Appendix — Model Deliverable — PS 8.5 (8-Minute Deck + One-Click Replication Packet)

Problem set: book/weeks/week-08/ps8.5.md (PS 8.5, Week 8). Chapter: Ch 8.5 — The 8-Minute Presentation & the Replication Packet.

PS 8.5 has no numeric answer key: it is the camp’s final project deliverable, and what is graded is not a finding but two crafts — whether eight minutes make one argument land, and whether a stranger can rebuild every claimed number with one command. So this appendix entry is not a worked solution but a model deliverable: a complete, A-grade slide-by-slide deck outline (seven slides — one title plus the six beats of Ch 8.5 §1 — with speaker notes) plus a completed replication-packet checklist — repo layout and make/run_all entry point included — for one cast project, followed by instructor grading notes that key each piece to the rubric and call out the moves that earn (and lose) points.

The project is Maya’s HMDA fair-lending difference-in-differences — the staggered DiD she carried from Lab 4 through her Week-7 pre-analysis plan and identification memo into the Week-8 paper: do state fair-lending examination programs reduce the county-level minority–white gap in mortgage-denial rates? Her primary estimate is a Callaway–Sant’Anna overall ATT against never- and not-yet-treated controls (Ch 4.2), clustered by state. Every magnitude below is illustrative — the figures (an ATT near −1.8 percentage points, a confidence interval, a baseline gap) are presented as a worked talk example to show what an A-grade slide says, not as a claimed empirical result; the real number comes out of make all on the real pinned HMDA vintage.

PART A — THE DECK (`paper/slides.pdf`, 7 slides: 1 title + 6 beats)

Below: each slide’s full-sentence title (the one-idea-per-slide title test, PS 8.5 Part 2a), what sits on the slide, and the speaker notes that would live in paper/slides-notes.md. Read the titles alone, top to bottom — they should tell the whole argument:

(title) Do Fair-Lending Exams Shrink the Mortgage Denial Gap? — Maya R., Capstone

Does adopting a state fair-lending examination program reduce the minority–white gap in mortgage-denial rates?

A mispriced or discriminatory denial gap locks families out of homeownership — and we don’t know if examinations actually move it.

Our estimate is credible if, absent the exams, treated and never-treated states’ denial gaps would have moved in parallel — which we defend with flat pre-trend leads and clean (never-/not-yet-treated) controls.

Adopting an exam program shrinks the denial gap by about 1.8 percentage points — roughly a fifth of the baseline gap.

The −1.8-point result is stable across clustering, control-group, and bandwidth choices, and survives a placebo on never-treated states.

Exams reduce the denial gap where the supervisor’s attention reaches; we do not show they change loan terms, only the approve/deny margin.

That is the argument in seven sentences. A reader who skims only the titles already knows the question, the stake, the design, the number, that it is robust, and where it stops. That is what one-idea-per-slide buys you.

Slide 0 — Title

On the slide: Title (Do Fair-Lending Exams Shrink the Mortgage Denial Gap?), name, the camp/conference, one line: “all results regenerate from raw data via make all — repo on the last slide.”

Speaker notes (~15 sec): “I’m Maya; I study whether a specific fair-lending policy actually changes who gets a mortgage. Everything you’ll see is reproducible — there’s a command at the end that rebuilds every number.” Get off this slide fast; the clock starts on slide 1.

Slide 1 — The question (beat 1)

Title: Does adopting a state fair-lending examination program reduce the minority–white gap in mortgage-denial rates?

On the slide: the question, large, and nothing else. Maybe one small map thumbnail of which states adopted, no labels.

Speaker notes (~40 sec): “Here’s the one question. Some states run examination programs that audit lenders for fair-lending compliance; others don’t, and they turned these on in different years. I ask whether turning one on shrinks the gap between how often minority and white applicants get denied — holding the applicant pool fixed as best I can.” Do not open with the history of the Fair Housing Act. The room cannot hold context it has no reason to want yet; the reason comes on slide 2.

Slide 2 — Why it matters (beat 2)

Title: A mispriced or discriminatory denial gap locks families out of homeownership — and we don’t know if examinations actually move it.

On the slide: one concrete stake, one number for scale (illustrative): “the raw minority–white denial gap runs near 9 points; exams cost states real money — do they work?”

Speaker notes (~25 sec): “Homeownership is how most American families build wealth, so a denial gap isn’t abstract — it compounds across a generation. States spend real money standing up these exam regimes on the theory that supervision shrinks the gap. Nobody has cleanly tested whether it does. That’s the gap I fill.” Concrete before abstract, a number before a Greek letter (CONVENTIONS §1). This is where you spend the motivation — after the question, not before.

Slide 3 — Design / identification (beat 3 — the spine)

Title: Our estimate is credible if, absent the exams, treated and never-treated states’ denial gaps would have moved in parallel — which we defend with flat pre-trend leads and clean (never-/not-yet-treated) controls.

On the slide: the identifying-assumption sentence in the Ch 7.5 contract form, plus one small picture — the cohort-mean denial-gap lines (treated cohorts tracking the never-treated line before adoption, peeling off after). No regression equation. A one-line spec strip at the bottom in plain words: “within-state-and-year comparison; clustered by state; Callaway–Sant’Anna against clean controls (Ch 4.2).”

Speaker notes (~90 sec — slow down here): “This is the slide that matters. My claim is causal only if one thing holds: that absent the exams, the states that adopted would have moved like the states that never did. I can’t prove that — it’s a statement about a counterfactual that never happened. But I defend it two ways. First, the picture: before each cohort adopts, its denial-gap line tracks the never-adopters — that’s the flat pre-trend you can see. Second, I never use an already-treated state as a control for a later one, because its own effect is still building and would contaminate the comparison — that’s the staggered-DiD trap from the literature, and the clean-control estimator avoids it. If you only remember one slide, remember this one.” This is where every hard question lands; if this slide is weak, the talk does not survive questions.

Slide 4 — Headline result (beat 4)

Title: Adopting an exam program shrinks the denial gap by about 1.8 percentage points — roughly a fifth of the baseline gap.

On the slide: one figure — the event-study plot: estimated effect by event-time, flat and hugging zero before adoption, dropping to about −1.8 after, with a shaded confidence band. One annotation: “overall ATT ≈ −1.8 pp [−2.8, −0.8].” (All illustrative.) No table.

