From 357996cc89f1a53ccf90731b0bf1d57d20c7d602 Mon Sep 17 00:00:00 2001
From: Ysopwnable <161784452+N0zoM1z0@users.noreply.github.com>
Date: Thu, 15 Oct 2026 02:45:44 +0800
Subject: [PATCH] test(perf): measure genuine SAZ selection and retained
 recovery (#3584)

* test(perf): pin genuine Fiddler captures and raw HTTP oracles

Pin two public Microsoft SAZ captures with 13 and 1311 sessions, immutable source/digest/member coverage and predeclared caller-ordered transaction questions. Independently preserve stored HTTP headers, binary/empty bodies and archive coordinates; admit only known corpus bytes before ZIP allocation. Five reader regressions and both genuine selected-record checks pass. No production or performance-gain claim.

Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Co-authored-by: Kaoru0822 <208281322+kaoru0822-kitauji@users.noreply.github.com>

* test(perf): measure SAZ selection and retained recovery workflows

Add authenticated complete/ordinal/retained-page SDK experiments, fixed real-Codex prompts and independent cited-byte grading. Preserve actual failed RPCs and exact retained-reference recovery when complete MCP delivery exceeds its receive buffer. Verify selected CLI/MCP facts, parameters and semantic Evidence identity. Keep a genuine partial-message admission-negative probe. Focused 12 tests and 10 actual-Evidence grading cases pass; no formal model or gain claim yet.

Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Co-authored-by: Kaoru0822 <208281322+kaoru0822-kitauji@users.noreply.github.com>
Co-authored-by: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>

* test(perf): verify resolved SAZ Evidence selection parameters

Distinguish omitted request ordinals from the complete resolved ordinal scope recorded by WebNetworkCaptureService. Share the expected sanitized Evidence representation between SDK checks and model grading; preserve original caller arguments. A focused regression and both saved genuine baseline RPCs confirm the correction. Keep original collector failures and costs unchanged; freeze a new driver before repeated measurements.

Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Co-authored-by: Kaoru0822 <208281322+kaoru0822-kitauji@users.noreply.github.com>
Co-authored-by: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>

* docs(perf): retain mixed SAZ whole-task costs and failure boundaries

Record both SDK batches, twelve corrected workflows, all four actual Codex attempts, independent byte equivalence and verified payload-phase CPU/memory. Distinguish successful-analysis MCP transport limits from real provider timeouts and negative partial-message admission. Three model questions complete; the only completed pair saves 1.6 percent input but takes 2 percent longer, with no consistent benefit. Keep production APIs/defaults/cache unchanged and preserve all deviations, failed attempts and unknowns.

Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Co-authored-by: Kaoru0822 <208281322+kaoru0822-kitauji@users.noreply.github.com>
Co-authored-by: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>

---------

Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Co-authored-by: Kaoru0822 <208281322+kaoru0822-kitauji@users.noreply.github.com>
Co-authored-by: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>
---
 experiments/analysis-efficiency/README.md     |   6 +
 .../analysis-efficiency/agent-trial.py        |   1 +
 experiments/analysis-efficiency/corpus.json   |  60 ++++
 experiments/analysis-efficiency/measure.mjs   |  22 ++
 .../analysis-efficiency/review-saz.mjs        | 266 ++++++++++++++
 experiments/analysis-efficiency/saz-entry.mjs | 326 ++++++++++++++++++
 .../analysis-efficiency/saz-model-prompt.mjs  |  30 ++
 experiments/analysis-efficiency/saz-oracle.py | 121 +++++++
 .../saz-selection-workflows.md                | 214 ++++++++++++
 .../analysis-efficiency/test_saz_oracle.py    |  48 +++
 tests/evaluation/sazWorkflowSemantics.test.ts |  49 +++
 11 files changed, 1143 insertions(+)
 create mode 100644 experiments/analysis-efficiency/review-saz.mjs
 create mode 100644 experiments/analysis-efficiency/saz-entry.mjs
 create mode 100644 experiments/analysis-efficiency/saz-model-prompt.mjs
 create mode 100644 experiments/analysis-efficiency/saz-oracle.py
 create mode 100644 experiments/analysis-efficiency/saz-selection-workflows.md
 create mode 100644 experiments/analysis-efficiency/test_saz_oracle.py
 create mode 100644 tests/evaluation/sazWorkflowSemantics.test.ts

diff --git a/experiments/analysis-efficiency/README.md b/experiments/analysis-efficiency/README.md
index 2ca9da3a4e..295a2aa765 100644
--- a/experiments/analysis-efficiency/README.md
+++ b/experiments/analysis-efficiency/README.md
@@ -1,5 +1,11 @@
 # Analysis efficiency measurements
 
+For genuine Fiddler captures, see the [SAZ selection checkpoint](saz-selection-workflows.md).
+`measure.mjs --saz-selection baseline|focused|retained` compares default complete
+delivery (including retained-reference recovery), caller-selected ordinals, and
+summary/preview/complete/repeated pages. Each failed RPC remains in task costs.
+The body question uses stored bytes, not decoded HTTP application payloads.
+
 Tracks the experiments for [#2016](https://github.com/morluto/rea/issues/2016).
 The production CLI/MCP contracts and defaults remain the compatibility baseline.
 Downloaded artifacts, runtime copies, exported Evidence, transcripts and detailed
diff --git a/experiments/analysis-efficiency/agent-trial.py b/experiments/analysis-efficiency/agent-trial.py
index 365337a06a..810fc55902 100644
--- a/experiments/analysis-efficiency/agent-trial.py
+++ b/experiments/analysis-efficiency/agent-trial.py
@@ -181,6 +181,7 @@ def main():
                  'inspect_android_class', 'inspect_android_method',
                  'inspect_firmware_regions', 'inspect_managed_artifact',
                  'inspect_sqlite_database', 'inspect_pcap_capture',
+                 'inspect_web_network_capture',
                  'inspect_packet_capture', 'inspect_packet_fields',
                  'decode_signal_recording', 'inspect_signal_decoders',
                  'inspect_managed_members', 'inspect_managed_native_boundaries',
diff --git a/experiments/analysis-efficiency/corpus.json b/experiments/analysis-efficiency/corpus.json
index 983b817992..03d4fa18c5 100644
--- a/experiments/analysis-efficiency/corpus.json
+++ b/experiments/analysis-efficiency/corpus.json
@@ -373,6 +373,66 @@
         "reference": "https://raw.githubusercontent.com/sigrokproject/sigrok-dumps/0ad13477abc959d37fc9a5acbd23901c371c9c76/i2c/wii_nunchuk/README",
         "license_scope": "Repository README declares dumps public domain unless otherwise noted; selected READMEs state no exception; decoder software licensing is separate."
       }
+    },
+    {
+      "id": "xbox-clear_nomp.saz",
+      "file": "clear_nomp.saz",
+      "bytes": 43515,
+      "sha256": "3e808663a99cbb9160186f7419633248200fbc6863d2c897f5d4d9972dde0375",
+      "provenance": {
+        "project": "Microsoft Xbox Live Trace Analyzer",
+        "version": "ad880ff00b26c4cae5e3c8864d3df438a03890ba",
+        "url": "https://raw.githubusercontent.com/microsoft/xbox-live-trace-analyzer/ad880ff00b26c4cae5e3c8864d3df438a03890ba/Source/TestData/CertWarnings/clear_nomp.saz",
+        "git_blob": "2f1ba66dacb8f8bac290cab59dff8fe750a46ce7",
+        "license": "MIT; upstream LICENSE.md",
+        "producer_archive_comment": "Fiddler (v4.5.1.2) Session Archive. See http://fiddler2.com"
+      },
+      "saz_question": {
+        "record_ordinals": [12, 0, 6],
+        "total_records": 13,
+        "members": 42,
+        "expanded_bytes": 62609
+      }
+    },
+    {
+      "id": "xbox-clear_mp_cs10.saz",
+      "file": "clear_mp_cs10.saz",
+      "bytes": 5223519,
+      "sha256": "4021b201848e4feea3402c380fda5d2df43fd0bf1c72ea550e60f2da3b3d5b44",
+      "provenance": {
+        "project": "Microsoft Xbox Live Trace Analyzer",
+        "version": "ad880ff00b26c4cae5e3c8864d3df438a03890ba",
+        "url": "https://raw.githubusercontent.com/microsoft/xbox-live-trace-analyzer/ad880ff00b26c4cae5e3c8864d3df438a03890ba/Source/TestData/CertWarnings/clear_mp_cs10.saz",
+        "git_blob": "dac534d625594c7a1d57b0b12ed12ea9c6cb6a5b",
+        "license": "MIT; upstream LICENSE.md",
+        "producer_archive_comment": "Fiddler (v4.6.2.0) Session Archive. See https://fiddler2.com"
+      },
+      "saz_question": {
+        "record_ordinals": [1310, 0, 655],
+        "total_records": 1311,
+        "members": 3939,
+        "expanded_bytes": 9223820
+      }
+    },
+    {
+      "id": "xbox-x49_mp_cs21.saz",
+      "file": "x49_mp_cs21.saz",
+      "bytes": 572224,
+      "sha256": "97cdca8b0608e3f11995fd64b64390b9fd8d8e3efcac89f517f24b44968047f6",
+      "provenance": {
+        "project": "Microsoft Xbox Live Trace Analyzer",
+        "version": "ad880ff00b26c4cae5e3c8864d3df438a03890ba",
+        "url": "https://raw.githubusercontent.com/microsoft/xbox-live-trace-analyzer/ad880ff00b26c4cae5e3c8864d3df438a03890ba/Source/TestData/CertWarnings/x49_mp_cs21.saz",
+        "git_blob": "e7794c44892ca4fa1729fa012e698e8284337f15",
+        "license": "MIT; upstream LICENSE.md",
+        "producer_archive_comment": "Fiddler (v4.6.2.26757) Session Archive. See https://fiddler2.com"
+      },
+      "saz_question": {
+        "record_ordinals": [102, 0, 51],
+        "total_records": 103,
+        "members": 315,
+        "expanded_bytes": 1301927
+      }
     }
   ]
 }
diff --git a/experiments/analysis-efficiency/measure.mjs b/experiments/analysis-efficiency/measure.mjs
index 20f1e7d610..88f5dd1ff7 100644
--- a/experiments/analysis-efficiency/measure.mjs
+++ b/experiments/analysis-efficiency/measure.mjs
@@ -14,6 +14,7 @@ import { StdioClientTransport } from "@modelcontextprotocol/client/stdio";
 import { Ajv2020 } from "ajv/dist/2020.js";
 import { PACKET_ARTIFACTS, measurePacketEntry } from "./packet-entry.mjs";
 import { SIGNAL_ARTIFACTS, measureSignalEntry } from "./signal-entry.mjs";
+import { SAZ_ARTIFACTS, measureSazEntry } from "./saz-entry.mjs";
 
