The evidence behind the claims.
This ledger maps the evidence anchors used across the site to their source type, population, period, outcome, method, and status. It includes currently approved claims only.
If a number elsewhere conflicts with this ledger, treat it as unverified until the underlying source record is reviewed and the conflict is corrected.
Separate what happened from what was estimated.
Observed records describe measured events. Validated findings test evidence against realized outcomes. Modeled findings estimate an association or counterfactual and are not realized savings.
Published research
The Decision Traces paper, arXiv:2604.19819, studies four years of records for 10,765 agents at one Fortune 500 insurance carrier. Its findings are retrospective and specific to the measured outcomes and populations described below.
Production records
The current production evidence comes from one Fortune 500 insurance carrier. The deployment has been live since January 2025. It is the only production-proven domain represented here.
Modeled analysis
Counterfactuals and fitted estimates answer what the historical record suggests, not what another company is guaranteed to realize. They require validation against each customer's own decision history before budgeting or production use.
Anchors for approved public claims.
Observed means measured in production records. Validated means tested against realized outcomes. Modeled means estimated by a fitted model or retrospective counterfactual.
Links such as /evidence#c01 point to the exact row a reader should inspect.
| Claim | Value | Evidence | Population | Period | Outcome | Method | Status |
|---|---|---|---|---|---|---|---|
| C01Candidates scored in the live deployment | 900,000+ | Production record | One Fortune 500 insurance carrier | Since Jan 2025 | Candidates scored | Customer production record | Observed |
| C05Published research cohort | 10,765 agents | Published research | Agents at one Fortune 500 insurance carrier | 2022 to 2025 | First production milestone | Retrospective observational study, arXiv:2604.19819 | Observed |
| C10Top performers removed by rigid filters | 80% by one filter; 98% cumulatively | Published research | Exact subset of n=50 top performers with the required text | Study window | RPA top-performer milestone | Retrospective filter counterfactual | Modeled |
| C11Exact deployment footprint | 215+ locations | Production record | One Fortune 500 insurance carrier | Since Jan 2025 | Operational footprint | Customer production record | Observed |
| C20Net savings, CFO-validated | $1.58M | Production record | One live deployment | Q1 2025 | Net savings | Customer record, validated by the customer's CFO | Observed |
| C21Annual production put at risk by one insurance-experience filter | $17.7M | Published research | Producing agents the filter would have rejected | 2022 to 2025 | Annual production at risk | Retrospective counterfactual on observed production | Modeled |
| C22Producing agents the filter would have rejected | 2,863 | Published research | One Fortune 500 insurance carrier | 2022 to 2025 | First production milestone | Retrospective counterfactual on observed production | Modeled |
| C23Planning coefficient for each day faster to the first production milestone | $54.35 per producing person | Published research | Producing-agent records at one carrier | 2022 to 2025 | Annual production | Fitted relationship in the reference record | Modeled |
| C25Median time from contract to first production milestone | 109 to 62 days, 47 days faster | Production record | Reference carrier cohorts | 2022 to 2025 | First production milestone | Historical median comparison | Observed |
| C26Requisition-to-hire time, distinct from C25 | 127 to 38 days | Production record | One live deployment | 2025 | Days from requisition to hire | Customer production record | Observed |
| C27Thirty-day ramp planning reference | $1,357 per producing agent per year | Published research | Producing-agent records at one carrier | 2022 to 2025 | Annual production | Reference estimate for a 30-day reduction | Modeled |
| C32Legal approval after six vendors had been rejected over 18 months | 17 days | Production record | One carrier vendor review | Before Jan 2025 launch | Legal approval time | Customer record | Observed |
| C33Contract to production | 34 days | Production record | One carrier deployment | Before Jan 2025 launch | Deployment time | Customer record | Observed |
| C40Resume keywords parsed and testable | 8,181 parsed; 3,597 testable | Published research | Applicant records at one carrier | 2022 to 2025 | Keyword testability | Keyword census, arXiv:2604.19819 | Observed |
| C41Predictive resume keywords after correction | 0 of 3,597 | Published research | Testable keyword set at one carrier | 2022 to 2025 | First production milestone | Hypothesis testing with Bonferroni correction | Validated |
| C42Resume keywords that were anti-predictive after correction | 30 | Published research | Testable keyword set at one carrier | 2022 to 2025 | First production milestone | Hypothesis testing with Bonferroni correction | Validated |
| C46AUC, resume keywords | 0.558 | Published research | One-carrier research cohort | 2022 to 2025 | Ranked separation on the defined outcome | AUC analysis | Validated |
| C47AUC, behavioral assessment alone | 0.647 | Published research | n=229, small one-carrier research sample | 2022 to 2025 | Ranked separation on the defined outcome | Cross-validated AUC | Validated |
| C48AUC, fused evidence | 0.735 | Published research | n=229, small one-carrier research sample | 2022 to 2025 | Ranked separation on the defined outcome | Cross-validated AUC | Validated |
| C49Ramp to production in the reference deployment | 8 to 12 months to 6 weeks | Production record | Reference carrier ramp cohorts | Reference customer records; exact period not published | Ramp to the defined production threshold | Historical cohort comparison | Observed |
The case study separates observed production outcomes from modeled counterfactuals. The research library explains each method and its limits.
Past data only. No live decisions.