Interval predictions from conformal prediction are usually read as "90% coverage", but on nested battery data that number promises the fleet average, not the individual cell. We measure the gap on 124 public LFP cells under a preregistered three-tier reporting protocol (fleet, batch, cell): pooled coverage looks healthy (0.875-0.906 across splits) while the worst cell is covered 0.1-2% of the time, with exact binomial tests against the nominal level. We then repair part of the gap with a hierarchically weighted conformal engine whose weights consume only side information frozen at prediction time - signature features, batch labels, cycle indices - and never touch residuals. The preregistered repair criterion passes: 10th-percentile cell coverage moves 0.750 to 0.830 (53% of the gap), the worst cell 0.642 to 0.694, with zero deferrals and a flat sensitivity spectrum. What the engine does not repair is localized rather than left open: under protocol drift the frozen weights harm monotonically on the SOH scale (the absolute scale is reported in full), a structural dead end we report instead of tuning away; and repair capacity is bounded by calibration-pool battery diversity, measured on both datasets. On a cross-chemistry external set the calibrated intervals collapse (0.286), while the within-domain skeleton replicates (pooled 0.933, worst cell 0.686). The deliverable is a reporting protocol - three tiers of promise anchored to measured regimes, worst-cell statistics, deferral share - that makes per-cell interval claims auditable before deployment.
Figures & tables
Figure 1: The object of the guarantee and the three-tier reading (concept figure; numbers from panel A, SOH scale, Table 1). (a) Calibration cycles nest inside cells; cells nest inside protocol batches. The conformal guarantee attaches to a draw from the pooled cycle set; maintenance decisions attach to one cell. (b) The same experiment read at three tiers: fleet pooled coverage 0.906, minimum batch coverage 0.787, worst single cell 0.642, against nominal 0.90; 28% of cells fall below 0.85. The headline tier passes while the decision-relevant tier fails.
scale
panel
pooled
min batch
worst cell
pbonf
frac <0.85
perm. p
SOH
A (random, 5-seed median)
0.906
0.787
0.642
4.2e-55
28%
—
SOH
C (official split)
0.875
0.790
0.014 (b2c1)
<10−300
35%
—
SOH
B (2017 → 2018 drift)
0.799
—
0.001 (b3c29)
—
—
<0.001
abs
A
0.897
0.790
—
6.1e-58
28%
—
abs
C
0.888
0.778
0.020 (b2c1)
<10−300
27%
—
abs
B
0.810
—
0.001 (b3c29)
—
—
<0.001
Table 1: Baseline split conformal, six gate cells (E1). Cell identifiers read batch, then cell within batch.
Figure 2: Per-cell coverage under baseline split conformal (SOH scale, nominal 90%). (a) Random split (panel A): five splits (thin) and their pointwise median (bold); the fleet-level number (0.906) sits on nominal while 28% of cells fall below 0.85 and the median worst cell is 0.642. (b) Official split (panel C): 83 test cells ranked by coverage; the worst cell (b2c1) is covered 1.4% of the time. A healthy bulk with a disastrous tail is the object the reporting protocol of §7 is built to expose.
arm
pooled
q10
min batch
worst
frac <0.85
deferral
baseline
0.906
0.750
0.787
0.642
28%
0%
engine
0.909
0.830
0.884
0.694
16%
0%
Table 2: Baseline vs. engine, panel A (SOH). Registered line: q10 ≥ 0.825.
Figure 3: Engine sensitivity and the drift dead end (SOH scale). (a, b) q10 and worst-cell coverage vs. the battery-level ESS target (log axis); gray lines mark the unweighted baseline. Panel A is flat (q10 0.804–0.830) with no knife-edge at the registered head value ESS_batt = 5; panel C peaks at ESS_batt = 1 with worst-cell repair 0.014 → 0.311. (c) Under protocol drift (panel B) the engine degrades monotonically toward the baseline from below — the harm is the mechanism, not the setting. Target 20 saturates at achieved ESS ≈ 18 (cap flag, §8.3).
arm
pooled
q10
min batch
worst
width q^
fallback
baseline
0.906
0.750
0.787
0.642
0.0019
—
engine (ours)
0.909
0.830
0.884
0.694
0.0021
0%
Mondrian (within-batch)
0.899
0.825
0.869
0.631
0.0022
0% (A)
GP posterior
0.965
0.913
0.932
0.786
0.0011
none possible
BHM-lite
0.953
0.863
0.910
0.831
0.0039 (2 × )
0% (A)
Table 3: Five arms, panel A (SOH), five-seed medians.
Figure 4: Feature families divide the labor across two axes (baseline arm, SOH; interval half-width on log scale). (a) Panel A: removing the trajectory family leaves q10 apparently better (0.870) at 29 × width — coverage surviving as an artifact of width. (b) Panel C: removing the environment/position family collapses q10 (0.726 → 0.543) at unchanged width — the cycle index carries batch signature, echoing §8.1. The electrochemical family is mild on both axes: its point sits essentially on the full-feature anchor.
arm
pooled
worst
reading
cross-domain plain (Severson pool → Oxford)
0.286
0.137 (oxc2)
pbonf 4.5e-36; 100% of cells below 0.85
cross-domain engine
0.376
0.216
+0.09; not saved
within-domain plain (LOO × 8)
0.933
0.686 (oxc2)
oxc2 binomial p 2.3e-05; Bonferroni 1.9e-04
within-domain engine
0.933
0.686
identical to plain; ESS_batt ceiling 3.73 (cap flag)
Table 4: Oxford, four arms (SOH, IR-free feature subset, n = 8 cells; full distributions, no percentiles).
School of Software Engineering, Sun Yat-sen University, Zhuhai 519082, China · Tianneng Battery Group Co., Ltd, Zhejiang, China · School of Communication Engineering, Hangzhou Dianzi University, China +1