PS 8.5 — Your 8-Minute Poster Talk + the One-Click Replication Packet

Course: ASSIP 8-Week Empirical Finance Camp · Week 8 · Problem Set 8.5 Covers: Ch 8.5 (The 8-Minute Poster Talk & the Replication Packet), standing on Lab 7 (the repository that is the packet), Lab 8 (the final poster compile and dry-run), Ch 7.5 (the threats table you defend under questions), Ch 8.2 (the robustness that fills your robustness panel), and Appendix D (the packet and style standard). Type: Project deliverable, not a numeric problem set. This is the last thing you build in ASSIP, and it is the same project you have carried since Week 7 — your question, your data, your pre-registered design — seen from its two public surfaces at the Aug 12, 2026 ASSIP Research Symposium: the 8-minute talk you give standing beside your poster (or on the symposium’s virtual-breakout track if you are one of Prof. Gao’s remote interns) and the folder a visitor clones from the QR code on the poster. There is no answer key with “the right number,” because what is graded is not a finding but two crafts: whether your eight minutes make one argument land in front of a symposium audience, and whether a stranger can rebuild every number you claimed by typing one command after scanning your QR code.

Total: 100 points. Point values are stated per part. A model deliverable — an A-grade panel-by-panel poster outline and a completed replication-packet checklist for one cast project (Maya’s HMDA fair-lending DiD), with instructor grading notes — is in Appendix E (E-w8-ps8.5-solutions.md). Read it after you draft your own, the way Ch 8.5 walked through the cast talks: not to copy, but to calibrate.

A note on what this assignment is for. Ch 8.5 stated the thesis as bluntly as it deserves, and this problem set is where you earn it: reproducibility is not a virtue you bolt on at the end — it is the form your credibility takes once you stop talking and the skeptic is alone with your folder. A correct result nobody can follow on a slide, or nobody can rebuild on their own machine, is — to the person in the room — indistinguishable from a guess. So the single most valuable outcome of this problem set is not a pretty deck. It is the moment, during the make clean && make all honesty test, when the rebuilt PDF comes back with a number that does not match the one on your slide — and you find the hand-edit before a reviewer does. If that happens, the assignment did its most important job.

What you submit

Three artifacts, all living inside the capstone repository you stood up in Lab 7 — there is no separate upload, because (Lab 7, Step 2) the repository is the submission:

The poster — paper/poster.pdf (48”×36” landscape, built with beamerposter or PowerPoint per Lab 8 §1.2), committed to your repo. Six content panels in the reading order of the six-beat arc (Ch 8.5 §1), with the JSSR-ready abstract \inputed into the title block (Lab 8 §1.1) and a scannable QR code resolving to your public repo in the contribution panel.
The talk notes — paper/talk-notes.md: your speaker notes, one short block per poster panel, including your pre-decided cut line (which panel you skim if you reach minute six still on the headline) and your prepared answers to the three hardest questions you expect. If you also use companion slides on the virtual-breakout track, ship paper/slides.pdf (≤ 7 slides) too — the same six-beat arc, same titles, just a screen-optimized layout.
The replication packet — your finished Lab 7 repository, with the six load-bearing parts assembled (Ch 8.5 §5) and a single make all / run_all.py entry point that regenerates every table, figure, the JSSR abstract, and the poster PDF from raw inputs. Plus REPLICATION-CHECKLIST.md at the repo root: the completed packet checklist (Part 4 below), each box checked with the command output or file that proves it.

The poster is the visual artifact and the talk notes are short by design; the packet is where most of the work already lives, finished. All three are graded together, because they are the same project: the poster + talk claim the numbers, the packet (reached via the QR code) proves them.

Part 1 — The 8-minute deck: the six-beat arc (35 points)

Build a deck that hits the six beats of Ch 8.5 §1, in order, one beat per slide, with a title slide making seven in total. Eight minutes has room for exactly one idea defended cleanly; the slides exist to keep you and the room on that one idea. Budget roughly one minute per slide and target a talk that fits in seven minutes in rehearsal, not eight (§3) — the eighth minute is insurance, not content.

Each beat below is one slide, and each is graded on whether it does its one job and no other:

(a) The question slide (5 pts). One sentence a smart sixteen-year-old could repeat back. Not the literature, not the motivation — the question. (Devon’s: “Does the launch of a spot Bitcoin ETF reduce the volatility of the underlying token?”) The title is the question. No three-paragraph history; the room cannot hold context it has not yet been given a reason to want.
(b) The why-it-matters slide (5 pts). One concrete stake a stranger feels in thirty seconds — concrete before abstract, a number before a Greek letter (CONVENTIONS §1). Priya’s opens not with the IPCC but with: “if wildfire risk is mispriced, insurers either go insolvent or abandon whole counties.” This comes after the question, never before.
(c) The design / identification slide (8 pts — the spine). Your identifying-assumption sentence from Ch 7.5, on the screen, in the contract form: “Our estimate is credible as long as [the assumption], which we defend by [the design].” This is not your regression equation — nobody reads a regression equation in sixty seconds. It is a slide about why the comparison is fair, ideally with one small picture. If the room leaves understanding only one slide, it must be this one; this is where every hard question lands.
(d) The headline-result slide (6 pts). One number, with its uncertainty, shown as a picture, not a table — and one sentence of interpretation in human units (“1.8 percentage points, about a fifth of the baseline gap,” not “\hat\beta = 0.018, t = 3.2”). A confidence interval is part of the result, not an afterthought; a point estimate with no interval reads as naïve or evasive.
(e) The robustness slide (6 pts). The visual answer to “is your result fragile?” — a specification curve or a small grid of alternative specs, all clustering near the headline (this is where Ch 8.2 pays off). You do not narrate every check; you show they exist and agree, in one sentence: “the result is stable across [the three choices I worried about most].” This pre-empts the §3 questions before anyone asks.
(f) The contribution slide (5 pts). One sentence naming what the room now knows that it did not eight minutes ago — and one clause naming what you are not claiming (“we show X, under assumption Y; we do not show Z”). That last clause is not weakness; it is the single most credibility-building sentence in the talk. Then stop.

Grading note on slide discipline. Every slide is also scored on the one-idea-per-slide rule (Part 2). A slide with two ideas, or a noun-phrase title (“Results”) instead of a full-sentence claim, loses points even if its content is correct — because in a live talk a crowded slide costs you the room.

Part 2 — One idea per slide, and the table-to-figure translation (20 points)

Two rules govern what goes on each slide; both respect the fact that a live audience reads with their ears, not their eyes.

(a) One idea per slide — and the title test (8 pts). A slide is a single claim with a single piece of evidence for it. The test is brutal and useful: if you cannot write the slide’s one idea as its title — a full sentence, not a noun phrase — the slide is doing too much. “Results” is not a title; “The denial gap falls 1.8 points after the regulation, and the fall is concentrated in algorithmic lenders” is a title, because it is the idea. A reader who reads only your titles should get the whole argument. For full marks, every one of your eight titles must pass this test; submit them as a list at the top of slides-notes.md so a grader can read the argument from titles alone.

(b) The table-to-figure translation (12 pts). This is the highest-leverage skill in the chapter. Your paper’s results live in a precise regression table (the four-decimal, standard-errors-in-parentheses object Appendix D specifies). That table is correct and unreadable from row eight of a lecture hall in four seconds. So you do not show it — you translate it into a figure for the talk, and leave the table in the paper and on a backup slide. The translation is mechanical (Ch 8.5 §2); pick the one your design demands and state, in slides-notes.md, which translation your headline and robustness slides use and why:

Your result is…	The talk figure is…	The eye reads…
a single coefficient + CI	a point with an error bar	“different from zero?” and “how precise?”
a coefficient across specs	a coefficient plot (dot-and-whisker per spec, reference line)	stability as a shape — a tight cluster
a difference-in-differences / event study	an event-study plot (effects by period: flat before, jump after)	the parallel-trends defense, made visible
a regression discontinuity	a binned scatter with the fitted jump at the cutoff	the discontinuity you can see
a subgroup pattern	a small-multiples bar chart	the gap is bigger here than there

The principle under all of them: a table is for a reader who can stop and study; a figure is for a listener who cannot. Same number, two encodings, chosen for the medium — this is not dumbing down; the error-bar slide is exactly as precise as the table. One mercy rule for the figures themselves, also graded: a slide figure needs axis labels readable from the back of the room and no more than is load-bearing — strip the gridlines, the dead legend, the third decimal on the ticks. Non-negotiable, and tied to Part 4: the figure on your slide must be the same file your code generated into paper/figures/, never a hand-drawn or hand-pasted picture. If your slide figure and your packet figure can drift apart, the chain is already broken.

Part 3 — Defending identification under questions (15 points)

The talk is eight minutes; the questions are where the talk is judged. A polished eight minutes followed by a flailing answer to the first question undoes the polish. The good news: you already did the work — the questions you will get are the rows of your Ch 7.5 threats-and-responses table. You are not improvising; you are reading from a script you authored when you were calm. In slides-notes.md, write out three prepared answers:

(a) Two prepared “what about [confound]?” answers (6 pts). Every hard empirical question has the shape “isn’t your estimate just picking up [some confounder]?” — which is your identifying assumption attacked at a named weak point you wrote yourself (column 1 of the threats table). For your two most likely such questions, write the four-part answer in order: the threat (their question), why it’s plausible (so you don’t dismiss it), what you did about it (column 3 — the test or design choice), and the residual concern (column 4 — what honestly remains). The template: “Yes, that’s the central worry — if [confounder] were driving this we’d see [signature]; we checked, and we see [reassuring result]; what I still can’t fully rule out is [residual], though [bound on how bad it could be].”

