Research is an investigation; a paper is a lossy compression of it, and much of the uncertainty disappears in the compression. Experiments have three outcomes, not two — the evidence supports the hypothesis, refutes it, or is insufficient to decide — but our publishing conventions file False and Unknown together under “didn’t work.” Both are knowledge. A refutation tells us something about the world, as the CAP theorem and the FLP impossibility result do, and nobody calls those failures. Unknown tells us something about the limits of our experiment: underpowered, confounded by the testbed, measuring an ill-defined quantity, run at insufficient scale. The real damage is done when Unknown is dressed as False and a question the evidence leaves open is reported closed. The talk works through a measurement system built to answer True, False, or Unknown — using latency constraints to decide when a path can be adjudicated at all — and argues for validating whether a system knows when to trust its answers rather than validating every answer. It closes on the boundary of this approach: coherent error, calibrated and wrong at once, of which the “power-law Internet” episode of 1999–2009 is the cautionary example. Consistency is not correctness, and detecting coherent error requires independent evidence.