FRAMING_GENERIC: &str = "You are a code reviewer. Read the file below and produce a \
 markdown review with findings.";
 
+const FRAMING_STAFFING: &str = "You are a senior staffing coordinator for a light-industrial \
+labor agency. You receive a fill request (role × count × city × deadline) and have access \
+to historical playbook patterns from prior fills in this geo, plus a corpus of candidate \
+workers + demand signals. Produce a markdown plan with: (1) one-line verdict (fillable | \
+contingent | unfillable), (2) ranked candidate list with name, city, role, distance, prior \
+fill citations from the playbook, (3) risks (double-booking, eligibility gaps, geo stretch) \
+with severity + confidence percent, (4) playbook reference IDs you used. Be precise — only \
+recommend candidates whose names appear in the matrix data; do NOT fabricate workers.";
+
 fn framing_text(f: ReviewerFraming) -> &'static str {
     match f {
         ReviewerFraming::Adversarial => FRAMING_ADVERSARIAL,
         ReviewerFraming::Generic => FRAMING_GENERIC,
+        ReviewerFraming::Staffing => FRAMING_STAFFING,
     }
 }
 
@@ -485,6 +524,9 @@ pub async fn execute(
         ));
     }
 
+    // Caller can override the matrix corpus per-call (Pass 2 corpus
+    // tightening). Falls back to modes.toml default.
+    let matrix_corpus = req.force_matrix_corpus.clone().or(matrix_corpus);
     let flags = flags_for_mode(&mode);
     let mut sources = EnrichmentSources {
         matrix_corpus: matrix_corpus.clone(),
@@ -621,7 +663,7 @@ pub async fn execute(
         let body = serde_json::json!({
             "focus_file": { "path": req.file_path, "content": file_content },
             "chunks": chunks_for_filter,
-            "threshold": 0.3,
+            "threshold": req.force_relevance_threshold.unwrap_or(0.3),
         });
         match client
             .post("http://localhost:3800/relevance")
@@ -723,7 +765,7 @@ pub async fn execute(
             { "role": "system", "content": framing_text(flags.framing) },
             { "role": "user", "content": user_prompt },
         ],
-        "temperature": 0.1,
+        "temperature": req.force_temperature.unwrap_or(0.1),
         "max_tokens": 4096,
     });
     let chat_client = match reqwest::Client::builder()

