Version 2.0, 16 September 2026. Retain the 106 audited pairs from DS162 (2022-05-24), DS421 Rev A (2020-07) and DS422 Rev D (2023-08). Add Lumileds DS562 revision 2025-04-08, Table 1 on PDF pages 4–6. Parsed 117 actual part-number rows, including 70-CRI entries; nomenclature placeholders are not products. All actual table part numbers are accounted for in extraction-audit.csv. Forty-five exact 80/90-CRI pairs are present in this newer catalogue. Unpaired and other-CRI rows remain in specifications.csv; no imputation.
For Lumileds, substitute only the CRI characters in the part number (80 → 90), then independently assert equal colour temperature, test current and junction temperature. For Bridgelux, substitute the documented CRI character and assert the same controls. Gen 5 part suffix x denotes the same range of chromaticity-bin options for both members; it is not a new measured option. The colour-tolerance wildcard does not multiply pair counts.
Reduction = 100 × (1 − efficacy90/efficacy80). Negative values indicate a higher listed efficacy at 90 CRI. Equal-lumen power difference = 100 × (efficacy80/efficacy90 − 1), a calculation under the rated efficacies, not a finished-lamp measurement. Use printed typical efficacy values rather than silently replacing them with flux divided by rounded voltage/power. Catalogue ranges are observed extrema, not confidence intervals. Stated flux tolerances and rounding mean very small differences should not be interpreted as measured advantages.
Do not pool manufacturer absolute efficacies: temperature and test conditions differ. Do not infer the causal effect of a generation upgrade from different catalogues' overall medians: product and colour-temperature mixes differ. All comparisons retain parts, document revisions, source pages, test conditions and URLs. The source matching is reproducible with analyze_led.py; older input tables are preserved in the package. Primary technical notes explain the mechanism, while the catalogue join is the study's original contribution.
Version 3.0 leads with the usual phosphor/efficacy trade-off, explaining it from manufacturer technical publications and quantifying it with the unchanged 151 pairs. All catalogues, the newer smaller penalty and the within-tolerance counterexample remain visible. No rows were removed to force agreement.