FastLED · zackees · Sep 18, 2026 · Sep 18, 2026 · Sep 18, 2026
diff --git a/.github/workflows/tests.yml b/.github/workflows/tests.yml
@@ -0,0 +1,46 @@
+name: tests
+
+# Runs the Python test suite and the checked-in artifact checks on every push
+# to main and every pull request. The colour tools compare against the
+# CIE 1931 2-degree CMF table, which is CC BY-SA 4.0 and not redistributed
+# here: it is fetched at run time and rejected unless its SHA-256 matches the
+# pin the tools themselves enforce.
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+
+permissions:
+  contents: read
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: astral-sh/setup-uv@v6
+
+      - name: Fetch the pinned CIE 1931 2-degree CMF table
+        run: |
+          set -euo pipefail
+          curl -sSfL --retry 3 -o "$RUNNER_TEMP/CIE_xyz_1931_2deg.csv" \
+            https://files.cie.co.at/Publications-datasets/CIE_xyz_1931_2deg.csv
+          echo "fa663e3535a7e0763a745993a1f0a192eb0275ac46ad2d1befd7626841e713c1  $RUNNER_TEMP/CIE_xyz_1931_2deg.csv" \
+            | sha256sum -c -
+
+      - name: Unit tests
+        env:
+          DATASHEETS_CIE_CSV: ${{ runner.temp }}/CIE_xyz_1931_2deg.csv
+        run: uv run --no-project --with jsonschema python -m unittest discover -s tests -v
+
+      - name: Checked-in artifacts are current
+        env:
+          CIE: ${{ runner.temp }}/CIE_xyz_1931_2deg.csv
+        run: |
+          set -euo pipefail
+          uv run --no-project --with jsonschema python tools/check_profile_artifacts.py
+          uv run --no-project python tools/generate_intensity_sensitivity.py --check
+          uv run --no-project python tools/bound_legacy_correction_accuracy.py "$CIE" --check
diff --git a/measured-profiles/LEGACY-ACCURACY-BOUND.md b/measured-profiles/LEGACY-ACCURACY-BOUND.md
@@ -0,0 +1,105 @@
+# Legacy-correction accuracy, bounded from datasheet data
+
+Generated by `tools/bound_legacy_correction_accuracy.py` for FastLED/FastLED#4035.
+**This is a bound, not a measurement.** The datasheets publish dominant-wavelength
+and luminous-intensity ranges, not chromaticity, so the ranges are sampled on a grid
+(each channel's dominant wavelength min/mid/max × FWHM low/mid/high × intensity
+min/mid/max) and every grid profile is scored, under two stated assumptions:
+
+- each emitter's spectrum is a Gaussian whose dominant wavelength (about equal-energy
+  white E) matches the datasheet, with FWHM red 15–25 nm,
+  green 25–40 nm, blue 18–30 nm (assumed, typical of
+  AlGaInP / InGaN emitters; not from the datasheets);
+- luminous intensity is proportional to Y at a common viewing geometry.
+
+Evaluation follows A1: an ordinary buffer is linear sRGB (B6/B8); the target is the
+sRGB colour Bradford-adapted from D65 to the device's full-drive white (Y=1); the
+legacy path drives each emitter at `code/255 × correction`. ΔE2000 over a black-safe
+neutral ramp (32–255), RGB primaries and CMY secondaries. Budget for a
+datasheet-derived profile: median ≤ 3, p95 ≤ 6.
+
+`nominal` is the grid's centre point: mid wavelength, mid FWHM, mid intensity. The
+ranges are the minimum and maximum of each statistic over the sampled grid. They are
+not a proof over the continuous ranges: a grid can miss a worse or better interior
+point, so each range is an inner bound on the true one.