scripts/mode_pass2_corpus_sweep.ts

@@ -0,0 +1,121 @@
+#!/usr/bin/env bun
+/**
+ * Pass 2: matrix corpus + relevance threshold sweep.
+ *
+ * For each (corpus, threshold) combination, run codereview_matrix_only
+ * on the same N files. Compares which corpus actually adds grounded
+ * findings vs codereview_isolation (matrix-off baseline).
+ *
+ * Output: data/_kb/mode_experiments.jsonl gets one row per call,
+ * tagged via the force_matrix_corpus + force_relevance_threshold
+ * fields visible in `sources`. Aggregator can then group by corpus.
+ *
+ * Usage: bun run scripts/mode_pass2_corpus_sweep.ts
+ */
+
+const GATEWAY = process.env.LH_GATEWAY ?? "http://localhost:3100";
+const MODEL = process.env.LH_MODEL ?? "openai/gpt-oss-120b:free";
+
+const FILES = (process.env.LH_FILES ?? [
+  "crates/queryd/src/delta.rs",
+  "crates/queryd/src/service.rs",
+  "crates/vectord/src/pathway_memory.rs",
+  "crates/gateway/src/v1/mode.rs",
+  "crates/aibridge/src/client.rs",
+].join(",")).split(",");
+
+const CORPORA = (process.env.LH_CORPORA ?? [
+  "distilled_procedural_v20260423102847",
+  "distilled_factual_v20260423095819",
+  "distilled_config_hint_v20260423102847",
+  "kb_team_runs_v1",
+].join(",")).split(",");
+
+const THRESHOLDS = (process.env.LH_THRESHOLDS ?? "0.2,0.3,0.4,0.5").split(",").map(Number);
+
+interface Result {
+  corpus: string;
+  threshold: number;
+  file: string;
+  ok: boolean;
+  matrix_kept?: number;
+  matrix_dropped?: number;
+  response_chars?: number;
+  latency_ms?: number;
+  error?: string;
+}
+
+async function runOne(corpus: string, threshold: number, file: string): Promise<Result> {
+  try {
+    const r = await fetch(`${GATEWAY}/v1/mode/execute`, {
+      method: "POST",
+      headers: { "content-type": "application/json" },
+      body: JSON.stringify({
+        task_class: "scrum_review",
+        file_path: file,
+        force_mode: "codereview_matrix_only",
+        force_model: MODEL,
+        force_matrix_corpus: corpus,
+        force_relevance_threshold: threshold,
+      }),
+      signal: AbortSignal.timeout(180_000),
+    });
+    if (!r.ok) {
+      const body = await r.text().catch(() => "");
+      return { corpus, threshold, file, ok: false, error: `HTTP ${r.status}: ${body.slice(0, 150)}` };
+    }
+    const j: any = await r.json();
+    return {
+      corpus, threshold, file, ok: true,
+      matrix_kept: j.sources?.matrix_chunks_kept,
+      matrix_dropped: j.sources?.matrix_chunks_dropped,
+      response_chars: (j.response ?? "").length,
+      latency_ms: j.latency_ms,
+    };
+  } catch (e: any) {
+    return { corpus, threshold, file, ok: false, error: e.message };
+  }
+}
+
+async function main() {
+  const total = CORPORA.length * THRESHOLDS.length * FILES.length;
+  console.log(`[pass2] corpora=${CORPORA.length} × thresholds=${THRESHOLDS.length} × files=${FILES.length} = ${total} runs`);
+  console.log(`[pass2] model=${MODEL}\n`);
+  let i = 0;
+  const results: Result[] = [];
+  for (const corpus of CORPORA) {
+    for (const threshold of THRESHOLDS) {
+      for (const file of FILES) {
+        i++;
+        process.stdout.write(`  [${i}/${total}] corpus=${corpus.slice(0, 30).padEnd(30)} thr=${threshold.toFixed(1)} ${file.slice(-32).padStart(32)} ... `);
+        const r = await runOne(corpus, threshold, file);
+        results.push(r);
+        if (r.ok) {
+          const total_chunks = (r.matrix_kept ?? 0) + (r.matrix_dropped ?? 0);
+          console.log(`✓ k=${r.matrix_kept}/${total_chunks} resp=${r.response_chars} ${((r.latency_ms ?? 0) / 1000).toFixed(1)}s`);
+        } else {
+          console.log(`✗ ${r.error}`);
+        }
+      }
+    }
+  }
+
+  console.log(`\n[pass2] complete · ${results.filter(r => r.ok).length}/${results.length} succeeded`);
+
+  // Per-corpus×threshold roll-up of kept-rate (the matrix usefulness proxy).
+  console.log(`\n[pass2] kept-rate by corpus × threshold (avg chunks kept per call):`);
+  console.log(`  ${"corpus".padEnd(40)} ${THRESHOLDS.map(t => `thr=${t.toFixed(1)}`).join(" ").padStart(35)}`);
+  for (const corpus of CORPORA) {
+    const cells = THRESHOLDS.map(t => {
+      const matched = results.filter(r => r.ok && r.corpus === corpus && r.threshold === t);
+      if (matched.length === 0) return "  —  ";
+      const avgKept = matched.reduce((s, r) => s + (r.matrix_kept ?? 0), 0) / matched.length;
+      return avgKept.toFixed(1).padStart(5);
+    }).join(" ");
+    console.log(`  ${corpus.slice(0, 40).padEnd(40)} ${cells}`);
+  }
+
+  console.log(`\n[pass2] aggregate findings/groundedness with: bun run scripts/mode_compare.ts`);
+}
+
+main().catch(e => { console.error(e); process.exit(1); });