Speaker notes (~80 sec): “Here’s the answer. Each dot is the effect on the denial gap in a year relative to when the exam turned on. Before — flat, sitting on zero; that’s the parallel-trends defense from the last slide, now as data. After — it drops and stays down, around 1.8 points. In human terms: the baseline gap is about nine points, so the exams close roughly a fifth of it. And the confidence interval — about minus-2.8 to minus-0.8 — doesn’t include zero, but notice I’m showing you the interval, not hiding behind a point estimate. The effect is real but not enormous.” The CI is part of the result, not an afterthought. The event-study plot is the table-to-figure translation (PS 8.5 Part 2b): a DiD result becomes an event-study plot, and the flat pre-period argues identification for free.

Slide 5 — Robustness (beat 5)

Title: The −1.8-point result is stable across clustering, control-group, and bandwidth choices, and survives a placebo on never-treated states.

On the slide: a specification curve / coefficient plot — one dot-and-whisker per alternative spec (state-clustered vs. wild-cluster-bootstrap SEs; never-treated vs. not-yet-treated controls; with/without the population floor; ±1 year of the adoption window), all clustering near −1.8, with a vertical reference line at the headline. A second small panel: the placebo distribution (random fake-adoption years) centered on zero with the real estimate far in the tail.

Speaker notes (~55 sec): “I tried to kill this number. Cluster by state, bootstrap the standard errors, swap the control group, raise the sample floor, shift the adoption window — it sits near minus-1.8 every time; that’s the tight vertical cluster. And the placebo: when I assign fake exam dates to states that never adopted, the effect vanishes to zero — so I’m not picking up some mechanical artifact of the estimator. The result didn’t depend on one lucky choice.” You don’t narrate all twelve specs; you show they exist and agree, in one sentence. This pre-empts the survivable critiques before anyone asks (Ch 8.5 §3).

Slide 6 — Contribution (beat 6)

Title: Exams reduce the denial gap where the supervisor’s attention reaches; we do not show they change loan terms, only the approve/deny margin.

On the slide: one sentence of what’s new, one clause of what’s not claimed. A pointer: “repo + one-click rebuild → make all · github.com/CAMP-ORG/capstone-2026-maya”.

Speaker notes (~40 sec): “So: a cleanly-identified estimate that fair-lending examinations actually shrink the denial gap, by about a fifth — under a parallel-trends assumption I’ve defended but can’t prove. What I’m not claiming: I look only at approve-versus-deny, not at the rate or terms applicants got, and only at mortgages, not other credit. And the exact code that produces every number is one command away — clone it and you’ll get my figure, seed and all. Thank you.” The “what we do not claim” clause is the single most credibility-building sentence in the talk — it signals you know exactly where your claim stops. Then stop; do not summarize the summary.

Backup slides (shown only if asked)

The precise regression table (Appendix D format, four decimals, SEs in parentheses — the object the slide-4 figure was translated from); the Goodman–Bacon decomposition; the full pre-trend leads with confidence bands; the imputation count by treatment status. These exist for the questioner who wants the fourth decimal — a table is for a reader who can stop and study; a figure is for a listener who cannot.

Speaker-notes appendix: timing, cut line, and prepared answers

Time budget (target 7:00, not 8:00). Title 0:15 · Q+why 1:05 · design 1:30 (slow) · headline 1:20 (let it land) · robustness 0:55 · contribution 0:40 = ~5:45 of content, ~1:15 of slack for the things that go wrong. Rehearsed timed ×4; longest run 7:10.

Cut line (pre-decided). If I’m at minute six and still on the headline, I drop the placebo panel of slide 5 and say one sentence (“it also survives a placebo — happy to show it”). I never cut the design slide or the headline. Deciding this in advance means I cut cleanly instead of panicking.

Prepared answer 1 — “Aren’t adopting states just on a better trajectory anyway?” (the fatal one). Threat: differential pre-trend — reform-minded legislatures might both pass exams and already be improving. Plausible: yes, entirely; this is the central worry. What I did: the event-study leads are flat and tight enough to rule out an economically meaningful pre-slope, and adoption timing tracks legislative calendars, not local lending conditions. Residual: a flat pre-trend fails to refute parallel trends but can’t confirm the post-period counterfactual, and with few treated states the leads have limited power. “So I’d say: consistent with parallel trends, not proof of it.”

Prepared answer 2 — “Why cluster by state, not county?” (survivable). This attacks a number, not the comparison. “Treatment turns on at the state level, so the errors correlate within state — that’s the right level (BDM). I also report a wild-cluster bootstrap because there are few treated states; the result holds. It’s on the robustness slide and in the backup table.” A survivable critique: the number might wiggle, it doesn’t vanish.

Prepared answer 3 — the located “I don’t know”: “Would this generalize to auto loans?” “My data is mortgages, so I genuinely don’t know. The mechanism I identify is examiner attention, and I’d expect it to matter wherever there’s a comparable exam regime — but auto lending has a different supervisor, so I wouldn’t extend the estimate without re-running it there.” Located, not a shrug: it admits the limit, explains why it’s a limit, and names what would resolve it — a better answer than a confident yes.

Fatal vs. survivable, and my retreat position. The fatal critique is a real pre-existing divergence in the denial gap (it attacks the comparison; no re-clustering saves me). The survivable ones attack numbers (clustering, winsorizing, controls) — “we checked; it doesn’t move.” Retreat position, written before I walk in: if someone shows the pre-trends genuinely tilt, I concede at that level — “you’re right I can’t rule that out with this design; the honest reading is a strong association consistent with the examiner-attention mechanism, which future work with [an RD on an exam-eligibility threshold] should test causally.” A conceded fatal critique is a wounded paper but an intact scientist.

PART B — THE REPLICATION PACKET (completed checklist)

Repo layout (the Lab 7 repository, finished)

