← demo
Learning · as of 2026-12-06
Backtest: 44 weekly origins (2026-01-05 → 2026-11-02), 28-day horizon, workload (volume × handle time) on every gate-channel-day. The cycle never reads the planted truth.
What the evidence says
- The record is hard to beat on a quiet estate. Workload WAPE is 11.01% for the record and 11.00% with every learned correction applied everywhere (+0.01 pts). Most gates have nothing real to learn, so applying corrections everywhere chases noise.
- Learning has to be earned, gate by gate. Under the walk-forward policy (a gate uses its corrections only after they beat the record in earlier windows) the estate scores 10.97% (+0.04 pts). That is the policy the proposal now uses, together with any confirmed shift in the last 70 days. 5 of 60 gates are learning today.
- The explicit overlays carry the big gains. Of the applied signals, 0 of 0 helped. Of the signals kept off the forecast, 1 would have hurt (kept off correctly) and 1 would have helped (held back at a cost). Migration ratios, calendar and client signals are modeled changes with a source. They are not inferred from noise.
- Human edits are scored like everything else. none to score.
- Best layer in the backtest: earned (walk-forward) (10.97%).
Backtest by layer (workload WAPE, %)
| Scope | record | + learned volume | + learned AHT | + migration ratio | earned (walk-forward) | shift-gated |
|---|
| All gates | 11.01 | 11.04 | 11.02 | 11.00 | 10.97 | 10.99 |
| Product A | 11.99 | 12.03 | 12.04 | 12.04 | 11.99 | 11.98 |
| Product C | 11.34 | 11.33 | 11.34 | 11.34 | 11.34 | 11.34 |
| Product D | 9.74 | 9.76 | 9.77 | 9.77 | 9.74 | 9.74 |
| Product E | 10.76 | 10.82 | 10.70 | 10.62 | 10.58 | 10.67 |
| Bias, all gates | -1.89 | -2.36 | -2.37 | -2.34 | -1.89 | -1.82 |
Forecast value added (FVA) = WAPE of the record − WAPE of the layer, in points; positive means the layer helped.
Forward check · the proposal of 2026-10-04 against actuals through 2026-12-05
| Layer | Volume WAPE | Bias |
|---|
| record | 10.75% | -2.36% |
| + learned level | 10.75% | -2.35% |
| + migration ratio | 10.69% | -2.13% |
| + applied signals | 10.5% | -2.13% |
| + human edits (published) | 10.5% | -2.13% |
Each signal, counterfactual and human edit, on the gate-days it touched
| Item | Kind | What | Gate-days | WAPE without | WAPE with | FVA | Verdict |
|---|
| SG-007-1 | counterfactual | had the proposed platform signal (I) been applied: +10% from 2026-10-20 | 282 | 11.1% | 9.95% | +1.14 | would have helped: held back at a cost |
| SG-004-1 | counterfactual | had the rumor client signal (U) been applied: -15% from 2026-10-12 | 55 | 9.43% | 18.3% | -8.87 | would have hurt: kept off correctly |
Auto-approval: earned by 54 of 60 gates
A gate may be approved without review when its proposal moves less than 1% and the proposal beat or matched the record in at least 80% of its last 12 windows, never losing one by more than 2 points.
| Gate | Product | Windows | Beat or matched | Worst loss | Record WAPE | Proposal WAPE | Status |
|---|
| G101 | E | 12 | 11 | 0.62 pts | 16.43% | 14.67% | earned |
| G102 | C | 12 | 12 | 0.4 pts | 15.19% | 15.29% | earned |
| G104 | D | 12 | 12 | 0.38 pts | 13.95% | 14% | earned |
| G105 | A | 12 | 12 | 0.32 pts | 11.75% | 11.83% | earned |
| G106 | A | 12 | 12 | 0.31 pts | 18.07% | 18.16% | earned |
| G107 | A | 12 | 12 | 0.22 pts | 14.48% | 14.54% | earned |
| G108 | A | 12 | 12 | 0.41 pts | 12.92% | 12.97% | earned |
| G109 | A | 12 | 12 | 0.25 pts | 17.27% | 17.37% | earned |
| G111 | A | 12 | 10 | 0.51 pts | 14.45% | 14.71% | earned |
| G112 | A | 12 | 12 | 0.35 pts | 13.76% | 13.9% | earned |
| G113 | A | 12 | 12 | 0.12 pts | 15.87% | 15.62% | earned |
| G114 | A | 12 | 10 | 0.6 pts | 15.16% | 15.19% | earned |
| G115 | A | 12 | 12 | 0.49 pts | 11.55% | 11.52% | earned |
| G116 | A | 12 | 10 | 0.8 pts | 12.87% | 13.16% | earned |
| G117 | A | 12 | 11 | 0.64 pts | 12.05% | 12.13% | earned |
| G118 | A | 12 | 12 | 0 pts | 12.06% | 11.84% | earned |