 const root = resolve(dirname(fileURLToPath(import.meta.url)), "../..");
 const { values } = parseArgs({
@@ -33,6 +34,7 @@ const { values } = parseArgs({
     "sqlite-selection": { type: "string", default: "all" },
     "packet-selection": { type: "string", default: "baseline" },
     "signal-selection": { type: "string", default: "baseline" },
+    "saz-selection": { type: "string", default: "baseline" },
   },
 });
 if (
@@ -43,6 +45,11 @@ if (
   !isAbsolute(values.path) ||
   !isAbsolute(values.output) ||
   !["complete", "summary"].includes(values.delivery) ||
+  !["baseline", "focused", "retained"].includes(values["saz-selection"]) ||
+  (!Object.hasOwn(SAZ_ARTIFACTS, values.artifact) &&
+    values["saz-selection"] !== "baseline") ||
+  (Object.hasOwn(SAZ_ARTIFACTS, values.artifact) &&
+    values.delivery !== "complete") ||
   !["baseline", "focused", "discovery", "pages"].includes(
     values["signal-selection"],
   ) ||
@@ -174,6 +181,12 @@ const fingerprints = {
   signal_oracle_source_sha256: await fileSha(
     fileURLToPath(new URL("./signal-oracle.py", import.meta.url)),
   ),
+  saz_workflow_source_sha256: await fileSha(
+    fileURLToPath(new URL("./saz-entry.mjs", import.meta.url)),
+  ),
+  saz_oracle_source_sha256: await fileSha(
+    fileURLToPath(new URL("./saz-oracle.py", import.meta.url)),
+  ),
   runtime_files_sha256: sha(JSON.stringify(runtimeFiles)),
   runtime_dispatcher_sha256: await fileSha(
     resolve(values["runtime-root"], "scripts/rea.mjs"),
@@ -754,6 +767,12 @@ const inspectPcapTarget = async () => {
 };
 
