SQL Change Guard ties scripts and production data requests bound for SQL Server, PostgreSQL and Oracle to approval and controlled execution. For every request it produces an evidence file that shows who asked, which rule applied, who approved and what ran. Where a signing key is configured, that file verifies without connecting to the product. It runs inside your own network.
Risk band: the worst finding decides
Non skippable rule: the approval step is locked
The ticket number is checked in your ticketing system
The script fingerprint goes on the record
Each tool knows one link of the chain; the link at the moment of execution is in none of their remits.
Knows what was asked for and who signed it off. It does not know the final text of the script that will run in production.
Knows the change it carried. It does not see work that goes outside the pipeline, such as an emergency fix at night.
Sees what ran. It does not know which approval and which justification it rests on.
When an auditor asks who approved this change and whether the text that ran was the text that was approved, the answer is pieced together by matching records across several systems by hand. Reading production data, in most organisations, is not tied to any request at all.
| Tool | Its job | What SQL Change Guard adds |
|---|---|---|
| Jira, ServiceNow | Holds the request and the business approval | Checks the ticket number in the ticketing system at save time and again at execution, and writes events back as notes. The connection is optional. |
| CI/CD, Liquibase, Flyway | Versions and ships schema change | Policy, approval and evidence for scripts that go outside the pipeline and for production data requests. There is no direct connection to delivery pipelines today. |
| DAM and SIEM | Monitors database activity and collects events | Links the work that ran to its approval and justification, and sends audit events to your SIEM over syslog, CEF, LEEF or JSON. It does not listen to live traffic. |
| Database tools | Direct access (SSMS, DBeaver, SQL Developer) | It does not close these paths. Control applies to work that goes through the product; your organisation narrows direct access with its own permission model. |
The script is parsed with the grammar of its database type and assessed against the rule catalog. The risk band equals the highest finding triggered. Approval steps follow from server, environment and band; a step tied to a rule marked non skippable cannot be passed by seniority. The rule set in force that day is frozen onto the request.
How a change request works →The fingerprint of the approved content goes on the record and is compared again before execution; if the content changed afterwards, execution stops and that event is recorded. Outcome, duration, errors and affected row counts are recorded. A run whose commit on the target is uncertain is not quietly counted as a success; it goes to reconciliation.
The three database engines →An evidence file (PDF and JSON) is produced per request: who asked, which rule applied, who approved, what ran and which exceptions occurred. Where a signing key is configured the package is signed, and it verifies with the script inside it without connecting to the product. The audit record is kept append only in a SQL Server ledger table; signed ledger digests can be stored outside the database.
Sample evidence file (PDF) →A data request opens with its justification. Detected sensitive columns are masked, and if detection runs out of time the remaining columns are masked too; a column requested unmasked needs a justification and an extra approval. The result is not shown on screen; it is delivered to the approved recipient in a password protected encrypted package, and who downloaded it is recorded.
How a data request works →Let us state the limit plainly. The product governs and evidences the work that goes through it. It does not give absolute protection against someone with full administrator access to the application database; the aim is to make unauthorised change harder and to make it detectable from independent points. How evidence is verified
The package below is the evidence file of a change request run end to end on a test installation: one requester, two approvers and a separate executor. Verifying it needs no connection to the product or to the internet.
Download the zip package and extract it to a folder. It holds the evidence file (PDF and JSON), a signed content list, the public key and a verification script.
Run Verify.ps1 with PowerShell in that folder. It checks the file digests and the package signature, and shows one by one that the save, approval and execution events are signed and all carry the same content fingerprint.
The script prints the fingerprint of the key that signed the package. This sample was signed with a trial key; its fingerprint is: c61b64ac6953715f21880178a04b8479ac8d8eb2cdd402b3efde6479678363e9. In your own installation the package is signed with your organisation's key, and the organisation publishes the fingerprint separately from the package.
What it shows and what it does not. The package shows that the record is unchanged from the way the product wrote it, and that the approved content is the content that ran. It does not show the approver's intent, or work done on the target database outside the product. Because the script ships inside the package it is not proof on its own; the same checks can be done with openssl.
The web application runs on IIS and the background service as a Windows service. Records stay in your own SQL Server database. It is not a cloud service; the licence is verified offline.
Target databases are reached from the application server through standard drivers. Nothing is installed on the target server; the permission needed is an account limited to the work the product will do there.
With corporate sign in (OpenID Connect, SAML) or LDAP the password is not stored by the product; local user passwords are hashed with BCrypt. Target connection passwords are stored encrypted with AES-256-GCM. The keys live in your configuration.
Only connections you switch on, to addresses you provide: email, ticketing, identity provider, SIEM. If timestamping is on, only a file digest goes out; one setting turns it off.
One server, a small team, inside your own network. At the end of the pilot you hold evidence files produced from your own environment, and you decide on those files rather than on a slide deck.