(b) One located “I don’t know” (5 pts). You will get a question you cannot answer; everyone does. Bluffing costs you more credibility than the gap ever would — a room of professionals smells it instantly. Write one honest, located “I don’t know”: not a bare shrug but “I don’t know — that’s outside what my design can speak to, because [reason]; my guess is [direction], but I’d want to [the specific check] before claiming it.” (Maya, asked whether her fair-lending effect generalizes to auto loans, does not invent a number; she names the mechanism — examiner attention — and says auto lending has a different supervisor, so she would re-run it there before extending the estimate.)

(c) Fatal vs. survivable, and your retreat position (4 pts). State, in one sentence, which single critique would be fatal to your causal claim — a fatal critique attacks the comparison; a survivable one attacks a number (Ch 8.5 §4, mapping to Ch 7.5 §4’s testable-vs-arguable). Re-clustering or winsorizing differently is survivable: “we checked; it doesn’t move.” A pre-existing divergence between treated and control, or an instrument that affects the outcome directly, is fatal: no re-clustering saves you. Then write your retreat position in advance: if the fatal critique lands and the attacker is right, you concede at the level it was made and retreat to what survives — almost always “this is a robust pattern that future work with [the better design] should test causally,” which is a real, honest, publishable contribution. Know this before you walk in; it is what the residual-concern column was for.

Part 4 — The one-click replication packet (30 points)

Now the second surface, the one the skeptic touches alone. The bar from Appendix D and Lab 7 is exact and unforgiving: a stranger, on a machine that has never seen your project, runs one command and watches every table and figure in your paper regenerate from raw data to final PDF. Not “mostly.” Not “after I email a fix.” One command, everything. Assemble the six load-bearing parts on top of your Lab 7 repo and verify each:

(a) The data — or a data-access script (5 pts). Public data ships in the packet, cached, with its vintage pinned. Licensed data (CRSP, Compustat, anything from WRDS) does not ship — the license forbids redistribution (CONVENTIONS §5) — so what ships instead is the data-access script in src/ (the bounded, pinned query that reconstructs it for anyone with the same access) plus the logs/pulls.jsonl log with content hashes. This is Lab 7’s asymmetry of reproducibility: public data fully re-runnable, licensed data recipe-reproducible. Your README states honestly which is which. (Maya’s HMDA is public, so it ships — many capstones’ licensed sources do not.)
(b) Code in run-order (5 pts). Scripts in src/ and notebooks in notebooks/, named or numbered so the order is unambiguous (01_pull.py, 02_build.py, 03_analysis.py). Lab 7’s rule holds: logic in src/, story in notebooks/. A reviewer should never have to guess which script runs first.
(c) A README (4 pts). The what / why / how-to-run document from Lab 7, but the “how to run” section has now collapsed to almost nothing, because the answer is one command. It still names the data sources, the secrets a user must export, and the licensed-vs-public split.
(d) A pinned environment file (4 pts). Both files from Lab 7: human-readable environment.yml (Python 3.11 + the CONVENTIONS §5 stack) and machine-exact environment.lock.yml. The sentence you must be able to say: “my results were produced under environment.lock.yml; rebuild that and you rebuild my numbers.”
(e) A fixed, named SEED (5 pts). Anywhere your code uses randomness — a bootstrap, a train/test split, a permutation or placebo test, a simulation — it must be seeded from a single named constant set once at the top of the run, not scattered ad-hoc. Without it, your bootstrap CI is different every run and your headline interval is, strictly, irreproducible. Use the modern generator, not the legacy global:
```
# config.py — one seed, imported everywhere randomness happens
SEED = 20260815  # the conference date; any fixed int, just FIXED and named

# at the top of every script/notebook that uses randomness:
import numpy as np
from config import SEED
rng = np.random.default_rng(SEED)   # a seeded Generator, not legacy np.random.seed
```
Record the seed in the README so a reviewer knows it was fixed before results, not tuned after. (A seed you searched over for a prettier interval is p-hacking with extra steps, Ch 7.3’s garden of forking paths in disguise; the pap-filed tag is what proves you didn’t.)
(f) A single make all / run_all.sh entry point (7 pts). The one-click. A single command that runs the whole pipeline in order — pull (or load cached), build the dataset, run every estimation, regenerate every figure and table, compile the paper to PDF — so a fresh clone goes from raw inputs to finished document with no human in the loop. The non-negotiable property: the figures and tables in your paper are generated by the code, never pasted by hand. 03_analysis.py writes paper/figures/event_study.pdf and paper/tables/main_results.tex, and the LaTeX \includegraphics/\inputs them. The instant a human copies a number from a notebook into the manuscript, the chain breaks and the packet is no longer one-click.

The honesty test, which you must run and record the result of: make clean && make all. Delete every derived file and rebuild from nothing. If the paper comes back byte-identical, the packet is real. If a number drifts, you just found a hand-edit — find it now.

