For technical teams
Administrator Guide
For the governance engineers who maintain Pramana's policy pack and the data stewards who work its queue. It assumes Pramana is installed and has run at least once; Install and Setup covers that.
A week with Pramana#
Pramana runs once a night. Everything else is people responding to what it found.
| When | Who | What |
|---|---|---|
| Every night | Pramana | Scans, proposes classifications, evaluates the controls, compares with the night before, writes the evidence packs |
| Most mornings | A data steward | Works the review queue; looks at any weakening drift and acknowledges it |
| When findings appear | A governance or platform engineer | Generates the statements that close them and puts them through change control |
| When the organisation changes its mind | An administrator | Updates the policy pack: a new category, a new binding, a new database in scope, an exemption |
| When asked | Audit, risk, compliance | Read the consoles and take evidence packs away |
The consoles#
Three Streamlit consoles, one per role. Inside an application a console cannot see who is looking at it, so which console a role may open is what decides what it can do.
| Console | Open to | Pages |
|---|---|---|
| PRAMANA_VIEWER (the default) | every role | Overview, Controls, Review queue, Drift, Evidence, Policy pack. All read-only |
| PRAMANA_REVIEW | stewards, admins | The same pages, plus: approve and reject proposals, acknowledge drift, add exemptions |
| PRAMANA_CONSOLE | admins | All of that, plus Setup, Remediation, Run and Sample |
| Page | What it is for |
|---|---|
| Overview | Overall status, failing controls, queue size, open weakening drift, coverage per database |
| Controls | Findings filtered by control and status; every sensitive column with the policy it should have and the one it has; add an exemption |
| Review queue | Approve or reject one proposal, or every proposal at or above a confidence |
| Drift | Open events, weakening first; acknowledge with a note |
| Evidence | Every pack: readable, as JSON, and the SQL to take it away |
| Policy pack | The ten tables as they stand, and every table with the governed tags it carries; validate |
| Setup | What Pramana may read, the scope, the classification tag, the governed tags, the frameworks |
| Remediation | Generate the statements, or the same as intents for Dhruva; the latest batch as one script |
| Run | Run any step; start and stop the schedule; the run log |
| Sample | Load, break, restore and clear the sample estate |
The sidebar of a console says which one it is: the edition, the schema it reads and a chip reading ADMIN, STEWARD or READ-ONLY. In every table a status is tinted: red for what fails or is missing, amber for what warns or waits, green for what passes, grey for what is exempt or was skipped. The consoles follow Snowsight's light or dark mode. Their fonts are loaded by your browser from fonts.gstatic.com; where that host is blocked they use the system font and nothing else changes. A console fetches nothing else from outside Snowflake.
The policy pack#
Your governance intent as data: ten tables in schema GOV. A fresh install has a taxonomy of nine categories, the thirteen controls, and the clause maps for UK GDPR, DORA and ISO 27001. Scope, bindings, the values of your governed tags and exemptions start empty.
| Table | What you say in it |
|---|---|
GOV_SETTING | The classification tag, the schedule, the frameworks, thresholds |
GOV_SCOPE | Which databases Pramana examines, and which schemas and tables it leaves out |
GOV_CATEGORY | Your taxonomy: the kinds of sensitive data, and how each is recognised |
GOV_BINDING | Which masking policy and row access policy protect each category |
GOV_TAG, GOV_TAG_VALUE | The table-level tags you govern by, such as retention, and the values each may carry |
GOV_CONTROL | Which controls are on, and whether each fails or warns |
GOV_FRAMEWORK | Which control is evidence for which clause |
GOV_EXEMPTION | The findings you have accepted, with reason, approver and expiry |
GOV_EVIDENCE_SOURCE | Other evidence to carry in each pack |
Every column of every table is described in the Reference.
Two ways to maintain it#
As rows. A PRAMANA_ADMIN holder has INSERT, UPDATE and DELETE on the ten tables:
INSERT INTO PRAMANA.GOV.GOV_BINDING (CATEGORY, MASKING_POLICY)
VALUES ('PII_EMAIL', 'GOVERNANCE.POLICIES.MASK_EMAIL');
INSERT INTO PRAMANA.GOV.GOV_BINDING (CATEGORY, MASKING_POLICY, ROW_ACCESS_POLICY, RAP_COLUMN)
VALUES ('HEALTH', 'GOVERNANCE.POLICIES.MASK_HEALTH', 'GOVERNANCE.POLICIES.RAP_REGION', 'REGION');
UPDATE PRAMANA.GOV.GOV_CONTROL SET SEVERITY = 'WARN' WHERE CONTROL_ID = 'ACC-02';
As CSV files in git. One file per table, named after it in lower case, headers equal to the column names. Download the template pack from the documentation page, edit it, and load it:
PUT file://gov_*.csv @PRAMANA.GOV.GOV_CONFIG_STAGE AUTO_COMPRESS = FALSE OVERWRITE = TRUE;
CALL PRAMANA.CORE.GOV_LOAD_CONFIG();
CALL PRAMANA.CORE.GOV_RUN('VALIDATE', NULL);
PUT sends files from your own machine, so it runs from the Snowflake CLI or SnowSQL and not from a Snowsight worksheet. To stay in a worksheet, upload the files to a stage of your own (Snowsight can do that from the stage's page), copy them across and load:
-- from a stage of yours that holds the gov_*.csv files, into the application's stage
COPY FILES INTO @PRAMANA.GOV.GOV_CONFIG_STAGE FROM @<your database>.<your schema>.<your stage>/pack/;
LIST @PRAMANA.GOV.GOV_CONFIG_STAGE;
CALL PRAMANA.CORE.GOV_LOAD_CONFIG();
CALL PRAMANA.CORE.GOV_RUN('VALIDATE', NULL);
The role that runs it needs PRAMANA_ADMIN and read access to your stage. LIST shows what arrived, and GOV_LOAD_CONFIG answers with the rows it read from each file.
