The most expensive errors in competitive intelligence are not analytical. They are mechanical: an untraced number, a mislabeled claim, a chart that says something the data does not. Analysis errors get debated; mechanical errors get remembered. One shipped mistake of the wrong kind, and every future deliverable you send gets read with a discount applied.
This checklist is the last hour before anything ships: a report, a slide, a market sizing, a two-paragraph escalation note. Ten checks, ordered by failure cost, not by workflow. The career-killers come first, because if time runs out, the bottom of the list is where you are allowed to cut corners. The top is not.
Each point carries a real, public example of the failure it exists to catch. None of these people were stupid. All of them skipped a check.
Band A: Career-killers. If one of these ships, the deliverable is dangerous
1. Trace every load-bearing fact to a source you opened yourself
Not a source someone cited. Not three articles that agree. The document itself. The specific failure this catches is false corroboration: multiple reports that feel independent but trace back to a single origin.
The example. The 2003 case for Iraqi mobile bioweapons labs rested heavily on one defector, codenamed Curveball. His claims appeared to be corroborated across multiple intelligence reports; in reality the "corroborating" streams traced back to the same man, relayed through different channels. The claims reached the UN Security Council as established fact. Three news articles citing the same wire story is the everyday CI version of the same disease. The check: for every fact under a conclusion, name the primary document and confirm you opened it.
2. Rebuild the arithmetic from raw inputs
Any number you calculated gets recalculated, from the source figures, ideally by a second person or at minimum by yourself in a fresh sheet. Formulas drift, ranges exclude rows, and models inherit errors silently.
The example. Reinhart and Rogoff's 2010 paper on debt and growth anchored austerity arguments across multiple governments. In 2013, a graduate student attempting to replicate it found, among other issues, an Excel range error that excluded five countries from a key average. The headline result weakened substantially once corrected. The paper had been cited for three years. Nobody had rebuilt the arithmetic.
3. Check units and definitions at every handoff
Whenever a number crosses a boundary, between analysts, between a source and your model, between currencies, between manufacturer revenue and end-user spend, verify that both sides mean the same thing. The 48-hour sizing treated this as the first two hours of work; QA re-tests it in the last hour.
The example. NASA's Mars Climate Orbiter was lost in 1999 because one engineering team delivered thruster data in pound-force seconds while the receiving system expected newton-seconds. Both teams' work was internally correct. The handoff was not, and review after review failed to test the interface. A $125 million spacecraft burned up over a unit label.
4. Verify the entity
Confirm that every company, ticker, subsidiary, and product name refers to the thing you think it does. Parent versus subsidiary, similarly named firms, and stale tickers are where this fails.
The example. In early 2020, shares of Zoom Technologies, a tiny OTC company with the ticker ZOOM, surged hundreds of percent as investors piled into what they believed was Zoom Video Communications, which trades as ZM. The SEC eventually suspended trading in the lookalike. Thousands of people committed real money to the wrong entity. A CI report can do exactly the same thing with a segment figure pulled from the wrong company's filing.
5. Audit the claim labels
Run the Fact vs. Inference Ladder over every analytical sentence. The specific check: is anything wearing a [F] that is actually a reported claim or an inference? Label inflation is the single fastest way to ship a falsehood while feeling rigorous.
The example. In April 2013, a hacked Associated Press account tweeted that explosions at the White House had injured the president. Automated and human traders treated a reported claim, from a normally reliable source, as verified fact. The S&P 500 shed roughly $130 billion in market value in about three minutes before the claim collapsed. The source's reputation promoted the claim one rung too high, which is precisely the promotion your labels exist to block.
Band B: Credibility erosion. These survive the meeting and damage you afterward
6. Stress-test the scope behind any market number
Every market size, share, or TAM figure gets one question: what exactly was counted, and would the reader count the same things? A defensible number with a stated scope beats a large number with a flattering one.
The example. WeWork's 2019 IPO filing presented a market opportunity ranging up to roughly $1.6 trillion, reached in its broadest cut by treating vast populations of urban desk workers as potential members. The arithmetic was fine; the scope was the fiction, and it became a public symbol of the whole prospectus's credibility problem. Reviewers may not catch a scope inflation in the room. They catch it later, and they remember whose slide it was on.
7. Date-stamp everything and hunt for staleness
Every figure carries an as-of date, and the check asks: is newer data available, and would it change the finding? Data does not announce its own expiry.
The example. "Dewey Defeats Truman," 1948. The famous headline traces to polling that had largely stopped weeks before the election, on the assumption that preferences were stable. The data was accurate when collected and wrong when used. A 2024 headcount figure in a 2026 deck is the same failure wearing business casual.
8. Run the chart integrity pass
Axes start where they claim to, categories run in a defensible order, baselines are honest, and the visual impression matches the underlying table. A chart is a claim; QA treats it like one.
The example. In May 2020, Georgia's Department of Public Health published a COVID-19 chart in which dates and counties were reordered non-chronologically, producing a clean visual decline that the underlying data did not show. The department apologized and corrected it, but the chart had already done its damage in public. Most chart crimes in CI are subtler and unintentional. The pass exists because intent does not change what the reader takes away.
Band C: Finish-line polish. Cheap to fix, embarrassing to ship
9. Sweep for template ghosts and internal notes
Search the document for placeholder text, tracked changes, comments, another client's name, and anything written for internal eyes. This is a two-minute mechanical sweep, and it catches the errors that make a deliverable look careless regardless of the analysis underneath.
The example. A 2016 peer-reviewed ecology paper shipped with an internal author note left in the published text, asking whether to cite a rival team's "crappy" paper. It survived the authors, the reviewers, and the editors, and it is what that paper is now remembered for. Every profession has its version. The sweep costs less than the anecdote.
10. The cold read
Someone who has not seen the deliverable reads only the title and the final recommendation, then tells you what they think it says and what they would do. If their answer does not match your intent, the deliverable is not done, no matter how good the middle is. If no second person is available, the fallback is a timed break and a printed read, but the independent reader is the real check.
The example. Hawaii's January 2018 false missile alert reached every phone in the state because the process allowed one operator to ship a state-wide message with no independent confirmation step. It took 38 minutes to correct. The lesson is not about interfaces; it is that anything important enough to send is important enough for a second pair of eyes before it goes.
Running the checklist
Three rules make it operational. First, run it top down: if the hour shrinks, Band C is sacrificed before Band B, and Band A is never sacrificed. Second, the checker signs. A named person confirming "traced, rebuilt, labeled" converts QA from a vibe into an accountability step, the same way the Ladder converts confidence into labels. Third, log what each pass catches. A running tally of near-misses tells you which checks your team actually needs, and it turns the checklist from ritual into feedback.
The checklist pairs with the rest of the system in one sentence: the Triage Matrix decides what gets analyzed, the Ladder and the Pyramid govern how the analysis is built, and this list is the gate it passes through on the way out the door. The analysis makes you smart. The last hour keeps you credible.