capstone-2026-maya/
├── README.md                  # what / why / how-to-run (one command) + public-vs-licensed split
├── Makefile                   # `make all` rebuilds EVERYTHING raw -> paper.pdf
├── run_all.sh                 # same pipeline, no make required
├── environment.yml            # human-readable: python=3.11 + CONVENTIONS §5 stack
├── environment.lock.yml       # machine-exact: every package at the solved version
├── config.py                  # the single fixed SEED, imported everywhere randomness happens
├── .gitignore                 # blocks data/, .env, *.parquet, keys (Lab 7 Step 5)
├── PRE-ANALYSIS-PLAN.md        # tagged `pap-filed` before any confirmatory regression
├── DEVIATIONS.md               # dated departures from the PAP, all after the tag
├── memo.md                     # the Ch 7.5 identification memo (the talk's design slide)
├── REPLICATION-CHECKLIST.md    # this checklist, each box proven
├── data/
│   ├── raw/                    # GIT-IGNORED; cached public HMDA pulls land here
│   │   └── .gitkeep
│   └── processed/              # GIT-IGNORED; built county-year panel
│       └── .gitkeep
├── src/
│   ├── 01_pull_data.py         # CFPB HMDA Data Browser pull (public) + FRED controls; logs hashes
│   ├── 02_build_dataset.py     # aggregate LAR -> county-year denial-gap panel; survivorship/look-ahead checks
│   └── 03_analysis.py          # CS estimator + robustness; WRITES tables/ and figures/ (seeded)
├── notebooks/
│   └── nb8.5_manuscript_build.ipynb   # story/narration; calls src/ logic, no analysis buried in cells
├── logs/
│   └── pulls.jsonl             # one line per pull: what, when, vintage, content hash
├── data-cards/
│   ├── hmda.md                 # public; CFPB vintage pinned
│   └── fred.md                 # public; controls
└── paper/
    ├── main.tex                # \input{tables/...} and \includegraphics{figures/...} — never pasted
    ├── slides.pdf              # the 8-slide deck (Part A)
    ├── slides-notes.md         # speaker notes, cut line, prepared answers
    ├── figures/                # event_study.pdf, spec_curve.pdf — GENERATED by 03_analysis.py
    └── tables/                 # main_results.tex — GENERATED by 03_analysis.py

The rule from Lab 7 holds visibly: logic in src/, story in notebooks/, the data/ tree git-ignored, the licensed-data wall enforced by .gitignore. Here HMDA is public, so it ships cached in data/raw/ with its vintage pinned — but the same layout would keep a CRSP-based capstone’s licensed bytes on GMU infrastructure and ship only the src/01_pull_data.py access script plus the hash log (Ch 8.5 §5, the asymmetry of reproducibility).

The single entry point — `Makefile`

# Makefile — `make all` regenerates EVERYTHING from raw to paper/main.pdf.
.PHONY: all data analysis paper clean

all: paper

data:                       ## pull (public HMDA + FRED) or load cache, then build the panel
    python src/01_pull_data.py
    python src/02_build_dataset.py

analysis: data              ## estimate CS ATT + robustness; write tables/ and figures/ (deterministic via SEED)
    python src/03_analysis.py

paper: analysis             ## compile the write-up, pulling in the just-built tables/figures
    cd paper && latexmk -pdf main.tex

clean:                      ## delete ALL derived outputs so a rebuild is honest
    rm -rf data/processed/* paper/tables/* paper/figures/* paper/main.pdf

The equivalent shell entry point, for a reviewer without make:

# run_all.sh — same pipeline: `bash run_all.sh`
set -euo pipefail                 # stop on the first error; never silently half-build
python src/01_pull_data.py
python src/02_build_dataset.py
python src/03_analysis.py         # tables/figures written here, deterministically (one SEED)
( cd paper && latexmk -pdf main.tex )
echo "Rebuilt paper/main.pdf from raw. Compare to the committed PDF."

The fixed, named SEED — `config.py`

# config.py — one seed, imported everywhere randomness happens
SEED = 20260815  # the conference date; any fixed int, just FIXED and named

# in 03_analysis.py, at the top, before any draw:
import numpy as np
from config import SEED
rng = np.random.default_rng(SEED)   # modern Generator, NOT legacy np.random.seed
# the placebo reassignment and the wild-cluster bootstrap both draw from this rng,
# passed in explicitly so a reader sees exactly what randomness each step consumes.

Maya’s analysis has two stochastic steps — the wild-cluster bootstrap on slide 5’s standard errors and the placebo that reassigns fake adoption years — and both draw from this one rng. That is why her headline interval [−2.8, −0.8] comes out identical on a reviewer’s machine; without the fixed seed it would drift run to run and the packet would, strictly, be irreproducible. The seed is recorded in the README and was fixed in the pap-filed commit (Ch 7.3), so it provably was not searched over for a prettier interval.

Completed `REPLICATION-CHECKLIST.md`

The deck - [x] 7 slides (1 title + 6 beats) in order (title · question · why · design · headline · robustness · contribution). → paper/slides.pdf - [x] Every title is a full-sentence claim; argument legible from titles alone. → list atop paper/slides-notes.md - [x] Design slide shows the identifying-assumption sentence (from memo.md), not the regression equation. - [x] Headline slide: one number + CI, as the event-study figure, human-units sentence (“about a fifth of the baseline gap”). - [x] Robustness slide: spec curve + placebo panel clustering near −1.8, one stability sentence. - [x] Contribution slide names what’s new and the “we do not show loan terms / non-mortgage credit” non-claim. - [x] Every slide figure is the file 03_analysis.py wrote into paper/figures/ — none hand-pasted. (verified: slides.pdf \includegraphicses figures/event_study.pdf)

The talk - [x] Rehearsed timed ×4; longest run 7:10, fits in seven. - [x] Cut line pre-decided (drop slide-5 placebo panel) — in slides-notes.md. - [x] Two “what about [confound]?” answers + one located “I don’t know” written out (above). - [x] Fatal critique named (pre-existing denial-gap divergence); retreat position written (concede → descriptive association → RD as future work).

The packet - [x] Six parts present: HMDA access script (public, ships cached) · src/01–03 in run-order · README · environment.yml+environment.lock.yml · config.py SEED · Makefile+run_all.sh. - [x] README states public HMDA / public FRED; no licensed data, no secrets in repo or history. → git ls-files | grep -E '\.env|\.parquet|crsp' returns nothing. - [x] SEED is one named constant via np.random.default_rng(SEED), threaded through bootstrap + placebo, recorded in README. - [x] Every paper/ table and figure is code-generated and \input/\includegraphics’d — none pasted. - [x] make clean && make all rebuilt paper/main.pdf; git diff --stat on the PDF and on tables//figures/ shows no change → no hand-edits. (the honesty test, passed) - [x] Verified on a fresh clone in a scratch dir, not the working copy: git clone … && conda env create -f environment.yml && conda activate capstone && make all → identical PDF.

INSTRUCTOR GRADING NOTES (keyed to the PS 8.5 rubric)

These notes show why this deliverable earns an A and where a weaker submission would lose points.