+
+| part | model | nominal median | nominal p95 | median range | median spread | p95 range | verdict |
+| --- | --- | ---: | ---: | --- | ---: | --- | --- |
+| SK6812 | UncorrectedColor | 6.91 | 10.90 | 3.60–10.50 | 6.91 | 7.69–15.41 | outside budget at every sample |
+| SK6812 | TypicalLEDStrip | 15.01 | 19.31 | 12.74–16.29 | 3.55 | 17.32–21.08 | outside budget at every sample |
+| SK6812 | Typical8mmPixel | 12.39 | 19.59 | 10.57–14.48 | 3.91 | 17.92–21.06 | outside budget at every sample |
+| WS2812B | UncorrectedColor | 1.22 | 14.45 | 1.07–3.69 | 2.62 | 12.64–16.52 | outside budget at every sample |
+| WS2812B | TypicalLEDStrip | 15.11 | 19.71 | 13.48–16.41 | 2.93 | 18.94–20.42 | outside budget at every sample |
+| WS2812B | Typical8mmPixel | 12.98 | 20.11 | 12.51–13.89 | 1.38 | 19.18–21.12 | outside budget at every sample |
+| WS2813A | UncorrectedColor | 4.40 | 9.33 | 1.48–5.52 | 4.04 | 8.38–10.40 | outside budget at every sample |
+| WS2813A | TypicalLEDStrip | 14.94 | 21.46 | 14.55–15.27 | 0.72 | 21.06–21.79 | outside budget at every sample |
+| WS2813A | Typical8mmPixel | 10.64 | 18.15 | 10.20–11.84 | 1.64 | 17.21–19.16 | outside budget at every sample |
+| WS2813B | UncorrectedColor | 3.99 | 9.34 | 1.32–5.31 | 3.99 | 8.36–10.44 | outside budget at every sample |
+| WS2813B | TypicalLEDStrip | 15.12 | 21.64 | 14.71–15.46 | 0.75 | 21.24–21.97 | outside budget at every sample |
+| WS2813B | Typical8mmPixel | 10.52 | 17.89 | 9.99–11.52 | 1.53 | 17.00–18.85 | outside budget at every sample |
+| WS2813C | UncorrectedColor | 2.24 | 6.68 | 1.74–3.43 | 1.68 | 6.00–7.99 | samples straddle the budget |
+| WS2813C | TypicalLEDStrip | 15.16 | 21.68 | 14.03–15.51 | 1.47 | 21.24–22.03 | outside budget at every sample |
+| WS2813C | Typical8mmPixel | 11.72 | 18.47 | 10.75–12.81 | 2.06 | 17.20–19.77 | outside budget at every sample |
+| WS2813D | UncorrectedColor | 2.72 | 7.05 | 2.04–4.03 | 1.99 | 6.22–8.42 | outside budget at every sample |
+| WS2813D | TypicalLEDStrip | 14.96 | 21.48 | 14.14–15.29 | 1.15 | 21.05–21.82 | outside budget at every sample |
+| WS2813D | Typical8mmPixel | 11.73 | 18.71 | 10.75–12.81 | 2.06 | 17.43–20.00 | outside budget at every sample |
+| WS2815 | UncorrectedColor | 5.11 | 10.41 | 1.76–7.85 | 6.09 | 8.42–16.26 | outside budget at every sample |
+| WS2815 | TypicalLEDStrip | 15.35 | 21.87 | 12.99–18.33 | 5.34 | 21.24–25.95 | outside budget at every sample |
+| WS2815 | Typical8mmPixel | 10.36 | 17.55 | 9.92–14.38 | 4.46 | 17.00–20.00 | outside budget at every sample |
+| WS2816B-2121 | UncorrectedColor | 6.67 | 15.44 | 2.64–11.21 | 8.57 | 7.77–26.70 | outside budget at every sample |
+| WS2816B-2121 | TypicalLEDStrip | 16.92 | 21.13 | 14.24–20.08 | 5.84 | 19.09–27.92 | outside budget at every sample |
+| WS2816B-2121 | Typical8mmPixel | 14.51 | 18.63 | 9.77–17.19 | 7.43 | 15.62–25.26 | outside budget at every sample |
+| GS8208 LED5050 option | UncorrectedColor | 2.88 | 9.53 | 2.17–8.50 | 6.33 | 8.11–18.52 | outside budget at every sample |
+| GS8208 LED5050 option | TypicalLEDStrip | 14.58 | 21.09 | 13.55–18.03 | 4.48 | 20.19–24.89 | outside budget at every sample |
+| GS8208 LED5050 option | Typical8mmPixel | 10.84 | 18.55 | 10.55–16.58 | 6.03 | 17.49–22.27 | outside budget at every sample |
+
+## What this establishes
+
+- **Datasheet ranges cannot pin accuracy to the A1 budget.** In 14 of 27 part/model
+  cells the median ΔE2000 over the sampled grid alone spans more than the whole
+  3 median budget. A sampled spread can only understate the continuous one,
+  so this holds for the ranges themselves: no single datasheet-derived profile could
+  be defended at that budget, and a runtime profile needs xy or spectral data the
+  PDFs do not publish.