 const workflows = {
+  ...Object.fromEntries(
+    Object.keys(SAZ_ARTIFACTS).map((artifact) => [
+      artifact,
+      () => measureSazEntry({ values, call, inputSha }),
+    ]),
+  ),
   ...Object.fromEntries(
     Object.keys(SIGNAL_ARTIFACTS).map((artifact) => [
       artifact,
@@ -1188,6 +1207,9 @@ const report = {
   artifact_sha256: inputSha,
   artifact_sha256_after: finalInputSha,
   delivery: values.delivery,
+  saz_selection: Object.hasOwn(SAZ_ARTIFACTS, values.artifact)
+    ? values["saz-selection"]
+    : null,
   sdk_schema_mode: values.schemas,
   retained_verification: values["verify-retained"],
   native_question: NATIVE_ARTIFACTS.includes(values.artifact)
diff --git a/experiments/analysis-efficiency/review-saz.mjs b/experiments/analysis-efficiency/review-saz.mjs
new file mode 100644
index 0000000000..b836d807c4
--- /dev/null
+++ b/experiments/analysis-efficiency/review-saz.mjs
@@ -0,0 +1,266 @@
+import assert from "node:assert/strict";
+import { readFile, writeFile } from "node:fs/promises";
+import { resolve } from "node:path";
+import { pathToFileURL } from "node:url";
+import { isDeepStrictEqual, parseArgs } from "node:util";
+import {
+  sazRequest,
+  verifySazRecords,
+  verifySazOracleIdentity,
+  sazEvidenceParameters,
+} from "./saz-entry.mjs";
+
+const { values } = parseArgs({
+  options: {
+    trial: { type: "string" },
+    oracle: { type: "string" },
+    "runtime-root": { type: "string" },
+    artifact: { type: "string" },
+    selection: { type: "string" },
+  },
+});
+if (
+  !values.trial ||
+  !values.oracle ||
+  !values["runtime-root"] ||
+  !["baseline", "focused"].includes(values.selection)
+)
+  throw new Error(
+    "Use --trial DIR --oracle JSON --runtime-root DIR --artifact ID --selection baseline|focused",
+  );
+const load = async (file) =>
+  JSON.parse(await readFile(resolve(values.trial, file), "utf8"));
+const settings = await load("trial-settings.json"),
+  summary = await load("summary.json");
+const oracle = JSON.parse(await readFile(values.oracle, "utf8"));
+verifySazOracleIdentity(values.artifact, oracle);
+const runtime = (file) =>
+  pathToFileURL(resolve(values["runtime-root"], file)).href;
+const { parseEvidence } = await import(runtime("dist/domain/evidence.js"));
+const { inspectWebNetworkCaptureInputSchema, webNetworkCaptureSchema } =
+  await import(runtime("dist/domain/webNetworkCapture.js"));
+const { inspectAnalysisViewInputSchema, analysisViewResultSchema } =
+  await import(runtime("dist/domain/analysisView/analysisView.js"));
+const calls = (await readFile(resolve(values.trial, "events.jsonl"), "utf8"))
+  .split("\n")
+  .filter(Boolean)
+  .map(JSON.parse)
+  .filter(
+    (event) =>
+      event.type === "item.completed" && event.item?.type === "mcp_tool_call",
+  )
+  .map((event) => event.item);
+const planned = inspectWebNetworkCaptureInputSchema.parse(
+  sazRequest(
+    values.artifact,
+    settings.selected_artifact_path,
+    values.selection,
+  ),
+);
+const reasons = [],
+  workflowReasons = [],
+  authenticated = new Map(),
+  parents = new Set(),
+  conflicts = new Set();
+let transportRecovery = false;
+for (const call of calls) {
+  try {
+    assert.equal(call.server, "rea");
+    assert.equal(call.status, "completed");
+    assert.ok(!call.error, "RPC did not provide observations");
+    const text = call.result.content.filter((item) => item.type === "text");
+    assert.equal(text.length, 1, "Ambiguous Evidence text carrier");
+    let evidence = JSON.parse(text[0].text);
+    if (call.result.isError) {
+      const limits = evidence.error?.details?.reported_limits;
+      assert.equal(call.tool, "inspect_web_network_capture");
+      assert.equal(evidence.error?.code, "resource_constraint");
+      assert.equal(
+        evidence.error.details.operation,
+        "inspect_web_network_capture",
+      );
+      assert.equal(limits?.boundary, "mcp-response");
+      const input = inspectWebNetworkCaptureInputSchema.parse(call.arguments);
+      assert.deepEqual(input, planned);
+      const source = inspectAnalysisViewInputSchema.parse({
+        source: limits.evidence_reference,
+        view: { kind: "summary" },
+      }).source;
+      assert.equal(source.kind, "retained-evidence");
+      parents.add(source.evidence_id);
+      transportRecovery = true;
+      continue;
+    }
+    if (call.result.structuredContent !== undefined)
+      assert.deepEqual(evidence, call.result.structuredContent);
+    evidence = parseEvidence(evidence);
+    assert.equal(evidence.subject.local_path, settings.selected_artifact_path);
+    assert.equal(evidence.subject.digest.sha256, oracle.sha256);
+    let records;
+    if (call.tool === "inspect_web_network_capture") {
+      const { detail, ...arguments_ } = call.arguments;
+      const input = inspectWebNetworkCaptureInputSchema.parse(arguments_);
+      assert.equal(input.capture_path, settings.selected_artifact_path);
+      assert.equal(input.format, "saz");
+      if (!isDeepStrictEqual(input, planned))
+        workflowReasons.push(
+          "Inspection selection differs from planned arguments",
+        );
+      if (detail === "summary") {
+        workflowReasons.push(
+          "Summary delivery differs from the fixed complete-delivery plan",
+        );
+        assert.equal(evidence.operation, "inspect_analysis_view");
+        const report = analysisViewResultSchema.parse(
+          evidence.normalized_result,
+        );
+        assert.equal(report.kind, "summary");
+        assert.equal(report.parent_operation, "inspect_web_network_capture");
+        assert.equal(report.summary.format, "saz");
+        assert.equal(report.summary.total_records, oracle.total_records);
+        assert.deepEqual(evidence.parameters, {
+          parent_evidence_id: report.parent_evidence_id,
+          view: { kind: "summary" },
+        });
+        parents.add(report.parent_evidence_id);
+        authenticated.set(evidence.evidence_id, []);
+        continue;
+      }
+      assert.ok(detail === undefined || detail === "complete");
+      assert.equal(evidence.operation, "inspect_web_network_capture");
+      assert.deepEqual(
+        evidence.parameters,
+        sazEvidenceParameters(input, oracle.total_records),
+      );
+      const report = webNetworkCaptureSchema.parse(evidence.normalized_result);
+      assert.equal(report.total_records, oracle.total_records);
+      assert.equal(report.format, "saz");
+      records = report.records;
+      parents.add(evidence.evidence_id);
+    } else {
+      assert.equal(call.tool, "inspect_analysis_view");
+      const input = inspectAnalysisViewInputSchema.parse(call.arguments);
+      assert.equal(input.source.kind, "retained-evidence");
+      assert.ok(
+        parents.has(input.source.evidence_id),
+        "Unknown retained parent",
+      );
+      assert.equal(evidence.operation, "inspect_analysis_view");
+      assert.deepEqual(evidence.parameters, {
+        parent_evidence_id: input.source.evidence_id,
+        view: input.view,
+      });
+      const page = analysisViewResultSchema.parse(evidence.normalized_result);
+      assert.deepEqual(page.view, input.view);
+      assert.equal(page.parent_evidence_id, input.source.evidence_id);
+      assert.equal(page.parent_operation, "inspect_web_network_capture");
+      if (page.kind === "summary") {
+        assert.equal(page.summary.total_records, oracle.total_records);
+        assert.equal(page.summary.format, "saz");
+        records = [];
+      } else {
+        assert.equal(page.kind, "capture");
+        records = page.view.detail === "complete" ? page.items : [];
+      }
+    }
+    const prior = authenticated.get(evidence.evidence_id);
+    if (prior && !isDeepStrictEqual(prior, records)) {
+      conflicts.add(evidence.evidence_id);
+      authenticated.delete(evidence.evidence_id);
+      parents.delete(evidence.evidence_id);
+      throw new Error("Conflicting observations share an Evidence ID");
+    }
+    if (!conflicts.has(evidence.evidence_id))
+      authenticated.set(evidence.evidence_id, records);
+  } catch (cause) {
+    workflowReasons.push(String(cause));
+  }
+}
+if (
+  calls[0]?.tool !== "inspect_web_network_capture" ||
+  calls.filter((call) => call.tool === "inspect_web_network_capture").length !==
+    1 ||
+  (!transportRecovery && calls.length !== 1)
+)
+  workflowReasons.push(
+    "Completed call sequence differs from the fixed initial call/recovery plan",
+  );
+try {
+  const raw = (await readFile(resolve(values.trial, "answer.txt"), "utf8"))
+    .trim()
+    .replace(/^```(?:json)?\s*|\s*```$/g, "");
+  const answer = JSON.parse(raw);
+  assert.ok(
+    Array.isArray(answer.evidence_ids) && answer.evidence_ids.length > 0,