Part 1 — The six-beat arc (35 pts) → Excellent

All six beats are present, in order, one per slide, and — the A-grade move — each slide does its one job and refuses the others. (a) Question (5): one sentence, no literature dump; the title is the question. (b) Why (5): a concrete stake with a scale number, placed after the question, not before — the single most common ordering error (opening abstract loses the room in thirty seconds) is avoided. (c) Design (8 — the spine): the slide shows the identifying-assumption sentence and a cohort-mean picture, not the regression equation — exactly right, because nobody reads a regression in sixty seconds; the spec is reduced to one plain-words strip. This is where weak decks collapse: a design slide that is a wall of \beta’s and fixed-effect notation cannot be read live, and the talk dies at the first hard question. (d) Headline (6): one number, with its confidence interval, as a figure, interpreted in human units (“a fifth of the baseline gap”). (e) Robustness (6): a spec curve that shows stability as a shape, plus a placebo — not a narrated list of twelve checks. (f) Contribution (5): the “what we do not claim” clause is present and load-bearing. Where points are lost on a typical submission: a “Results” slide that is the paper’s regression table screenshotted (unreadable from row eight — auto-deduction on Part 2b); a headline with a point estimate and no interval (reads as evasive); a design slide showing the equation instead of the assumption; motivation before the question. Each is scored down on the specific beat.

Part 2 — One idea per slide + table-to-figure (20 pts) → Excellent

(a) Title test (8): every title is a full-sentence claim, and the argument is legible from the titles alone — the deliverable even prints the seven-title list so a grader can verify it in ten seconds. A deck with noun-phrase titles (“Background,” “Results,” “Robustness”) loses this part wholesale, because the titles carry no argument. (b) Translation (12): the student names which translation each slide uses and why — the DiD result becomes an event-study plot (flat-then-jump, which argues parallel trends for free), the robustness becomes a coefficient/spec plot (stability as a tight cluster), and the precise table is demoted to a backup slide for the fourth-decimal questioner. The figures are stripped to load-bearing elements. Critically, the deliverable verifies the slide figure is the file the code generated — closing the loop with Part 4 so the slide and the packet cannot drift apart. Where points are lost: showing the table on the main slide instead of translating it; a figure with unreadable ticks, dead legends, and gridlines; or — the silent failure that surfaces in Part 4 — a slide figure that was exported by hand and no longer matches what make all produces.

Part 3 — Defending identification (15 pts) → Excellent

(a) Two prepared confound answers (6): the four-part structure (threat → plausible → what-I-did → residual) is followed exactly, drawn straight from the Ch 7.5 threats table — the student is reading from a script authored when calm, not improvising. The pre-trend answer concedes the worry is central and plausible before answering, which reads as more credible than a confident dismissal. (b) Located “I don’t know” (5): the auto-loan answer is the chapter’s gold standard — it admits the limit, explains why (different supervisor), and names what would resolve it (re-run there), rather than bluffing a number. (c) Fatal vs. survivable + retreat (4): the student correctly identifies that attacks on the comparison are fatal (a real pre-trend) while attacks on a number are survivable (clustering), maps this to Ch 7.5 §4’s testable-vs-arguable, and — the mature move — writes the retreat to a descriptive claim before walking in. Where points are lost: answering a confound question by dismissing it (“we control for that”); bluffing the “I don’t know”; or misclassifying a fatal critique as survivable and trying to re-cluster their way out of a broken comparison in front of a room that can see it is broken.

Part 4 — The one-click packet (30 pts) → Excellent

All six parts present and each verified, not merely asserted. (a) Data (5): correctly ships public HMDA cached with a pinned vintage and states the public/licensed split honestly. (b) Run-order (5): 01→02→03 numbering makes the order unambiguous; logic in src/, story in notebooks/. (c) README (4): the how-to-run has collapsed to one command. (d) Env (4): both environment.yml and environment.lock.yml committed; the student can say the “rebuild that and you rebuild my numbers” sentence. (e) SEED (5): a single named SEED in config.py, threaded through both stochastic steps (bootstrap and placebo) via np.random.default_rng, recorded in the README, fixed in the pap-filed commit so it provably was not searched over. (f) Entry point (7): a Makefile and a run_all.sh, with the figures/tables generated by 03_analysis.py and \input/\includegraphics’d — never pasted. The decisive A-grade evidence is the honesty test, run and recorded: make clean && make all rebuilt the PDF byte-identical, and the student re-verified on a fresh clone in a scratch directory — the true stranger’s-machine test, not the working copy where stale derived files hide hand-edits. Where points are lost: the single most common — and most penalized — failure is a number on a slide that the rebuild does not reproduce, because it was hand-copied from a notebook into the manuscript and has since drifted from the code. That breaks the one-click chain and fails the part regardless of how polished the deck is. Also penalized: a seed scattered ad-hoc (or the legacy np.random.seed), a missing lock file (“it worked on my machine”), or licensed data / a .env committed to history (a Lab 7 Step-5 violation that is serious independent of the grade).

One-line grading heuristic

The two highest-signal questions for any PS 8.5 submission: (1) Can a reader reconstruct the entire argument from the slide titles alone, with the design slide showing the assumption (not the equation) and the headline showing a number with its interval as a figure? (2) Does make clean && make all, on a fresh clone, regenerate the exact number the headline slide claims? A student who clears both has internalized the camp’s thesis — reproducibility is the form credibility takes — because the talk invites verification (“here’s the command that checks me”) instead of asking to be believed, and the packet makes good on the invitation. This exemplar clears both cleanly, which is what an A looks like.

End of model deliverable for PS 8.5. The exemplar is Maya’s HMDA fair-lending DiD — the project she carried from Lab 4 through the Week-7 PAP, identification memo (Ch 7.5), and robustness battery (Ch 8.2) into this final talk and packet. All magnitudes (the ≈ −1.8 pp ATT, the [−2.8, −0.8] interval, the ~9-point baseline gap, the “fifth of the gap”) are illustrative, presented as a worked talk example, not a claimed estimate; the real numbers come out of make all on the pinned, public CFPB HMDA vintage recorded in data-cards/hmda.md. The packet standard is Appendix D + Lab 7; the staggered-DiD machinery is Callaway & Sant’Anna (2021) and Goodman-Bacon (2021), as built in Ch 4.2 and Lab 4; the SEED/make/one-command discipline is Ch 8.5 §5 and Gentzkow & Shapiro (2014). No secrets or keys appear; the SEED = 20260815 constant and config.py pattern are illustrative.

