A Policy-as-Code Investigation · Florida Private Schools

The law is
written in
plain text.
We made it
executable.

We gave an AI a single research brief — carefully engineered across 8 iterative prompts — and it returned the complete regulatory schema for Florida private schools: 72 requirements, 11 agencies, every validation rule expressed as boolean code ready for a compliance engine to run.

8
Prompts to get there
72
Regulations extracted
11
Gov. agencies
1
Afternoon
Nord Anglia Education · Florida North Broward Prep · Windermere Prep Gemini Deep Research

A school that doesn't know
what it doesn't know

A private international school in Florida is simultaneously subject to Florida state statutes, federal immigration law, labor regulations, IRS tax code, EPA asbestos rules, OSHA standards, FTC data privacy guidelines, and Department of Health inspection requirements — each enforced by a different agency, each on a different cycle. Nobody in the school has a complete picture. The regulations aren't secret. They're just scattered.

For Nord Anglia Education — which operates schools across 35 countries — the question was pointed: do we actually know everything we're required to comply with in Florida? And more pressingly: is there a way to verify compliance automatically, against actual school documents, rather than relying on institutional memory?

The answer required a prior step. Before any software could check compliance, every regulation had to be extracted from legal text and expressed as machine-readable logic. That extraction is what this story is about.

"The output format is as important as the question. Ask for a summary and you get prose. Ask for a Data Schema Table and you get a database."
The core insight · this project

The method was 8 prompts — a deliberate, iterative sequence run in a single Gemini Deep Research session. Each prompt had a specific purpose. The result was not a report. It was a schema: 72 rows of compliance logic, ready to be loaded into a software engine and run against school documents in real time.

72
Regulations — from legal text to boolean logic — in one session
The Method

Eight prompts.
A deliberate sequence.

Most AI research is a single query. This was a structured methodology — each prompt building on the last, each one serving a specific function in a pipeline designed to turn statute language into executable code.

