New Research · Q2 2026 10,765 agents · 4 yrs production data AI live

98% of your top performers would never have been interviewed.

We tested every screening filter a Fortune 500 insurance carrier runs — keywords, industry experience, assessments — against 10,765 hires and four years of actual production data. The filters don't predict. We built the infrastructure that does.

Dataset
10,765 agents
Lift, p=0.006
2.47× top performer
Status
VPC · live
fig 01 · decision trace · live artifact
customer-owned · reproducible
decision_trace · v0472.json
req: rv-2412-08831
MR
M. Rojas — Producer, Commercial Lines
req · austin-tx · screened 2026-04-14
87
fit score
Evidence chain
01 ATS filter
Industry experience ≥ 3 yrs retail banking · adjacent
would reject
02 Keyword
"licensed producer" · "P&C" not in resume
would reject
03 Assessment
Conscientiousness 82 · Coachability 79 dims: 28
above threshold
04 AI interview
Structured · 4 of 6 signals present 12 min
pattern match
05 Production
Predicted 18-mo output cohort p74 · $54/day lift
top-quartile
model: carrier-hire-2025.11 · calibrated on 4,214 hires
open trace →
Trusted inside regulated enterprises Deployed in customer-owned environments
The screening gap

Keyword filters don't predict production. Ours do.

We ran 8,181 candidate keywords against four years of post-hire production at a Fortune 500 carrier. After Bonferroni correction, none predicted sustained performance. Thirty were anti-predictive — correlated with lower output.

The standard ATS funnel at this customer eliminated 98% of their eventual top performers cumulatively. The industry-experience filter alone eliminated 80% of them.

80%
of top performers were filtered out by the "relevant industry experience" requirement alone.
$17.7M in annual production at a 2,000-hire volume that this single filter would have excluded from the pipeline.

Source: 4-year retrospective on 10,765 agents at a Fortune 500 carrier. Top performers defined as sustained-production cohort (p75+ over 18 months).
Keywords tested: 8,181
0 predict production (Bonferroni-corrected)
30 anti-predictive
AUC 0.512 → 0.618
Fig 02 · Production by Fit Score decile
5.37× production gap between top-scored and bottom-scored segments.
Avg. daily production ($) by Nodes Fit Score decile, 18-month sustained cohort. n=10,765. p=0.0007.
Bottom decile Top decile
decile 010203040506070809decile 10
Core IP · Decision Traces

Every hire, from filter to field, as queryable evidence.

A Decision Trace connects what the ATS screened on, what the assessment measured, what the interview surfaced, and what actually happened in production — for every candidate, indefinitely.

Institutional knowledge becomes auditable. When a hiring manager leaves, their judgment doesn't leave with them.

  • 01 · Ingest
    Every candidate event from your ATS, HRIS, assessments, and CRM zero data egress
  • 02 · Attribute
    Every decision — screen, advance, reject, hire — linked to its originating signal reversible
  • 03 · Measure
    Each signal scored against real production outcomes, not interviewer gut feel compound over time
  • 04 · Replay
    Query "show me every hire where the industry filter would have rejected a top performer" auditable
candidate #8871 of 10,765
Signal captured
What it measured
Production outcome
ATS keyword: "producer"
src: workday · stage: req-level
Resume string match
present: no · confidence 1.00
excluded
Would-have-been top quartile
counterfactual · monte carlo · n=312
Industry tag: "insurance ≥ 3y"
src: recruiter_notes · stage: screen
Prior industry tenure
value: 0 yrs · flag: adjacent
excluded
Top-performer pattern match
dims: 28/28 · score 87
overrode filter
Behavioral assessment
src: assessio · stage: mid-funnel
Conscientiousness, coachability
c: 82 · co: 79
advanced
Correlates with 18-mo output
r: 0.31 · p: 0.004
AI interview (structured)
src: nodes.interview · 12 min
4/6 top-performer signals present
transcript · sig hash 0a91f…
advanced
Hired · promoted at month 11
output: $54.35/day over cohort
validated
trace_id · dc_8871_2026-04-20
signed · reproducible · customer-owned
Platform

One model. The entire employee lifecycle.