Submission checklist — `REPLICATION-CHECKLIST.md`

Paste this into REPLICATION-CHECKLIST.md at your repo root and check a box only when you can point to the command output or file that proves it — this is exactly the audit a referee (or Prof. Gao, at the conference) would run.

The deck - [ ] Exactly 7 slides (1 title + 6 beats) hitting the beats in order (title · question · why · design · headline · robustness · contribution), one beat per slide. - [ ] Every slide title is a full-sentence claim, not a noun phrase; the argument is legible from the titles alone (list pasted atop slides-notes.md). - [ ] The design/identification slide shows the identifying-assumption sentence (Ch 7.5), not the regression equation. - [ ] The headline slide shows one number with its confidence interval, as a figure, with a human-units interpretation sentence. - [ ] The robustness slide is a specification curve or coefficient grid clustering near the headline (Ch 8.2), with a one-sentence stability claim. - [ ] The contribution slide names what is now known and what you are not claiming. - [ ] Every slide figure is the same file your code generated into paper/figures/ — none hand-drawn or hand-pasted.

The talk - [ ] Rehearsed timed at least three times; fits in seven minutes, not eight. - [ ] A cut line is pre-decided and written in slides-notes.md (which slide you skip if you are behind — never the design or headline slide). - [ ] Two prepared “what about [confound]?” answers (threat → plausible → what you did → residual) and one located “I don’t know,” written out. - [ ] Your one fatal critique is named, and your retreat position is written before you walk in.

The packet - [ ] All six parts present: data/access-script · code in run-order · README · pinned env (environment.yml + environment.lock.yml) · fixed named SEED · single make all/run_all.sh. - [ ] README states the public-vs-licensed split honestly; no licensed data and no secrets anywhere in the repo or its history (git ls-files is clean — Lab 7, Step 5). - [ ] The SEED is a single named constant, threaded through every stochastic step via np.random.default_rng(SEED), and recorded in the README. - [ ] Every table and figure in paper/ is generated by code and \input/\includegraphics’d — none pasted by hand. - [ ] make clean && make all rebuilds the paper from nothing and the PDF is identical to the committed one. (If not: you found a hand-edit — fix it, then re-run.) - [ ] Verified on a fresh clone, not just your working copy — the true “stranger’s machine” test (Lab 8).

How this is graded

This problem set is graded on the two crafts, weighted as the parts show: the deck (Parts 1–2, 55 pts) on whether one argument lands cleanly under slide discipline; question-handling (Part 3, 15 pts) on whether you defend identification honestly rather than bluff; and the packet (Part 4, 30 pts) on whether make clean && make all actually rebuilds your claims. The highest-signal single check is the honesty test: a packet whose rebuild does not reproduce the slide’s headline number fails the part regardless of how polished the deck is, because — Ch 8.5 §6 — a result a stranger cannot reconstruct is not yet a result; it is a claim about a result. The model deliverable in Appendix E shows the A-grade standard for one cast project, annotated with the rubric levels each piece hits.

The bridge — and the close of the camp

This is the last deliverable, and it closes the arc the whole camp walked. Look back at Week 1: the first thing you did was compute an average return of about eight percent, and it felt like an answer. It was the first question. You learned the eight percent had a standard error, that the standard error rested on assumptions you could not see, that a “significant” sign could be an artifact of how many specifications you quietly tried. Week 2 turned correlation into regression and the ghost of omitted-variable bias; Weeks 3–4 turned “we control for X” into designs — IV, RD, DiD, synthetic control — each a specific argument for why a comparison is fair; Week 5 taught you to read a paper as a skeptic; Week 6 sharpened the inference; Week 7 made it yours — an idea, data pulled reproducibly, a pre-analysis plan tagged before you could fool yourself, an identification memo that named every threat. Week 8 made it communicable and checkable: written up with calibrated verbs (Ch 8.3), stress-tested for robustness (Ch 8.2), peer-reviewed (Ch 8.4), and now — today — defended in eight minutes and packaged so a stranger can rebuild every claim with one command.

The single sentence this whole camp existed to let you say truthfully: “Here is what I found, here is exactly why you should believe it, and here is the command that lets you check me.” When you build this deck and this packet, you can say it. The eight percent was the toe in the water. This is the swim.

End of PS 8.5. The model deliverable is in book/appendices/E-solutions-manual/E-w8-ps8.5-solutions.md. Build your own deck and finish your own packet first; read the exemplar to calibrate, not to copy. The skeptic who clones a stranger’s repo and types make all, and the author who makes their own rebuild honestly, are the same person — this assignment is you doing to your own work, on purpose, what a referee will do to it next.