1
Prompt 1 · The Meta-Move
Ask the AI to critique the prompt before running it
→ Gemini Deep Research
what would an expert change in this prompt that a beginner would miss? [ + original compliance research brief attached ]
Don't run your first draft. Feed it to the AI and ask what an expert would change. This meta-prompting technique surfaces blind spots in seconds that domain expertise would take days to identify. The AI returned 6 structural gaps in the original prompt — most critically: it was written from a public-school mindset. Private school law in Florida lives in Chapter 1002 F.S. Treating them identically produces hundreds of false-positive regulations.
6 structural gaps identified. Key finding: the original prompt conflated Florida's public school statute (Chapter 1003) with private school law (Chapter 1002). The AI flagged that this would return the wrong regulatory universe entirely.
2
Prompt 2 · Get the Better Prompt Written
"Give me the full prompt I can pass to Gemini Deep Research"
→ Gemini Deep Research
give me the full prompt that I can pass to Gemini Deep Research
After the AI described what to change, the next move was simple: ask it to write the improved prompt. Critique → implementation. Two steps, not one. The AI produced a completely restructured brief with a new role framing ("Senior Compliance Architect"), a strict scope (Chapter 1002 only, exclude Chapter 1003), and a Data Schema Table as the output format — replacing the original legal memo structure.
Expert-rewritten prompt returned. The output format changed from prose to a structured table with four columns: Artifact Name | Trigger Condition | Validation Logic (The Code) | Citation. This column structure is what makes the output machine-executable.
3
Prompt 3 · The Core Research Run
Run the expert-designed prompt — ~20 minutes, 24+ sources
→ Gemini Deep Research · ~20 min · 24+ primary sources
Role: Act as a Senior Compliance Architect and Legal Researcher. Context: "Policy as Code" engine for Nord Anglia (North Broward + Windermere). Problem: Most data conflates Public (Chapter 1003) vs Private (Chapter 1002) rules. Scope: State: Chapter 1002 ONLY — EXCLUDE Chapter 1003 Scholarship: Step Up For Students (FES-UA, FES-EO, FTC) via Provider Handbooks Federal: DHS/ICE SEVIS for F-1 visa students Accreditation: Where does FCIS or Cognia supersede state inspection? Output Format — Data Schema Table (NOT a legal memo): Artifact Name | Trigger Condition | Validation Logic | Citation // "Do not provide generic summaries." // Validation Logic must be written as boolean code, not prose. // Example: Tdap_Date > DOB + 10 years — not "must have Tdap vaccine"
The role framing ("Senior Compliance Architect") and the instruction to write Validation Logic as boolean code — not prose — are the two design decisions that separate this from a typical AI search. Without them, you get summaries. With them, you get logic gates. "Must be dated within 12 months" becomes "Exam_Date > (Entry_Date - 12 months)". That's the difference between a document you read and code you run.
28 regulations returned across 5 state categories. Each row includes an exact form number, a trigger condition, and boolean validation logic. State layer complete.
4
Prompt 4 · Scope Expansion
Add buildings and infrastructure
→ Gemini Deep Research
Are there any building and infrastructure requirements — can you add that too
Expanding scope within the same session preserves context. The AI already knew the school type, the output format, and the regulatory universe. A fresh session would reset all of that. This one built on it — adding 12 regulations that a state-only search would never surface, including federal AHERA asbestos rules which apply to all private schools regardless of funding.
Added 12 building & safety regulations — Certificate of Occupancy, Radon Testing (Form DH 1777, 5-year cycle, stops after 2 tests), AHERA Asbestos Plan, OSHA 300 Injury Log, Chemical Hygiene Plan for science labs, ADA barrier documentation.
5
Prompt 5 · The Hidden Layer
Unlock the federal overlay — six words, 32 regulations
→ Gemini Deep Research
while these are state level requirements are there any federal requirements?
Six words. An entirely new regulatory dimension. Federal rules for international private schools carry the highest-consequence requirements in the entire matrix — including criminal liability for IRS cash reporting failures and the permanent loss of F-1 visa issuance rights for SEVIS non-compliance. A state-only research frame would miss all of it.
Added 32 federal regulations across 7 agencies: DHS/ICE (Form I-20, I-9, SEVIS logs), DOL (FLSA, FMLA, PUMP Act 2023), FTC (COPPA, Vendor DPAs), IRS (Form 8300 cash reporting — federal felony if missed, BOI FinCEN 2024), Copyright Act (movie licensing), OSHA/AHERA overlays.
6
Prompt 6 · The Self-Audit
Feed the full list back — ask what's been missed
→ Gemini Deep Research
review these listed and let me know if anyone has been missed in your evaluation [ + entire compiled regulation list pasted in ]
The AI's initial research runs are forward-looking — scanning for things to include. This backward pass is a different cognitive task: checking a completed list for what's absent. Pasting the full compiled list back in and asking for a gap check reliably surfaces regulations that fell through the cracks of the category-by-category approach.
4 gaps found. Added: Nursing Mother's Break Log (PUMP Act 2023 — newly enacted, widely missed), SEVIS Status Update Log with the specific 21-day reporting window for dropout/address changes, English Proficiency documentation for international student admissions, Form I-901 SEVIS Fee Receipt.
7
Prompt 7 · Lock the Schema
Enforce one consistent format across all 72 rows
→ Gemini Deep Research
can you reorganise [ full list ] in this format: State/Federal | Category | Artifact Name | Trigger Condition | Validation Logic (The "Code") | Citation
After 6 passes of expansion and cross-checking, the accumulated list was inconsistently formatted. This prompt enforced a uniform 6-column schema across every row. The result is not a document — it is a database. Every row becomes a checklist item the software engine can evaluate: load this schema, scan the school's documents against each row, return compliant / partial / non-compliant with evidence.
72 regulations unified into one schema: State/Federal | Category | Artifact Name | Trigger Condition | Validation Logic (boolean code) | Citation. This is checklist.json — the file that powers the compliance engine.
8
Prompt 8 · The Honest Failure — and the smarter fix
When AI can't do something, find the workaround
→ Gemini Deep Research
can you give me the specific URLs for each citation as reference → First attempt: URLs returned but led to catalog landing pages → AI acknowledged: "previous links were citation landing pages like a library card catalog" → Second attempt: still opened incorrect pages // Resolution: Convert each citation to a Google search query instead // → "s. 1002.42(2)(b) F.S." → Google: "Florida Statute 1002.42 private school" // → Always resolvable · Never rots · Always finds current text
AI tools have a known limitation with live URL verification. The solution — converting every citation to a pre-filled Google search query — is actually better than hardcoded URLs. Statute pages move. Search results always find the current text. The failure led to a more robust design than the original request would have produced.
Final schema delivered with Google search query links for all 72 citations. Every link opens a targeted search for the exact statute — always live, always current, never a dead URL.
The Difference

The same regulation.
Two outputs.

