retention-risk

Example: Retention Risk — Committee View (governed glass-box ML)

The retention-arm marquee: one dark, one-page committee dashboard that renders a segment-first glass-box hazard model as a planning instrument — and ships only behind a human publish gate.

Its design thesis: every "AI for HR" dashboard claims to be trustworthy; this one performs it.

• The published model, reconstructed — not re-fit. Every number comes from the engine (foundation/compute/retention.py). The agent rebuilds the pinned trained artifact from its manifest (model_from_manifest — coefficients, standardizer, Platt calibration, band thresholds) and scores the exact published model, so the dashboard renders fast and reproduces the committed metrics in CI.
• Disagreement as a first-class visual. Segment risk is computed two ways — the model's calibrated bottom-up and the observed Kaplan-Meier top-down — and the gap is plotted in red, not averaged away. Customer Success reads 18.3% where history shows 3.3% (+15.0pp); a governed glass-box shows that and asks the committee to interrogate it.
• Every metric beside its no-skill baseline. On a ~1.8%-per-month event, honest numbers look small — so the trust panel prints the baseline next to each one (PR-AUC 0.078 vs 0.018; precision@decile 8.4% vs 1.8% → 4.6× lift, 76 of 900 flags real). The realism guard fails the build if the model looks too good, and prints the ceiling it refuses to cross right on the KPI.
• Glass-box, exact, associational. The driver panel is the model's additive log-odds per feature — unvested equity and comp-ratio protect; stale raises/promotions and manager-team churn push — the exact decomposition, labeled associational, not causal. Three planted decoy features rank #15/#17/#23 of 23: a legible leakage tripwire.
• Segment-first, never a person. No individual score, no name, no leaderboard anywhere; regions are broad-only; small segments are suppressed (floor n ≥ 30). Fairness is shipped as a visibly unchecked checklist — a governed system shows its unfinished edges. The model card is the binding contract: this is a planning signal, never an adverse-action input.
• Drawn deterministically. Every chart is server-rendered inline SVG — no JavaScript, no CDN — so the committed HTML/digest are byte-diffed in CI. The PNG below is an illustrative snapshot (browser rendering isn't part of the deterministic gate).

All data is synthetic. It demonstrates the mechanics + governance of a glass-box retention model, not real-world predictive accuracy. Voluntary exits only.

Sample output

Run it

cd examples/retention-risk
python3 run.py                                          # draft only
python3 run.py --publish                                # refused: needs a valid approver
python3 run.py --publish --approved-by "People Analytics Lead"

Test it

python3 evals/test_retention_agent.py

The eval proves the composer is presentation-only over the engine (segment/company numbers equal the engine's), the charts are deterministic and injection-safe, no per-employee identifier or country-level region leaks, the honesty rails render (guardrails, no-skill baselines, decoy readout, unchecked fairness boxes), the publish gate refuses an invalid approver, and it fails closed when the engine is unavailable or the realism guard trips. See SPEC.md.