scripts/mode_pass3_variance.ts

@@ -0,0 +1,109 @@
+#!/usr/bin/env bun
+/**
+ * Pass 3: variance test.
+ *
+ * Runs codereview_lakehouse on the SAME file N times at each of M
+ * temperatures. Measures run-to-run stability of grounded finding
+ * count, response size, and latency. Anything <100% groundedness
+ * is a leak; track which symbols got hallucinated.
+ *
+ * Output appends to data/_kb/mode_experiments.jsonl. The aggregator
+ * can group by ts and identify variance buckets.
+ *
+ * Usage: bun run scripts/mode_pass3_variance.ts
+ */
+
+const GATEWAY = process.env.LH_GATEWAY ?? "http://localhost:3100";
+const MODEL = process.env.LH_MODEL ?? "openai/gpt-oss-120b:free";
+
+const FILES = (process.env.LH_FILES ?? [
+  "crates/queryd/src/delta.rs",
+  "crates/vectord/src/pathway_memory.rs",
+  "crates/gateway/src/v1/mode.rs",
+].join(",")).split(",");
+
+const TEMPS = (process.env.LH_TEMPS ?? "0.0,0.1,0.3").split(",").map(Number);
+const REPS = Number(process.env.LH_REPS ?? 5);
+
+interface Result {
+  file: string;
+  temp: number;
+  rep: number;
+  ok: boolean;
+  response_chars?: number;
+  latency_ms?: number;
+  error?: string;
+}
+
+async function runOne(file: string, temp: number, rep: number): Promise<Result> {
+  try {
+    const r = await fetch(`${GATEWAY}/v1/mode/execute`, {
+      method: "POST",
+      headers: { "content-type": "application/json" },
+      body: JSON.stringify({
+        task_class: "scrum_review",
+        file_path: file,
+        force_mode: "codereview_lakehouse",
+        force_model: MODEL,
+        force_temperature: temp,
+      }),
+      signal: AbortSignal.timeout(180_000),
+    });
+    if (!r.ok) {
+      const body = await r.text().catch(() => "");
+      return { file, temp, rep, ok: false, error: `HTTP ${r.status}: ${body.slice(0, 150)}` };
+    }
+    const j: any = await r.json();
+    return {
+      file, temp, rep, ok: true,
+      response_chars: (j.response ?? "").length,
+      latency_ms: j.latency_ms,
+    };
+  } catch (e: any) {
+    return { file, temp, rep, ok: false, error: e.message };
+  }
+}
+
+async function main() {
+  const total = FILES.length * TEMPS.length * REPS;
+  console.log(`[pass3] files=${FILES.length} × temps=${TEMPS.length} × reps=${REPS} = ${total} runs`);
+  console.log(`[pass3] model=${MODEL}\n`);
+  let i = 0;
+  const results: Result[] = [];
+  for (const file of FILES) {
+    for (const temp of TEMPS) {
+      for (let rep = 1; rep <= REPS; rep++) {
+        i++;
+        process.stdout.write(`  [${i}/${total}] temp=${temp.toFixed(1)} rep=${rep}/${REPS} ${file.slice(-32).padStart(32)} ... `);
+        const r = await runOne(file, temp, rep);
+        results.push(r);
+        if (r.ok) {
+          console.log(`✓ resp=${r.response_chars} ${((r.latency_ms ?? 0) / 1000).toFixed(1)}s`);
+        } else {
+          console.log(`✗ ${r.error}`);
+        }
+      }
+    }
+  }
+
+  console.log(`\n[pass3] complete · ${results.filter(r => r.ok).length}/${results.length} succeeded`);
+
+  // Per-file × temp variance summary (response_chars stddev as a quick
+  // proxy for output instability).
+  console.log(`\n[pass3] response_chars variance (mean ± stddev) by file × temp:`);
+  console.log(`  ${"file".padEnd(40)} ${TEMPS.map(t => `temp=${t.toFixed(1)}`.padStart(20)).join(" ")}`);
+  for (const file of FILES) {
+    const cells = TEMPS.map(t => {
+      const xs = results.filter(r => r.ok && r.file === file && r.temp === t).map(r => r.response_chars ?? 0);
+      if (xs.length === 0) return "          —          ";
+      const mean = xs.reduce((s, x) => s + x, 0) / xs.length;
+      const sd = Math.sqrt(xs.reduce((s, x) => s + Math.pow(x - mean, 2), 0) / xs.length);
+      return `${Math.round(mean).toString().padStart(7)} ± ${Math.round(sd).toString().padEnd(6)}`.padStart(20);
+    }).join(" ");
+    console.log(`  ${file.slice(0, 40).padEnd(40)} ${cells}`);
+  }
+
+  console.log(`\n[pass3] grounding variance via: bun run scripts/mode_compare.ts (look for grounded-N column drift)`);
+}
+
+main().catch(e => { console.error(e); process.exit(1); });

