With no recorded baseline, how do you defend a before-and-after number to a behavioral interviewer?

## Why this question exists Asking where a number came from is the cheapest calibration an interviewer has. Candidates who ran their own change usually know the before-value, where it was displayed and why it was that bad. Candidates who absorbed a figure from a status update usually know only the after-number and the adjective attached to it. One follow-up separates the two groups, which is why the probe is so common and so often decisive. The failure it catches is specific: an impressive after-number quoted with no baseline, followed by a collapse when asked how it was measured. The collapse is worse than never quoting the figure, because it retroactively puts every other claim on the page in question. ## The three moves ### 1. Name the source before the value Leading with provenance changes the register of the whole answer. *The queue exposed backlog age, and we alerted on it, so the before-value is the alert history* is a sentence that pre-empts the probe. Sources that hold up: metric dashboards and their retention window, alert or paging history, ticket and incident timelines with timestamps, a job's own run log, a cost or billing report, release notes with dates attached. Sources that do not: recollection, a number someone quoted in a meeting, a figure from a document you never verified. ### 2. Split measured from inferred, explicitly Most real claims are a mixture, and saying so is a strength. A truthful decomposition sounds like this: > The nine minutes is measured — every run since the change is in the log. The forty-one is the worst case across the runs we captured beforehand; the typical run was better than that, and I quote the worst case because that is the one that woke people up. Notice what that does. It defines which statistic is being used, admits which half is sampled rather than continuous, and gives the reason for the choice. An interviewer who wanted to probe now has nothing left to probe — you already did it. ### 3. Bound instead of inventing precision When the underlying data is genuinely gone, estimate openly and give a range or a floor. *Above forty minutes on the runs I saw* is defensible; a confident single figure produced from nothing is not. Two honest estimation techniques: - **Reconstruct from an artefact that still exists.** Ticket close times, alert timestamps, a change log, calendar entries for the incident reviews. These are weaker than a metric series but they are real, and you can say which one you used. - **Reason from a known constant.** If the job ran on a fixed schedule and overlapped the next run before your change but no longer does, the schedule interval bounds the before-value from below. State the reasoning, not just the conclusion. What you must not do is round an estimate into something that sounds authoritative. Tidy figures read as invented, and non-tidy figures read as measured, which is one reason a real drop from 41 minutes to 9 is more persuasive than any smoothed version of itself. ## The forward fix This question is much easier to answer if you never had to reconstruct anything. When you start work you expect to improve, record the current value, the date-free description of where you read it, and the statistic you are using — worst case, median, ninety-fifth percentile. Keep that in your own notes rather than only in an employer's system, because access ends when employment does. Keep the record proportionate and mind your obligations: many employers treat operational and financial detail as confidential, and what you may carry away varies by employer and by contract, so check your own situation and prefer relative or de-identified forms where there is any doubt. ## What the scope caveat buys you The same honesty applies to coverage. If the improvement landed on three of seventeen services, saying so up front removes the most obvious challenge to the claim. Candidates fear that the caveat shrinks the achievement. In practice it does the opposite: an interviewer who hears an unprompted limitation stops auditing and starts asking about the mechanism, which is the conversation you wanted. ## When the honest answer is that you do not know Sometimes there is no source, no artefact and no sound reconstruction. Say that, then say what you can support: what you changed, why it was slow before in mechanical terms, and how you would measure it if you were doing it again. That last part is the senior signal in the whole exchange — knowing what instrumentation you would put in first is worth more than any recovered figure.