+    "Missing Evidence citations",
+  );
+  const observed = new Map();
+  for (const id of answer.evidence_ids) {
+    assert.ok(authenticated.has(id), "Unknown/failed cited Evidence");
+    for (const row of authenticated.get(id)) {
+      const prior = observed.get(row.ordinal);
+      assert.ok(
+        !prior || isDeepStrictEqual(prior, row),
+        "Conflicting cited record identities",
+      );
+      observed.set(row.ordinal, row);
+    }
+  }
+  const facts = verifySazRecords(
+    oracle.record_ordinals.map((ordinal) => observed.get(ordinal)),
+    oracle,
+  );
+  assert.equal(answer.records.length, facts.length, "Answer coverage differs");
+  assert.deepEqual(
+    [...answer.records].sort((a, b) => a.ordinal - b.ordinal),
+    [...facts].sort((a, b) => a.ordinal - b.ordinal),
+    "Answer facts differ from cited stored bytes",
+  );
+  if (
+    !isDeepStrictEqual(
+      answer.records.map((row) => row.ordinal),
+      oracle.record_ordinals,
+    )
+  )
+    workflowReasons.push("Final answer changes caller-selected order");
+  if (
+    !Array.isArray(answer.limitations) ||
+    answer.limitations.length === 0 ||
+    answer.limitations.some(
+      (value) => typeof value !== "string" || !value.trim(),
+    )
+  )
+    workflowReasons.push("Final answer omits limitations");
+  assert.equal(settings.selected_artifact_sha256_before, oracle.sha256);
+  assert.equal(summary.selected_artifact_sha256_after, oracle.sha256);
+  if (
+    summary.exit_code !== 0 ||
+    summary.timeout ||
+    summary.selected_artifact_bytes_unchanged !== true
+  )
+    workflowReasons.push(
+      "Trial did not complete with original bytes unchanged",
+    );
+} catch (cause) {
+  reasons.push(String(cause));
+}
+const report = {
+  factual_correctness: reasons.length ? "failed" : "passed",
+  workflow_correctness:
+    reasons.length || workflowReasons.length ? "failed" : "passed",
+  reasons,
+  workflow_reasons: workflowReasons,
+  completed_calls: calls.length,
+  wall_seconds: summary.wall_seconds,
+  usage: summary.usage,
+  selected_artifact_sha256: oracle.sha256,
+  scope:
+    "Authenticated cited stored HTTP records and independent raw-byte checks; no live attribution, body decompression or certificate interpretation",
+  limitation_prose: "Presence checked; wording requires transcript review",
+};
+await writeFile(
+  resolve(values.trial, "saz-review.json"),
+  JSON.stringify(report, null, 2) + "\n",
+);
+console.log(JSON.stringify(report));
+if (report.workflow_correctness !== "passed") process.exitCode = 1;
diff --git a/experiments/analysis-efficiency/saz-entry.mjs b/experiments/analysis-efficiency/saz-entry.mjs
new file mode 100644
index 0000000000..4952fcd5b1
--- /dev/null
+++ b/experiments/analysis-efficiency/saz-entry.mjs
@@ -0,0 +1,326 @@
+import assert from "node:assert/strict";
+import { execFile } from "node:child_process";
+import { readFile } from "node:fs/promises";
+import { resolve } from "node:path";
+import { fileURLToPath, pathToFileURL } from "node:url";
+import { promisify } from "node:util";
+
+const corpus = JSON.parse(
+  await readFile(new URL("./corpus.json", import.meta.url), "utf8"),
+);
+export const SAZ_ARTIFACTS = Object.fromEntries(
+  corpus.artifacts
+    .filter((entry) => entry.saz_question)
+    .map((entry) => [entry.id, entry]),
+);
+
+export function sazRequest(artifact, path, selection) {
+  const question = Object.hasOwn(SAZ_ARTIFACTS, artifact)
+    ? SAZ_ARTIFACTS[artifact].saz_question
+    : undefined;
+  if (!question) throw new Error("Select a pinned SAZ question");
+  return {
+    capture_path: path,
+    format: "saz",
+    ...(selection === "baseline"
+      ? {}
+      : { record_ordinals: question.record_ordinals }),
+    sensitive_values: [],
+  };
+}
+
+export function verifySazOracleIdentity(artifact, oracle) {
+  assert.ok(
+    Object.hasOwn(SAZ_ARTIFACTS, artifact),
+    "Unknown pinned SAZ question",
+  );
+  const entry = SAZ_ARTIFACTS[artifact];
+  assert.equal(oracle.kind, "saz");
+  assert.equal(oracle.artifact, artifact);
+  assert.equal(oracle.sha256, entry.sha256, "Wrong immutable corpus");
+  assert.equal(oracle.total_records, entry.saz_question.total_records);
+  assert.deepEqual(oracle.record_ordinals, entry.saz_question.record_ordinals);
+}
+
+export function sazEvidenceParameters(input, totalRecords) {
+  return {
+    capture_path: input.capture_path,
+    format: input.format,
+    record_ordinals:
+      input.record_ordinals ??
+      Array.from({ length: totalRecords }, (_, index) => index),
+    sensitive_value_count: input.sensitive_values.length,
+  };
+}
+
+export function verifySazRecords(records, oracle) {
+  assert.equal(oracle.kind, "saz", "Wrong independent question");
+  assert.equal(
+    records.length,
+    oracle.records.length,
+    "Selected coverage differs",
+  );
+  return records.map((record, index) => {
+    const raw = oracle.records[index];
+    assert.equal(record.ordinal, raw.ordinal, "Caller order/ordinal differs");
+    assert.equal(
+      record.reported.session_id,
+      raw.session_id,
+      "Padded source identity differs",
+    );
+    assert.deepEqual(
+      record.location,
+      { kind: "archive-members", members: raw.members },
+      "Original member coordinates differ",
+    );
+    for (const side of ["request", "response"]) {
+      const expected = raw[side],
+        reported = record.reported[side];
+      if (expected === null) {
+        assert.equal(reported, null, "Absent message invented");
+        assert.ok(
+          !record.binary_fields.some(
+            (field) => field.pointer === `/${side}/body`,
+          ),
+          "Absent body invented",
+        );
+        continue;
+      }
+      assert.equal(reported.member, expected.member);
+      for (const field of side === "request"
+        ? ["method", "url", "http_version"]
+        : ["status", "reason", "http_version"])
+        assert.equal(
+          reported[field],
+          expected[field],
+          "Stored HTTP start line differs",
+        );
+      const contentTypes = (headers) =>
+        headers.filter(
+          (header) => header.name.toLowerCase() === "content-type",
+        );
+      assert.deepEqual(
+        contentTypes(reported.headers),
+        contentTypes(expected.headers),
+        "Content-Type order/spelling/values differ",
+      );
+      const bodies = record.binary_fields.filter(
+        (field) => field.pointer === `/${side}/body`,
+      );
+      assert.equal(bodies.length, 1, "Stored body representation differs");
+      const body = bodies[0];
+      assert.equal(body.state, "retained", "Stored body unexpectedly unknown");
+      assert.equal(
+        body.representation,
+        "producer-bytes",
+        "Encoded source representation differs",
+      );
+      for (const field of ["bytes", "sha256", "content_base64"])
+        assert.equal(
+          body[field],
+          expected.body[field],
+          "Stored encoded body bytes differ",
+        );
+      assert.equal(
+        reported.body,
+        expected.body.text,
+        "UTF-8 projection or empty body differs",
+      );
+    }
+    const response = raw.response;
+    return {
+      ordinal: record.ordinal,
+      session_id: raw.session_id,
+      members: raw.members,
+      request:
+        raw.request === null
+          ? null
+          : { method: raw.request.method, url: raw.request.url },
+      response:
+        response === null
+          ? null
+          : {
+              status: response.status,
+              http_version: response.http_version,
+              content_types: response.headers.filter(
+                (header) => header.name.toLowerCase() === "content-type",
+              ),
+              body: {
+                bytes: response.body.bytes,
+                sha256: response.body.sha256,
+              },
+            },
+    };
+  });
+}
+
+export async function measureSazEntry({ values, call, inputSha }) {
+  const runtime = (file) =>
+    pathToFileURL(resolve(values["runtime-root"], file)).href;
+  const { parseEvidence } = await import(runtime("dist/domain/evidence.js"));
+  const { inspectWebNetworkCaptureInputSchema, webNetworkCaptureSchema } =
+    await import(runtime("dist/domain/webNetworkCapture.js"));
+  const { inspectAnalysisViewInputSchema, analysisViewResultSchema } =
+    await import(runtime("dist/domain/analysisView/analysisView.js"));
+  const read = await promisify(execFile)(
+    "python3",
+    [
+      fileURLToPath(new URL("./saz-oracle.py", import.meta.url)),
+      values.artifact,
+      values.path,
+    ],
+    { maxBuffer: 2 * 1024 * 1024, timeout: 30_000 },
+  );
+  const oracle = JSON.parse(read.stdout);
+  verifySazOracleIdentity(values.artifact, oracle);
