The Honest Blank

Two papers argue that the most informative output a system can produce is sometimes nothing at all — and both provide formal frameworks for when silence beats speculation.

Bröcker and Schultz (arXiv: 2604.02187) develop a scorecard for possibilistic weather forecasts — predictions that explicitly represent ignorance rather than forcing all uncertainty into probability distributions. Their five-number diagnostic separates forecast performance dimensions that probability-only methods conflate. A possibilistic forecast can say “temperatures between 5°C and 15°C, but I can’t tell you the distribution within that range” — preserving the information about what’s known while honestly marking what isn’t.

Singhal et al. (arXiv: 2604.01849) find that 61% of AI code completion suggestions are edited or rejected by developers. They propose a cost-theoretic framework where, above a critical entropy threshold, the model should insert explicit placeholders rather than guessing. The result: 19-50% reduction in editing costs. The blank isn’t a failure to predict — it’s a prediction that the model’s uncertainty exceeds the threshold at which guessing creates more work than it saves.

The structural claim: an honest blank is more informative than a confident wrong answer. The possibilistic forecast that says “I don’t know the distribution” preserves the user’s ability to apply their own judgment. The code placeholder that says “fill this in” preserves the developer’s ability to write what they actually need. In both cases, the system’s silence carries information: specifically, the information that this is a region where the system’s model is unreliable.

This is counterintuitive for system designers trained to maximize coverage. A weather forecast that sometimes says “I don’t know” looks less capable than one that always provides a probability distribution. A code completion that sometimes produces blanks looks less useful than one that always suggests something. The metrics reward completeness, not honesty.

But Singhal et al. provide the economic argument: when the cost of editing a wrong suggestion exceeds the cost of writing from scratch, the suggestion has negative value. The breakeven point is the critical entropy threshold. Below it, guess — the expected cost is positive. Above it, leave it blank — the expected cost of guessing is negative. This is a precise, quantitative version of “if you don’t know, say so.”

Bröcker and Schultz provide the epistemological argument: probability theory requires more knowledge than we sometimes have. When you force an unknown into a probability distribution, you manufacture precision that doesn’t exist. The possibilistic framework preserves the distinction between “I think the probability is 0.3” and “I think the value lies somewhere in this range but I can’t assign probabilities within it.” Both carry information. The second carries the additional information that no finer resolution is justified.

The connection between these papers is the recognition that most prediction systems are designed to always say something, and that this design choice has costs. Weather forecasts that always provide distributions sometimes provide false precision. Code completions that always provide suggestions sometimes create editing burden. In both cases, the system would serve users better by occasionally admitting ignorance — but the current evaluation metrics punish silence and reward verbosity.

The deeper question: if the honest blank is more valuable than the confident guess above a threshold, why don’t more systems implement it? The answer might be that users have been trained to expect completeness, and silence feels like failure even when it’s the correct output. The possibilistic forecast looks worse in a dashboard. The code placeholder looks worse in a demo. But in practice — when the user has to act on the forecast or work with the code — the honest blank is more helpful precisely because it doesn’t waste attention on low-confidence output that will need to be revised anyway.