Conditional Calibration Contracts under Coupled Load: A Branch-Aware Audit of Switched Inference
Organizations: Independent research
Abstract
Bursty serving changes both queue states and the population selected for each model. We study conditional calibration contracts as a specification to be validated, rather than a certified property of an existing router. CA-MHD combines load-coupled distillation, closed-loop replay and a cost-informed Schmitt router in a discrete-event simulator with a hardware service table. Across 40 cells (ten training seeds and four content-load coupling levels), we distinguish the original teacher, frozen wide heads on an adapted trunk, and the narrow student. A common teacher-trajectory audit has median absolute student-teacher deficit 0.06 percentage points, but does not certify the student-selected deployment support. Historical results use a kernel correctness gap, not the published SmoothECE; we separately prove selected-branch composition for a residual functional at a common fixed scale. The deployed wide branch has lower ordinal accuracy than the original teacher in all 40 cells. Same-checkpoint student-only serving has lower historical ordinal gap in 27/40 cells; routing also exposes violation tails up to 17.5 percent. Student p99 latency remains above the 1.5 ms target. Retrospective temporal recalibration improves the ordinal gap in 39/40 cells, but causal delayed-label availability is unverified and no incremental per-mode benefit is detected after multiplicity adjustment. An in-sample CUSUM detects 30/40 checkpoint changes at five historical false alarms, with no prospective false-alarm guarantee. These findings support a branch-aware audit and a scoped negative result, not demonstrated routing superiority. Independent budgets, selected-support validation and net utility under a satisfied SLA remain open.
Figures & tables
| Method | viol% | p99 ms | switches | cov | |
|---|---|---|---|---|---|
| B1: teacher (saturated) | 70.6 (0.2) | 168.3 | 0 | 0.00 | |
| B1’: student-only (LCD) | 1.8 (0.6) | 1.6 | 0 | 1.00 | |
| Schmitt·LCD+DAgger (ours) | 4.3 (5.2) | 13.0 | 4 | 0.67 | |
| Schmitt·LCD (ours) | 4.3 (5.2) | 13.0 | 4 | 0.67 | |
| B4: fixed-threshold (best-q) | 1.9 (0.6) | 1.7 | 1,279 | 0.69 | |
| B5: random matched | 64.4 (4.0) | 24.3 | 16,676 | 0.68 |
| head | LCD | B3a | B3b | LCD/B3a | LCD/B3b | |
|---|---|---|---|---|---|---|
| gate | 1.09 | 1.08 | 1.21 | -0.8% | +10.3% | |
| choice | 1.13 | 1.13 | 1.31 | -0.2% | +13.8% | |
| corn | 2.12 | 2.27 | 3.56 | +6.5% | +40.4% | |
| gate | 1.54 | 1.52 | 1.72 | -1.3% | +10.6% | |
| choice | 1.66 | 1.66 | 1.97 | +0.1% | +15.6% | |
| corn | 1.22 | 1.24 | 4.73 | +2.0% | +74.2% |
| state.head | -acc | S-acc | -H | S-H | (pts) | ||
|---|---|---|---|---|---|---|---|
| calm/gate | 97.1 | 97.1 | 1.02 | 1.20 | 0.18 | 100758 | |
| calm/choice | 97.1 | 97.1 | 1.07 | 1.23 | 0.16 | 100758 | |