The most important upgrade across 8 prompts wasn't the scope — it was the format. The same regulatory content, expressed two different ways, produces a document versus a database. Only one of those can power software.

Prompt V1 — prose output (unexecutable)
Teacher requirements: Private schools must ensure teachers are qualified. Florida law allows alternatives to state certification for private school educators. Vague
Background checks: All staff working with students must undergo background screening, renewed periodically as required. No deadline
Bus inspections: School vehicles must be regularly inspected to ensure student safety. No interval
Scholarship audits: Schools receiving scholarship funds may face financial oversight depending on the amount received. No threshold
Radon testing: Schools should test for radon gas as required by Florida law, at appropriate intervals. No form number
Prompt V3 — boolean logic (executable)
Has_Degree == TRUE OR Exp_Years >= 3 OR Special_Skills == TRUE → s.1002.421(1)(h) Runs as code
Status == "Eligible" AND Date_Screened > (Today - 5yr) → VECHS → s.1002.42(2)(c) Has deadline
School_Days_Since(Inspection) <= 30 → Form 2020-IF → Rule 6A-3.0171 School days, not calendar
IF Revenue > $250,000: CPA_Opinion == "Unmodified" AND Filed <= Sept 15 → s.1002.421(1)(q) Has threshold
IF Tests_Done >= 2: Required = FALSE (unless Remodel == TRUE) → Form DH 1777 → s.404.056 Has form + stop rule
What AI surfaced

The rules schools
didn't know existed

The structured prompt approach didn't just retrieve known regulations faster. It surfaced regulatory nuances that a manual research process — or a generic AI search — would likely miss entirely. Here are the five most consequential findings.

01
The For-Profit Switch
Nord Anglia schools are corporate subsidiaries of a private equity firm — not nonprofits. Every compliance framework built for 501(c)(3) schools is wrong for them.
No Form 990. Instead: mandatory Beneficial Ownership Information (BOI) filing with FinCEN (Corporate Transparency Act 2024), Form 8300 for any cash tuition payment over $10,000 filed within 15 days — criminal felony if missed. W-8BEN for any foreign contractor or guest speaker. The AI identified this switch immediately. Most generic frameworks miss it entirely.
02
The Teacher OR Gate
Private schools in Florida don't require state teacher certification. The law is an OR gate, not a minimum standard.
Under s.1002.421(1)(h), a teacher qualifies if they have any one of: a bachelor's degree in any field, 3 or more years of teaching experience, or documented special skills (a native French speaker teaching French, a professional musician teaching band). A compliance engine checking for "state certification" on every teacher would flag false non-compliance for every private school in Florida.
03
The Radon Stop Rule
Radon testing in Florida schools is not annual. It is mandatory exactly twice — then it stops.
Under s.404.056, schools must test once upon occupancy and once 5 years later. After that, testing stops unless the building is remodeled. Most compliance frameworks treat this as an ongoing annual requirement. The correct logic is: IF Tests_Done >= 2: Required = FALSE (unless Remodel == TRUE). Schools are either over-testing or forgetting the second test entirely.
04
The FERPA Trap
Florida scholarship programs (FTC, FES) are state tax-credit funds — not federal funds. FERPA doesn't apply. But Florida's own SOPIPA does.
Schools that assume FERPA triggers automatically are applying the wrong framework. The actual exposure is Florida's SOPIPA (s.1006.1494) — which prohibits EdTech vendors from selling or profiling student data regardless of federal funding status. Every vendor contract must include specific SOPIPA compliance language. The AI surfaced this distinction; a federal-law-first research approach would have missed the state-level exposure.
05
The 30-Day School Clock
Bus safety inspections run every 30 school days — not 30 calendar days. The distinction closes one inspection cycle entirely.
Rule 6A-3.0171 F.A.C. specifies school days. A 180-school-day year requires 6 inspections annually, not 12. A compliance tool checking on a calendar-month basis would flag false violations through summer and school breaks. More importantly, a school tracking on calendar months might miss that their bus needs inspection after returning from a long break.
The Output

The actual schema
the session produced

Below is a direct sample from the final compliance matrix — exactly as it emerged from the 8-prompt Gemini session, organized into the uniform 6-column schema enforced in Step 7. Each row is loaded into the compliance engine as a discrete check.

