How Centific regrades frontier AI work at three levels of specificity, and what our finance pilot found

Connect with Centific to discover what's next in AI.

See where to meet us

Connect with Centific.

Find an event

Platforms

Data Marketplace

Data Canvas

AI Data Foundry

OneForma

AI Localization

Expert Network

Join our Expert Network

Build & Train AI

RL Environments

Data Collection & Creation

RLHF & Preference Optimization

Supervised Fine Tuning

Model Safety & Evaluation

Internationalization

Vertical AI

Physical AI

Healthcare

Vision AI

Explore our full suite of AI platforms, data marketplaces, and expert services designed to build, train, fine-tune, and deploy reliable, production-grade AI systems at scale.

Platforms

Data Marketplace

Data Canvas

AI Data Foundry

OneForma

AI Localization

Expert Network

Join our Expert Network

Build & Train AI

RL Environments

Data Collection & Creation

RLHF & Preference Optimization

Supervised Fine Tuning

Model Safety & Evaluation

Internationalization

Vertical AI

Physical AI

Healthcare

Vision AI

Explore our full suite of AI platforms, data marketplaces, and expert services designed to build, train, fine-tune, and deploy reliable, production-grade AI systems at scale.

Book a Demo

Research insight

How Centific regrades frontier AI work at three levels of specificity, and what our finance pilot found

Learn how enterprise AI evaluation should measure both the quality of AI outputs and the reasoning behind them, using evidence-based grading and human expert validation to assess real-world business performance.

Published on Jul 7, 2026

•

5 min read time

Table of contents

Summarize

AI Summary by Centific

Turn this article into insights

with AI-powered summaries

Summarize article

Give me key takeaways

Topics

AI Evaluation

Enterprise AI

LLM Benchmarks

AI Data

AI Evaluation

Enterprise AI

LLM Benchmarks

AI Data

Author(s)

Parth Kulshreshtha

Shivali Dalmia

Prasanna Desikan

The top AI models have closed the distance between each other. On professional tasks, financial models, research memos, executive reports, they now produce deliverables that are difficult to tell apart on quality, and increasingly difficult to distinguish from a human expert's work.

Closing that distance creates a problem for evaluation: a benchmark built to score the finished deliverable has little to say once every deliverable looks equally polished. What it needs instead is a rubric with tiers. Instead of one flat checklist, Centific built a rubric that grades every task at three levels of specificity: what any competent finance professional must get right, what the particular occupation demands, and what this single task requires.

Tiering fixes how sharply we grade. It does not settle what we grade. Our answer was to score each task on two separate questions: whether the deliverable was correct, and whether the model reached it through sound reasoning, both judged against work produced by human experts in the same role. A model can land the right number by an unsound route, and on professional work that distinction matters as much as the answer. Grading both is where AI benchmarking has to go next, because a correct output is no longer evidence that the model knew what it was doing.

Most AI benchmarks grade whether work looks right, not whether it is right

Most AI benchmarks score the shape of an answer: whether it has the expected sections, covers the expected points, or resembles a reference answer. None of those checks tell you whether the substance underneath is correct. Two shortcomings stand out:

Presence, not correctness. A deliverable can contain every section a reviewer expects and still rest on a flawed calculation or an invented figure. Two answers can look equally complete while one is right and the other is wrong.
Outcome, not process. Benchmarks grade the final artifact, not how the model arrived at it. A right answer reached through a broken process scores the same as one reached soundly, but only one of them will hold up on the next task.

For enterprise work, closing that space is the whole point. You need to know that the underlying work is truly correct and produced by a process you can rely on.

Centific’s approach: grade the work, not its appearance

Centific built an evaluation approach that checks the work itself, on two levels:

Outcome-based evaluation. Grade the final deliverable: is it correct, complete, and grounded in the source material? Each answer is checked against a layered answer key: general professional standards first, then the demands of the specific profession, then the exact task.
Trajectory-based evaluation. Grade how the model got there: the steps, the reasoning, and the use of the provided data, catching the failure that surface scoring misses: a right answer reached the wrong way, or a number that appears without ever being derived.