# PS 8.5 — Your 8-Minute Poster Talk + the One-Click Replication Packet **Course:** ASSIP 8-Week Empirical Finance Camp · Week 8 · Problem Set 8.5 **Covers:** Ch 8.5 (The 8-Minute Poster Talk & the Replication Packet), standing on Lab 7 (the repository that *is* the packet), Lab 8 (the final poster compile and dry-run), Ch 7.5 (the threats table you defend under questions), Ch 8.2 (the robustness that fills your robustness panel), and Appendix D (the packet and style standard). **Type:** Project deliverable, not a numeric problem set. This is the last thing you build in ASSIP, and it is the same project you have carried since Week 7 — your question, your data, your pre-registered design — seen from its two public surfaces at the **Aug 12, 2026 ASSIP Research Symposium**: the **8-minute talk you give standing beside your poster** (or on the symposium's virtual-breakout track if you are one of Prof. Gao's remote interns) and the **folder a visitor clones from the QR code on the poster**. There is no answer key with "the right number," because what is graded is not a finding but two *crafts*: whether your eight minutes make one argument land in front of a symposium audience, and whether a stranger can rebuild every number you claimed by typing one command after scanning your QR code. **Total: 100 points.** Point values are stated per part. A model deliverable — an A-grade panel-by-panel poster outline and a completed replication-packet checklist for one cast project (Maya's HMDA fair-lending DiD), with instructor grading notes — is in Appendix E (`E-w8-ps8.5-solutions.md`). Read it *after* you draft your own, the way Ch 8.5 walked through the cast talks: not to copy, but to calibrate. A note on what this assignment is for. Ch 8.5 stated the thesis as bluntly as it deserves, and this problem set is where you earn it: **reproducibility is not a virtue you bolt on at the end — it is the form your credibility takes once you stop talking and the skeptic is alone with your folder.** A correct result nobody can follow on a slide, or nobody can rebuild on their own machine, is — to the person in the room — indistinguishable from a guess. So the single most valuable outcome of this problem set is *not* a pretty deck. It is the moment, during the `make clean && make all` honesty test, when the rebuilt PDF comes back with a number that does not match the one on your slide — and you find the hand-edit *before* a reviewer does. If that happens, the assignment did its most important job. --- ## What you submit Three artifacts, all living inside the capstone repository you stood up in Lab 7 — there is no separate upload, because (Lab 7, Step 2) the repository *is* the submission: 1. **The poster** — `paper/poster.pdf` (48"×36" landscape, built with `beamerposter` or PowerPoint per Lab 8 §1.2), committed to your repo. Six content panels in the reading order of the six-beat arc (Ch 8.5 §1), with the JSSR-ready abstract `\input`ed into the title block (Lab 8 §1.1) and a scannable QR code resolving to your **public** repo in the contribution panel. 2. **The talk notes** — `paper/talk-notes.md`: your speaker notes, one short block per poster panel, including your **pre-decided cut line** (which panel you skim if you reach minute six still on the headline) and your **prepared answers** to the three hardest questions you expect. If you also use companion slides on the virtual-breakout track, ship `paper/slides.pdf` (≤ 7 slides) too — the same six-beat arc, same titles, just a screen-optimized layout. 3. **The replication packet** — your finished Lab 7 repository, with the **six load-bearing parts** assembled (Ch 8.5 §5) and a single `make all` / `run_all.py` entry point that regenerates every table, figure, the JSSR abstract, **and the poster PDF** from raw inputs. Plus `REPLICATION-CHECKLIST.md` at the repo root: the completed packet checklist (Part 4 below), each box checked with the command output or file that proves it. The poster is the visual artifact and the talk notes are short by design; the packet is where most of the work already lives, finished. All three are graded together, because they are the same project: the poster + talk *claim* the numbers, the packet (reached via the QR code) *proves* them. --- ## Part 1 — The 8-minute deck: the six-beat arc (35 points) Build a deck that hits the six beats of Ch 8.5 §1, in order, **one beat per slide**, with a title slide making seven in total. Eight minutes has room for exactly one idea defended cleanly; the slides exist to keep you and the room on that one idea. Budget roughly one minute per slide and target a talk that fits in **seven** minutes in rehearsal, not eight (§3) — the eighth minute is insurance, not content. Each beat below is one slide, and each is graded on whether it does *its* one job and no other: - **(a) The question slide (5 pts).** One sentence a smart sixteen-year-old could repeat back. Not the literature, not the motivation — the question. (Devon's: *"Does the launch of a spot Bitcoin ETF reduce the volatility of the underlying token?"*) The title *is* the question. No three-paragraph history; the room cannot hold context it has not yet been given a reason to want. - **(b) The why-it-matters slide (5 pts).** One concrete stake a stranger feels in thirty seconds — *concrete before abstract, a number before a Greek letter* (CONVENTIONS §1). Priya's opens not with the IPCC but with: *"if wildfire risk is mispriced, insurers either go insolvent or abandon whole counties."* This comes *after* the question, never before. - **(c) The design / identification slide (8 pts — the spine).** Your identifying-assumption sentence from Ch 7.5, on the screen, in the contract form: *"Our estimate is credible as long as [the assumption], which we defend by [the design]."