Algorithmic Policy & Governance · Capstone brief

POL-08 · Shadow vs. production

Argue when a model should stay in shadow mode; define exit criteria with metrics the school could actually measure.

The question

Argue when a model should stay in shadow mode; define exit criteria with metrics the school could actually measure.

Policy / materials

One proposed use case. Metric definitions. Kill criteria as well as promote criteria.

Expected failure modes

Only upside metrics. No owner for the shadow review. Eternal shadow as avoidance.

Done looks like

Shadow protocol; promote/kill criteria; review cadence; owner role.

Five C's

CT: evidence before scale. CR: measurable exits. CO: peer red-teams metrics. CM: protocol page. CZ: people affected if promoted early.

Mentor role

Teacher or tech coach reviews measurability. School-supervised.

Rubric calibration

R1: one use case. R2: metrics defined. R3: shadow vs prod compared. R4: kill criteria. R5: one-pager. R6: refuses vanity metrics.

Two ways this goes wrong

(a) Promote on vibes. (b) Shadow forever with no review.

Credit lane fit

Lane A immediately. No verified credit claim.