Scoring logic calibrated against your validated top performers flows through every stage — sourcing, screening, interviewing, ramping, retaining, promoting. The model compounds.

01 · Fit Scoring

A 0–100 score, written back to your ATS as a native field.

Calibrated per role and per location against your validated top-performer pattern. 28+ behavioral, skill, and cultural dimensions. Every score ships with a plain-English rationale.

87
Producer — Commercial Linescalibrated · austin-tx · 28 dims
p-74
ATS-nativePer-role calibrationRationale included
02 · AI Interviews

Structured interviews conducted mid-funnel.

Scores auto-update against production outcomes — not interviewer gut feel. Full transcripts, signal timestamps, and signed audit trails available to hiring managers.

Structured interview · 12 min4 of 6 signals · transcript included
rec
advance
Async or liveCalibrated scoringFull transcripts
03 · Persona-based Sourcing

Find passive candidates that match what actually predicts production.

Identifies passive candidates across LinkedIn, GitHub, public portfolios, and industry networks using the same patterns the model learned from your own wins — not inferred boilerplate.

LinkedIn · GitHub · portfoliosLearned from internal winsNo boilerplate heuristics
04 · Post-hire Intelligence

Ten agents, one model, a compounding scoring substrate.

The same model that screens your pipeline runs ramp acceleration, retention risk, attrition modeling, internal mobility, succession planning, manager intelligence, and career pathing.

Same model end-to-endCustomer-owned weightsVPC-resident
Agent · 01
Ramp Acceleration
Role-calibrated onboarding signal
Agent · 02
Retention Risk
Early-warning signal stream
Agent · 03
Attrition Modeling
Cohort-level forecasting
Agent · 04
Internal Mobility
Cross-role pattern matching
Agent · 05
Succession Planning
Key-role readiness scoring
Agent · 06
Manager Intelligence
Team-level production trace
Agent · 07
Career Pathing
Next-role fit & readiness
Agent · 08
Compensation Fit
Pay-for-production calibration
Agent · 09
Territory Match
Geo-calibrated placement
Agent · 10
Referral Signal
Who's likely to refer a top performer
Deployment

Inside your VPC. Zero data egress. Legal in weeks.

Single-tenant deployment in customer-owned infrastructure. Fine-tuned, open-source models — no third-party AI in the chain. SOC 2 Type II. Reviewable by your security team the same way Snowflake is.

Legal approval
17days
After six prior vendors were rejected over eighteen months at the same carrier.
Contract → production
34days
VPC-deployed, integrated with the customer's ATS and HRIS, scoring live pipeline.
Third-party AI in the chain
0
Open-source foundation models, fine-tuned inside the customer boundary. Weights owned by the customer.
Compliance posture
SOC 2 Type II
Single-tenant. No shared inference. No shared training. Customer-owned fine-tuned models.
deployment.topology operational
Customer VPC · single-tenant
ATSworkday · live
HRISukg · hourly sync
CRMsalesforce · prod
Nodes modelfine-tuned · customer-owned
Trace storepostgres · encrypted
Inferencein-cluster · gpu
egress: 0 bytes · no third-party AI
External surface
Admin console only · SSO · SCIM · audit log stream
30 minutes · Your data · No pitch deck

See what the last four years of your hires actually predicted.

We'll run our backtest against a sample of your production data, in your environment. You'll see — numerically — which of your filters worked, which didn't, and what a calibrated Fit Score would have done differently.

01
30-minute call, your research & security team invited
02
We backtest against a de-identified sample of your data
03
You see — numerically — what your filters missed
04
If the math works, we scope a VPC pilot (median: 34 days to prod)