feat: add verified industry reference examples

.coderabbit.yaml

@@ -0,0 +1,20 @@
+# yaml-language-server: $schema=https://coderabbit.ai/integrations/schema.v2.json
+
+reviews:
+  auto_review:
+    enabled: true
+    drafts: false
+  path_filters:
+    - "examples/**"
+    - "!examples/**/data/**"
+    - "!examples/**/fixtures/**"
+    - "examples/**/fixtures/SOURCES.md"
+    - "!examples/**/verified-run/**"
+    - "examples/**/verified-run/README.md"
+    - "!examples/**/uv.lock"
+  path_instructions:
+    - path: "examples/**"
+      instructions: |
+        Review runnable source, tests, configuration, and documentation.
+        Check that commands match the implementation and that tests cover failure paths.
+        Flag client-side model-name translation and undocumented synthetic results.

examples/README.md

@@ -25,6 +25,9 @@ service keys.
 | [Multi-model contract review with the OpenAI Agents SDK](./contract-review-agent) | Running an OpenAI Agents SDK agent whose every model call (triage, orchestration, vision, OCR, embeddings, rerank, entity extraction, text-to-SQL, reasoning, and a safety guardrail) is served by one SIE cluster, each step on the right catalog model, with per-model observability | `chat/completions`, `encode`, `score`, `extract` | GPU SIE deployment required; standalone `uv` project; real contracts fetched from CUAD (CC BY 4.0) | Runnable demo |
 | [Turn difficult PDFs into Markdown](./document-to-markdown) | Preserving tables, reading order, headings, and form labels across real financial, academic, and government PDFs | `extract` | SIE endpoint with `docling`; standalone `uv` project; source PDFs fetched at run time | Runnable evaluation example |
 | [Review a flood claim packet](./insurance-claims-agent) | Finding an unsigned form and a $400 evidence mismatch across a policy, estimate, inventory, and damage photograph | `extract`, `score`, `chat/completions` | GPU SIE deployment; standalone `uv` project; public FEMA documents and public-domain photograph fetched at run time | Runnable agent example |
+| [Trace a restated filing figure](./financial-filing-agent) | Following one reported figure through an original filing, corrective notice, and restatement while preserving source status | `extract`, `encode`, `score` | SIE endpoint; standalone `uv` project; public SEC facts and saved verified evidence | Runnable agent example |
+| [Reproduce CMS's L1851 documentation finding](./prior-authorization-review-agent) | Tracing a published six-month requirement against a seven-month face-to-face encounter and CMS's recoupment result | `extract`, `encode`, `score` | SIE endpoint; standalone `uv` project; exact CMS published example | Runnable agent example |
+| [Reconstruct a bearing failure](./maintenance-triage-agent) | Turning the NTSB's three East Palestine detector readings into a cited temperature and alert sequence without adding a new causal claim | `extract`, `encode`, `score` | SIE endpoint; standalone `uv` project; exact NTSB illustrated report spread | Runnable agent example |
 
 For docs publishing, lead with the quickest runnable demos, then use the
 benchmark and evaluation examples for deeper technical users.

examples/financial-filing-agent/.env.example

@@ -0,0 +1,2 @@
+SIE_CLUSTER_URL=http://localhost:8080
+SIE_API_KEY=

examples/financial-filing-agent/.gitignore

@@ -0,0 +1,7 @@
+.env
+.venv/
+.pytest_cache/
+.ruff_cache/
+__pycache__/
+*.pyc
+runs/

examples/financial-filing-agent/README.md

@@ -0,0 +1,84 @@
+# Trace a restated figure back to the controlling filings
+
+This example follows one Pathward Financial fact through exact table rows from
+an original Form 10-Q and a restated Form 10-K/A, plus exact sentences from an
+Item 4.02 Form 8-K. It returns both versions, calculates the change, and keeps
+the source status and company caveat attached.
+
+The result is narrow on purpose. It does not classify the restatement as fraud
+or misconduct, and it does not make an investment recommendation.
+
+## What runs on SIE
+
+| Step | Model | Output |
+|---|---|---|
+| Parse the source packet | `docling` | Markdown with source headings |
+| Retrieve candidate passages | `BAAI/bge-m3` | Dense vectors and cosine ranking |
+| Rerank against the exact question | `Qwen/Qwen3-Reranker-4B` | Ordered evidence with scores |
+| Verify the cited spans | `urchade/gliner_multi-v2.1` | Exact source spans |
+| Recover exact source spans | `fastino/gliner2-large-v1` | Table values, period, company and reliance-status entities |
+
+Each stage consumes the prior stage. The result fails closed if either entity
+model omits a required source span. The original values come from the original
+Form 10-Q table. The restated values come from the Form 10-K/A table, and the
+pipeline checks that its “As Previously Reported” column matches the Form 10-Q.
+The company caveat is retained verbatim only when the reranked evidence contains
+its exact source sentence. Ordinary Python code calculates the difference with
+decimal arithmetic.
+
+## Verified result
+
+The recorded component calls ran through the public SIE server on an NVIDIA L4.
+The restated passage ranked first for the question about the restated Q3 FY2023
+figure.
+
+```text
+Net income attributable to parent   $45.096M -> $36.080M
+Change                               -$9.016M (-20.0%)
+Diluted EPS                          $1.68 -> $1.34
+Source status                        earlier affected filings should no longer be relied upon
+```
+
+Pathward's exact statement stays in the result: the accounting change does not
+impact net income over the life of the portfolio, but changes when elements of
+the programs are recognized.
+
+## Run it
+
+```bash
+cd examples/financial-filing-agent
+cp .env.example .env
+uv sync
+
+uv run review-filing --run-id local
+uv run eval-filing runs/local
+```
+
+Set `SIE_CLUSTER_URL` and `SIE_API_KEY` to use SIE Cloud. The default points to
+a local server at `http://localhost:8080`.
+
+## Evidence bundle
+
+Every run writes:
+
+```text
+runs/<run-id>/manifest.json       endpoint, model IDs, fixture hashes, latency
+runs/<run-id>/raw/parse.json      complete Docling response
+runs/<run-id>/raw/retrieve.json   embeddings and cosine ranking
+runs/<run-id>/raw/rerank.json     complete reranker response
+runs/<run-id>/raw/entities.json   combined entity spans
+runs/<run-id>/raw/gliner2-*.json raw GLiNER2 source-span responses
+runs/<run-id>/raw/mapped.json     validated record mapped from Docling table coordinates
+runs/<run-id>/parsed.md           parsed packet used downstream
+runs/<run-id>/review.json         source-versioned result and calculated delta
+runs/<run-id>/evaluation.json     deterministic checks
+```
+
+`verified-run/` contains the saved evidence used to design the example. It is
+not a latency benchmark. The first request included model provisioning.
+Its manifest preserves the upstream-style Docling name configured during
+acquisition and the canonical `docling` ID that SIE actually served. New runs
+use the canonical ID directly.
+
+See [fixtures/SOURCES.md](fixtures/SOURCES.md) for accession numbers, URLs, and
+source checksums.