Settings and controls are merged, so a file may carry only the keys you change. The other seven tables are replaced by their file. A file that is absent, or has no rows, leaves its table as it is.
If you keep the pack in git, treat the files as the master copy. An exemption added in a console exists only in the application until you add it to gov_exemption.csv as well; the next load of a file that has rows replaces the table with that file.
Tuning what counts as sensitive#
A column reaches the review queue in one of two ways: Snowflake's classifier examined its table and assigned a category your taxonomy maps, or the column's name matches a category's NAME_REGEX.
- Too many proposals. Tighten the name rule of the category that produces them, or raise
CLASSIFY_MIN_CONFIDENCEso that low-confidence classifier results stop counting. A single column that is not what its name suggests is better rejected in the queue than excluded by a rule. - Too few. Tables the classifier has never examined produce proposals only from name rules. Run Snowflake's sensitive data classification on them and scan again. CLS-04 lists the sensitive tables with no recent classifier result.
- A new kind of data. Add a row to
GOV_CATEGORY, bind its policies inGOV_BINDING, and runVALIDATE.
Bindings#
A binding says which policy protects a category. A category with no binding is still governed, in that the classification controls apply to it, but no policy is required for it. VALIDATE warns about each unbound category, so that leaving one unprotected is a decision and not an oversight.
Working the review queue#
Each proposal is a column that looks like a category and is not labelled with it. Pramana applies nothing; a steward decides. In PRAMANA_REVIEW → Review queue, or:
CALL PRAMANA.CORE.GOV_DECIDE('<proposal id>', 'APPROVE', 'confirmed with the data owner');
CALL PRAMANA.CORE.GOV_DECIDE('<proposal id>', 'REJECT', 'a merchant, not a person');
CALL PRAMANA.CORE.GOV_DECIDE_ALL('APPROVE', 'HIGH');
A rejection is remembered: Pramana treats that column as exempt from then on, and the table stops counting as sensitive because of it. An approval turns into a SET TAG statement the next time remediation is generated. Once a scan sees that tag on the column, Pramana closes the proposal (its status becomes APPLIED) and stops writing the statement.
Exemptions#
An exemption accepts one finding, for one object and one control, until a date. Add it from the Controls page of PRAMANA_REVIEW, or:
CALL PRAMANA.CORE.GOV_ADD_EXEMPTION(
'FINANCE_DB.CORE.CUSTOMER.EMAIL_HASH', 'CLS-01',
'One-way hash of the address. Reviewed by the DPO.', '2027-06-30'::DATE);
-- withdraw it: the next evaluation raises the finding again
CALL PRAMANA.CORE.GOV_CLOSE_EXEMPTION('FINANCE_DB.CORE.CUSTOMER.EMAIL_HASH', 'CLS-01');
The user who makes the call is recorded as the approver. A target can be a column or a whole table; a table-level exemption covers its columns. Every exemption in force is printed in every evidence pack, so write the reason for the person who will read it there.
Closing findings#
Every approved proposal and every fixable failure becomes an ALTER TABLE … SET TAG, SET MASKING POLICY, ADD ROW ACCESS POLICY or REVOKE statement. Pramana records them; you run them through your change process.
CALL PRAMANA.CORE.GOV_RUN('REMEDIATE', NULL);
SELECT STEP, STATUS, CONTROL_ID, TARGET, ACTION, STATEMENT FROM PRAMANA.CORE.GOV_REMEDIATION_OPEN_V ORDER BY STEP;
Each row has one of three statuses, and STEP puts them in the order a change would take them:
| Status | What it is | What you do |
|---|---|---|
PROPOSED | A statement that can run as written | Put it through your change process |
TEMPLATE | A CREATE MASKING POLICY or CREATE ROW ACCESS POLICY for a policy your pack binds and the account does not have. The name is right, the signature comes from the columns waiting for it, and the body is a placeholder that does not run as written | Complete the body: who may see clear values, and what everyone else sees. That rule is yours to decide, so Pramana does not guess it |
BLOCKED | The statement that attaches that policy to a column or table. It is correct, and it cannot run until the policy exists | Run it after the completed template |
Who runs them. Not Pramana, and not just any role. A statement changes a table, so it needs a role that may: the owner of the table, or a role that holds the account-level privileges for the job (APPLY TAG, APPLY MASKING POLICY, APPLY ROW ACCESS POLICY, and MANAGE GRANTS for a REVOKE), with USAGE on the database and schema. Run by a role with no access to the database, a statement fails with "Database … does not exist or not authorized". That is Snowflake refusing the role, not a fault in the statement.