The grading runs in two passes. A first pass pairs deterministic rubric checks, objective items verified straight from the artifact, with AI-assisted judging, where an independent panel of models (drawn from different families than the systems under test) rules on the judgment calls, each verdict backed by evidence rather than impression. A subject-matter expert then validates the results before they count. Throughout, every answer is measured against a deliverable from a human expert in the same role.

Outcome and trajectory checks run through a first pass of deterministic rubrics plus AI-assisted judging, then a subject-matter expert validates the results.

Testing the approach on a finance pilot

To pressure-test the approach, Centific applied it to finance, a domain we chose because of its difficulty level. The testbed was GDPval, OpenAI’s benchmark of economically valuable knowledge work drawn from professional occupations. Finance is an unforgiving case: most deliverables are models built on shifting, assumption-driven inputs, where a single early slip cascades through everything downstream.

Centific rebuilt the finance rubrics into the three-tier structure, put the evidence-bound panel to work, and graded leading frontier models against deliverables produced by human experts in the same roles.

Explore the three-tier rubric applied to a sample finance task pilot in the Centific benchmark viewer.

Frontier models matched human experts on most finance tasks

One finding stood out, and it was not about which model won. On these finance tasks, today’s frontier models matched or beat the expert human deliverables on the majority of tasks, and the top models were so close to one another that the outcome-only evaluation could no longer clearly separate the top models.

The scores also showed where the difficulty still lives: the occupation-specific tier was the soft spot. The frontier models cleared the general and task tiers in the low-to-mid 90s but slipped to around 85% and 82% on occupation-specific expectations. And because that tier is the hardest to clear, it is exactly where the evaluation still separates strong work from merely adequate: role-specific professional judgment, not task completion, is where the shortfall remains.

The instrument had become sharper than the tasks it was measuring. When the best systems all do near-perfect work on the deliverable, the deliverable alone can no longer separate them. Grading how the work was done becomes the difference that still counts. Tellingly, when a model did edge past a human, it was often because the human had made a small, checkable error the model avoided.

Grading the reasoning, not just the outcome, is what earns enterprise trust

An evaluation that checks the actual work, outcome and trajectory, with every verdict tied to evidence, lets an organization deploy AI on professional tasks with confidence, and tell the difference between a system that is reliably right and one that is merely convincing. Because the approach grades against professional standards rather than finance-specific details, it can also be applied to other enterprise functions where companies want that same confidence in AI’s work.

To learn more about Centific’s Agentic and RL capabilities, explore our benchmarks or contact us to request a demonstration.

Are your ready to get

modular

AI solutions delivered?

Centific offers a plugin-based architecture built to scale your AI with your business, supporting end-to-end reliability and security. Streamline and accelerate deployment—whether on the cloud or at the edge—with a leading frontier AI data foundry.

Start Building

Connect data, models, and people — in one enterprise-ready platform.

Latest Insights

Ideas, insights, and

research from our team

From original research to field-tested perspectives—how leading organizations build, evaluate, and scale AI with confidence.

Explore

Research insight

Can voice agents help you get through the day?

Aug 13, 2026

Research insight

Why AI agent benchmark scores need a trust layer

Aug 7, 2026

Press release

Centific Announces Leadership Transition: Chuen Hong Lew to Succeed Venkat Rangapuram as Chief Executive Officer

Aug 3, 2026

Connect with Centific

Stay ahead of what’s next

Stay ahead

Updates from the frontier of AI data.

Receive updates on platform improvements, new workflows, evaluation capabilities, data quality enhancements, and best practices for enterprise AI teams.

Book a Demo

Get a live walkthrough

Talk to our team

Careers

See all our open positions

Turn data into AI that works

Book a demo