Level
ID
Artifact Name
Validation Logic — The Code
Citation
State · Administrative & Financial Compliance
STATE
ADM-01
Notarized Annual Survey
Submission_Date <= May 1st AND Notarized == TRUE
s. 1002.42(2)(b), F.S.
STATE
ADM-02
Attendance Register
Sum(Days) >= 170 OR Sum(Hours) >= Grade_Threshold
Rule 6A-1.09512, F.A.C.
STATE
ADM-03
Scholarship Compliance Form SCF-1
Submission_Date <= March 1st AND Notarized == TRUE
s. 1002.421, F.S.
State · Personnel & Hard Legal Requirements
STATE
PER-01
FDLE VECHS Clearance
Status == "Eligible" AND Date_Screened > (Today - 5 years)
s. 1002.421(1)(m)
STATE
PER-02
Teacher Qualification Record
Has_Degree == TRUE OR Exp_Years >= 3 OR Special_Skills == TRUE
s. 1002.421(2)(h)
Federal · DHS & ICE — Immigration / SEVIS
FEDERAL
IMM-01
Form I-20 (Certificate of Eligibility)
IF Citizenship != "USA": Status == "Active" AND End_Date > Today
8 CFR 214.2(f)
FEDERAL
IMM-02
SEVIS Status Update Log
Reported_To_SEVIS_Date <= (Event_Date + 21 days)
8 CFR 214.2(f)
FEDERAL
IMM-03
Form I-9 (Employment Eligibility)
Section_2_Signed <= (Start_Date + 3 days) AND Docs_Valid == TRUE
8 CFR 274a.2
State · Financial & Scholarship Compliance
STATE
FIN-01
AUP Audit (Agreed-Upon Procedures)
IF Revenue > $250,000: CPA_Opinion == "Unmodified" AND Filed <= Sept 15
s. 1002.421(1)(q)
Federal · IRS & Treasury — For-Profit Corporate
FEDERAL
IRS-01
Form 8300 (Cash Payments)
IF Cash > $10,000: Filed_Within_15_Days == TRUE // felony if not
26 U.S.C. 6050I
FEDERAL
IRS-02
Beneficial Ownership Info (BOI)
IF Entity == "Corp": Report_Filed_FinCEN == TRUE // CTA 2024
31 CFR 1010.380
State · Transportation — Rule 6A-3
STATE
TRN-01
Form 2020-IF (Bus Safety Inspection)
School_Days_Since(Last_Inspection) <= 30 // not calendar days
Rule 6A-3.0171, F.A.C.
STATE
TRN-02
Form MCSA-5876 (Medical Certificate)
Expiration_Date > Today AND Max_Validity <= 2 years
Rule 6A-3.0141, F.A.C.
Federal · Building & Safety — AHERA / OSHA
FEDERAL
BSF-01
AHERA Asbestos Management Plan
IF Built_Year < 1989: AMP_On_Site == TRUE AND Reinspection < 3yr ago
40 CFR 763
FEDERAL
BSF-02
Radon Testing Report (DH 1777)
IF Tests_Done < 2: Date > (Today - 5yr). IF Tests_Done >= 2: Required = FALSE
s. 404.056, F.S.
Replicate it

The full prompt —
ready to run

Below is the expert-engineered research brief from Step 3 — the one that did the core work. Copy it into Gemini Deep Research and follow it with the four expansion prompts listed at the bottom. Adapt the school names and state for your context.