| G119 | A | 12 | 10 | 0.77 pts | 15.1% | 15.39% | earned |
| G120 | A | 12 | 12 | 0.28 pts | 12.41% | 12.47% | earned |
| G121 | A | 12 | 12 | 0.46 pts | 16.61% | 16.75% | earned |
| G122 | A | 12 | 12 | 0.14 pts | 15.13% | 15.05% | earned |
| G123 | A | 12 | 12 | 0.17 pts | 13.2% | 13.25% | earned |
| G124 | A | 12 | 12 | 0.33 pts | 12.33% | 12.39% | earned |
| G126 | A | 12 | 12 | 0.12 pts | 10% | 10.05% | earned |
| G127 | A | 12 | 11 | 0.62 pts | 18.76% | 18.83% | earned |
| G128 | A | 12 | 10 | 0.53 pts | 14.55% | 14.79% | earned |
| G129 | A | 12 | 12 | 0.35 pts | 13.72% | 13.87% | earned |
| G130 | A | 12 | 12 | 0.2 pts | 11.64% | 11.74% | earned |
| G132 | A | 12 | 12 | 0.4 pts | 9.58% | 9.7% | earned |
| G133 | A | 12 | 12 | 0.25 pts | 13.64% | 13.69% | earned |
| G134 | A | 12 | 12 | 0.21 pts | 17.06% | 17.11% | earned |
| G135 | A | 12 | 12 | 0.36 pts | 13.27% | 13.14% | earned |
| G136 | C | 12 | 12 | 0.36 pts | 8.98% | 8.97% | earned |
| G137 | C | 12 | 12 | 0.12 pts | 13.31% | 13.34% | earned |
| G138 | C | 12 | 12 | 0.17 pts | 14.66% | 14.69% | earned |
| G139 | C | 12 | 11 | 0.51 pts | 15.96% | 16.05% | earned |
| G140 | C | 12 | 11 | 0.52 pts | 16.23% | 16.3% | earned |
| G141 | C | 12 | 12 | 0.41 pts | 16.81% | 16.95% | earned |
| G142 | C | 12 | 10 | 0.55 pts | 14.21% | 14.35% | earned |
| G143 | C | 12 | 12 | 0.24 pts | 9.22% | 9.25% | earned |
| G144 | C | 12 | 12 | 0.09 pts | 17.36% | 17.29% | earned |
| G145 | C | 12 | 11 | 0.59 pts | 21.82% | 22.01% | earned |
| G146 | C | 12 | 12 | 0.1 pts | 14.84% | 14.83% | earned |
| G147 | C | 12 | 12 | 0.02 pts | 9.35% | 9.06% | earned |
| G148 | D | 12 | 10 | 0.71 pts | 17.6% | 17.74% | earned |
| G149 | D | 12 | 12 | 0.13 pts | 11.27% | 11.26% | earned |
| G150 | D | 12 | 12 | 0.16 pts | 10.72% | 10.74% | earned |
| G151 | D | 12 | 12 | 0.49 pts | 10.81% | 10.6% | earned |
| G152 | D | 12 | 12 | 0.2 pts | 11.22% | 11.16% | earned |
| G153 | D | 12 | 12 | 0.02 pts | 9.54% | 9.5% | earned |
| G154 | D | 12 | 12 | 0.12 pts | 10.08% | 10% | earned |
| G155 | D | 12 | 12 | 0.2 pts | 9.6% | 9.56% | earned |
| G156 | E | 12 | 11 | 0.53 pts | 10.71% | 10.89% | earned |
| G158 | E | 12 | 12 | 0.44 pts | 12.29% | 12.42% | earned |
| G160 | E | 12 | 12 | 0.47 pts | 10.35% | 10.54% | earned |
| G103 | A | 12 | 7 | 1.64 pts | 13.27% | 13.68% | won only 7/12 windows |
| G110 | A | 12 | 0 | 2.04 pts | 20.77% | 21.95% | won only 0/12 windows |
| G125 | A | 12 | 9 | 1.07 pts | 18.14% | 18.44% | won only 9/12 windows |
| G131 | A | 12 | 9 | 0.93 pts | 16.15% | 16.44% | won only 9/12 windows |
| G157 | E | 12 | 9 | 0.65 pts | 11.09% | 11.06% | won only 9/12 windows |
| G159 | E | 12 | 8 | 0.88 pts | 11% | 11.23% | won only 8/12 windows |
Event-effect library (median effect on the day, logged events)
| Event | Product | Events | Volume, day of | Volume, day after | Handle time, day of |
|---|
| weather | A | 18 | +18.92% | +3.13% | +10.54% |
| weather | C | 18 | +20.00% | +3.59% | +9.92% |
| weather | D | 18 | +12.10% | +0.47% | +7.05% |
| weather | E | 18 | +16.47% | +0.67% | +9.83% |
| outage | A | 1 | -41.67% | +39.24% | +34.91% |
| outage | E | 1 | -52.08% | +25.73% | +32.31% |
What the learning phase changed in the cycle
- Learning rate. The everyday rate dropped to 0.01. At 0.04, corrections chased noise and made the estate worse than the record.
- Shift confirmation. A detector alarm must hold for 7 days (mean error ≥ 0.8σ on the same side) before it counts. Across about 130 series, bare alarms were mostly noise.
- Shrinkage. Migration ratios are shrunk toward plan in proportion to their uncertainty, so a noisy channel does not carry its overshoot into the forecast.
- Earned learning. Learned corrections apply only to gates that earned them (walk-forward, lookback 8 windows, margin 0.25 pts) or had a confirmed shift in the last 70 days.
Synthetic data. In production the forward check runs every day as actuals arrive, and the earning table updates with it.