PBL Slope Monitoring
Per-session slope of Δ₄₇,raw vs δ⁴⁷ and Δ₄₈,raw vs δ⁴⁸ for ETH-1 and ETH-2 (both production batches) combined. A slope near zero indicates no residual PBL effect. The Δ₄₇ slopes confirm negligible PBL effects on m/z 47. The Δ₄₈ slopes show a small but consistent non-zero offset, for which a constant correction (median slope, blue dotted line) is applied to raw δ⁴⁸ before Δ₄₈ standardization. Δ₄₈ slopes shown only post–31 Oct 2024 (pre-amp bin pump + collector upgrade). Non-zero slopes indicate compositional non-linearity from secondary electrons (Bernecker et al. 2026; Bernecker et al. 2023).
Δ₄₇,raw vs δ⁴⁷ Slope per Session (ETH-1 + ETH-2)
Δ₄₈,raw vs δ⁴⁸ Slope per Session (ETH-1 + ETH-2)
Transducer Pressure–δ⁴⁷ Bias Assessment
Three complementary analyses test whether transducer pressure biases raw δ⁴⁷, using re-read raw data files across the full pressure range (including below the current 15 mbar threshold). B1 (Within-Session): each replicate's δ⁴⁷ deviation from its session×standard mean — WG cancels exactly, so all standards are valid regardless of temporal coverage or bottle changes. B2 (Cross-Session Per-Standard): pools all analyses of each standard; shown for reference but confounded by WG drift for standards with partial temporal coverage. B3 (Fixed-Effects): δ⁴⁷ ~ transducer pressure + session + standard — the most powerful pooled test, restricted to ETH-1, ETH-3, ETH-4 (the only standards with single composition spanning the full instrument record). ETH-2 and IAEA-C2 excluded (old/new bottle); CIT, GU-1, MERCK, IAEA-C1 excluded (partial temporal coverage creates spurious pressure–WG correlations). Generated by audit_nu_versions.py.
B3: Fixed-Effects δ⁴⁷ vs Transducer Pressure (session + standard controlled)
B2: Per-Standard Raw δ⁴⁷ vs Transducer Pressure
B1: Within-Session δ⁴⁷ Deviation vs Transducer Pressure
Threshold Zoom: Within-Session δ⁴⁷ Deviation (Transducer Pressure 8–16 mbar)
Cumulative Sum & EWMA Process Control
CUSUM (Cumulative Sum) and EWMA (Exponentially Weighted Moving Average) charts detect small persistent shifts in anchor standard residuals before conventional threshold-based alerts trigger. Applied to ETH-1/2 Δ₄₇ residuals from nominal. CUSUM accumulates sequential deviations — a monotonic upward or downward trend signals a sustained bias. EWMA smooths with a decay factor (λ = 0.2) and flags breaches of ±3σ control limits. Both are standard in analytical chemistry QC (IUPAC guidelines).
CUSUM — ETH-1/2 Δ₄₇ Residuals
EWMA — ETH-1/2 Δ₄₇ Residuals (λ = 0.2)
Standards Outside the Reference Frame
IAEA-C1, GU-1, and MERCK are standardized through D47crunch using the ETH-1–4 / IAEA-C2 / CIT anchor set but are not used to define the reference frame. Their Δ₄₇ values serve as independent accuracy checks: agreement with published nominals confirms the reference frame is well-calibrated. MERCK has no published I-CDES Δ₄₇ nominal, so its IQR-screened long-term mean is computed from lab data and used as a reference for reproducibility tracking (dotted line). IAEA-C1 nominal: 0.3018 ‰; GU-1 nominal: 0.2254 ‰ (Fiebig et al. 2024).
Δ₄₇ Time Series — External Standards
Δ₄₇ Residuals — External Standards
IAEA-C1 & GU-1: residuals from published nominals. MERCK: residuals from lab mean.
δ¹³C — External Standards
δ¹⁸O — External Standards
ETH-3 Reference Frame Evaluation
ETH-3 is used in the I-CDES reference frame but its long-term stability is under scrutiny. These plots track whether ETH-3 is drifting relative to nominal and how it behaves in Δ₄₇–Δ₄₈ space. Data restricted to post–31 Oct 2024 (pre-amp bin pump + collector upgrade) for dual clumped diagnostics.
Δ₄₇ vs Δ₄₈ Cross-Plot (ETH-3 highlighted)
ETH-3 Δ₄₇ vs Δ₄₈ — Zoom + NO₂ Vector
ETH-3 Δ₄₇ Drift Over Time
ETH-3 CUSUM — Δ₄₇ Residuals from Nominal
ETH-3 Δ₄₇ Session-Mean Residual from Nominal (0.6132 ‰)
Methodological Choice Comparison
Each tab isolates a single methodological choice — pre-crunch filtering, anchor set, WG constraint, or post-crunch screening threshold — and compares 2–3 variants. IAEA-C1 and GU-1 serve as external litmus tests (never used as anchors). Only the last 2 years of data are included.