Make the pilot auditable before the claims get bigger.
AI grading pilots need evidence before claims scale. Planisphere tests tutor and grader workflows with synthetic tasks, measures cohort behavior, keeps student records out of the audit, and emits signed records for Title VI, FERPA, and compliance review.
You can’t govern what you can’t measure.
A hosted grader is a black box. It scores, it scaffolds, it corrects — and every one of those outputs runs through rules that put the responsibility on the district, not the vendor. Title VI demands the grader not disadvantage students across protected classes. §504 and IDEA demand it not penalize an accommodated or IEP-tuned response. FERPA demands education records stay protected, and COPPA demands no data is collected from a child. The catch: a grader that scaffolds differently for one cohort is a Title VI question hiding inside an answer key.
None of those questions can be answered by a vendor’s fairness whitepaper. They require a measurement of how the grader actually behaves — taken on your fleet against a synthetic probe set, repeatable by an adversary, and produced without a single student record ever entering the audit surface.
What the School Board measures.
The School Board points the Astrolabe at your fleet of grader and tutor tools and grades each one on three axes. The same measurement saves money and survives a review.
The smaller bill. Districts rarely run one AI tool — they run several, often with overlapping mandates bought by different schools or departments. When the School Board clusters them by behavior, the tools that produce the same scores collapse into the same group. Every duplicate is a license you can drop without losing a capability. The measurement names which ones.
The defensible civil-rights file. The tools that survive get graded against the doctrine — and every grade is signed. When the grader scaffolds differently for one cohort, the audit names the cluster it collapses and flags the Title VI / §504 / IDEA question. When OCR, an accreditor, or a parent asks how you knew the grader was fair, you have a record, not an assurance.
Proof for every output.
A grader scores an output. The district may later need to show what happened, how the tool behaved, whether the result was consistent across cohorts, and whether student records stayed protected. Planisphere tests the grader with synthetic tasks, produces a signed record of the behavior, and lets that record cross the boundary without carrying a student's work. When a parent, a board member, or the Office for Civil Rights asks how the district supervised its AI graders, this is the record you hand them.
The doctrine it maps to.
The education version uses the same Planisphere measurement engine as the rest of the system. It shapes the evidence around K-5 duties: runtime contracts, civil-rights anchors, and reproducible record fields a third party can check byte-for-byte.
The audit happens in code. audit_k5(responses, grade_band=…) asserts FERPA and COPPA before it returns, scores every pedagogical axis, runs the Title VI / §504 / IDEA civil-rights report over the cohort-divergent outliers, and emits the record through the Port — the same boundary the Law Board, the War Board, and the Med Board ship through. Want the full corpus? The complete pin-and-state reference lives at the Education corpus surface.
Next step.
The first engagement is a scoping conversation. Tool inventory, hosting topology, the grade band (k-2 or 3-5), the doctrine anchor (Title VI / §504 / IDEA / FERPA / COPPA / curriculum-specific), and the target review date. We return a Phase 0 scope and an evidence-record delivery date. No student record changes hands — none ever enters the audit.