When a statement leaves the list. Pramana runs nothing, so it learns that you have run a statement only when a scan shows it the result. A batch is written from the latest evaluation, and an evaluation goes by the latest scan. A statement you ran an hour ago is therefore still in the next batch if no scan has seen the change, and ACCOUNT_USAGE runs up to two or three hours behind your account. Once it has caught up, scan and evaluate again (the nightly run does both) and generate the statements once more. What has been done is no longer listed, and an approved proposal whose tag is now on the column is closed. When nothing is left the batch is empty, GOV_REMEDIATION_OPEN_V returns no rows and the Remediation page says so. Generate after an evaluation, not straight after a scan: a snapshot that has not been evaluated has no findings to write statements from, and the run tells you.
While the built-in sample is loaded, the statements are for the sample's fictional databases. They show what remediation looks like and cannot be run: there is no such database in your account.
If the columns waiting for one policy have more than one data type, the template says so: a masking policy takes one type, so the category needs a policy per type or a narrower binding.
What stays with a person: the body of a policy that does not exist yet (POL-04), a tag value the taxonomy does not know (CLS-02), which value of a governed tag a table should carry (RET-01, RET-02 and the controls of your own tags), a table owned by the wrong role (ACC-02), and the judgement calls (CLS-03, POL-03).
Drift#
From the second run on, each snapshot is compared with the one before it. Every difference is recorded as an event; the ones that weaken your position are flagged and stay on the Drift page until someone acknowledges them.
SELECT DRIFT_ID, EVENT_TYPE, TARGET, BEFORE_VALUE, AFTER_VALUE, WEAKENING FROM PRAMANA.CORE.GOV_DRIFT_OPEN_V;
CALL PRAMANA.CORE.GOV_ACKNOWLEDGE_DRIFT('<drift id>', 'Change CHG-1042, approved by the data owner');
Acknowledging records who looked and what they concluded. It does not undo the change, and the event stays in every evidence pack that covers that night. The Reference lists each event and when it counts as weakening.
Controls and frameworks#
The control catalogue is Pramana's; how strictly each control is graded is yours.
-- report a control as a warning, not a failure
UPDATE PRAMANA.GOV.GOV_CONTROL SET SEVERITY = 'WARN' WHERE CONTROL_ID = 'ACC-02';
-- switch a control off
UPDATE PRAMANA.GOV.GOV_CONTROL SET ENABLED = FALSE WHERE CONTROL_ID = 'CLS-04';
Evidence packs are built for the frameworks named in the FRAMEWORKS setting. UK GDPR and DORA are on from the start; ISO 27001 is mapped and waiting:
UPDATE PRAMANA.GOV.GOV_SETTING SET SETTING_VALUE = 'GDPR_UK,DORA,ISO27001' WHERE SETTING_KEY = 'FRAMEWORKS';
A framework of your own is a set of rows saying which control is evidence for which clause, and its name added to the same setting:
INSERT INTO PRAMANA.GOV.GOV_FRAMEWORK (FRAMEWORK, CONTROL_ID, CLAUSE, CLAUSE_TITLE) VALUES
('INTERNAL_POLICY', 'POL-01', 'DP-7', 'Personal data is masked outside production support'),
('INTERNAL_POLICY', 'ACC-01', 'DP-9', 'No general read access to personal data');
Recipes#
A new source system lands. Add its database to the scope, run, work the queue, generate the statements. It appears in the next pack with its coverage.
A retired tag value. CLS-02 lists every column carrying a value the taxonomy does not know. Either add the value as a category, or retag: the queue proposes the taxonomy's category for such columns.
Masking applied through the tag. Pramana checks what reaches each column, however it got there. Bind that policy to the category so Pramana knows it is the expected one.
A column that is not personal data despite its name. Reject the proposal (it will not come back), or add an exemption with the reason. The second shows in every pack.
Hand remediation to the platform team. GOV_REMEDIATION_OPEN_V is the script, in STEP order; the Remediation page shows the latest batch as one block to copy. Ready statements come first, so the block runs up to the first policy template and stops there until someone has completed it.
Prove a point in time.
SELECT PACK_MD, PACK_SHA256
FROM PRAMANA.GOV.GOV_EVIDENCE_PACK
WHERE GENERATED_AT::DATE = '2026-09-30' AND FRAMEWORK = 'GDPR_UK';
Carry Vigil's reconciliation results. If you also run Vigil, Pravidha's data quality and reconciliation application:
GRANT APPLICATION ROLE VIGIL.VIGIL_EVIDENCE_READER TO APPLICATION PRAMANA;
UPDATE PRAMANA.GOV.GOV_EVIDENCE_SOURCE SET ENABLED = TRUE WHERE SOURCE_NAME = 'VIGIL';
Pramana stands alone without Vigil; DQ-01 is then skipped.
Problems and requests: support@pravidha.com. General questions: info@pravidha.com.