* This is **not** your regression equation — nobody reads a regression equation in sixty seconds. It is a slide about *why the comparison is fair*, ideally with one small picture. If the room leaves understanding only one slide, it must be this one; this is where every hard question lands. - **(d) The headline-result slide (6 pts).** **One** number, with its uncertainty, shown as a *picture*, not a table — and one sentence of interpretation in human units ("1.8 percentage points, about a fifth of the baseline gap," not "$\hat\beta = 0.018$, $t = 3.2$"). A confidence interval is part of the result, not an afterthought; a point estimate with no interval reads as naïve or evasive. - **(e) The robustness slide (6 pts).** The visual answer to "is your result fragile?" — a specification curve or a small grid of alternative specs, all clustering near the headline (this is where Ch 8.2 pays off). You do not narrate every check; you show they exist and agree, in one sentence: *"the result is stable across [the three choices I worried about most]."* This pre-empts the §3 questions before anyone asks. - **(f) The contribution slide (5 pts).** One sentence naming what the room now knows that it did not eight minutes ago — *and* one clause naming what you are **not** claiming ("we show X, under assumption Y; we do not show Z"). That last clause is not weakness; it is the single most credibility-building sentence in the talk. Then stop. **Grading note on slide discipline.** Every slide is also scored on the **one-idea-per-slide** rule (Part 2). A slide with two ideas, or a noun-phrase title ("Results") instead of a full-sentence claim, loses points even if its content is correct — because in a live talk a crowded slide costs you the room. --- ## Part 2 — One idea per slide, and the table-to-figure translation (20 points) Two rules govern what goes *on* each slide; both respect the fact that a live audience reads with their ears, not their eyes. **(a) One idea per slide — and the title test (8 pts).** A slide is a single claim with a single piece of evidence for it. The test is brutal and useful: if you cannot write the slide's one idea as its *title* — a full sentence, not a noun phrase — the slide is doing too much. "Results" is not a title; *"The denial gap falls 1.8 points after the regulation, and the fall is concentrated in algorithmic lenders"* is a title, because it *is* the idea. A reader who reads only your titles should get the whole argument. For full marks, every one of your eight titles must pass this test; submit them as a list at the top of `slides-notes.md` so a grader can read the argument from titles alone. **(b) The table-to-figure translation (12 pts).** This is the highest-leverage skill in the chapter. Your paper's results live in a precise regression table (the four-decimal, standard-errors-in-parentheses object Appendix D specifies). That table is *correct* and *unreadable from row eight of a lecture hall in four seconds.* So you do not show it — you **translate it into a figure** for the talk, and leave the table in the paper and on a backup slide. The translation is mechanical (Ch 8.5 §2); pick the one your design demands and state, in `slides-notes.md`, *which* translation your headline and robustness slides use and *why*: | Your result is… | The talk figure is… | The eye reads… | |---|---|---| | a single coefficient + CI | a **point with an error bar** | "different from zero?" and "how precise?" | | a coefficient across specs | a **coefficient plot** (dot-and-whisker per spec, reference line) | stability as a *shape* — a tight cluster | | a difference-in-differences / event study | an **event-study plot** (effects by period: flat before, jump after) | the parallel-trends defense, made visible | | a regression discontinuity | a **binned scatter with the fitted jump at the cutoff** | the discontinuity you can *see* | | a subgroup pattern | a **small-multiples bar chart** | the gap is bigger here than there | The principle under all of them: *a table is for a reader who can stop and study; a figure is for a listener who cannot.* Same number, two encodings, chosen for the medium — this is not dumbing down; the error-bar slide is exactly as precise as the table. One mercy rule for the figures themselves, also graded: a slide figure needs axis labels readable from the back of the room and *no more than is load-bearing* — strip the gridlines, the dead legend, the third decimal on the ticks. **Non-negotiable, and tied to Part 4:** the figure on your slide must be the *same file your code generated* into `paper/figures/`, never a hand-drawn or hand-pasted picture. If your slide figure and your packet figure can drift apart, the chain is already broken. --- ## Part 3 — Defending identification under questions (15 points) The talk is eight minutes; the questions are where the talk is judged. A polished eight minutes followed by a flailing answer to the first question undoes the polish. The good news: you already did the work — **the questions you will get are the rows of your Ch 7.5 threats-and-responses table.** You are not improvising; you are reading from a script you authored when you were calm. In `slides-notes.md`, write out three prepared answers: **(a) Two prepared "what about [confound]?" answers (6 pts).** Every hard empirical question has the shape *"isn't your estimate just picking up [some confounder]?"