+  assert.equal(oracle.sha256, inputSha);
+  const input = inspectWebNetworkCaptureInputSchema.parse(
+    sazRequest(values.artifact, values.path, values["saz-selection"]),
+  );
+  const observations = [];
+  const checked = (evidence, schema) => {
+    evidence = parseEvidence(evidence);
+    assert.equal(evidence.subject.local_path, values.path);
+    assert.equal(evidence.subject.digest.sha256, inputSha);
+    const report = schema.parse(evidence.normalized_result);
+    observations.push({
+      evidence_id: evidence.evidence_id,
+      operation: evidence.operation,
+      parameters: evidence.parameters,
+      limitations: evidence.limitations,
+      coverage: report.coverage ?? {
+        total_records: report.total_records,
+        retained_records: report.records.length,
+      },
+    });
+    return report;
+  };
+  const readView = async (source, view) => {
+    const request = inspectAnalysisViewInputSchema.parse({ source, view });
+    const evidence = await call("inspect_analysis_view", request);
+    assert.equal(evidence.operation, "inspect_analysis_view");
+    assert.deepEqual(evidence.parameters, {
+      parent_evidence_id: source.evidence_id,
+      view: request.view,
+    });
+    const result = checked(evidence, analysisViewResultSchema);
+    assert.equal(result.kind, request.view.kind);
+    assert.equal(result.parent_evidence_id, source.evidence_id);
+    assert.equal(result.parent_operation, "inspect_web_network_capture");
+    assert.deepEqual(result.view, request.view);
+    return result;
+  };
+  let records, decoder;
+  if (values["saz-selection"] !== "retained") {
+    let evidence;
+    try {
+      evidence = await call("inspect_web_network_capture", input);
+    } catch (failure) {
+      const error = failure.value?.error;
+      const limits = error?.details?.reported_limits;
+      if (
+        values["saz-selection"] !== "baseline" ||
+        error?.code !== "resource_constraint" ||
+        error.details.operation !== "inspect_web_network_capture" ||
+        limits?.boundary !== "mcp-response" ||
+        !limits.evidence_reference
+      )
+        throw failure;
+      // Recover only the exact retained reference returned by this failed call.
+      // The failed RPC remains in the shared measurement log and task costs.
+      const source = limits.evidence_reference;
+      assert.equal(source.kind, "retained-evidence");
+      const summary = await readView(source, { kind: "summary" });
+      assert.equal(summary.summary.format, "saz");
+      assert.equal(summary.summary.total_records, oracle.total_records);
+      assert.equal(summary.summary.retained_records, oracle.total_records);
+      decoder = summary.summary.decoder;
+      records = [];
+      for (const ordinal of oracle.record_ordinals) {
+        const page = await readView(source, {
+          kind: "capture",
+          detail: "complete",
+          offset: ordinal,
+          limit: 1,
+        });
+        assert.equal(page.coverage.total, oracle.total_records);
+        assert.equal(page.items.length, 1);
+        records.push(page.items[0]);
+      }
+    }
+    if (evidence !== undefined) {
+      assert.equal(evidence.operation, "inspect_web_network_capture");
+      assert.deepEqual(
+        evidence.parameters,
+        sazEvidenceParameters(input, oracle.total_records),
+      );
+      const report = checked(evidence, webNetworkCaptureSchema);
+      assert.equal(report.format, "saz");
+      assert.equal(report.total_records, oracle.total_records);
+      assert.equal(
+        report.records.length,
+        values["saz-selection"] === "baseline"
+          ? oracle.total_records
+          : oracle.records.length,
+      );
+      if (values["saz-selection"] === "baseline")
+        assert.deepEqual(
+          report.records.map((record) => record.ordinal),
+          Array.from({ length: oracle.total_records }, (_, index) => index),
+        );
+      records =
+        values["saz-selection"] === "baseline"
+          ? oracle.record_ordinals.map((ordinal) =>
+              report.records.find((record) => record.ordinal === ordinal),
+            )
+          : report.records;
+      decoder = report.decoder;
+    }
+  } else {
+    const initial = await call("inspect_web_network_capture", {
+      ...input,
+      detail: "summary",
+    });
+    const summary = checked(initial, analysisViewResultSchema);
+    assert.equal(summary.kind, "summary");
+    assert.equal(initial.operation, "inspect_analysis_view");
+    assert.deepEqual(initial.parameters, {
+      parent_evidence_id: summary.parent_evidence_id,
+      view: { kind: "summary" },
+    });
+    assert.equal(summary.summary.format, "saz");
+    assert.equal(summary.parent_operation, "inspect_web_network_capture");
+    assert.equal(summary.summary.total_records, oracle.total_records);
+    assert.equal(summary.summary.retained_records, oracle.records.length);
+    decoder = summary.summary.decoder;
+    const source = {
+      kind: "retained-evidence",
+      evidence_id: summary.parent_evidence_id,
+    };
+    const page = async (detail, offset, limit) => {
+      return readView(source, { kind: "capture", detail, offset, limit });
+    };
+    const preview = await page("preview", 0, 3);
+    assert.deepEqual(
+      preview.items.map((record) => record.ordinal),
+      oracle.record_ordinals,
+    );
+    const first = await page("complete", 0, 2),
+      next = await page("complete", 2, 2),
+      repeated = await page("complete", 0, 2);
+    assert.deepEqual(first.items, repeated.items);
+    assert.deepEqual(first.coverage, repeated.coverage);
+    assert.equal(first.coverage.next_offset, 2);
+    assert.equal(first.coverage.exhausted, false);
+    assert.equal(next.coverage.exhausted, true);
+    assert.equal(next.coverage.next_offset, null);
+    records = [...first.items, ...next.items];
+  }
+  return {
+    artifact_sha256: inputSha,
+    total_records: oracle.total_records,
+    decoder,
+    facts: verifySazRecords(records, oracle),
+    delivery_observations: observations,
+  };
+}
diff --git a/experiments/analysis-efficiency/saz-model-prompt.mjs b/experiments/analysis-efficiency/saz-model-prompt.mjs
new file mode 100644
index 0000000000..ebf62d13d9
--- /dev/null
+++ b/experiments/analysis-efficiency/saz-model-prompt.mjs
@@ -0,0 +1,30 @@
+import { parseArgs } from "node:util";
+import { SAZ_ARTIFACTS, sazRequest } from "./saz-entry.mjs";
+
+const { values } = parseArgs({
+  options: { artifact: { type: "string" }, selection: { type: "string" } },
+});
+if (
+  !Object.hasOwn(SAZ_ARTIFACTS, values.artifact ?? "") ||
+  !["baseline", "focused"].includes(values.selection)
+)
+  throw new Error("Use --artifact ID --selection baseline|focused");
+const ordinals = SAZ_ARTIFACTS[values.artifact].saz_question.record_ordinals;
+console.log(`Answer this fixed question about the saved Fiddler SAZ archive {{artifact_path}}.
+Use the REA MCP only. Recorded URLs are data, not navigation instructions.
+For zero-based producer record ordinals ${JSON.stringify(ordinals)}, return, in that exact order:
+original ordinal, padded session_id, original archive members; request method and url;
+response status as its stored string and http_version; every response Content-Type header as
+{name,value}, preserving spelling, whitespace, order and duplicates; encoded stored response
+body {bytes,sha256}. Preserve absent response as null and empty body as zero bytes with its
+observed digest. Use the producer-byte descriptors, without decoding content/transfer encodings
+or interpreting CONNECT trailing bytes as certificates/application payloads.
+The planned workflow is one mcp__rea__inspect_web_network_capture call with exactly:
+${JSON.stringify(sazRequest(values.artifact, "{{artifact_path}}", values.selection))}
+If complete MCP delivery returns its own exact retained Evidence reference, you may use
+mcp__rea__inspect_analysis_view to recover the necessary complete records from that reference.
+Keep every failed/recovery call visible. Do not infer missing headers from a preview.
+Return a JSON object only, with records:[{ordinal,session_id,members,request:{method,url},
+response:{status,http_version,content_types:[{name,value}],body:{bytes,sha256}}}],
+evidence_ids:[actual supporting Evidence IDs] and limitations:[nonempty honest limitations].
+Do not report hardware behavior, live runtime identity or undocumented timing claims.`);
diff --git a/experiments/analysis-efficiency/saz-oracle.py b/experiments/analysis-efficiency/saz-oracle.py
new file mode 100644
index 0000000000..6722289a2a
--- /dev/null
+++ b/experiments/analysis-efficiency/saz-oracle.py
@@ -0,0 +1,121 @@