# Model Deliverable — PS 8.5 (8-Minute Deck + One-Click Replication Packet) **Problem set:** `book/weeks/week-08/ps8.5.md` (PS 8.5, Week 8). **Chapter:** Ch 8.5 — The 8-Minute Presentation & the Replication Packet. PS 8.5 has no numeric answer key: it is the camp's final project deliverable, and what is graded is not a finding but two *crafts* — whether eight minutes make one argument land, and whether a stranger can rebuild every claimed number with one command. So this appendix entry is not a worked solution but a **model deliverable**: a complete, A-grade slide-by-slide deck outline (seven slides — one title plus the six beats of Ch 8.5 §1 — with speaker notes) plus a completed replication-packet checklist — repo layout and `make`/`run_all` entry point included — for one cast project, followed by **instructor grading notes** that key each piece to the rubric and call out the moves that earn (and lose) points. The project is **Maya's HMDA fair-lending difference-in-differences** — the staggered DiD she carried from Lab 4 through her Week-7 pre-analysis plan and identification memo into the Week-8 paper: *do state fair-lending examination programs reduce the county-level minority–white gap in mortgage-denial rates?* Her primary estimate is a Callaway–Sant'Anna overall ATT against never- and not-yet-treated controls (Ch 4.2), clustered by state. Every magnitude below is **illustrative** — the figures (an ATT near −1.8 percentage points, a confidence interval, a baseline gap) are presented as a worked *talk* example to show what an A-grade slide says, not as a claimed empirical result; the real number comes out of `make all` on the real pinned HMDA vintage. --- ## PART A — THE DECK (`paper/slides.pdf`, 7 slides: 1 title + 6 beats) Below: each slide's full-sentence title (the one-idea-per-slide title test, PS 8.5 Part 2a), what sits *on* the slide, and the speaker notes that would live in `paper/slides-notes.md`. Read the **titles alone, top to bottom** — they should tell the whole argument: > 0. *(title)* Do Fair-Lending Exams Shrink the Mortgage Denial Gap? — Maya R., Capstone > 1. **Does adopting a state fair-lending examination program reduce the minority–white gap in mortgage-denial rates?** > 2. **A mispriced or discriminatory denial gap locks families out of homeownership — and we don't know if examinations actually move it.** > 3. **Our estimate is credible if, absent the exams, treated and never-treated states' denial gaps would have moved in parallel — which we defend with flat pre-trend leads and clean (never-/not-yet-treated) controls.** > 4. **Adopting an exam program shrinks the denial gap by about 1.8 percentage points — roughly a fifth of the baseline gap.** > 5. **The −1.8-point result is stable across clustering, control-group, and bandwidth choices, and survives a placebo on never-treated states.** > 6. **Exams reduce the denial gap where the supervisor's attention reaches; we do not show they change loan *terms*, only the approve/deny margin.** That is the argument in seven sentences. A reader who skims only the titles already knows the question, the stake, the design, the number, that it is robust, and where it stops. That is what one-idea-per-slide buys you. --- ### Slide 0 — Title **On the slide:** Title (*Do Fair-Lending Exams Shrink the Mortgage Denial Gap?*), name, the camp/conference, one line: "all results regenerate from raw data via `make all` — repo on the last slide." **Speaker notes (~15 sec):** "I'm Maya; I study whether a specific fair-lending policy actually changes who gets a mortgage. Everything you'll see is reproducible — there's a command at the end that rebuilds every number." *Get off this slide fast; the clock starts on slide 1.* --- ### Slide 1 — The question (beat 1) **Title:** *Does adopting a state fair-lending examination program reduce the minority–white gap in mortgage-denial rates?* **On the slide:** the question, large, and nothing else. Maybe one small map thumbnail of which states adopted, no labels. **Speaker notes (~40 sec):** "Here's the one question. Some states run examination programs that audit lenders for fair-lending compliance; others don't, and they turned these on in different years. I ask whether turning one on shrinks the gap between how often minority and white applicants get denied — holding the applicant pool fixed as best I can." *Do not open with the history of the Fair Housing Act. The room cannot hold context it has no reason to want yet; the reason comes on slide 2.* --- ### Slide 2 — Why it matters (beat 2) **Title:** *A mispriced or discriminatory denial gap locks families out of homeownership — and we don't know if examinations actually move it.* **On the slide:** one concrete stake, one number for scale (illustrative): "the raw minority–white denial gap runs near 9 points; exams cost states real money — do they work?" **Speaker notes (~25 sec):** "Homeownership is how most American families build wealth, so a denial gap isn't abstract — it compounds across a generation. States spend real money standing up these exam regimes on the *theory* that supervision shrinks the gap. Nobody has cleanly tested whether it does. That's the gap I fill." *Concrete before abstract, a number before a Greek letter (CONVENTIONS §1). This is where you spend the motivation — after the question, not before.* --- ### Slide 3 — Design / identification (beat 3 — the spine) **Title:** *Our estimate is credible if, absent the exams, treated and never-treated states' denial gaps would have moved in parallel — which we defend with flat pre-trend leads and clean (never-/not-yet-treated) controls.* **On the slide:** the **identifying-assumption sentence** in the Ch 7.5 contract form, plus one small picture — the cohort-mean denial-gap lines (treated cohorts tracking the never-treated line *before* adoption, peeling off *after*). **No regression equation.** A one-line spec strip at the bottom in plain words: "within-state-and-year comparison; clustered by state; Callaway–Sant'Anna against clean controls (Ch 4.2)." **Speaker notes (~90 sec — slow down here):** "This is the slide that matters. My claim is causal *only if* one thing holds: that absent the exams, the states that adopted would have moved like the states that never did. I can't prove that — it's a statement about a counterfactual that never happened. But I defend it two ways. First, the picture: before each cohort adopts, its denial-gap line tracks the never-adopters — that's the flat pre-trend you can see. Second, I never use an already-treated state as a control for a later one, because its own effect is still building and would contaminate the comparison — that's the staggered-DiD trap from the literature, and the clean-control estimator avoids it. If you only remember one slide, remember this one." *This is where every hard question lands; if this slide is weak, the talk does not survive questions.* --- ### Slide 4 — Headline result (beat 4) **Title:** *Adopting an exam program shrinks the denial gap by about 1.8 percentage points — roughly a fifth of the baseline gap.