examples/financial-filing-agent/config.yaml

@@ -0,0 +1,16 @@
+cluster:
+  url: "http://localhost:8080"
+  api_key: ""
+  provision_timeout_s: 900
+
+models:
+  parse: "docling"
+  retrieve: "BAAI/bge-m3"
+  rerank: "Qwen/Qwen3-Reranker-4B"
+  entities: "urchade/gliner_multi-v2.1"
+  extract: "fastino/gliner2-large-v1"
+
+review:
+  query: "Compare Pathward's original and restated Q3 FY2023 net income and diluted EPS, then identify the superseded source and company caveat."
+  candidate_chunks: 8
+  top_k: 8

examples/financial-filing-agent/financial_filing/__init__.py

@@ -0,0 +1 @@
+"""Source-versioned financial filing review."""

examples/financial-filing-agent/financial_filing/evaluate.py

@@ -0,0 +1,102 @@
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+from dataclasses import asdict, dataclass
+from pathlib import Path
+from typing import Any
+
+from rich.console import Console
+from rich.table import Table
+
+console = Console()
+
+
+@dataclass(frozen=True)
+class Check:
+    name: str
+    passed: bool
+    detail: str
+
+
+def evaluate_review(review: dict[str, Any]) -> list[Check]:
+    change = review.get("change", {})
+    excluded = set(review.get("claims_excluded", []))
+    evidence = review.get("ranked_evidence", [])
+    evidence_text = "\n".join(str(row.get("text", "")) for row in evidence).casefold()
+    return [
+        Check("route", review.get("route") == "superseded_figure", str(review.get("route"))),
+        Check("net-income-delta", abs(float(change.get("value_millions", 0)) + 9.016) < 0.0001, str(change)),
+        Check("eps-delta", abs(float(change.get("diluted_eps", 0)) + 0.34) < 0.0001, str(change)),
+        Check("percent-delta", abs(float(change.get("percent", 0)) + 20.0) < 0.05, str(change)),
+        Check(
+            "source-lineage", len(review.get("controlling_sources", [])) == 3, str(review.get("controlling_sources"))
+        ),
+        Check(
+            "caveat-preserved",
+            "over the life of the portfolio" in review.get("company_caveat", ""),
+            review.get("company_caveat", ""),
+        ),
+        Check(
+            "ranked-source-evidence",
+            bool(evidence)
+            and all(
+                token in evidence_text
+                for token in (
+                    "pathward financial",
+                    "$45,096",
+                    "$36,080",
+                    "$1.68",
+                    "$1.34",
+                    "no longer",
+                    "over the life of the portfolio",
+                )
+            ),
+            str([row.get("chunk_id") for row in evidence]),
+        ),
+        Check(
+            "unsafe-claims-excluded",
+            {"fraud", "misconduct", "investment recommendation"} <= excluded,
+            ", ".join(sorted(excluded)),
+        ),
+    ]
+
+
+def evaluate_run(run_dir: Path) -> bool:
+    review = json.loads((run_dir / "review.json").read_text(encoding="utf-8"))
+    checks = evaluate_review(review)
+    passed = all(check.passed for check in checks)
+    evaluation_path = run_dir / "evaluation.json"
+    evaluation_path.write_text(
+        json.dumps({"passed": passed, "checks": [asdict(check) for check in checks]}, indent=2) + "\n",
+        encoding="utf-8",
+    )
+    manifest_path = run_dir / "manifest.json"
+    if manifest_path.exists():
+        manifest = json.loads(manifest_path.read_text(encoding="utf-8"))
+        artifacts = [
+            artifact for artifact in manifest.get("artifacts", []) if artifact.get("path") != "evaluation.json"
+        ]
+        artifacts.append(
+            {"path": "evaluation.json", "sha256": hashlib.sha256(evaluation_path.read_bytes()).hexdigest()}
+        )
+        manifest["artifacts"] = artifacts
+        manifest_path.write_text(json.dumps(manifest, indent=2) + "\n", encoding="utf-8")
+    table = Table("Check", "Result", "Detail")
+    for check in checks:
+        table.add_row(check.name, "[green]pass[/]" if check.passed else "[red]fail[/]", check.detail)
+    console.print(table)
+    return passed
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Evaluate a saved filing review")
+    parser.add_argument("run_dir", type=Path)
+    args = parser.parse_args()
+    if not evaluate_run(args.run_dir):
+        raise SystemExit(1)
+
+
+if __name__ == "__main__":
+    main()