| calm/corn | 67.0 | 66.4 | 1.73 | 1.58 | -0.15 | 100758 | |
| warm/gate | 97.1 | 97.1 | 1.04 | 1.04 | -0.00 | 77072 | |
| warm/choice | 97.1 | 97.1 | 1.07 | 1.07 | -0.00 | 77072 | |
| warm/corn | 67.0 | 66.6 | 1.52 | 1.67 | 0.15 | 77072 |
| gate | choice | corn (attacks) | |||||||
|---|---|---|---|---|---|---|---|---|---|
| W | cf-S | cf-L | W | cf-S | cf-L | W | cf-S | cf-L | |
| 94.5 | 97.1 | 97.1 | 94.7 | 97.1 | 97.1 | 64.8 | 65.9 | 66.9 | |
| 95.4 | 97.1 | 97.1 | 95.6 | 97.1 | 97.1 | 59.8 | 66.8 | 67.1 | |
| 95.2 | 97.1 | 97.1 | 95.8 | 97.1 | 97.1 | 57.0 | 67.0 | 67.3 | |
| 95.2 | 97.1 | 97.1 | 95.6 | 97.1 | 97.1 | 53.4 | 66.7 | 66.8 | |
| served | mixt.-Platt | mode-Platt | mode-iso | same-S+recal | |
|---|---|---|---|---|---|
| 2.95 (0.71–4.45) | 1.92 (0.95–2.79) | 1.64 (0.70–2.94) | 2.28 (1.12–3.45) | 1.88 (1.17–2.42) | |
| 2.03 (1.35–3.16) | 1.04 (0.77–3.99) | 0.91 (0.73–1.14) | 1.18 (0.78–1.79) | 0.97 (0.65–1.13) | |
| 1.65 (0.94–2.98) | 1.00 (0.79–1.28) | 0.96 (0.73–1.11) | 1.25 (0.72–1.47) | 0.91 (0.68–1.09) | |
| 1.68 (0.86–3.91) | 0.79 (0.68–1.21) | 0.78 (0.63–1.01) | 0.86 (0.75–1.14) | 0.78 (0.66–1.01) |
| sw | |||||||||
|---|---|---|---|---|---|---|---|---|---|
| viol% | corn | viol% | corn | viol% | corn | viol% | corn | ||
| b1_samenet | 0 | 1.84 | 2.01 | 1.51 | 1.83 | 1.64 | 1.58 | 1.34 | 1.58 |
| asis | 4 | 1.94 | 3.19 | 1.66 | 2.14 | 1.63 | 1.71 | 1.15 | 1.66 |
| jit5 | 4 | 1.94 | 3.19 | 1.81 | 2.44 | 1.63 | 1.79 | 1.15 | 1.66 |
| jit15 | 4 | 1.95 | 3.14 | 2.10 | 2.56 | 1.87 | 1.79 | 1.42 | 1.86 |
| delay0.5 | 4 | 38.17 | 2.94 | 39.70 | 2.61 | 39.65 | 2.21 | 39.94 | 2.55 |
| Historical criterion / later sensitivity | Evidence | Status |
|---|---|---|
| T1 historical-gap temperature fit | gate/choice 1.3–1.5% vs goal 1%; corn varies by coupling | ADOPTED; absolute calibration goal not universally met |
| T2a LCD vs B3a ( 30% hot) | corn +6.5/+2.0/–0.9/–5.8% by (10 seeds); heads 0–2% | FAILED magnitude criterion; descriptive direction is mixed |
| T2a’ LCD vs B3b (blind) | corn +40–77%; heads 5–17% | DESCRIPTIVE comparison with damaged B3b, not general superiority |
| T2b TOST ( 0.3 pts) | small gate/choice differences; corn pts, seeds favorable (10 seeds) | PENDING test/interval reconciliation; differences alone are not TOST |
| T3 router ( 1/3 B4, viol parity) | 12-cell sweep, seeds 0–2: 4–6 vs 200–2044 switches; sealed gamma 0.5 outside band | DESCRIPTIVE test sweep; no 40-cell validation of selected replacement. Nominal B4 switch cost 0.057–0.211% of horizon |
| T4 predictive transient bound | corn drain medians 3.6–5.1 pts overall; 5.5–7.8 on the teacher-served subset (163 windows, 40 cells) | PENDING independent RHS validation; 200 event-block draws descriptive |
Adaptive Model Routing
Adaptive model routing selects models or specialized experts per input to balance prediction quality, computational cost, latency, and resource constraints.
See all 101 papers