What makes AI ready for real work?

Begin with the work you need done.

A benchmark is a repeatable test with a defined task and scoring rule. A reasoning test, a document comparison and a tool-use task each observe a different part of an AI system. Their scores answer different questions.

For a document workflow, success means the requested file can be opened, its evidence can be checked and its content meets the brief. For data analysis, the output must also be reproducible from the supplied data. A general model ranking does not resolve either decision on its own.

We use public results to narrow candidates and identify failure modes worth testing. To decide whether a method belongs in LabChin, we then need a matched task using our own sources, permissions, tools and output checks. This report follows that order.[1]

Judge the deliverable against the actual brief.

GDPval makes professional deliverables the object of comparison. Its original Gold study used 220 tasks across 44 occupations, with experts comparing outputs without knowing who produced them. The chart reproduces three archived results from that study.

A win-or-tie rate describes those comparisons. It is not the proportion of jobs an organization can hand over to AI. The task set, reviewer judgment and model configurations define how far the result can be interpreted.[2]

For HDATF, the practical question is how much work remains after the file is produced. A review should retain missing requirements, unsupported numbers and the revisions a person had to make. A polished layout is one part of quality; it cannot compensate for incorrect evidence or an unusable file.

Finding an answer and supporting a report are separate checks.

BrowseComp tests whether a system can locate difficult facts. Parallel's published comparison used a random sample of 100 questions from the 1,266-question set. It reported both accuracy and request cost. The results below therefore describe a sampled, provider-run experiment.[3][4]

Ultra8x reported 58% accuracy at $2.40 per request; Ultra reported 45% at $0.30. Their budgets differ. These figures make the accuracy-cost tradeoff visible, but do not isolate the effect of the underlying model or prove an advantage at equal cost.

A research report needs additional checks. DeepResearch Bench II and ResearchRubrics examine report content against explicit criteria. For LabChin, we plan to retain the passages supporting each major claim, identify contradictory evidence and record which parts of the question remain unanswered.

The model works inside a system. Evaluate that system.

An agent harness is the software that supplies a model with tools, context and control over a task. LangChain held GPT-5.2-Codex fixed while changing this surrounding software. Its Terminal-Bench 2.0 result rose from 52.8% to 66.5% across 89 tasks.[5][6]

The study changed instructions, tools and middleware together, including checks before completion and guidance about the environment. The 13.7-point difference belongs to that combined experiment. It does not establish the contribution of each individual change or predict HDATF's improvement.

This suggests a testable choice for our own work: hold the task and model fixed, then compare a workflow with and without output verification. Keep the produced file, verification result, retries and elapsed time. The test should explain both successful completion and the failures that required intervention.

Memory needs a separate test. Recalling an old fact is useful only if the system also handles corrections and keeps information within its permitted scope. The archived LoCoMo result covers single-hop recall; it does not verify these work-level requirements. We therefore keep it as supporting evidence in the appendix.

The next evidence must be our own work, measured.

The public studies identify useful questions; they do not supply a score for our product. Our next step is a recorded comparison on representative tasks. We plan to assess evidence retrieval, analysis and usable delivery separately, so a strong result in one area cannot hide a failure in another.

For research and data work, preserve the question, source versions and checks on the final output. A report should expose unsupported claims. An analysis should retain the input data, execution record and reproducible outputs. The following criteria define what we intend to inspect, not completed tests.[1][7][8][9]