* — which is your identifying assumption attacked at a named weak point you wrote yourself (column 1 of the threats table). For your **two most likely** such questions, write the four-part answer in order: the threat (their question), why it's plausible (so you don't dismiss it), what you did about it (column 3 — the test or design choice), and the residual concern (column 4 — what honestly remains). The template: *"Yes, that's the central worry — if [confounder] were driving this we'd see [signature]; we checked, and we see [reassuring result]; what I still can't fully rule out is [residual], though [bound on how bad it could be]."* **(b) One located "I don't know" (5 pts).** You will get a question you cannot answer; everyone does. Bluffing costs you more credibility than the gap ever would — a room of professionals smells it instantly. Write one honest, *located* "I don't know": not a bare shrug but *"I don't know — that's outside what my design can speak to, because [reason]; my guess is [direction], but I'd want to [the specific check] before claiming it."* (Maya, asked whether her fair-lending effect generalizes to auto loans, does not invent a number; she names the mechanism — examiner attention — and says auto lending has a different supervisor, so she would re-run it there before extending the estimate.) **(c) Fatal vs. survivable, and your retreat position (4 pts).** State, in one sentence, which single critique would be **fatal** to your causal claim — *a fatal critique attacks the comparison; a survivable one attacks a number* (Ch 8.5 §4, mapping to Ch 7.5 §4's testable-vs-arguable). Re-clustering or winsorizing differently is survivable: "we checked; it doesn't move." A pre-existing divergence between treated and control, or an instrument that affects the outcome directly, is fatal: no re-clustering saves you. Then write your **retreat position** in advance: if the fatal critique lands and the attacker is right, you concede at the level it was made and retreat to what survives — almost always *"this is a robust pattern that future work with [the better design] should test causally,"* which is a real, honest, publishable contribution. Know this before you walk in; it is what the residual-concern column was *for*. --- ## Part 4 — The one-click replication packet (30 points) Now the second surface, the one the skeptic touches alone. The bar from Appendix D and Lab 7 is exact and unforgiving: **a stranger, on a machine that has never seen your project, runs one command and watches every table and figure in your paper regenerate from raw data to final PDF.** Not "mostly." Not "after I email a fix." One command, everything. Assemble the **six load-bearing parts** on top of your Lab 7 repo and verify each: - **(a) The data — or a data-access script (5 pts).** Public data ships *in* the packet, cached, with its vintage pinned. Licensed data (CRSP, Compustat, anything from WRDS) does *not* ship — the license forbids redistribution (CONVENTIONS §5) — so what ships instead is the **data-access script** in `src/` (the bounded, pinned query that reconstructs it for anyone with the same access) plus the `logs/pulls.jsonl` log with content hashes. This is Lab 7's *asymmetry of reproducibility*: public data fully re-runnable, licensed data recipe-reproducible. Your README states honestly which is which. *(Maya's HMDA is public, so it ships — many capstones' licensed sources do not.)* - **(b) Code in run-order (5 pts).** Scripts in `src/` and notebooks in `notebooks/`, named or numbered so the order is unambiguous (`01_pull.py`, `02_build.py`, `03_analysis.py`). Lab 7's rule holds: *logic in `src/`, story in `notebooks/`.* A reviewer should never have to guess which script runs first. - **(c) A README (4 pts).** The *what / why / how-to-run* document from Lab 7, but the "how to run" section has now collapsed to almost nothing, because the answer is one command. It still names the data sources, the secrets a user must `export`, and the licensed-vs-public split. - **(d) A pinned environment file (4 pts).** Both files from Lab 7: human-readable `environment.yml` (Python 3.11 + the CONVENTIONS §5 stack) and machine-exact `environment.lock.yml`. The sentence you must be able to say: *"my results were produced under `environment.lock.yml`; rebuild that and you rebuild my numbers."* - **(e) A fixed, named SEED (5 pts).** Anywhere your code uses randomness — a bootstrap, a train/test split, a permutation or placebo test, a simulation — it must be seeded from a *single named constant* set once at the top of the run, not scattered ad-hoc. Without it, your bootstrap CI is *different every run* and your headline interval is, strictly, irreproducible. Use the modern generator, not the legacy global: ```python # config.py — one seed, imported everywhere randomness happens SEED = 20260815 # the conference date; any fixed int, just FIXED and named # at the top of every script/notebook that uses randomness: import numpy as np from config import SEED rng = np.random.default_rng(SEED) # a seeded Generator, not legacy np.random.seed ``` Record the seed in the README so a reviewer knows it was fixed *before* results, not tuned after. (A seed you *searched over* for a prettier interval is p-hacking with extra steps, Ch 7.3's garden of forking paths in disguise; the `pap-filed` tag is what proves you didn't.) - **(f) A single `make all` / `run_all.sh` entry point (7 pts).** The one-click. A single command that runs the whole pipeline *in order* — pull (or load cached), build the dataset, run every estimation, regenerate every figure and table, compile the paper to PDF — so a fresh clone goes from raw inputs to finished document with no human in the loop. The non-negotiable property: **the figures and tables in your paper are *generated by the code*, never pasted by hand.** `03_analysis.py` writes `paper/figures/event_study.pdf` and `paper/tables/main_results.tex`, and the LaTeX `\includegraphics`/`\input`s them. The instant a human copies a number from a notebook into the manuscript, the chain breaks and the packet is no longer one-click. The honesty test, which you must run and record the result of: **`make clean && make all`.