Gemini Deep Research · Expert Compliance Extraction Prompt
Role: Act as a Senior Compliance Architect and Legal Researcher for an EdTech startup. Context: I am building a "Policy as Code" compliance engine for private, international K-12 schools in Florida (specifically Nord Anglia schools like North Broward Prep and Windermere Prep). The software will ingest PDFs (forms, logs, certificates) and structured data to automatically flag non-compliance. The Problem: Most compliance data conflates Public school rules (highly regulated, Chapter 1003) with Private school rules (Chapter 1002, deregulated but contractually bound). Do not mix them. Your Task: Conduct deep research to identify the specific audit constraints. Do not provide generic summaries. Identify: — The specific Document Artifacts (what exact file does an auditor look for?) — The Validation Logic (what boolean rule makes it valid or invalid?) Scope of Research: Florida State: Chapter 1002 ONLY — EXCLUDE Chapter 1003 (public schools) Scholarship: Step Up For Students (FES-UA, FES-EO, FTC) Use Provider Handbooks, not just statutes Federal/Intl: DHS/ICE SEVIS record-keeping requirements for F-1 visa students Accreditation: Note where FCIS or Cognia accreditation supersedes state inspection Research Categories: 1. Personnel & Hard Legal Requirements — Teacher Certification: exact OR-gate logic for non-certified private school teachers? — Background Checks: FDLE Level 2 — valid duration before re-screening? What Clearinghouse result must be in the file? ("Eligible" vs "Not Eligible"?) — Ethics Training: s. 1012.796 — one-time or annual renewal? 2. International Student Compliance (SEVIS / F-1) — Form I-20: What Financial Solvency evidence must be retained? — Reporting: Deadlines for status changes (e.g., within 21 days of dropout?) — English Proficiency: what documentation must the school retain? 3. Health, Safety & Physical Plant CONSTRAINT: EXCLUDE SREF. Private schools use Florida Building Code + FFPC only. — DH 680: Specific 7th Grade validation trigger (Tdap booster date vs DOB)? — DH 3040: Freshness rule — exam must be dated within 12 months of start date? — Fire: Annual inspection by local Fire Marshal — what does "clear" look like? — Radon: Re-testing interval for existing buildings? Does testing ever stop? — Food: DOH inspection frequency for private school cafeterias (quarterly?)? 4. Financial & Scholarship Compliance — AUP Audit: At what revenue threshold does attestation become a full CPA Audit? — Surety Bond: When is one required? What amount? 5. Transportation (Private School — s. 316.615) — Driver files: CDL endorsements required (P and S?), medical cert frequency — Vehicle: Inspection measured in school days or calendar days? Output Format — Data Schema Table (NOT a legal memo): Category | Artifact Name | Trigger Condition | Validation Logic (The Code) | Citation // Column definitions: // Artifact Name: Exact filename or form number (e.g., "Form DH 680", "Form MCSA-5876") // Trigger Condition: When does this rule activate? (New Hire / Annually / Citizenship != USA) // Validation Logic: Boolean rule written as code, not prose // e.g.: Tdap_Date > DOB + 10 years // e.g.: Status == "Eligible" AND Date_Screened > (Today - 5 years) // Citation: Specific Statute, F.A.C. Rule, or Handbook Section
Run these 4 follow-up prompts in order after the main research run:
"Are there any building and infrastructure requirements — can you add that too"
"While these are state level requirements, are there any federal requirements?"
"Review these listed and let me know if anyone has been missed in your evaluation [paste full list]"
"Can you reorganise in format: State/Federal | Category | Artifact Name | Trigger Condition | Validation Logic (The Code) | Citation"
For source links: use Google search query URLs instead of hardcoded URLs — statute pages move, search results don't
Conclusion

The law was always
readable. Now it's
executable.

The regulations governing a Florida private school were never secret. They are published in Florida Statutes, the Code of Federal Regulations, Department of Health guidelines, and scholarship provider handbooks. The problem was never access — it was structure. Every rule existed in legal prose. No software can evaluate legal prose directly. Software needs logic gates.

The 8-prompt methodology above is a translation pipeline. It takes statute language — "schools must maintain a register of daily attendance" — and returns code: Sum(Instructional_Days) >= 170 OR Sum(Net_Hours) >= Grade_Threshold. That code is what a compliance engine runs against a school's uploaded documents. Every row of the schema is a test. Every test returns compliant, partial, or non-compliant — with evidence cited from the document itself.

"We didn't build a faster way to read regulations. We built a way to stop reading them at all — and start running them instead."
ComplianceIQ · Policy as Code · Florida Private Schools

The same methodology applies to any jurisdiction, any school system, any regulatory domain. The prompts change. The output format stays the same. And the output format is the whole game.