A Jev that can ask Opus for help gives Opus accuracy at 4x lower cost

Hightouch evaluated whether lightweight decision models can audit customer identity resolution systems and found that combining small models with frontier engines achieves top-tier accuracy at a fraction of the cost. According to hightouch.com's report, pairing Jev with Claude Opus 5.5 delivered frontier accuracy while cutting operating costs by more than fourfold. Standalone decision models like Jev and GPT-6 Luna achieved 88.8 to 97.3 percent accuracy, running ten times faster and costing ninety times less than frontier models.

This report examines how automated language model judges identify record clustering mistakes and over-merged profiles across enterprise customer databases. It covers benchmark results, latency metrics, cost structures, and confidence thresholds across Jev, GPT-6 Luna, Claude Opus 5.5, Gemini 3.1 Pro, and GPT-6 Sol. The analysis is designed for data engineers, analytics leads, and database architects seeking to verify customer graphs without manual audits. Readers will find exact benchmarks and routing strategies required to evaluate identity graphs across enterprise warehouses.

A Jev that can ask Opus for help gives Opus: what actually changed

Hightouch established that data teams no longer need to choose between slow manual reviews, costly frontier models, and unreliable heuristics when auditing customer identities. In benchmark tests, Claude Opus 5.5 reached 99.0 percent accuracy with three false positives and seven false negatives, taking 2.2 seconds per identity at $3.86 per thousand records. Decision models like Jev ran in 0.15 seconds at $0.05, while GPT-6 Luna took 0.17 seconds at $0.10. A hybrid harness pairing Jev with Opus delivered frontier accuracy at four times lower cost.

The decisive breakthrough stemmed from differences in error calibration between competing decision models. Out of the box, GPT-6 Luna posted 97.3 percent accuracy compared to Jev at 95.6 percent under a default fifty percent threshold, but Luna exhibited high confidence when wrong. Conversely, Jev displayed superior calibration because its mistakes accompanied low confidence scores. Raising Jev's threshold to sixty percent achieved 95.7 percent accuracy. Because Jev signals when uncertain, an automated harness routes only ambiguous records to Claude Opus, preserving accuracy while cutting expenses.

A Jev that can ask Opus for help gives Opus: how it works

Researchers tested systems using a benchmark created with pseudopeople, a simulation tool from the Institute for Health Metrics and Evaluation at the University of Washington. The software modeled a fictional town using census records, household surveys, WIC files, Social Security rosters, and tax forms spanning 2020 through 2040. From this data, Hightouch built one thousand sample identities, splitting them between single individuals and over-merged profiles. Prompts instructed models to evaluate records within designated tags, determine individual counts, and supply justifications only when an over-merge occurred.

Prompts incorporated explicit noise examples, such as typos changing Valentino to Valentinp or 1978 to 1973, character misreads like Sam becoming 5am, and phonetic spellings like Phillips turning into Filips. Prompts also covered swapped month and day values, sibling nicknames, placeholder female child names, blank fields, and copied household details. While error examples improved performance across all models, supplying contextual descriptions of government sources added at most one point to frontier models and provided zero benefit to decision models, which score conditions rather than conducting reasoning.

A Jev that can ask Opus for help gives Opus: why it matters

Identity resolution maps records to real people, forming the foundation for customer lifetime value tracking and enterprise analytics. Over-merging represents the most dangerous failure mode because it blends distinct individuals into a single profile. Measuring over-merging on production datasets is difficult because real data lacks an answer key. Teams often spend months reviewing statistical heuristics or sampling records manually. Because human auditors review only fifty identities per hour, gathering over one thousand samples to achieve precision within plus or minus one percent is painfully slow.

Automating audits with language models enables organizations to grade hundreds of thousands of identity clusters in parallel within seconds. While frontier alternatives like Gemini 3.1 Pro reached 98.3 percent accuracy at $4.48 per thousand records and GPT-6 Sol achieved 98.3 percent at $2.50, relying exclusively on large models remains cost-prohibitive. Inexpensive decision models priced between $0.05 and $0.10 enable continuous monitoring, but their six-point accuracy drop on difficult cases risks undetected merges. Tiering Jev and Claude Opus eliminates this financial bottleneck, allowing teams to run regression audits across entire identity graphs.

A Jev that can ask Opus for help gives Opus: who is affected

Data engineers, customer data platform administrators, and warehouse architects managing identity graphs directly benefit from these findings. Consumer platforms across retail, financial services, and healthcare continuously encounter dirty warehouse data where simplistic rules break down. For example, simpler over-merges involving Daniel Z Gardner at 11648 Main Street and Jenna K Gardner at 8203 West Farwell Avenue across census, W-2, and 1040 records are easily caught by standard decision models. However, difficult cases involving shared surnames such as Christopher Brentley Carpenter and Christopher B Carpenter require the deeper discriminatory capabilities of frontier engines.

Marketing leaders and compliance officers are also impacted because inaccurate identity resolution corrupts customer attribution and creates data governance risks. Over-merged profiles aggregate purchases across separate customers, distorting lifetime value calculations and causing marketing workflows to send irrelevant recommendations to merged users. Furthermore, mixing personal records creates severe privacy violations by exposing sensitive account information. Implementing tiered automated auditing architectures allows platform teams to reliably quantify error rates across identity graphs, ensuring customer profiles accurately reflect real-world individuals before data feeds downstream operational systems, marketing clouds, and executive dashboards.

A Jev that can ask Opus for help gives Opus: what to watch

Teams implementing decision models must monitor confidence calibration curves as models update and enterprise record distributions evolve. In benchmark tests, adjusting GPT-6 Luna's threshold from fifty percent to seventy percent reduced false positives from seventeen to fourteen while increasing false negatives from ten to thirteen, keeping accuracy flat at 97.3 percent. By comparison, tuning Jev to a sixty percent cutoff lowered false positives from thirty-one to twenty-two, which is vital for filtering ambiguous records. Engineers must verify whether model providers modify scoring distributions, because unannounced calibration shifts could trigger excessive calls to Claude Opus, eroding savings.

Engineers should observe whether future decision models can handle complex contextual dependencies without requiring external generative reasoning support. Because decision models gained nothing from detailed narratives about public agency sources, developers should concentrate prompt optimization on real error examples rather than lengthy background context. Additionally, production deployments must measure how hybrid harnesses perform when evaluating millions of records across international character sets, diverse naming conventions, and non-traditional address structures. Tracking whether next-generation decision models can independently untangle near-identical family profiles without frontier assistance will reveal whether tiered routing remains a permanent requirement.