03 — Integration Plan (P0–P5 + C0–C3)

Sequenced capabilities, one branch + one focused PR each (04_contribution_plan.md). Effort = S (<1 day) / M (1–3 days) / L (>3 days) / XL (>1 week). Risk = Low / Med / High.

Sequencing principle: the closure track (C0–C3) is independent of the outcomes machinery and is the timely deliverable, so it runs in parallel with P0–P1 and ships first. Everything else follows Philly's order.

P0 ──► P1 ──► P2 ──► P3 ──► P4 ──► P5
 └───► C0 ──► C1 ──► C2 ──► C3   (ships before P3)

Status 2026-09-15: P0–P5 and C0–C3 shipped, plus the residual-z engine (pipeline_miami/analysis/residual_z.py) and the climate-survey track. What remains is listed under Not built in README.md. Runbook: 07_runbook.md.

P0. Repo setup + scaffolding — S · Low

prisma/schema.prisma: MiamiSchool, MiamiMetricDefinition, MiamiSchoolYearMetric, MiamiComparisonGroup, MiamiDataLoad, MiamiSchoolEvent; enums MiamiSchoolType, MiamiAdmissionCategory, MiamiSubgroup, MiamiEventAction; MetricUnit gains USD. Migration is additive.
pipeline_miami/ (copy of pipeline_philly/ skeleton): discover.py with (a) www.fldoe.org → origin.fldoe.org rewrite recorded as fetched_from, (b) an MSID fetch that primes the session cookie, (c) inspectors for .xls (xlrd), PDF (pdfplumber page text + table extraction), ArcGIS JSON.
scripts/loaders/miami/_lib.ts: readLatestProvenance, startDataLoad, padSchoolNumber, suppressionToken census helper, schoolYearFromSuffix.
pipeline_miami/config/{outcomes,covariates,comparability}.yaml populated from 00_overview.md M.5/M.9.
.gitignore rules already added for data/cities/miami/sources/**.

Gate: python -m pipeline_miami.discover lists every sources.yaml entry; --source fldoe_school_grades_2024-25 re-fetches and reproduces the committed sha256.

C0. Closure-track sources — M · Med

Fetch the full FLDOE membership series (2016-17 → 2025-26, by grade and by grade × race), FISH LOS Dade (2022-23 → 2024-25 and any earlier editions found under /core/fileparse.php/20080/), ESSA PPE (2018-19 → 2024-25), DEFP 2024 + 2025, COFTE, the ABC books 2019-20 → 2026-27, Statistical Highlights 2013-14 → 2025-26; CCD membership 2014-15 → 2022-23 (the two years already fetched bracket them).
PDF-table parsers for FISH LOS (row = facility; multi-line names) and ABC capacity blocks; each parser's output is checked against a second extraction (camelot or per-page regex) and the ABC numbers already read by hand for the nine closed schools.
Findings markdown per source under pipeline_miami/findings/.

Gate: every FISH facility row parses to 10 numeric fields; ABC blocks for all targeted schools parse; both parsers agree cell-for-cell.

P1. Source ingest — directory, accountability, tests, graduation, absenteeism, safety — L · Med

Loaders under scripts/loaders/miami/:

msid_directory.ts + ccd_directory.ts → miami_schools (three-way charter check; level from grade flags; region/board district from GIS; lat/lng from MSID with GIS fallback).
school_grades.ts → school_grade per year (1999 →), components as metrics, title_i, pct_econ_disadv (covariate), roster history.
srs_assessments.ts — walks each results year page, maps file prefixes to (subject, grade, administration, regime), loads proficiency + mean scale score + level distribution per school × grade.
learning_gains.ts.
graduation.ts — race and category files; cohort-year indexing documented.
absenteeism.ts — both definitions; ADA/ADM; the 2016-17 inaccuracy flag stored as a load note.
sesir.ts → incident counts and per-100 rates by category (mapping documented in methodology).
metric_definitions.ts.

Gate: row counts per (source, year) match the findings; smoke query returns FAST ELA-3 proficiency for ≥ 290 Dade schools in 2024-25.

C1. Closure-track loaders — M · Low-Med

membership.ts (FLDOE by grade; by grade × race), ccd_membership.ts (cross-publisher copy), fish_los.ts, abc_book.ts (capacity blocks + projections), essa_ppe.ts, ccd_staff.ts, school_events.ts (hand-curated CSV → MiamiSchoolEvent, one row per board action with the PDF URL and page).
Derived metrics: enrollment_change_5yr_pct, k_to_g12_ratio, utilization_permanent_pct, utilization_school_capacity_pct (COFTE-based, FISH LOS), students_per_teacher.

Gate: every district-run E/M/S school has enrollment for every year 2016-17 → 2025-26 and capacity for 2022-23 → 2024-25; the nine closed schools' stored values equal the ABC book.

P2. Climate-survey ingest — M · Low

climate_survey.ts: students / parents / staff workbooks 2018-19 → 2024-25 → miami_survey_responses (school × year × item × respondent: agree/disagree/DK %, N, suppressed).
Rollups: survey_student_safety, survey_student_teaching, survey_staff_leadership, survey_parent_recommend, survey_student_grade_a_b etc. — first-class set decided after item-text continuity check across years.

Gate: rollups present for ≥ 450 schools in 2024-25; item wording diffed across years and recorded in a survey-regime config.

C2. Closure-track validation — M · Low

pipeline_miami/verify/: (i) completeness vs MSID; (ii) base case vs FLDOE xlsx / FISH parse / ESSA xlsx; (iii-a) ABC book + dashboards; (iii-b) CCD membership and staff, GIS capacity; (iv) packet recomputation. Full-QA JSON committed under docs/qa_reports/miami/.

Gate: PASS on all sections; QA report drives the page's noindex flag.

C3. Closure-track analysis + page — L · Med

Packet: scripts/analysis/miami-enrollment-descriptives.ts → data/analysis/miami-enrollment.json (trend by sector, grade pipeline, school-level change distribution, utilization distribution, lift table vs the 2026 lists, PPE vs size, decomposition, ablations).
Page app/miami/methodology/analysis/enrollment/page.tsx in the working-analysis idiom; figures per 06_…md §6; map component (attendance boundaries + points).
Proof trees for every figure and stat (components/proof/* reused; lib/miami/lineage.ts).
Methodology page app/miami/methodology/metrics/enrollment-capacity/page.tsx with the three utilization definitions.

Gate: full site walkthrough; editorial passes per docs/editorial-process.md.

P3. Peer groups + percentiles — S · Low

derive_peer_groups.ts and compute_percentiles.ts copied from Philly with Miami features (M.9); flip-diff at k = 20/30/40.

Gate: Outliers renders; magnet schools do not dominate the top decile once choice status is a hard filter.

P4. Validation harness (outcomes) — M · Low-Med

components.yaml for every P1 metric; (iii-a) School Grades components at ±2 pp; (iii-b) EDFacts absenteeism at ±1.5 pp on the 10% metric, CRDC 15+ days as a correlation check only (different definition).

P5. Site + stories — XL · Med

Routes under app/miami/* mirroring app/philly/*; navbar tab; search bucket; ~20 story stubs; methodology hub; tests/unit/routes.test.ts additions.

Effort

Item	Effort
P0	S
C0 + C1 + C2 + C3	M + M + M + L ≈ 3–4 weeks
P1	L
P2	M
P3	S
P4	M
P5	XL

Total ≈ 9–12 weeks for one engineer who knows the Philly build; the closure page can be live in the first month. Risk concentrates in C0 (PDF parsing) and P1 (yearly assessment file drift).