HR, Finance, and Engineering ask "how many employees do we have?" and get three different numbers — from the exact same underlying data. Workforce Astra encodes each definition once, as governed Snowflake Semantic Views across five domains, so every team gets the identical, SQL-transparent answer.
Ask "how many employees do we have?" two ways. Same 150-row dataset, same moment in time — the numbers genuinely disagree, because one query is governed and one isn't.
Six illustrative queries, written the way the Semantic View answers them. Click one — the
answer and its SQL are computed in your browser from the same 150 rows Snowflake holds. Inside
Snowflake these same questions resolve through employee_360 with the SQL visible.
Comp-ratio below 0.85 and a rating of 4 or above — underpaid high performers, the real retention risk. Click a name for the cited review evidence, then preview the payload the mock MCP server in this repo would draft. Nothing is sent.
Finance and HR count "management overhead" differently. org_health_360 fixes the
definitions of span of control and overspan once, from the reporting lines already in the data.
Reorgs are previewed through a guarded simulate_reorg procedure that rejects self-moves, circular reporting lines and overspan breaches. It never changes the org.
The board and legal ask different questions about the same pay data. pay_equity_360
answers both, with each definition fixed once.
Stay and exit interview transcripts are scored inside Snowflake by a Snowpark Python
procedure (SCORE_VOICE_TRANSCRIPTS) using Cortex sentiment and multi-label reason
classification against a fixed taxonomy — not keyword matching.
Negative interviews from people the structured metric already flags as flight risks (underpaid, rated 4+). Expand for the transcript and the review behind it.
Disclosure: 5 of the 18 synthetic interviews were deliberately placed on employees who already meet the flight-risk rule, so the cross-signal is demonstrable. The sentiment scores and reason labels are Cortex's own output.
Bands get set once and drift. band_health_360 audits whether people are actually
paid inside their own published ranges — and whether the ranges separate the levels at all.
0% means the ranges are disjoint; 100% means identical. At 71%, pay alone can't tell an M1 manager from an IC5.
Guardrail: the auditor and its audit_comp_bands MCP tool report findings only. There is no proposed-pay column and no way to ask for one — a human sets the band.
Every governed metric has an owner, a certified definition and a golden-value test. The suite runs on a schedule, so a changed definition — or a changed Cortex model — fails loudly instead of silently moving a number in a board pack.
The suite passes with HIGH alerts open: those aren't test failures, they're the schedulers correctly escalating a real pay-equity breach and people paid below their band minimum.
employee_360, org_health_360, pay_equity_360, employee_voice_360, band_health_360. Every governed question resolves through a view, never a raw table, and each answers in Cortex Analyst.
SCORE_VOICE_TRANSCRIPTS runs inside Snowflake, calling Cortex sentiment and AI_CLASSIFY on 18 transcripts and writing governed results.
$workforce-astra-data-gen generates, verifies and loads the data; $workforce-astra-governance-check runs the regression suite and reports tasks and alerts.
Six tools — manager alert, comp-adjustment draft, promotion nomination, requisition, guarded reorg simulator, and a report-only band auditor. dry_run defaults to true; nothing is sent.
Live service, indexing and serving both active over 127 review notes (arctic-embed-m-v1.5), returning the cited review text behind a flight-risk call.
Three Snowflake Tasks: a daily 25-test regression run, a weekly pay-equity drift check, and a quarterly band review. Streamlit-in-Snowflake portal on top.