* **On the slide:** **one figure** — the event-study plot: estimated effect by event-time, flat and hugging zero before adoption, dropping to about −1.8 after, with a shaded confidence band. One annotation: "overall ATT ≈ −1.8 pp [−2.8, −0.8]." *(All illustrative.)* **No table.** **Speaker notes (~80 sec):** "Here's the answer. Each dot is the effect on the denial gap in a year relative to when the exam turned on. Before — flat, sitting on zero; that's the parallel-trends defense from the last slide, now as data. After — it drops and stays down, around 1.8 points. In human terms: the baseline gap is about nine points, so the exams close roughly a fifth of it. And the confidence interval — about minus-2.8 to minus-0.8 — doesn't include zero, but notice I'm *showing* you the interval, not hiding behind a point estimate. The effect is real but not enormous." *The CI is part of the result, not an afterthought. The event-study plot is the table-to-figure translation (PS 8.5 Part 2b): a DiD result becomes an event-study plot, and the flat pre-period argues identification for free.* --- ### Slide 5 — Robustness (beat 5) **Title:** *The −1.8-point result is stable across clustering, control-group, and bandwidth choices, and survives a placebo on never-treated states.* **On the slide:** a **specification curve / coefficient plot** — one dot-and-whisker per alternative spec (state-clustered vs. wild-cluster-bootstrap SEs; never-treated vs. not-yet-treated controls; with/without the population floor; ±1 year of the adoption window), all clustering near −1.8, with a vertical reference line at the headline. A second small panel: the placebo distribution (random fake-adoption years) centered on zero with the real estimate far in the tail. **Speaker notes (~55 sec):** "I tried to kill this number. Cluster by state, bootstrap the standard errors, swap the control group, raise the sample floor, shift the adoption window — it sits near minus-1.8 every time; that's the tight vertical cluster. And the placebo: when I assign fake exam dates to states that never adopted, the effect vanishes to zero — so I'm not picking up some mechanical artifact of the estimator. The result didn't depend on one lucky choice." *You don't narrate all twelve specs; you show they exist and agree, in one sentence. This pre-empts the survivable critiques before anyone asks (Ch 8.5 §3).* --- ### Slide 6 — Contribution (beat 6) **Title:** *Exams reduce the denial gap where the supervisor's attention reaches; we do not show they change loan terms, only the approve/deny margin.* **On the slide:** one sentence of what's new, one clause of what's *not* claimed. A pointer: "repo + one-click rebuild → `make all` · github.com/CAMP-ORG/capstone-2026-maya". **Speaker notes (~40 sec):** "So: a cleanly-identified estimate that fair-lending examinations actually shrink the denial gap, by about a fifth — under a parallel-trends assumption I've defended but can't prove. What I'm *not* claiming: I look only at approve-versus-deny, not at the rate or terms applicants got, and only at mortgages, not other credit. And the exact code that produces every number is one command away — clone it and you'll get my figure, seed and all. Thank you." *The "what we do not claim" clause is the single most credibility-building sentence in the talk — it signals you know exactly where your claim stops. Then stop; do not summarize the summary.* --- ### Backup slides (shown only if asked) The precise regression **table** (Appendix D format, four decimals, SEs in parentheses — the object the slide-4 figure was *translated from*); the Goodman–Bacon decomposition; the full pre-trend leads with confidence bands; the imputation count by treatment status. These exist for the questioner who wants the fourth decimal — *a table is for a reader who can stop and study; a figure is for a listener who cannot.* --- ### Speaker-notes appendix: timing, cut line, and prepared answers **Time budget (target 7:00, not 8:00).** Title 0:15 · Q+why 1:05 · **design 1:30 (slow)** · **headline 1:20 (let it land)** · robustness 0:55 · contribution 0:40 = ~5:45 of content, ~1:15 of slack for the things that go wrong. Rehearsed timed ×4; longest run 7:10. **Cut line (pre-decided).** If I'm at minute six and still on the headline, I **drop the placebo panel of slide 5** and say one sentence ("it also survives a placebo — happy to show it"). I never cut the design slide or the headline. Deciding this in advance means I cut cleanly instead of panicking. **Prepared answer 1 — "Aren't adopting states just on a better trajectory anyway?" (the fatal one).** *Threat:* differential pre-trend — reform-minded legislatures might both pass exams *and* already be improving. *Plausible:* yes, entirely; this is the central worry. *What I did:* the event-study leads are flat and tight enough to rule out an economically meaningful pre-slope, and adoption timing tracks legislative calendars, not local lending conditions. *Residual:* a flat pre-trend fails to *refute* parallel trends but can't *confirm* the post-period counterfactual, and with few treated states the leads have limited power. "So I'd say: consistent with parallel trends, not proof of it." **Prepared answer 2 — "Why cluster by state, not county?" (survivable).** *This attacks a number, not the comparison.* "Treatment turns on at the state level, so the errors correlate within state — that's the right level (BDM). I also report a wild-cluster bootstrap because there are few treated states; the result holds. It's on the robustness slide and in the backup table." *A survivable critique: the number might wiggle, it doesn't vanish.* **Prepared answer 3 — the located "I don't know": "Would this generalize to auto loans?"** "My data is mortgages, so I genuinely don't know. The mechanism I identify is *examiner attention*, and I'd expect it to matter wherever there's a comparable exam regime — but auto lending has a different supervisor, so I wouldn't extend the estimate without re-running it there." *Located, not a shrug: it admits the limit, explains why it's a limit, and names what would resolve it — a better answer than a confident yes.* **Fatal vs. survivable, and my retreat position.** The *fatal* critique is a real pre-existing divergence in the denial gap (it attacks the comparison; no re-clustering saves me). The *survivable* ones attack numbers (clustering, winsorizing, controls) — "we checked; it doesn't move." **Retreat position, written before I walk in:** if someone shows the pre-trends genuinely tilt, I concede at that level — "you're right I can't rule that out with this design; the honest reading is a strong association consistent with the examiner-attention mechanism, which future work with [an RD on an exam-eligibility threshold] should test causally." A conceded fatal critique is a wounded paper but an intact scientist. --- ## PART B — THE REPLICATION PACKET (completed checklist) ### Repo layout (the Lab 7 repository, finished) ``` capstone-2026-maya/ ├── README.md # what / why / how-to-run (one command) + public-vs-licensed split ├── Makefile # `make all` rebuilds EVERYTHING raw -> paper.pdf ├── run_all.sh # same pipeline, no make required ├── environment.yml # human-readable: python=3.11 + CONVENTIONS §5 stack ├── environment.lock.yml # machine-exact: every