scripts/mode_pass4_staffing.ts

@@ -0,0 +1,127 @@
+#!/usr/bin/env bun
+/**
+ * Pass 4: staffing_inference_lakehouse cross-domain validation.
+ *
+ * Runs the staffing-domain mode against synthetic fill requests.
+ * Validates that the modes-as-prompt-molders architecture generalizes
+ * beyond code review — the composer pattern (file_content + bug
+ * fingerprints + relevance-filtered matrix + domain framing) should
+ * produce grounded staffing recommendations the same way it produces
+ * grounded code reviews.
+ *
+ * Each fill request is posted as `file_content` (since the runner's
+ * shape expects file content; for staffing it's the request payload).
+ * file_path is set to a synthetic path under requests/ so pathway
+ * memory bucketing groups requests by geo+role.
+ *
+ * Usage: bun run scripts/mode_pass4_staffing.ts
+ */
+
+const GATEWAY = process.env.LH_GATEWAY ?? "http://localhost:3100";
+const MODEL = process.env.LH_MODEL ?? "openai/gpt-oss-120b:free";
+
+interface FillRequest {
+  city: string;
+  state: string;
+  role: string;
+  count: number;
+  deadline: string;
+  notes?: string;
+}
+
+const REQUESTS: FillRequest[] = [
+  { city: "Toledo", state: "OH", role: "Welder", count: 2, deadline: "2026-04-29", notes: "OSHA 10 required" },
+  { city: "Nashville", state: "TN", role: "Forklift Operator", count: 3, deadline: "2026-05-01" },
+  { city: "Chicago", state: "IL", role: "Assembler", count: 5, deadline: "2026-04-30", notes: "second shift" },
+  { city: "South Bend", state: "IN", role: "Electrician", count: 1, deadline: "2026-04-28", notes: "journeyman license" },
+  { city: "Murfreesboro", state: "TN", role: "Packaging Operator", count: 4, deadline: "2026-05-02" },
+];
+
+function requestToPayload(req: FillRequest): string {
+  return [
+    `# Fill Request`,
+    `Role: ${req.role} × ${req.count}`,
+    `Location: ${req.city}, ${req.state}`,
+    `Deadline: ${req.deadline}`,
+    req.notes ? `Notes: ${req.notes}` : "",
+    "",
+    "Recommend candidates from the matrix data. Cite playbook references.",
+  ].filter(Boolean).join("\n");
+}
+
+interface Result {
+  req: FillRequest;
+  ok: boolean;
+  response_chars?: number;
+  bug_fingerprints?: number;
+  matrix_kept?: number;
+  matrix_dropped?: number;
+  latency_ms?: number;
+  error?: string;
+  preview?: string;
+}
+
+async function runOne(req: FillRequest): Promise<Result> {
+  const payload = requestToPayload(req);
+  const file_path = `requests/${req.role.toLowerCase().replace(/\s+/g, "_")}_${req.city.toLowerCase().replace(/\s+/g, "_")}_${req.state}.md`;
+  try {
+    const r = await fetch(`${GATEWAY}/v1/mode/execute`, {
+      method: "POST",
+      headers: { "content-type": "application/json" },
+      body: JSON.stringify({
+        task_class: "staffing_inference",
+        file_path,
+        file_content: payload,
+        force_mode: "staffing_inference_lakehouse",
+        force_model: MODEL,
+      }),
+      signal: AbortSignal.timeout(180_000),
+    });
+    if (!r.ok) {
+      const body = await r.text().catch(() => "");
+      return { req, ok: false, error: `HTTP ${r.status}: ${body.slice(0, 200)}` };
+    }
+    const j: any = await r.json();
+    return {
+      req, ok: true,
+      response_chars: (j.response ?? "").length,
+      bug_fingerprints: j.sources?.bug_fingerprints_count,
+      matrix_kept: j.sources?.matrix_chunks_kept,
+      matrix_dropped: j.sources?.matrix_chunks_dropped,
+      latency_ms: j.latency_ms,
+      preview: (j.response ?? "").slice(0, 400),
+    };
+  } catch (e: any) {
+    return { req, ok: false, error: e.message };
+  }
+}
+
+async function main() {
+  console.log(`[pass4] requests=${REQUESTS.length} model=${MODEL} mode=staffing_inference_lakehouse\n`);
+  let i = 0;
+  const results: Result[] = [];
+  for (const req of REQUESTS) {
+    i++;
+    process.stdout.write(`  [${i}/${REQUESTS.length}] ${req.role.padEnd(22)} × ${req.count} in ${req.city}, ${req.state} ... `);
+    const r = await runOne(req);
+    results.push(r);
+    if (r.ok) {
+      console.log(`✓ resp=${r.response_chars} bug=${r.bug_fingerprints ?? 0} mtx=${r.matrix_kept ?? 0}/${(r.matrix_kept ?? 0) + (r.matrix_dropped ?? 0)} ${((r.latency_ms ?? 0) / 1000).toFixed(1)}s`);
+    } else {
+      console.log(`✗ ${r.error}`);
+    }
+  }
+
+  console.log(`\n[pass4] complete · ${results.filter(r => r.ok).length}/${results.length} succeeded\n`);
+
+  // Show first successful response head to verify the framing actually
+  // produced staffing-style output (verdict + ranked candidates) not
+  // generic prose.
+  const first = results.find(r => r.ok && r.preview);
+  if (first) {
+    console.log(`[pass4] first successful response preview (${first.req.city} ${first.req.role}):`);
+    console.log(first.preview!.split("\n").map(l => "  | " + l).join("\n"));
+  }
+}
+
+main().catch(e => { console.error(e); process.exit(1); });