** Delete every derived file and rebuild from nothing. If the paper comes back byte-identical, the packet is real. If a number drifts, you just found a hand-edit — find it now. --- ## Submission checklist — `REPLICATION-CHECKLIST.md` Paste this into `REPLICATION-CHECKLIST.md` at your repo root and check a box only when you can point to the command output or file that proves it — this is exactly the audit a referee (or Prof. Gao, at the conference) would run. **The deck** - [ ] Exactly 7 slides (1 title + 6 beats) hitting the beats in order (title · question · why · design · headline · robustness · contribution), one beat per slide. - [ ] Every slide title is a *full-sentence claim*, not a noun phrase; the argument is legible from the titles alone (list pasted atop `slides-notes.md`). - [ ] The design/identification slide shows the *identifying-assumption sentence* (Ch 7.5), not the regression equation. - [ ] The headline slide shows *one* number with its *confidence interval*, as a **figure**, with a human-units interpretation sentence. - [ ] The robustness slide is a specification curve or coefficient grid clustering near the headline (Ch 8.2), with a one-sentence stability claim. - [ ] The contribution slide names what is now known *and* what you are *not* claiming. - [ ] Every slide figure is the *same file your code generated* into `paper/figures/` — none hand-drawn or hand-pasted. **The talk** - [ ] Rehearsed timed at least three times; fits in **seven** minutes, not eight. - [ ] A **cut line** is pre-decided and written in `slides-notes.md` (which slide you skip if you are behind — never the design or headline slide). - [ ] Two prepared "what about [confound]?" answers (threat → plausible → what you did → residual) and one located "I don't know," written out. - [ ] Your one **fatal** critique is named, and your **retreat position** is written before you walk in. **The packet** - [ ] All six parts present: data/access-script · code in run-order · README · pinned env (`environment.yml` + `environment.lock.yml`) · fixed named SEED · single `make all`/`run_all.sh`. - [ ] README states the public-vs-licensed split honestly; **no licensed data and no secrets** anywhere in the repo or its history (`git ls-files` is clean — Lab 7, Step 5). - [ ] The SEED is a single named constant, threaded through every stochastic step via `np.random.default_rng(SEED)`, and recorded in the README. - [ ] Every table and figure in `paper/` is *generated by code* and `\input`/`\includegraphics`'d — none pasted by hand. - [ ] **`make clean && make all`** rebuilds the paper from nothing and the PDF is *identical* to the committed one. (If not: you found a hand-edit — fix it, then re-run.) - [ ] Verified on a **fresh clone**, not just your working copy — the true "stranger's machine" test (Lab 8). --- ## How this is graded This problem set is graded on the two crafts, weighted as the parts show: the deck (Parts 1–2, 55 pts) on whether one argument lands cleanly under slide discipline; question-handling (Part 3, 15 pts) on whether you defend identification honestly rather than bluff; and the packet (Part 4, 30 pts) on whether `make clean && make all` actually rebuilds your claims. The highest-signal single check is the honesty test: a packet whose rebuild does not reproduce the slide's headline number fails the part regardless of how polished the deck is, because — Ch 8.5 §6 — *a result a stranger cannot reconstruct is not yet a result; it is a claim about a result.* The model deliverable in Appendix E shows the A-grade standard for one cast project, annotated with the rubric levels each piece hits. --- ## The bridge — and the close of the camp This is the last deliverable, and it closes the arc the whole camp walked. Look back at Week 1: the first thing you did was compute an average return of about eight percent, and it felt like an answer. It was the *first question*. You learned the eight percent had a standard error, that the standard error rested on assumptions you could not see, that a "significant" sign could be an artifact of how many specifications you quietly tried. Week 2 turned correlation into regression and the ghost of omitted-variable bias; Weeks 3–4 turned "we control for X" into *designs* — IV, RD, DiD, synthetic control — each a specific argument for why a comparison is fair; Week 5 taught you to read a paper as a skeptic; Week 6 sharpened the inference; Week 7 made it *yours* — an idea, data pulled reproducibly, a pre-analysis plan tagged before you could fool yourself, an identification memo that named every threat. Week 8 made it *communicable and checkable*: written up with calibrated verbs (Ch 8.3), stress-tested for robustness (Ch 8.2), peer-reviewed (Ch 8.4), and now — today — defended in eight minutes and packaged so a stranger can rebuild every claim with one command. The single sentence this whole camp existed to let you say truthfully: *"Here is what I found, here is exactly why you should believe it, and here is the command that lets you check me."* When you build this deck and this packet, you can say it. The eight percent was the toe in the water. This is the swim. *End of PS 8.5. The model deliverable is in `book/appendices/E-solutions-manual/E-w8-ps8.5-solutions.md`. Build your own deck and finish your own packet first; read the exemplar to calibrate, not to copy. The skeptic who clones a stranger's repo and types `make all`, and the author who makes their own rebuild honestly, are the same person — this assignment is you doing to your own work, on purpose, what a referee will do to it next.*