33 Poster Criticism
This chapter is part of a book in active development and has not yet been through the author’s review. Content may change as the review advances.
The research decision. If your venue is a poster, run the gallery walk and the print-lock audit before the artifact leaves your control. Whatever your format, you decide whether the draft is done or needs one more fix, and you defend the lock call out loud.
33.1 Why this decision matters
The decision on the table: whether this draft is honest enough to become permanent, or needs one more fix first.
Picture the last hour before a poster prints. A research-integrity judge, the person at a conference who probes whether a poster’s claims are actually backed, walks your row and stops at one number on your board. “Where did this come from?” If you can walk that judge to the exact line of code that produced it, the whole poster earns trust. If you answer “I think that came from the analysis somewhere,” the judge stops believing every other number too. Once the file goes to the printer you cannot patch it, so the honesty has to be built in before it leaves your hands. This chapter is the last audit between a draft poster and a permanent one.
33.2 The concept
Three passes turn a fragile poster into one that survives a stranger, and each ends in a decision only you can make.
The first pass is the gallery-walk audit, a structured walk where you stop at a poster and run the same fixed checklist instead of glancing and moving on. Four questions form its core. The claim-boundary audit asks whether the headline stays inside its evidence, where a claim boundary is the line between what your evidence licenses and what it does not, sitting in the headline verb: an observational field trial may say “is associated with,” never “causes.” The figure-honesty audit asks whether the picture matches the numbers, where figure honesty means a chart cannot say more than its data, so an axis starting just below the data draws a small gap as a canyon. The read-path audit asks whether a stranger finds the headline in ninety seconds (Erren and Bourne 2007). The uncertainty audit asks whether the interval sits beside the claim, not buried in a footnote. A fifth check, accessibility, asks whether every reader can receive the claim, and its fix has a name: redundant encoding means carrying a distinction in more than one channel, such as a value label plus a hatch pattern, so color is never the only signal a color-blind reader can lose (Crameri et al. 2020).
The second pass is the oral defense, the live conversation at your board where you answer questions you cannot edit afterward. Three lenses cover almost every hard question: an outsider testing whether your claim survives translation, a specialist probing a design choice, and a skeptic offering a rival explanation for the same evidence. The judgment underneath is defend-versus-concede: you defend when the objection asks for something your evidence already licenses, and you concede, narrowing the claim on the spot, when it asks for a crossing your design never paid for. Your compass position, the kind and reach of the question your project answered, tells you which call to make.
The third pass is the lock. Claim-evidence traceability means every number on the poster traces to the exact cell that produced it, so you can show that line when a judge points at it. The terminal lock is the moment the poster becomes final and unchangeable, which is why the audit before it is the most consequential one you run on the whole project.
33.3 A worked example
A plant-biology poster carries the headline “Compost RAISES corn yield.” Its center figure is a bar chart of mean grain yield for plots that received a compost amendment against plots that did not, drawn as a red bar and a green bar. The means are 185 and 191 bushels per acre, a gap of 6. The y-axis starts at 180, and no interval is shown. The plots were not randomized; the grower chose which fields got compost.
Walk the gallery audit across it. The claim-boundary audit catches the verb first: “raises” is causal, but nothing was randomized, so a naturally richer field could have received both the compost and the higher yield. The honest headline is “compost is associated with about 6 more bushels per acre.” The figure-honesty audit catches the axis, floored at 180, which draws a 6-bushel gap several times taller than it truly is, so run it from 0. The uncertainty audit catches the missing interval: a 95% interval of roughly ±9 bushels per acre swamps a 6-bushel gap, so it has to ride beside the headline. The accessibility audit catches the red-green pair, which a reviewer who is red-green color-blind, roughly one man in twelve, cannot tell apart, so you add value labels and a hatch pattern.
You cannot fix everything before printing, so you decide. The claim-boundary hit travels farthest, because a reviewer who catches one causal overreach distrusts the whole board, so you lock the associational verb first. That single decision, made by you and defended out loud, is the work of this chapter.
The audit’s accessibility item is not optional politeness: color-only encoding loses readers whose vision it never reaches (Crameri et al. 2020).
The block below rebuilds the poster’s data, prints the interval the board left off, and redraws the figure from zero. The last printed line is the one the headline’s verb depends on.
import numpy as np, pandas as pd
import matplotlib.pyplot as plt
SEED = 464
rng = np.random.default_rng(SEED)
# Fields the GROWER chose. Richer fields got the compost.
n_plots = 60
soil_quality = rng.normal(0, 1, size=n_plots)
composted = soil_quality > np.median(soil_quality) # not randomized
yield_bu = 184 + 4.5 * soil_quality + rng.normal(0, 11, size=n_plots)
means = np.array([yield_bu[~composted].mean(), yield_bu[composted].mean()])
ci = np.array([1.96 * yield_bu[g].std(ddof=1) / np.sqrt(g.sum())
for g in (~composted, composted)])
print(f"no compost : {means[0]:.0f} bu/ac ± {ci[0]:.0f}")
print(f"composted : {means[1]:.0f} bu/ac ± {ci[1]:.0f}")
print(f"gap : {means[1]-means[0]:.0f} bu/ac")
print(f"mean soil quality, composted vs not : "
f"{soil_quality[composted].mean():+.1f} vs {soil_quality[~composted].mean():+.1f}")
fig, ax = plt.subplots(figsize=(4.2, 3.2))
ax.bar(["no compost", "composted"], means, yerr=ci, capsize=6,
color=["#4a4a4a", "#a8a8a8"], hatch=["", "//"], edgecolor="black")
ax.set_ylim(0, 230)
ax.set_ylabel("grain yield (bu/ac)")
ax.set_title("Yield by compost status (grower-chosen fields)")
plt.tight_layout()
plt.show()33.4 An AI failure case
You describe your compost figure to your AI and ask for the hardest skeptic questions. Back comes a crisp six-item list: the sample size, the single growing season, the soil-test method, generalizability to other regions, cost, and the definition of “yield.” It reads as thorough, and the temptation is to treat the poster as fully audited.
