The Spectrometer That Cancelled Itself Out
A hospital laboratory runs a batch of 240 serum samples through a single spectrometer. Midway through the run the optical filter degrades, adding +0.4 mg/dL to every reading. Unknown to anyone, that morning's calibration standard had itself been prepared 0.4 mg/dL too high, so the instrument's firmware subtracts 0.4 from every output. The two errors cancel exactly. Technician Priya reads sample #117 at 5.4 mg/dL, within the normal range, and records "no repeat needed." The sample's true value is 5.4 mg/dL. She followed the standard protocol, the calibration log was signed, and nothing in the run triggered a quality flag. Six months later an audit reconstructs both errors. Priya never learns of the audit and continues to believe her result was correct. A colleague argues that because her true belief was, in retrospect, a coincidence, she cannot have known the value, while a supervisor insists that she plainly did everything an epistemically responsible technician could do.
1. Does Priya know the sample's value? Assess the case under the justified-true-belief analysis and under a safety condition on knowledge.
Read the worked answer
Under the tripartite analysis all three conditions look satisfied: the reading (5.4 mg/dL) is true, she believes it, and she is justified in the sense of having followed an approved method with signed calibration and no anomaly. That is exactly why the case is instructive - JTB is met and yet the truth looks accidental.
Add a safety condition: S knows that p only if S's belief could not easily have been false. Test the nearest possible worlds. The margin is razor-thin: the filter drift is +0.4 and the firmware offset is −0.4. Had the filter degraded slightly more, or had the standard been prepared a fraction differently, the cancellation fails and the displayed value is 5.8 or 5.0 - a false belief or a different true one. Because the truth depends on two independent errors aligning almost exactly, her belief is not robust across nearby worlds. On a safety or sensitivity account she does not know; she has a justified true belief that is true by epistemic luck of the environmental kind. The Gettier structure here is not a false lemma but a double coincidence, which is why the JTB framing alone cannot resolve it.
2. The audit later reveals the cancellation, but Priya never hears of it. Does this discovery change whether she knew at the time? Distinguish justification from knowledge.
Read the worked answer
We must separate two temporal questions. Synchronic justification is assessed on the evidence available to the agent at the time. Priya had a signed calibration log, a clean quality run and no reason to suspect drift; her credence was appropriate to her evidence, and an internalist will say she was fully justified - even epistemically blameless. The audit does not retroactively worsen her evidence, because she never acquires it.
Knowledge is different because it is factive and, on externalist views, depends on the actual reliability of the process that produced the belief. The audit reveals that the token process in that run was not reliable at that moment: the reading was correct only through compensation. So the externalist denies knowledge while continuing to grant justification. This is not a paradox; it is the ordinary gap between "did the right epistemic thing" and "got it right non-accidentally."
The practical upshot is institutional rather than personal. The audit is a reason to change the lab's calibration-checking regime, not a reason to censure Priya, and it shows why hospitals track process reliability separately from individual blame.
3. Suppose Priya had in fact read the audit before recording her result. What would change epistemically, and what would not?
Read the worked answer
Reading the audit supplies a defeater. She would now know that in this run the filter and the calibration cancelled, which is evidence that her method's output on this occasion is not a reliable indicator of the true value. Even if the number happens to be correct, she should suspend judgement and re-run the sample, because her evidence no longer supports the belief.
What would not change is the truth of the proposition or the fact that the belief, if retained, would still be true. Belief is not automatically revised by evidence; she might dogmatically keep the original figure, in which case she is now unjustified as well as unknowing.
It matters how specific the defeater is. "A calibration error is possible somewhere in this lab" is a general statistical defeater and may leave her particular justification intact if the protocol already accommodates that risk. "The error occurred in this run and cancelled here" is a specific defeater and defeats this belief. The distinction between global and instance-specific defeaters is the crux of the case.