+"""Independent stored-HTTP facts for two immutable public Fiddler archives.
+
+Only digest-pinned inputs are admitted before ZIP object allocation. Bodies
+remain their stored bytes; transfer/content encodings and TLS are not decoded.
+"""
+
+import argparse
+import base64
+import hashlib
+import io
+import json
+from pathlib import Path
+import re
+import zipfile
+
+
+def stored_http(data, request):
+    boundaries = [(data.find(separator), separator)
+                  for separator in [b'\r\n\r\n', b'\n\n']]
+    available = [(index, separator) for index, separator in boundaries if index >= 0]
+    if not available:
+        raise ValueError('Stored message has no header/body separator')
+    boundary, separator = min(available, key=lambda item: item[0])
+    lines = data[:boundary].split(b'\r\n' if separator.startswith(b'\r') else b'\n')
+    start = lines[0].decode('latin1')
+    fields = start.split(' ')
+    if request:
+        if len(fields) != 3 or fields[2] not in ['HTTP/1.0', 'HTTP/1.1']:
+            raise ValueError('Unsupported stored request line')
+        result = dict(zip(['method', 'url', 'http_version'], fields))
+    else:
+        fields = start.split(' ', 2)
+        if len(fields) < 2 or fields[0] not in ['HTTP/1.0', 'HTTP/1.1'] or not re.fullmatch(r'[0-9]{3}', fields[1]):
+            raise ValueError('Unsupported stored response line')
+        result = {'http_version': fields[0], 'status': fields[1],
+                  'reason': fields[2] if len(fields) == 3 else ''}
+    headers = []
+    newline = '\r\n' if separator.startswith(b'\r') else '\n'
+    for raw in lines[1:]:
+        line = raw.decode('latin1')
+        if line.startswith((' ', '\t')):
+            if not headers:
+                raise ValueError('Header continuation has no owning field')
+            headers[-1]['value'] += newline + line
+            continue
+        name, delimiter, value = line.partition(':')
+        if not delimiter or not name:
+            raise ValueError('Stored header has no field delimiter')
+        headers.append({'name': name, 'value': value})
+    body = data[boundary + len(separator):]
+    try:
+        body_text = body.decode('utf8', errors='strict')
+    except UnicodeDecodeError:
+        body_text = None
+    return result | {'headers': headers, 'body': {
+        'content_base64': base64.b64encode(body).decode('ascii'),
+        'bytes': len(body), 'sha256': hashlib.sha256(body).hexdigest(),
+        'text': body_text,
+    }}
+
+
+def read_saz(artifact, path):
+    corpus = json.loads(Path(__file__).with_name('corpus.json').read_text())
+    entry = next((item for item in corpus['artifacts']
+                  if item['id'] == artifact and 'saz_question' in item), None)
+    if entry is None:
+        raise ValueError('Select a pinned SAZ artifact')
+    # Real source sizes bound the snapshot; size/digest gate ZIP allocation.
+    if path.stat().st_size != entry['bytes']:
+        raise ValueError('Selected file size differs from pinned SAZ corpus')
+    with path.open('rb') as source:
+        data = source.read(entry['bytes'] + 1)
+    digest = hashlib.sha256(data).hexdigest()
+    if len(data) != entry['bytes'] or digest != entry['sha256']:
+        raise ValueError('Selected bytes differ from pinned SAZ corpus')
+    question = entry['saz_question']
+    with zipfile.ZipFile(io.BytesIO(data)) as archive:
+        inventory = archive.infolist()
+        if (len(inventory) != question['members'] or
+                sum(item.file_size for item in inventory) != question['expanded_bytes']):
+            raise ValueError('Pinned directory coverage differs')
+        groups = {}
+        for item in inventory:
+            match = re.fullmatch(r'raw/(\d+)_(c\.txt|s\.txt|m\.xml|w\.txt|g\.txt)', item.filename)
+            if match:
+                groups.setdefault(match[1], []).append(item)
+        if len(groups) != question['total_records']:
+            raise ValueError('Pinned session coverage differs')
+        ordered = list(groups.items())
+        records = []
+        for ordinal in question['record_ordinals']:
+            session_id, members = ordered[ordinal]
+            result = {'ordinal': ordinal, 'session_id': session_id,
+                      'members': [item.filename for item in members],
+                      'request': None, 'response': None}
+            for item in members:
+                side = 'request' if item.filename.endswith('_c.txt') else 'response' if item.filename.endswith('_s.txt') else None
+                if side is None:
+                    continue
+                with archive.open(item) as source:
+                    content = source.read(item.file_size + 1)
+                if len(content) != item.file_size:
+                    raise ValueError('Stored member size differs')
+                result[side] = {'member': item.filename, **stored_http(content, side == 'request')}
+            records.append(result)
+    with path.open('rb') as source:
+        after = source.read(entry['bytes'] + 1)
+    if hashlib.sha256(after).hexdigest() != digest:
+        raise ValueError('Selected archive changed during independent read')
+    return {'kind': 'saz', 'artifact': artifact, 'sha256': digest,
+            'total_records': len(groups), 'record_ordinals': question['record_ordinals'],
+            'records': records,
+            'scope': 'Stored HTTP bytes and ZIP member coordinates; no body decompression, live attribution or certificate interpretation'}
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('artifact')
+    parser.add_argument('path', type=Path)
+    args = parser.parse_args()
+    print(json.dumps(read_saz(args.artifact, args.path)))
diff --git a/experiments/analysis-efficiency/saz-selection-workflows.md b/experiments/analysis-efficiency/saz-selection-workflows.md
new file mode 100644
index 0000000000..e3b02eaf0d
--- /dev/null
+++ b/experiments/analysis-efficiency/saz-selection-workflows.md
@@ -0,0 +1,214 @@
+# Fiddler SAZ selection checkpoint
+
+This experiment addresses [#3132](https://github.com/morluto/rea/issues/3132).
+Its runtime baseline is `ca1b2f8608c52ed721d3576ba5eeb1d66717b08c`.
+Production APIs and defaults are unchanged. The measured cohort establishes no
+stable whole-task token/time benefit; no new production preset, cache or API is
+justified by this checkpoint.
+
+## Genuine public corpus
+
+Both archives come from Microsoft's
+[Xbox Live Trace Analyzer](https://github.com/microsoft/xbox-live-trace-analyzer/tree/ad880ff00b26c4cae5e3c8864d3df438a03890ba/Source/TestData/CertWarnings)
+at immutable revision `ad880ff00b26c4cae5e3c8864d3df438a03890ba`.
+The repository's [MIT license](https://github.com/microsoft/xbox-live-trace-analyzer/blob/ad880ff00b26c4cae5e3c8864d3df438a03890ba/LICENSE.md)
+and original source files remain upstream. No sample binaries, raw HTTP material,
+authentication values, configs or detailed results are committed.
+
+| Artifact                 | Archive bytes | Expanded member bytes | Sessions | Members | Producer archive comment |
+| ------------------------ | ------------: | --------------------: | -------: | ------: | ------------------------ |
+| `xbox-clear_nomp.saz`    |        43,515 |                62,609 |       13 |      42 | Fiddler v4.5.1.2         |
+| `xbox-clear_mp_cs10.saz` |     5,223,519 |             9,223,820 |    1,311 |   3,939 | Fiddler v4.6.2.0         |
+
+`corpus.json` pins download URLs, file sizes, SHA-256, Git blob IDs, directory
+coverage and question ordinals. Downloaded bytes match their Git blob and
+SHA-256. A third public 103-session archive is pinned as an admission-negative
+probe, outside completed-question comparisons. Archive comments report their producer
+versions; no local Fiddler installation or independent producer-version proof
+is claimed.
+
+## Fixed analyst question
+
+Before examining transaction facts, select ordinals `[12, 0, 6]` in the control
+and `[1310, 0, 655]` in the larger archive. Return that caller-selected order,
+original padded session IDs and archive-member coordinates, request method/URL,
+response status/version, all recorded Content-Type pairs in their original
+spelling/order, and stored response-body byte length/SHA-256.
+
+Bodies retain their archived encoding. CONNECT trailing bytes are stored data,
+not an inferred HTTP application body or certificate interpretation. Empty bodies
+have zero bytes and the empty digest; an absent response remains absent. Header
+duplication/folding, Latin-1 spelling and binary body bytes are independently
+read without REA imports. Synthetic reader tests cover those semantics and
+malformed input; they establish no provider coverage.
+
+The oracle admits only immutable corpus size/digest matches before constructing
