Restricted Access
The Model QA view is limited to finance model administrators. Your signed-in account isn't on the access list — contact Derek Jensen if you believe you should have access.
Admin — Forecast Trust Monitor
Is The Model Still Telling The Truth?
Week-1 projections graded against GL actuals, drift tripwires per driver, and the live throughput governors that anchor the model to demonstrated behavior. Snapshots are captured nightly; a status of drift means the model and reality have parted ways and assumptions need review.
Loading model QA…
Weekly Reasonability Check AP vs AR
A week where we collect heavily and pay out very little is usually the AP side being understated, not a real cash windfall. Any week under 50% is flagged for someone to confirm before the number is trusted. Scenario base_v2.
| Week | AR in | AP out | Payroll out | AP / AR | Job AP deferred | Status |
|---|
AP Deferral Queue Oldest first
The waterfall defers job AP as one aggregate number. This walks the funded dollars across the actual vouchers oldest-to-newest, so the deferral is a list of invoices rather than a total. Sub-labor never appears — it is must-pay.
| Vendor | Invoices deferred | Amount deferred | Longest wait |
|---|
Rule Validation Regression suite
The specific numbers Lucas and Jenifer quoted in review, asserted against live data. These are the cases the corrections were built to satisfy — if a later change breaks one, it shows up here rather than in a meeting. Evaluated on cache refresh, not on page load.
| Case | Assertion | Expected | Actual | Result |
|---|
AR Model Backtest v3 gate
The promotion gate for the behavior-curve AR engine (base_v3). Each of the last 12 Mondays is replayed with curves trained only on data available before the replay window: predict the next 4 weeks of collections from that day's open invoices, then grade against what those same invoices actually did. Expected flow (dollar-weighted payment hazards) is the live v3 line; P50 placement (each invoice whole at its most-likely week) is shown for contrast — it failed calibration and is used only for the per-invoice drill, never the line. Bias 1.00 = perfectly calibrated.
| Horizon | Actual collected | Flow predicted | Flow bias | Flow MAE / wk | Placement bias | Verdict |
|---|
Drift Monitor Trailing 8 wks
Average week-ahead projection vs what actually happened, per driver. ok < 20% bias · watch 20–40% · drift ≥ 40%. Snapshots taken before a model change keep grading the OLD model — expect statuses to clear as new snapshots accumulate.
| Driver | Weeks | Avg projected / wk | Avg actual / wk | Bias (actual − proj) | Bias % | Status |
|---|
Throughput Governors
The model cannot project cash converting faster than we have demonstrated. Capacities re-measure nightly from trailing actuals — if behavior changes, the model follows automatically and the drift monitor shows the transition.
Week-1 Accuracy History
Each row: the projection made at the start of that week vs the week's actuals (net cash change).
| Week | Projected net | Actual net | Miss | Projected AR in | Actual AR in | Projected AP out | Actual AP out |
|---|