Chapter 13 — Reps

These reps move your Discussion section forward. Every one of them produces a paragraph, a table, or a verdict you will paste into this week’s deliverable. There is no toy data here — bring your real results from Chapter 11 and your real analysis from Chapter 12.

Ground rules.

• Use your own project’s numbers. If a run is missing, note it as a gap — do not invent a number to fill a cell. Fabrication is the failure this whole chapter exists to prevent.
• The AI policy this week: use an LLM for phrasing and structure, never for claims. Every claim-bearing sentence must be claim-evidence-mapped against a real result before it stays. Verify every cited comparison in dblp or Semantic Scholar (Appendix C).
• Keep a running discussion-notes.txt in your portfolio repo; each rep appends to it. Commit after each rep so the provenance is real.
• Where a rep says “label your assumption,” mark it explicitly as an assumption and write how you’d confirm it.

Rep 1 — Add yourself to the matrix

Open code/comparison-matrix.csv. Fill the top rows with three real prior-work papers from your literature review, then add your own work as the bottom row, with the same columns (method / dataset / metric / result / limitation).

# from your portfolio repo
cp chapters/13-discussion-and-interpretation/code/comparison-matrix.csv portfolio/discussion/matrix.csv
# edit, then sanity-check it opens cleanly
python -c "import csv; print(*[r for r in csv.reader(open('portfolio/discussion/matrix.csv'))], sep='\n')"

Write: the three comparison sentences that fall out of reading your result column against the prior rows. At least one must name an asymmetry (e.g., they reported one run, you reported five seeds — so the comparison is not clean).

Rep 2 — The “because” rewrite

Take your draft Results-adjacent prose and find every sentence that merely restates a number. Rewrite each so it begins with interpretation: because, however, this suggests, in contrast to.

Write: three before/after pairs. For each “after,” note in one phrase what new information the interpretation adds that the table did not already carry.

Rep 3 — Claim-evidence map your own draft

Run your draft Discussion through the Claim–Evidence Mapper widget (or fill in code/claim-evidence-map.txt by hand). Break it into atomic claims, attach the specific result for each, and label every link Supported, Over-reach, or Unsupported.

| Claim sentence | Specific evidence (table/fig/test) | Verdict |
|---|---|---|
| ...            | Table 2, row 3                     | Supported |
| ...            | (none)                             | Unsupported |

Write: the count of each verdict, and — for every Over-reach — the narrowed sentence; for every Unsupported — your decision (measure-and-move-to-Results, or delete). No Unsupported claim may survive.

Rep 4 — Snap the scope on your headline claim

Take your single most important claim and stress-test its scope. List every condition under which you actually measured it (which dataset, which language, which model version, how many seeds). Then list what your current wording implicitly promises.

Write: the gap between the two, and the final narrowed claim sentence that the evidence fully covers. Label any generalization you believe but did not test as an assumption for future work.

Rep 5 — Three ways to say one result

Pick your headline result. Write it three ways using code/three_ways.py as a guide.

python chapters/13-discussion-and-interpretation/code/three_ways.py \
  --mean-a 0.410 --mean-b 0.382 --sd 0.021 --n 5
# prints the p-value form, the CI form, and the Cohen's d form for your numbers

Write: all three sentences for your real numbers, then a one-line argument for which form is the most honest for your reader, citing the ASA 2019 reasoning (effect size + CI over bare p).

Rep 6 — Interpret an unexpected result

Find the result that surprised you — the ablation that didn’t move, the split where the baseline won, the metric that went the “wrong” way. Write the honest three-part discussion: (1) report it plainly, (2) offer candidate explanations labeled as untested hypotheses, (3) state what you’d measure to test them.

Write: the paragraph. Then check it against the HARKing trap — does any sentence present a post-hoc explanation as if it were your original hypothesis? If so, rewrite to separate preregistered question from after-the-fact suspicion.

Rep 7 — One limitation per validity category

Using code/limitations-checklist.txt, write one concrete, specific limitation for each Cook & Campbell category: statistical conclusion, internal, construct, external.

Statistical conclusion: ...
Internal:               ...
Construct:              ...
External:               ...

Write: the four limitations. Each must name a specific uncontrolled threat and say how it bounds your conclusion. “More experiments could be done” is banned — be concrete.

Rep 8 — The contamination construct check

If your project used a public benchmark as a metric, address contamination as a construct-validity threat. Was the test set plausibly in your model’s training data? Note your model’s version and date (it matters — closed models drift).

Write: one paragraph stating whether contamination is a credible threat for your benchmark/model pair, what (if anything) you can say about decontamination, and how it bounds the “our method is better” claim. If contamination is not a concern for your setup, say why in one sentence.

Rep 9 — Red-team your own Discussion

Put on the hostile-reviewer hat. Read your draft and find the single sentence you would attack first if you wanted to reject the paper. Write the attack as a real reviewer would phrase it.

Write: the targeted sentence, the reviewer’s attack, and your fix (narrow it, support it, or cut it). This is the sentence most likely to draw fire — fix it before submission, not after.

Rep 10 — Verify every comparison citation

List every prior-work paper your Discussion compares against. For each, verify it exists and your numbers are right by checking dblp or Semantic Scholar — not your memory, and not an LLM’s.

# example: confirm a paper's metadata against dblp before you cite its result
# (open https://dblp.org and search the title; confirm authors, venue, year)

Write: a checklist of each cited comparison with a “verified ✓” against its source. Flag any number you quoted that you could not confirm at the source — and either fix it or drop the comparison.

Done? One Last Thing.

Assemble the reps into a first full Discussion Section draft in your acmart/IEEEtran Overleaf project. Order: interpret your headline finding → compare against the matrix (Rep 1) → discuss the surprise honestly (Rep 6) → limitations by validity category (Rep 7, Rep 8) → one sentence of what this opens up next.

Then do the final pass that makes it a deliverable: run the whole assembled draft back through the Claim–Evidence Mapper one more time. Zero Unsupported claims. Every Over-reach narrowed. Every comparison verified (Rep 10). Export the map and drop it in your portfolio next to the draft — it is your evidence to a future reviewer (and to your future self) that you tested everything before you held it fast.

That assembled, mapped draft is the spine of P13.

Up next: Project 13