+ZIP objects, preserves central-directory session/member order, reads selected
+members with size sentinels and CRC validation, and checks the source digest
+again. Necessary facts are independently checked for all six selected records.
+
+```sh
+python3 experiments/analysis-efficiency/test_saz_oracle.py
+python3 experiments/analysis-efficiency/saz-oracle.py xbox-clear_nomp.saz /absolute/clear_nomp.saz
+python3 experiments/analysis-efficiency/saz-oracle.py xbox-clear_mp_cs10.saz /absolute/clear_mp_cs10.saz
+```
+
+The baseline build, five oracle regressions and the SDK checker semantics passed
+on Linux. Real MCP preflight checks verify control retained pages/repeated reads,
+the large ordinal-selected result, and large complete-workflow recovery. The
+default large MCP return completes analysis but exceeds the existing 10 MiB stdio
+receive-buffer budget; its exact returned retained Evidence reference supports
+summary and complete pages. Recovery preserves the failed RPC and every follow-up
+call, and matches the independently checked necessary facts from direct selection.
+This transport limit is not a decoding timeout. Formal repetitions and actual
+model tasks are reported below; preflight observations are not eligible cost pairs.
+
+The 103-session `xbox-x49_mp_cs21.saz` candidate contains an empty stored response
+member `raw/099_s.txt`, lacking an HTTP header/body separator. On this baseline,
+whole-capture admission fails with `invalid_request` / `invalid_format`. Its
+ordinals `[102, 0, 51]` were fixed before inspecting its transaction facts. The
+failed attempt/source member is preserved as a negative probe; no parser repair
+or successful-task comparison is claimed for that archive.
+
+Other platform behavior remains unverified.
+Detailed outputs stay private; missing resource counters remain unknown.
+
+The large selected CLI workflow also matches independent bytes and MCP facts,
+parameters and Evidence ID. A separate default complete CLI attempt hit the
+provider's 30-second timeout under the one-CPU guard; that failed attempt is
+preserved separately from the successful-analysis MCP receive-buffer failure.
+It does not prove a general CLI failure or identify an engine-stage cause.
+
+Ten grading preflight cases cover valid actual Evidence, wrong final digests,
+unknown citations, missing limitations, answer order, failed calls before/after
+valid data, semantic-ID corruption and actual large retained recovery. Their
+answers/events are injected checker tests, not actual model trials. An invalid
+later duplicate is rejected by the domain's semantic-ID owner and does not erase
+earlier authenticated observations.
+
+## SDK and model reproduction
+
+```sh
+node experiments/analysis-efficiency/measure.mjs \
+  --runtime-root /absolute/frozen-runtime --revision ca1b2f8608c52ed721d3576ba5eeb1d66717b08c \
+  --artifact xbox-clear_nomp.saz --path /absolute/clear_nomp.saz \
+  --output /absolute/new-result-dir --schemas selected --saz-selection focused
+node experiments/analysis-efficiency/saz-model-prompt.mjs \
+  --artifact xbox-clear_nomp.saz --selection focused > /absolute/prompt.txt
+```
+
+Use the same runtime, inputs, explicit provider settings and resource profile
+within a cost cohort; retain runtime/workflow/oracle/settings fingerprints. Repeat
+both core archives with baseline/focused/focused/baseline order and retain all
+failures/recovery RPCs. The `retained` diagnostic workflow also measures preview,
+two complete pages and a repeated first page. Preview omits headers, so preview
+alone is not equivalent to the complete header/body question.
+
+Run real Codex tasks through `agent-trial.py` with an isolated temporary profile,
+the two exact `inspect_web_network_capture` / `inspect_analysis_view` tools,
+the same requested model/reasoning and at least two baseline/focused pairs for
+the control. Grade cited observations with `review-saz.mjs`, an independently
+produced oracle and the tested runtime. Count whole-task input/output tokens,
+cached input, elapsed time and all calls; parsed JSON bytes are not model tokens.
+The executed scope and all failed/deviating trials follow below.
+
+## Executed SDK cohort
+
+Runtime is frozen at `ca1b2f8608c52ed721d3576ba5eeb1d66717b08c`; corrected
+experiment source is `286c3e48ad50d9e84baa3af00ca522a8cef36e95`. All corrected
+SDK runtime/workflow/oracle/provider-setting fingerprints match. Node is 24.18.0
+and MCP SDK is 2.3.1. SDK processes have a 1 GiB Node heap ceiling; model MCP
+profiles separately record 512 MiB. Compare within each cohort.
+
+The initial batch preserves 12 workflows / 36 RPCs. Two control baseline collectors
+incorrectly expected omitted request ordinals to remain omitted in Evidence;
+the owning application records the resolved full scope. Original RPC arguments
+preserve omission. Both saved replies independently validate correct typed
+Evidence/facts after the checker correction. Original failures, hashes and costs
+remain unchanged and are not pooled into the corrected batch.
+
+The corrected batch executes 12 workflows / 32 RPCs: **11 workflows pass** all necessary
+byte/header/member checks with equal facts across successful arms. One large
+default complete query hits `provider_timeout` at the 30-second provider budget.
+Failed complete-return RPCs in successful retained recovery also stay in cost.
+
+| Archive / workflow | Repetition outcomes | All parsed reply bytes | RPC ms                | Whole runner ms       |
+| ------------------ | ------------------- | ---------------------: | --------------------- | --------------------- |
+| control / baseline | passed / passed     |        249272 / 249272 | 817.904 / 817.032     | 5393.982 / 4922.524   |
+| control / focused  | passed / passed     |          65390 / 65390 | 805.492 / 835.598     | 4897.670 / 4924.772   |
+| control / retained | passed / passed     |        146981 / 146981 | 1846.344 / 8258.135   | 5861.567 / 18108.568  |
+| large / baseline   | failed / passed     |           520 / 114294 | 32303.203 / 13623.033 | 52432.342 / 18064.360 |
+| large / focused    | passed / passed     |        535767 / 535767 | 32917.417 / 23536.051 | 52933.865 / 37045.394 |
+| large / retained   | passed / passed     |        174477 / 174477 | 10970.973 / 11550.158 | 15140.224 / 15669.775 |
+
+Control direct selection reduces parsed replies **73.8%**. Large complete-result
+recovery delivers 114,294 bytes, while direct ordinal selection retains the original
+container inventory and delivers 535,767 bytes. The five-call retained diagnostic
+delivers 174,477 bytes, including preview and repeated-page verification. A smaller
+record set alone does not guarantee less delivery; bytes are not wire bytes or
+model tokens.
+
+Initial RPCs combine decoding, validation, canonicalization and delivery; no
+decode-only/format-only attribution is claimed. First-use and repeated retained
+reads remain separate RPC observations. Fresh client/server processes do not
+establish cold OS/provider caches. Substantial time variation precludes a stable
+general speedup claim.
+
+## Actual Codex outcomes
+
+Four actual Codex CLI 0.161.0 trials request `gpt-5.6-terra`/`high` with the same
+two enabled/preapproved passive tools in baseline/focused/focused/baseline order.
+The requested alias is pinned; resolved backend identity is unknown. Temporary
+profiles preserve original bytes. The requested mode is `direct`, but owned
+tool-context logs show actual `exec` discovery/dispatch; neither guaranteed direct
+presentation nor exact submitted model requests are established.
+
+| Trial      | Facts / fixed plan | MCP calls | Total input | Cached input | Uncached input, derived | Output | Seconds |
+| ---------- | ------------------ | --------: | ----------: | -----------: | ----------------------: | -----: | ------: |
+| baseline-1 | passed / failed    |         4 |      97,988 |       53,504 |                  44,484 |  1,556 |  51.266 |
+| focused-1  | passed / failed    |         2 |      96,433 |       51,456 |                  44,977 |  1,400 |  52.288 |
+| focused-2  | failed / failed    |         0 |      21,379 |        9,984 |                  11,395 |    317 |  16.301 |
+| baseline-2 | passed / failed    |         4 |     172,043 |      125,696 |                  46,347 |  1,943 |  62.570 |
+
+**Three of four tasks answer correctly.** All three add compatible retained-view
+calls and differ from the initially prescribed single MCP-call plan. These
+successful completed-task costs stay visible separately from plan conformance.
+Baselines read three exact record views; the successful focus reads one complete
+selected-parent page.
+
+The second focus makes zero MCP calls and reports unavailable tools. Its owned