package at the solved version ├── config.py # the single fixed SEED, imported everywhere randomness happens ├── .gitignore # blocks data/, .env, *.parquet, keys (Lab 7 Step 5) ├── PRE-ANALYSIS-PLAN.md # tagged `pap-filed` before any confirmatory regression ├── DEVIATIONS.md # dated departures from the PAP, all after the tag ├── memo.md # the Ch 7.5 identification memo (the talk's design slide) ├── REPLICATION-CHECKLIST.md # this checklist, each box proven ├── data/ │ ├── raw/ # GIT-IGNORED; cached public HMDA pulls land here │ │ └── .gitkeep │ └── processed/ # GIT-IGNORED; built county-year panel │ └── .gitkeep ├── src/ │ ├── 01_pull_data.py # CFPB HMDA Data Browser pull (public) + FRED controls; logs hashes │ ├── 02_build_dataset.py # aggregate LAR -> county-year denial-gap panel; survivorship/look-ahead checks │ └── 03_analysis.py # CS estimator + robustness; WRITES tables/ and figures/ (seeded) ├── notebooks/ │ └── nb8.5_manuscript_build.ipynb # story/narration; calls src/ logic, no analysis buried in cells ├── logs/ │ └── pulls.jsonl # one line per pull: what, when, vintage, content hash ├── data-cards/ │ ├── hmda.md # public; CFPB vintage pinned │ └── fred.md # public; controls └── paper/ ├── main.tex # \input{tables/...} and \includegraphics{figures/...} — never pasted ├── slides.pdf # the 8-slide deck (Part A) ├── slides-notes.md # speaker notes, cut line, prepared answers ├── figures/ # event_study.pdf, spec_curve.pdf — GENERATED by 03_analysis.py └── tables/ # main_results.tex — GENERATED by 03_analysis.py ``` The rule from Lab 7 holds visibly: *logic in `src/`, story in `notebooks/`*, the `data/` tree git-ignored, the licensed-data wall enforced by `.gitignore`. Here HMDA is **public**, so it ships cached in `data/raw/` with its vintage pinned — but the same layout would keep a CRSP-based capstone's licensed bytes on GMU infrastructure and ship only the `src/01_pull_data.py` access script plus the hash log (Ch 8.5 §5, the asymmetry of reproducibility). ### The single entry point — `Makefile` ```makefile # Makefile — `make all` regenerates EVERYTHING from raw to paper/main.pdf. .PHONY: all data analysis paper clean all: paper data: ## pull (public HMDA + FRED) or load cache, then build the panel python src/01_pull_data.py python src/02_build_dataset.py analysis: data ## estimate CS ATT + robustness; write tables/ and figures/ (deterministic via SEED) python src/03_analysis.py paper: analysis ## compile the write-up, pulling in the just-built tables/figures cd paper && latexmk -pdf main.tex clean: ## delete ALL derived outputs so a rebuild is honest rm -rf data/processed/* paper/tables/* paper/figures/* paper/main.pdf ``` The equivalent shell entry point, for a reviewer without `make`: ```bash # run_all.sh — same pipeline: `bash run_all.sh` set -euo pipefail # stop on the first error; never silently half-build python src/01_pull_data.py python src/02_build_dataset.py python src/03_analysis.py # tables/figures written here, deterministically (one SEED) ( cd paper && latexmk -pdf main.tex ) echo "Rebuilt paper/main.pdf from raw. Compare to the committed PDF." ``` ### The fixed, named SEED — `config.py` ```python # config.py — one seed, imported everywhere randomness happens SEED = 20260815 # the conference date; any fixed int, just FIXED and named # in 03_analysis.py, at the top, before any draw: import numpy as np from config import SEED rng = np.random.default_rng(SEED) # modern Generator, NOT legacy np.random.seed # the placebo reassignment and the wild-cluster bootstrap both draw from this rng, # passed in explicitly so a reader sees exactly what randomness each step consumes. ``` Maya's analysis has two stochastic steps — the wild-cluster bootstrap on slide 5's standard errors and the placebo that reassigns fake adoption years — and **both** draw from this one `rng`. That is why her headline interval `[−2.8, −0.8]` comes out identical on a reviewer's machine; without the fixed seed it would drift run to run and the packet would, strictly, be irreproducible. The seed is recorded in the README and was fixed in the `pap-filed` commit (Ch 7.3), so it provably was not searched over for a prettier interval. ### Completed `REPLICATION-CHECKLIST.md` **The deck** - [x] 7 slides (1 title + 6 beats) in order (title · question · why · design · headline · robustness · contribution). → `paper/slides.pdf` - [x] Every title is a full-sentence claim; argument legible from titles alone. → list atop `paper/slides-notes.md` - [x] Design slide shows the identifying-assumption sentence (from `memo.md`), not the regression equation. - [x] Headline slide: one number + CI, as the event-study **figure**, human-units sentence ("about a fifth of the baseline gap"). - [x] Robustness slide: spec curve + placebo panel clustering near −1.8, one stability sentence. - [x] Contribution slide names what's new *and* the "we do not show loan terms / non-mortgage credit" non-claim. - [x] Every slide figure is the file `03_analysis.py` wrote into `paper/figures/` — none hand-pasted. *(verified: `slides.pdf` `\includegraphics`es `figures/event_study.pdf`)* **The talk** - [x] Rehearsed timed ×4; longest run 7:10, fits in seven. - [x] Cut line pre-decided (drop slide-5 placebo panel) — in `slides-notes.md`. - [x] Two "what about [confound]?" answers + one located "I don't know" written out (above). - [x] Fatal critique named (pre-existing denial-gap divergence); retreat position written (concede → descriptive association → RD as future work). **The packet** - [x] Six parts present: HMDA access script (public, ships cached) · `src/01–03` in run-order · README · `environment.yml`+`environment.lock.yml` · `config.py` SEED · `Makefile`+`run_all.sh`. - [x] README states public HMDA / public FRED; **no licensed data, no secrets** in repo or history. → `git ls-files | grep -E '\.env|\.parquet|crsp' ` returns nothing. - [x] SEED is one named constant via `np.random.default_rng(SEED)`, threaded through bootstrap + placebo, recorded in README. - [x] Every `paper/` table and figure is code-generated and `\input`/`\includegraphics`'d — none pasted. - [x] **`make clean && make all`** rebuilt `paper/main.pdf`; `git diff --stat` on the PDF and on `tables/`/`figures/` shows **no change** → no hand-edits. *(the honesty test, passed)* - [x] Verified on a **fresh clone** in a scratch dir, not the working copy: `git clone … && conda env create -f environment.yml && conda activate capstone && make all` → identical PDF. --- ## INSTRUCTOR GRADING NOTES (keyed to the PS 8.5 rubric) These notes show *why* this deliverable earns an A and where a weaker submission would lose points. ### Part 1 — The six-beat arc (35 pts) → **Excellent** All six beats are present, in order, one per slide, and — the A-grade move — each slide does *its* one job and refuses the others. **(a) Question (5):** one sentence, no literature dump; the title *is* the question. **(b) Why (5):** a concrete stake with a scale number, placed *after* the question, not before — the single most common ordering error (opening abstract loses the room in thirty seconds) is avoided. **(c) Design (8 — the spine):** the slide shows the *identifying-assumption sentence* and a cohort-mean