When your AI pipeline hits garbage output, should it quietly plug in a default and carry on, or stop?

This explores the choice a pipeline faces when an output can't be parsed or validated: patch in a placeholder and keep going, or throw the output out and stop or retry. The corpus has no single rule for this, but it gives a sharp test: what happens downstream to the value you substituted?

The clearest case is a harness that caught parsing errors and quietly fell back to a default rating. An optimizer downstream was ranking outputs, so the default turned a detected failure into a valid-looking candidate. The safety check became fail-open Does a default fallback defeat a safety check?. The note's point is that the failure path, not the check itself, decides whether a guardrail works. A default is tolerable when nothing downstream selects, ranks, or acts on the value. It is dangerous when something does, because selection pressure will find the placeholder.

Rejection is the natural choice when dropping a sample is cheap and the pipeline can regenerate. Self-improving transformers generate solutions, filter for correctness, and retrain on what survives. The recipe is to filter, not patch, and it kept improving across rounds without saturating Can transformers improve exponentially by learning from their own correct solutions?. Nothing needs a default because supply is abundant. Checking also doesn't have to be slow. Asynchronous verifiers can run alongside generation and intervene only on violations, with near-zero latency cost on correct runs Can verifiers monitor reasoning without slowing generation down?.

When you can't drop the output, don't trust the substitution or transformation to be clean. A gate that compresses text should re-score what it keeps against the same threshold. ChannelGuard proposes this to close a 23.4% leak, though the corpus notes it is untested in real-world use Should sanitizers re-score their compressed output before passing it?. Rejecting at the action boundary is the other option. An authorization layer with signed tokens blocked every unsafe action even though memory poisoning had compromised the model's judgment Can memory poisoning compromise decision-making even with authorization layers?. That layer blocks the action and does not repair the bad decision, which is a sign that rejection is the sturdier default where consequences are real.

Two cautions apply if you do substitute. First, don't let the source of the error grade its own repair. Models over-trust answers they generated themselves, and comparing against broader alternatives breaks that loop Why do models trust their own generated answers?. Second, errors that stay in context degrade later steps sharply, and scaling the model doesn't fix that Do models fail worse when their own errors fill the context?. The corpus doesn't test defaults directly here, but a silently substituted value that flows into history looks like the same kind of contamination. Per-item checks also can't see a pattern building up, such as fallbacks quietly becoming the norm, because sequence-level constraints need a stateful monitor Can stateless checks ever catch sequence-level constraint violations?. If you substitute defaults, count them.