A checker who forms a judgment only after seeing the first person's answer isn't really checking
Aliases: independence in dual verification · functional safety
What it is
Independence in two-person verification requires the checker to have already formed their own judgment about the target, prerequisites, and parameters before seeing the first person's conclusion, and to hold a clearly granted right to challenge that conclusion. Independence is not a slogan; it breaks down into four concrete dimensions — separate information access, separate timing, the authority to say no, and the organizational standing to actually say it. If any one of the four fails, independence exists in name only.
Why it happens
Seeing someone else's answer first produces anchoring, a cognitive effect that is hard to overcome by will alone; an authority gradient — a junior person facing a senior one — and shared training that instills the same habits of thought both make two people's errors statistically correlated: a mistake one person is prone to make becomes more likely for the other too, which exactly cancels the point of two-person verification — lowering the chance of a simultaneous error through two independent samples. Blind entry (not letting the checker see the other person's answer beforehand), rotating checking partners regularly, and explicitly granting either party the right to pause the process at any time without being blamed for it are specific measures aimed at weakening this pressure to conform. But even with all of that in place, if both people are looking at the same sensor reading, that shared evidence source is still shared and cannot count as two independent observations.
Where it stops holding
Complete independence is not always feasible or even desirable for complex tasks — some judgments genuinely require the two people to communicate and exchange context to reach a sound conclusion, and insisting on total isolation can leave a checker without the context they need. So once the independent phase ends and each side has submitted a judgment, there still needs to be a stage where both sides' context is brought together to discuss any disagreement — independence applies to forming the initial judgment, not to the entire process staying siloed forever. Also, two people coming from different departments or different employers is only organizational separation, and does not automatically produce cognitive independence — if they were trained on the same material and absorbed the same assumptions, the independence that organizational boundary suggests can be superficial.
Applying it
Design the checking process as three clearly separated stages — independent judgment, automatic system comparison, and discussion of any disagreement — hiding each party's input from the other until both have submitted, and granting either party the right to pause the process when uncertain, with no penalty.
- Independent entry must also use a common unit and a common data timestamp, or an apparent "disagreement" may just reflect different units or different moments of data rather than a genuine difference in judgment, contaminating later statistics.
- How to check: audit not just the agreement rate between the two people's final conclusions, but also the detection rate of seeded errors and how genuine disagreements were actually resolved — a high agreement rate with an empty record of disagreement handling is itself a sign of conformity rather than true independence.