picture, **not the regression equation** — exactly right, because nobody reads a regression in sixty seconds; the spec is reduced to one plain-words strip. This is where weak decks collapse: a design slide that is a wall of $\beta$'s and fixed-effect notation cannot be read live, and the talk dies at the first hard question. **(d) Headline (6):** one number, *with its confidence interval*, as a figure, interpreted in human units ("a fifth of the baseline gap"). **(e) Robustness (6):** a spec curve that shows stability as a *shape*, plus a placebo — not a narrated list of twelve checks. **(f) Contribution (5):** the "what we do *not* claim" clause is present and load-bearing. *Where points are lost on a typical submission:* a "Results" slide that is the paper's regression table screenshotted (unreadable from row eight — auto-deduction on Part 2b); a headline with a point estimate and no interval (reads as evasive); a design slide showing the equation instead of the assumption; motivation before the question. Each is scored down on the specific beat. ### Part 2 — One idea per slide + table-to-figure (20 pts) → **Excellent** **(a) Title test (8):** every title is a full-sentence claim, and the argument is legible from the titles alone — the deliverable even prints the seven-title list so a grader can verify it in ten seconds. A deck with noun-phrase titles ("Background," "Results," "Robustness") loses this part wholesale, because the titles carry no argument. **(b) Translation (12):** the student names *which* translation each slide uses and *why* — the DiD result becomes an **event-study plot** (flat-then-jump, which argues parallel trends for free), the robustness becomes a **coefficient/spec plot** (stability as a tight cluster), and the precise table is demoted to a backup slide for the fourth-decimal questioner. The figures are stripped to load-bearing elements. Critically, the deliverable verifies the slide figure *is the file the code generated* — closing the loop with Part 4 so the slide and the packet cannot drift apart. *Where points are lost:* showing the table on the main slide instead of translating it; a figure with unreadable ticks, dead legends, and gridlines; or — the silent failure that surfaces in Part 4 — a slide figure that was exported by hand and no longer matches what `make all` produces. ### Part 3 — Defending identification (15 pts) → **Excellent** **(a) Two prepared confound answers (6):** the four-part structure (threat → plausible → what-I-did → residual) is followed exactly, drawn straight from the Ch 7.5 threats table — the student is reading from a script authored when calm, not improvising. The pre-trend answer concedes the worry is *central and plausible* before answering, which reads as more credible than a confident dismissal. **(b) Located "I don't know" (5):** the auto-loan answer is the chapter's gold standard — it admits the limit, explains *why* (different supervisor), and names what would resolve it (re-run there), rather than bluffing a number. **(c) Fatal vs. survivable + retreat (4):** the student correctly identifies that *attacks on the comparison are fatal* (a real pre-trend) while *attacks on a number are survivable* (clustering), maps this to Ch 7.5 §4's testable-vs-arguable, and — the mature move — writes the retreat to a descriptive claim *before* walking in. *Where points are lost:* answering a confound question by dismissing it ("we control for that"); bluffing the "I don't know"; or misclassifying a fatal critique as survivable and trying to re-cluster their way out of a broken comparison in front of a room that can see it is broken. ### Part 4 — The one-click packet (30 pts) → **Excellent** All six parts present and each *verified*, not merely asserted. **(a) Data (5):** correctly ships public HMDA cached with a pinned vintage and states the public/licensed split honestly. **(b) Run-order (5):** `01→02→03` numbering makes the order unambiguous; logic in `src/`, story in `notebooks/`. **(c) README (4):** the how-to-run has collapsed to one command. **(d) Env (4):** both `environment.yml` and `environment.lock.yml` committed; the student can say the "rebuild that and you rebuild my numbers" sentence. **(e) SEED (5):** a single named `SEED` in `config.py`, threaded through *both* stochastic steps (bootstrap and placebo) via `np.random.default_rng`, recorded in the README, fixed in the `pap-filed` commit so it provably was not searched over. **(f) Entry point (7):** a `Makefile` *and* a `run_all.sh`, with the figures/tables **generated by `03_analysis.py`** and `\input`/`\includegraphics`'d — never pasted. The decisive A-grade evidence is the **honesty test, run and recorded**: `make clean && make all` rebuilt the PDF byte-identical, *and* the student re-verified on a **fresh clone** in a scratch directory — the true stranger's-machine test, not the working copy where stale derived files hide hand-edits. *Where points are lost:* the single most common — and most penalized — failure is a number on a slide that the rebuild does not reproduce, because it was hand-copied from a notebook into the manuscript and has since drifted from the code. That breaks the one-click chain and fails the part regardless of how polished the deck is. Also penalized: a seed scattered ad-hoc (or the legacy `np.random.seed`), a missing lock file ("it worked on my machine"), or licensed data / a `.env` committed to history (a Lab 7 Step-5 violation that is serious independent of the grade). ### One-line grading heuristic The two highest-signal questions for any PS 8.5 submission: **(1)** Can a reader reconstruct the entire argument from the slide *titles* alone, with the design slide showing the *assumption* (not the equation) and the headline showing a number *with its interval as a figure*? **(2)** Does **`make clean && make all`, on a fresh clone**, regenerate the exact number the headline slide claims? A student who clears both has internalized the camp's thesis — *reproducibility is the form credibility takes* — because the talk *invites* verification ("here's the command that checks me") instead of asking to be believed, and the packet makes good on the invitation. This exemplar clears both cleanly, which is what an A looks like. --- *End of model deliverable for PS 8.5. The exemplar is Maya's HMDA fair-lending DiD — the project she carried from Lab 4 through the Week-7 PAP, identification memo (Ch 7.5), and robustness battery (Ch 8.2) into this final talk and packet. All magnitudes (the ≈ −1.8 pp ATT, the [−2.8, −0.8] interval, the ~9-point baseline gap, the "fifth of the gap") are **illustrative**, presented as a worked talk example, not a claimed estimate; the real numbers come out of `make all` on the pinned, public CFPB HMDA vintage recorded in `data-cards/hmda.md`. The packet standard is Appendix D + Lab 7; the staggered-DiD machinery is Callaway & Sant'Anna (2021) and Goodman-Bacon (2021), as built in Ch 4.2 and Lab 4; the SEED/`make`/one-command discipline is Ch 8.5 §5 and Gentzkow & Shapiro (2014). No secrets or keys appear; the `SEED = 20260815` constant and `config.py` pattern are illustrative.*

PART A — THE DECK (paper/slides.pdf, 7 slides: 1 title + 6 beats)