Build a review workflow for synthetic private-equity distribution data. Start with deterministic reconciliation checks, expose the evidence to a Cortex Agent, and optionally run a managed evaluation. Inspect the trace before claiming that the agent selected a skill or invoked a tool.

This tutorial covers one reconciliation workflow, not a complete contractual waterfall, investment recommendation, or production financial system. All amounts are synthetic. Balanced totals do not establish recipient entitlements.

The companion project contains the notebooks, SQL, Python, Rust, tests, and project-local Cortex Code skills under this guide's assets/pe-agentic-system/ directory. Non-image assets are available from the Git repository, not from the published guide website.

Clone the Quickstarts repository and enter the companion project directory:

The verifier should report 35 artifact hashes and the skill hash verified. A matching hash proves file integrity relative to the manifest, not correctness or authorization to execute the file.

Review these files before continuing:

Do not use Run All. The notebook contains multiple exploratory tracks outside the tutorial path. Every approval flag starts as False. Any cell that creates, replaces, deletes, or bills raises an error until you set its APPROVE_* flag to True in a separate cell. Use only the reconciliation block identified below, in a fresh kernel. Review each cell before executing it. Other cells are not prerequisites.

In the companion project directory, replace the five placeholder values in the following script with your approved targets and run it locally. Use simple uppercase, unquoted identifiers for database, schema, role, and warehouse. The schema must be new for fresh setup; the database should already exist and be approved.

Review the resulting changes, then refresh the integrity manifest:

Do not commit your account-specific bundle back to the public repository. Refresh hashes only after reviewing the edits; do not use it to bypass an unexplained hash mismatch.

Upload finview_verification_demo.ipynb and the configured generated/ directory to the same folder in your Snowflake Workspace, preserving the bundle's nested directories. Open the notebook with the required packages and a fresh kernel.

Locate the reconciliation configuration cell containing both BUNDLE_DIR and SETUP_MODE (zero-based cell 28 in the companion version prepared for this guide). The tutorial runs cells 28-38 only. Identify cells by their content as well as their position; cell numbers can change when you edit a notebook.

Before running that configuration cell:

1. Change BUNDLE_DIR to Path('generated').
2. Set EXPECTED_ACCOUNT, DATABASE, SCHEMA, ROLE, and WAREHOUSE to the same approved values used in the manifest.
3. Set SETUP_MODE = 'fresh' for a new demo schema. The existing default, reuse_finview, expects previously created fixtures and is not a first-run mode.
4. Keep APPROVE_SETUP, APPROVE_SKILL_UPLOAD_AND_AGENT, APPROVE_ONE_BILLABLE_EVALUATION, and APPROVE_CANCEL set to False.

Run only this configuration cell. It verifies bundle hashes and account identity, then explicitly selects the configured role and warehouse. Confirm the displayed context before proceeding. A connection alias is not an account identity.

Run the next cell, beginning with execute_file('00_preflight.sql'). It displays the setup SQL, support SQL, agent and evaluation definitions, and runtime skill. Confirm the database is STANDARD and the proposed fresh schema does not exist. Ask your administrator to verify required privileges, including evaluation task privileges and Cortex model access if you will use the optional evaluation.

The approval flags start as False. The first screenshot shows the gates before approval, the second after you set them in a separate cell:

After reviewing the exact statements, set APPROVE_SETUP = True in a separate cell and run the setup cell beginning with if not APPROVE_SETUP: once. The gate resets to False before execution. Do not rerun the configuration cell to approve setup; doing so resets the gates.

Fresh setup can create a database if it is absent. Use an approved existing database and verify its name first. Setup is not transactional. If any statement fails, inspect the partially created state; do not rerun blindly or execute the notebook's separate deletion/redeployment cells.

Run the cell that calls execute_file('02_checks.sql') and checks expected_status. It verifies the five synthetic case outcomes:

The direct DMF call must return 4 intentionally failed cases. This is expected for these fixtures; an acceptance rule requiring zero failures would reject them. Do not reinterpret missing allocations as zero. The shortfall case represents a synthetic $63 million discrepancy, not a rerun of a production financial algorithm.