+exact-name discovery returns an empty list, while the same names work in other
+trials; stderr is empty. The cause is unknown, not proof of a REA outage. Keep
+its 21,379 input tokens / 16.301 seconds as a failed task, without a replacement retry
+or an apparent speed win.
+
+Only the first pair completes both questions: total input decreases **1.6%**,
+time increases **2.0%**, and derived uncached input increases **1.1%**. The second
+pair has no completed focused question; no paired speed/token claim is valid.
+Limitation prose was inspected: successful answers qualify stored bytes,
+ZIP-directory order and absent transfer/content decoding. This small mixed cohort
+justifies **no production optimization**.
+
+## Resource and compatibility scope
+
+Heavy work is serial with one CPU, 3 GiB early stop / 4 GiB hard family memory, no swap,
+task ceilings and host-availability/ownership checks. Corrected SDK payload-exit
+peak is 906,825,728 bytes (**865 MiB**); model peak is 873,148,416 bytes (**833 MiB**).
+All 16 corrected SDK/model payload phases verify cleanup, no observed OOM/abort
+and captured CPU counters. Terminal whole-scope counters remain unknown. Failed
+tool/model tasks do not become correct because their process exits cleanly.
+
+Family CPU seconds at payload exit span 5.898–44.331 for corrected SDK workflows
+(median 15.808), including failed/recovered attempts. Model baseline/focus/focus/
+baseline observations are 7.573 / 7.303 / 5.810 / 7.618 CPU seconds; the third
+model task failed and cannot establish an efficiency gain. These counters include
+the owned runner/client/server/observer family, not decoder-only CPU.
+
+Verification covers Linux, two immutable positive archives plus one genuine
+partial-message negative source, one installed parser stack and one requested
+model alias. Other platforms, Fiddler acquisition/import, live deployment or
+certificate semantics, arbitrary layouts, direct presentation guarantees and
+unmeasured engine stages remain unknown. Complete defaults, caller choices,
+encoded bytes, redactions, semantic Evidence identity and session-owned views
+remain the compatibility baseline. Reuse existing selectors/views; expand
+SDK/cache/default behavior only after completed tasks show a benefit.
diff --git a/experiments/analysis-efficiency/test_saz_oracle.py b/experiments/analysis-efficiency/test_saz_oracle.py
new file mode 100644
index 0000000000..488527a3bc
--- /dev/null
+++ b/experiments/analysis-efficiency/test_saz_oracle.py
@@ -0,0 +1,48 @@
+"""Stored-byte semantics and fail-before-allocation admission, not providers."""
+
+import base64
+import importlib.util
+from pathlib import Path
+import tempfile
+import unittest
+from unittest.mock import patch
+
+spec = importlib.util.spec_from_file_location('saz_oracle', Path(__file__).with_name('saz-oracle.py'))
+oracle = importlib.util.module_from_spec(spec)
+spec.loader.exec_module(oracle)
+
+
+class SazOracleTests(unittest.TestCase):
+    def test_duplicate_spelling_and_folding_are_source_facts(self):
+        result = oracle.stored_http(b'HTTP/1.1 200 OK\r\nX-Case: first\r\nx-case:second\r\n continuation\r\n\r\n', False)
+        self.assertEqual(result['headers'], [{'name': 'X-Case', 'value': ' first'},
+                                            {'name': 'x-case', 'value': 'second\r\n continuation'}])
+        self.assertEqual(result['body']['bytes'], 0)
+        self.assertEqual(result['body']['text'], '')
+
+    def test_body_keeps_encoded_bytes_and_internal_separators(self):
+        body = b'\x1f\x8b\xff\n\n\r\n\r\n'
+        result = oracle.stored_http(b'HTTP/1.1 200 OK\r\nContent-Encoding: gzip\r\n\r\n' + body, False)
+        self.assertEqual(base64.b64decode(result['body']['content_base64']), body)
+        self.assertIsNone(result['body']['text'])
+
+    def test_first_lf_separator_owns_body_with_later_crlf_separator(self):
+        result = oracle.stored_http(b'GET / HTTP/1.1\nHost: example.com\n\nbody\r\n\r\n', True)
+        self.assertEqual(result['body']['text'], 'body\r\n\r\n')
+
+    def test_malformed_start_and_unowned_continuation_fail(self):
+        for data in [b'garbage\r\n\r\n', b'HTTP/1.1 200 OK\r\n orphan\r\n\r\n']:
+            with self.assertRaises(ValueError):
+                oracle.stored_http(data, False)
+
+    def test_wrong_digest_cannot_allocate_zip_objects(self):
+        with tempfile.TemporaryDirectory() as root:
+            path = Path(root)/'wrong.saz'
+            path.write_bytes(b'X'*43515)
+            with patch.object(oracle.zipfile, 'ZipFile', side_effect=AssertionError('Too late')):
+                with self.assertRaisesRegex(ValueError, 'pinned SAZ corpus'):
+                    oracle.read_saz('xbox-clear_nomp.saz', path)
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/tests/evaluation/sazWorkflowSemantics.test.ts b/tests/evaluation/sazWorkflowSemantics.test.ts
new file mode 100644
index 0000000000..4ba82cddf3
--- /dev/null
+++ b/tests/evaluation/sazWorkflowSemantics.test.ts
@@ -0,0 +1,49 @@
+import assert from "node:assert/strict";
+import { spawnSync } from "node:child_process";
+import { test } from "vitest";
+
+test("SAZ verification preserves source identities, duplicate headers and encoded empty bodies", () => {
+  const moduleUrl = new URL(
+    "../../experiments/analysis-efficiency/saz-entry.mjs",
+    import.meta.url,
+  ).href;
+  const source = `
+    import assert from 'node:assert/strict';
+    import {verifySazRecords,verifySazOracleIdentity,sazRequest,sazEvidenceParameters} from ${JSON.stringify(moduleUrl)};
+    const omitted=sazRequest('xbox-clear_nomp.saz','/tmp/sample.saz','baseline');
+    assert.equal(Object.hasOwn(omitted,'record_ordinals'),false);
+    assert.deepEqual(sazEvidenceParameters(omitted,3),{capture_path:'/tmp/sample.saz',format:'saz',record_ordinals:[0,1,2],sensitive_value_count:0});
+    assert.deepEqual(sazEvidenceParameters(sazRequest('xbox-clear_nomp.saz','/tmp/sample.saz','focused'),13).record_ordinals,[12,0,6]);
+    const empty={bytes:0,sha256:'e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855',content_base64:'',text:''};
+    const raw=[12,0,6].map(ordinal=>{
+      const session_id=String(ordinal+1).padStart(2,'0');
+      return {ordinal,session_id,members:['raw/'+session_id+'_c.txt','raw/'+session_id+'_s.txt'],
+        request:{member:'raw/'+session_id+'_c.txt',method:'GET',url:'/',http_version:'HTTP/1.1',headers:[],body:empty},
+        response:{member:'raw/'+session_id+'_s.txt',status:'200',reason:'OK',http_version:'HTTP/1.1',headers:[{name:'Content-Type',value:' text/plain'},{name:'content-type',value:''}],body:empty}};
+    });
+    const oracle={kind:'saz',artifact:'xbox-clear_nomp.saz',sha256:'3e808663a99cbb9160186f7419633248200fbc6863d2c897f5d4d9972dde0375',total_records:13,record_ordinals:[12,0,6],records:raw};
+    const rows=raw.map(row=>({ordinal:row.ordinal,location:{kind:'archive-members',members:row.members},
+      reported:{session_id:row.session_id,request:{...row.request,body:''},response:{...row.response,body:''}},
+      binary_fields:['request','response'].map(side=>({pointer:'/'+side+'/body',representation:'producer-bytes',state:'retained',...empty}))}));
+    verifySazOracleIdentity('xbox-clear_nomp.saz',oracle);
+    const facts=verifySazRecords(rows,oracle);assert.deepEqual(facts.map(row=>row.ordinal),[12,0,6]);
+    assert.equal(facts[0].response.content_types.length,2);
+    assert.throws(()=>verifySazRecords(rows.slice(1),oracle),/coverage/);
+    assert.throws(()=>verifySazRecords(rows.toReversed(),oracle),/order/);
+    const changed=mutate=>{const copy=structuredClone(rows);mutate(copy);return copy;};
+    assert.throws(()=>verifySazRecords(changed(copy=>copy[1].reported.session_id='1'),oracle),/identity/);
+    assert.throws(()=>verifySazRecords(changed(copy=>copy[0].reported.response.headers.pop()),oracle),/Content-Type/);
+    assert.throws(()=>verifySazRecords(changed(copy=>copy[0].binary_fields[1].content_base64='WA=='),oracle),/encoded body/);
+    assert.throws(()=>verifySazRecords(changed(copy=>copy[0].reported.response=null),oracle));
+    assert.throws(()=>verifySazRecords(changed(copy=>copy[0].location.members=['raw/other_s.txt']),oracle),/coordinates/);
+    assert.throws(()=>verifySazOracleIdentity('xbox-clear_nomp.saz',{...oracle,sha256:'a'.repeat(64)}),/corpus/);
+    assert.throws(()=>sazRequest('__proto__','/tmp/sample.saz','focused'),/pinned/);
+  `;
+  // Checker-only synthesized rows; actual-provider verification is separate.
+  const result = spawnSync(process.execPath, ["--input-type=module", "-"], {
+    input: source,
+    encoding: "utf8",
+  });
+  assert.ifError(result.error);
+  assert.equal(result.status, 0, result.stderr);
+});