Reduce downside and recover value hidden by rigid filters

Nodes connects historical hiring decisions to the production outcomes that followed. Before the next governed decision, it shows the evidence, uncertainty, value at risk, and recommendation. A named person approves, edits, delays, or declines.

The current production proof is insurance talent at one Fortune 500 carrier. The deployment has been live since January 2025 across 200+ locations nationwide. The research and production records are customer-specific; another carrier must validate the result against its own roles, definitions, policies, and outcomes.

Published method: Four years of records covering 10,765 agents at one carrier. Read Decision Traces on arXiv.

Explore the inputs: Use the insurance hiring ROI calculator, then replace the reference assumptions with your approved numbers in a Decision Replay.

What the historical record found

• The study parsed 8,181 skills and could test 3,597 keywords against the defined production milestone. After Bonferroni correction, zero predicted that milestone and 30 were anti-predictive. Review the keyword analysis.
• One insurance-experience filter would have rejected 2,863 producing agents associated with $17.7M in annual production. This is a retrospective counterfactual, not realized savings or a forecast. Review the counterfactual.
• Full evidence fusion reached AUC 0.735 in an evaluable n=229 small sample from one-carrier research. It measures ranked separation within that sample and does not predict certainty for an individual. Review the model comparison.
• Median time to the first production milestone moved from 109 to 62 days in the reference record. The 47-day observed difference is separate from requisition-to-hire time, which moved from 127 to 38 days.
• The customer's finance team validated $1.58M in Q1 net savings. This is customer attribution within one deployment and does not prove Nodes caused every operational change.

How the insurance workflow is governed

Nodes reads approved context from the systems the carrier already runs. A workflow defines its trigger, evidence sources, outcome, policies, human approval path, bounded action, and later measurement. The system proposes the next action with supporting evidence and the value of acting or waiting.

A named reviewer makes the decision. Approved work can then move across the permitted systems. A signed Decision Trace records the evidence, model and policy versions, human input, approved action, and later outcome. The human choice remains context; the governed outcome supplies evidence for later evaluation.

When a later calibration candidate is proposed, it runs in shadow against the incumbent. A cleared threshold makes it eligible for review. A named human must promote it before it can serve production recommendations.

Customer-controlled deployment

Production inference is VPC-resident and single-tenant. Customer records, inputs, outputs, outcomes, and Decision Traces stay inside the customer's approved environment. The customer owns its weights. Nodes holds SOC 2 Type I and Type II certifications.

A separate release path may move a customer-approved weight artifact after two independent PII checks. Any returned upgrade is tested on synthetic data, evaluated locally in shadow, and promoted only by a named human. Review the VPC boundary.

Frequently asked questions

Does Nodes automatically hire or reject an applicant? No. Nodes shows evidence, counterevidence, uncertainty, and a recommendation. A named human makes the final decision. Regulated execution requires the approved human gate.

Does AUC 0.735 mean 73.5% accuracy? No. It is a ranked-separation measure from an n=229 small one-carrier research sample. It is neither an individual probability nor a universal production rate.

What is observed versus modeled? The timing and customer-record savings figures are observed or customer-attributed within the reference deployment. The $17.7M filter analysis is a retrospective counterfactual. The $54.35 daily coefficient and $1,357 annual planning reference are modeled and require local validation.

Can another carrier use the same model without validation? No. A new carrier, role, or outcome begins with a read-only historical replay and its own proof threshold.

Bring one repeated insurance talent decision and the outcome that followed. Request a Decision Replay.