The notebook also runs input checks on valuation assumptions and a Monte Carlo sensitivity on WACC. Both use synthetic inputs:

The following plotting cell is optional. Stop before agent creation if the fixture status or DMF assertions fail. These direct queries do not establish that an agent or skill executed.

Review the agent definition and runtime skill. The instruction-only reconciliation skill uses the review_reconciliation tool.

Set APPROVE_SKILL_UPLOAD_AND_AGENT = True in a separate cell. Run the cell starting with if not APPROVE_SKILL_UPLOAD_AND_AGENT: once. It resets the gate, uploads the skill and evaluation configuration, verifies downloaded stage bytes, and executes the agent-creation SQL. This creates configuration; it does not prove that the agent invoked a skill.

An example answer, in which the agent blocks the shortfall case and cites the reconciliation evidence:

If runtime skills or the definition are unsupported, stop rather than silently omitting the skill. Hash-addressed stage paths aid reproducibility but do not enforce immutability.

Skip this step if you do not have approval for billable execution. Review the evaluation configuration and confirm the created agent version, judge availability, and six-row input dataset. The prompts include the five cases and an override request to approve the known shortfall.

Set APPROVE_ONE_BILLABLE_EVALUATION = True in a separate cell, then run the cell starting with if not APPROVE_ONE_BILLABLE_EVALUATION: once. It resets the gate and records that a start was attempted. A failed start can still create evaluation state; inspect status before considering a retry.

The next status/results cell may be rerun to check progress. Do not rerun START to poll for results. Review failed records rather than tuning and repeating the evaluation solely to obtain a favorable screenshot.

Screenshots illustrate prior runs; your account, model availability, and outputs may differ. A small synthetic evaluation is a smoke test, not evidence of production accuracy or calibrated financial risk.

Select a real RECORD_ID from the evaluation output and set it in the next trace cell. Inspect the trace for skill selection/loading and the reconciliation tool invocation with returned evidence. Report uncertainty if the trace does not establish that the skill ran; agent configuration and aggregate scores are not substitutes for execution evidence.

You can also open the agent's Evaluations tab in Snowsight, select the run and record, and inspect Trace details.

Run the evidence-export cell only if you intend to retain the results locally. Review exported rows and query identifiers before sharing them. If an evaluation is still active and you need to cancel it, review the final cancellation cell and explicitly enable APPROVE_CANCEL. Cancellation does not delete demo objects. This walkthrough does not attach scheduled DMFs. Obtain separate approval for object cleanup; do not run the notebook's unrelated deletion cells.

The repository also includes exploratory MCP, retrieval/judge, financial modeling, and Rust examples. They are not required for this tutorial and are not covered by the reconciliation block's approvals or validation. Each track has its own APPROVE_* flags. Review their requirements, feature availability, and side effects independently. The screenshots below come from earlier runs.

The notebook can expose the reconciliation as a managed MCP tool, call it for each synthetic case, and export every request and response as evidence. A TruLens run then scores those calls with a Cortex judge model.

A separate track builds semantic views over SEC filing data and two analyst agents for diligence and capital-markets disclosure. Run its preflight cell before creating anything.

A standalone synthetic track reranks passages with ColBERT MaxSim, fuses rankings with reciprocal rank fusion, and judges answers with AI_COMPLETE and the snowflake-decision model. The model is in public preview. The judge scores two synthetic cases with a Brier score. Treat it as a smoke test, not a calibration result.

Do not automatically execute model-generated code or treat illustrative DCF/LBO examples as validated financial models.

You prepared a reviewed execution bundle, checked deliberately flawed synthetic financial data, and configured a reconciliation agent. If you completed the optional evaluation, you also inspected its outputs and trace evidence. Validate financial conclusions independently; this reference implementation is not investment advice.