It is not. The list never mentions the truncated axis, and never mentions the red-green bars. Those are the two most concrete, most fixable flaws on the board, and the polished list quietly skipped both. This is the illusion of completeness: a long, organized answer that omits the thing that matters most.
Here is how you catch it. You map each question to one audit. Four land on claim boundary and uncertainty, none on figure honesty, none on accessibility. Two whole audits are missing, which tells you the list is incomplete no matter how confident it reads. Then you check the survivors against the exaggeration factor your own notebook cell printed, and cut any question that assumes a gap other than the 6 bushels you computed. The completeness check is yours, not the tool’s.
A boundary note before the practice. The walk, the board, and the lock below are poster procedure: a gallery walk tests what a viewer notices from across the room, and a print lock tests what fails when the file becomes a physical object. A manuscript needs a different cold read, and a talk needs a timed rehearsal; do not rename either one a gallery walk. What travels to any format is the audit’s spine: read the claim against its boundary, rank the hits by how far the error travels, take the three lenses, revise once, then make the lock call. The Studio 10 milestone shows where that spine lands when your venue wants a talk, a deck, or a page.
33.5 It is your turn
You are working inside Studio 10: Prepare to publish or present. This lesson adapts your work to the poster format. Work the format your venue actually asks for; the studio page routes you.
You have a draft page and a headline figure; this step is the audit that decides whether either one survives a hostile stranger.
The hands-on half of this section lives in the chapter’s companion notebook: open it in Colab with the badge at the top, and work the steps there.
Commit your own audit first, then delegate. Each prompt hands the tool a job you can check, with a verify note naming the failure it defends against.
Treat each prompt as an opening move. When the answer comes back thin, say what it missed and ask again. Reviewers get sharper on the second pass, human and machine alike. Some tools now run that loop internally and return only the final critique, which is convenient and also hides the reasoning you would otherwise get to question.
These stay yours, however fluent the tool sounds. The words of your headline claim, and the verb that keeps it inside your compass position. The call, made out loud, to defend a point or concede it. And the decision that the poster is honest enough to lock. The tool can generate hard questions and scan a figure for color traps; deciding what your evidence licenses, and answering for it at the board, is yours.
Run the gallery-walk audit on your own draft as if you had never seen it: claim boundary, figure honesty, read path, uncertainty, accessibility. Write down every hit, located precisely enough to fix.
Generate the hostile questions (you map each to an audit).
Act as a hostile poster reviewer. My headline figure is a bar chart comparing mean corn yield for compost-amended versus control plots on a y-axis that starts at 180, titled "Compost RAISES corn yield," with no interval shown and the two bars told apart by red versus green only. Generate the six hardest skeptic questions you would ask, then name the single flaw this figure cannot answer.After running, verify (counters illusion of completeness): map each question to one audit (claim boundary, figure honesty, read path, uncertainty, accessibility). An entire audit the tidy list skips is the gap. Then check the survivors against the real gap and exaggeration factor your own notebook cell printed.
A second angle, optional:
List the color encodings to verify (you check against the rendered figure).
Here is the code that draws my poster's headline figure: [paste the cell]. List every place it encodes a meaningful distinction with color, and for each, say what a red-green color-blind reader and a screen-reader user would fail to receive. Then propose one redundant encoding that carries the same distinction.After running, verify (counters plausible-but-wrong-method): read each encoding it names against your rendered figure. Generic accessibility advice that does not match how your figure is actually drawn is wrong for your case, so trust the picture you can see, and re-render after applying one proposed fix.
Rank the hits by how far the error travels. A causal verb your design never bought travels farthest, because a reader who catches one overreach stops trusting everything else.
Field the three lenses out loud: an outsider, a specialist, and a skeptic with a rival explanation. For each question, decide defend or concede, and write the narrower wording you would concede to.
Red-team your spoken answer (you keep the words).
Here is my written headline: "[paste it]". Here is what I plan to say when a skeptic asks whether something else could explain it: "[paste it]". Find every place my spoken answer claims more than my written headline. Do not rewrite it; list the exact words.After running, verify (counters silent scope change): read each flagged phrase next to your compass position. Keep every catch of a spoken “causes” your written poster never said, and reject any push to concede a point your design does license.
Revise once, fixing the hits in ranked order, then make the lock call. Every number traced and every citation opened, or not yet.
Log it in your AI Research Ledger, and verify at least one fixed number with a named method from the Verification Guide. An AI reviewer may run the check with you; the decision to accept or reject stays yours.