Bounded output screening: Practice Library

What it changes

dampenedFailures adopted by people or agents(bounded automated screen on model output)

increasedOperator deference drift(automation complacency in the staff it feeds)

Who can pull it

Harness builderDeveloperDeploying organization

What it looks like institutionally

A bounded output-screening gate sits between the model and the people who act on its output, screening for the failure classes it was built to catch. It is genuinely useful: it holds output flow below the rate at which a human queue saturates, and it stops the obvious failures automatically.

Its honest limit is a ceiling. Measured automated-catch fractions top out well below completeness in optimistic settings and fall sharply on hard or adversarial content — so a gate that "passes" is not a gate that "verified". Treated as a substitute for human judgment it becomes worse than nothing, because the staff it feeds relax into automation complacency: the very people whose second look the system still depends on stop looking.

Use it as one bounded layer: monitor its miss rate as a health signal, keep a human check on the decisions that can least afford to be wrong, and pace the load so the gate itself never saturates (see Pace the pipeline).

Addresses: Unscreened model output · Overtrust in a bounded gate · Automation complacency.

Sources & Evidence

Claims made on this page and what supports them. The full registry lives in Evidence.

EmpiricalClinical assessors bound by algorithmic allocation with limited override capacity form a documented constraine…

Clinical assessors bound by algorithmic allocation with limited override capacity form a documented constrained-judgment pattern in home-care assessment.

sutton2020AcademicSave

Sutton, R. T., Pincock, D., Baumgart, D. C., Sadowski, D. C., Fedorak, R. N., & Kroeker, K. I. (2020). An overview of clinical decision support systems: Benefits, risks, and strategies for success. NPJ Digital Medicine, 3(1), 17. https://doi.org/10.1038/s41746-020-0221-y

doi.org/10.1038/s41746-020-0221-y

Appears in: 10²³AI authored research

upturnGroundingAdvocacySave

Upturn, Calculated Need: automated home-care hour allocation https://www.upturn.org/work/calculated-need/

https://www.upturn.org/work/calculated-need/

Appears in: PAN framework development

Grounds: deployment audit: Arkansas ARChoices / Idaho Medicaid

EmpiricalAutomated output checks are partial, not complete — measured detector-accuracy bands sit well below completene…

Automated output checks are partial, not complete — measured detector-accuracy bands sit well below completeness, especially on hard or adversarial content.

theillusionofprogressGroundingPreprintSave

'The Illusion of Progress' (arXiv:2508.08285) — LLM-as-Judge Precision 0.736 / Recall 0.957 / F1 0.832 vs human consensus on QA. https://arxiv.org/abs/2508.08285

https://arxiv.org/abs/2508.08285

Grounds: empirical cap: catch_at_generation (max)

halogenGroundingPeer-reviewedSave

HALoGEN (arXiv:2501.08292) — best models hallucinate 4%-86% of generated facts depending on domain. https://arxiv.org/abs/2501.08292

https://arxiv.org/abs/2501.08292

Grounds: empirical cap: model_error_base (min)

datadogllmasajudge2025GroundingIndustrySave

Datadog LLM-as-a-judge (2025) — detection F1 drops substantially from HaluBench to the harder RAGTruth; harder hallucinations are harder to catch.

Grounds: empirical cap: catch_at_generation (max)

mentalhealthchatbotdetectionGroundingPreprintSave

Mental-health chatbot detection (arXiv:2604.06216) — GPT judges 54.6% accuracy, 9.3% recall (miss 90.7% of hallucinations); traditional methods F1<0.30 on subjective content. https://arxiv.org/abs/2604.06216

https://arxiv.org/abs/2604.06216

Grounds: empirical cap: catch_at_generation (max)

EmpiricalAutomated catch fractions cap out below completeness — around 84% balanced accuracy in optimistic settings ver…

Automated catch fractions cap out below completeness — around 84% balanced accuracy in optimistic settings versus about 55% on hard content and 9.3% recall in worst-case measurements.

faithfulragleaderboardGroundingPreprintSave

Faithful RAG leaderboard (arXiv:2505.04847) — FaithJudge with o3-mini-high reaches ~84% balanced accuracy / ~82% F1 on FaithBench (optimistic ceiling). https://arxiv.org/abs/2505.04847

https://arxiv.org/abs/2505.04847

Grounds: empirical cap: catch_at_generation (max)

theillusionofprogressGroundingPreprintSave

'The Illusion of Progress' (arXiv:2508.08285) — LLM-as-Judge Precision 0.736 / Recall 0.957 / F1 0.832 vs human consensus on QA. https://arxiv.org/abs/2508.08285

https://arxiv.org/abs/2508.08285

Grounds: empirical cap: catch_at_generation (max)

mentalhealthchatbotdetectionGroundingPreprintSave

https://arxiv.org/abs/2604.06216

Grounds: empirical cap: catch_at_generation (max)

datadogllmasajudge2025GroundingIndustrySave

Datadog LLM-as-a-judge (2025) — detection F1 drops substantially from HaluBench to the harder RAGTruth; harder hallucinations are harder to catch.

Grounds: empirical cap: catch_at_generation (max)

samedetectionaccuracyliteratGroundingPeer-reviewedSave

Same detection-accuracy literature as catch_at_generation (FaithBench arXiv:2410.13210; arXiv:2508.08285); audit-time detection is bounded by the same hallucination-detection ceiling.

Grounds: empirical cap: catch_at_generation (max); empirical cap: decontaminate (max)

Bounded output screening

What it changes

Who can pull it

What it looks like institutionally

Documented cases behind this pattern

Michigan MiDAS

Deciding whether this lever fits your deployment?

Sources & Evidence