Methods + caveats

Tier-α prediction stack

• Input: GC×GC composition as a (carbon × class) HC distribution (mass %). Tier-α auto-detects raw APR / summarized APR / HC-distribution xlsx.
• Random isomer per (C, class) bin sampled from the bundled NIST DB (704 molecules; densities, viscosities, BP, DCN, surface tension, flash, freeze, LHV, MW, TSI; per-molecule T-dependent ρ, η).
• Mixture properties via established blending rules (Boehm 2024): density volume-fraction; DCN/LHV mass-fraction; flash via Wickey FPI; surface tension Macleod-Sugden ST^0.25; viscosity Walther/D341 2-point.
• Smoke point from mole-fraction-weighted TSI via Olson 1985 form Sp = 5.27 · MW / (TSI + 1.5). Constants are first-cut and need refit against measured Sp; current cohort bias +13.6 mm.
• Cp(T), k(T), ε from per-class correlations (RSP, Latini, Maxwell ε≈n²) in ffp_estimators.py; class-deterministic, so MC CIs ≈ 0 by construction.
• Electrical conductivity reported as pure-HC baseline only (~3 pS/m). Real fuel σ is dominated by Stadis-450 SDA additives not present in composition input. Cannot be predicted from composition alone.
• AIT from class + carbon correlations; mole-fraction weighted. Approximate; calibration against measured AIT not yet done.

Isomer prior

When VUV+Bell-method identifies a peak with score ≥ 5/7 statistics, the matched isomer (and any constrained alternates) replace the uniform random draw in the (C, class) bin. Bins without a confident ID fall back to uniform sampling.

WJFS-11 sensitivity demo (synthetic Q25 prior): density bias dropped from +9.89 to +4.55 kg/m³ (54 % reduction) by constraining iso-paraffin sampling to the lower-quartile densities — the kind of constraint a real Bell ID would impose.

Vetting

All parity plots are reproducible from a clean checkout of ~/heyne/vuv_results/; the WJFS-2024 cohort holdout is the canonical validation. Smoke-point and electrical-conductivity caveats above are not bugs — they're known structural limitations.