+- **Legacy models under A1:** 26 of 27 cells are outside budget at every
+  sampled profile, 0 within budget at every one, and the rest straddle it. These
+  verdicts are over the grid, not proven over the continuous ranges.
+  `TypicalLEDStrip` and `Typical8mmPixel` score worse than `UncorrectedColor` under A1
+  because A1 is relative colorimetry: the target is adapted to the device's own
+  full-drive white, so a correction that tints that white is penalized. The
+  supplementary absolute table below scores the goal those corrections were built for.
+
+## Supplementary: absolute colorimetry (not A1)
+
+Target and reference white are D65 with no adaptation; nominal profile only. Context
+for the legacy corrections' design intent, not an A1 acceptance number.
+
+| part | model | median | p95 |
+| --- | --- | ---: | ---: |
+| SK6812 | UncorrectedColor | 24.48 | 30.97 |
+| SK6812 | TypicalLEDStrip | 25.86 | 33.79 |
+| SK6812 | Typical8mmPixel | 19.48 | 25.86 |
+| WS2812B | UncorrectedColor | 22.75 | 28.84 |
+| WS2812B | TypicalLEDStrip | 26.63 | 32.48 |
+| WS2812B | Typical8mmPixel | 18.09 | 24.44 |
+| WS2813A | UncorrectedColor | 16.32 | 23.27 |
+| WS2813A | TypicalLEDStrip | 19.51 | 26.25 |
+| WS2813A | Typical8mmPixel | 12.21 | 16.21 |
+| WS2813B | UncorrectedColor | 15.51 | 22.05 |
+| WS2813B | TypicalLEDStrip | 18.10 | 24.93 |
+| WS2813B | Typical8mmPixel | 12.70 | 16.50 |
+| WS2813C | UncorrectedColor | 19.48 | 26.26 |
+| WS2813C | TypicalLEDStrip | 20.88 | 28.01 |
+| WS2813C | Typical8mmPixel | 16.44 | 22.84 |
+| WS2813D | UncorrectedColor | 19.80 | 26.78 |
+| WS2813D | TypicalLEDStrip | 21.71 | 28.72 |
+| WS2813D | Typical8mmPixel | 15.80 | 22.18 |
+| WS2815 | UncorrectedColor | 15.31 | 21.70 |
+| WS2815 | TypicalLEDStrip | 18.17 | 22.83 |
+| WS2815 | Typical8mmPixel | 14.76 | 18.79 |
+| WS2816B-2121 | UncorrectedColor | 6.55 | 15.59 |
+| WS2816B-2121 | TypicalLEDStrip | 17.48 | 21.43 |
+| WS2816B-2121 | Typical8mmPixel | 13.95 | 17.85 |
+| GS8208 LED5050 option | UncorrectedColor | 17.29 | 24.27 |
+| GS8208 LED5050 option | TypicalLEDStrip | 21.30 | 27.66 |
+| GS8208 LED5050 option | Typical8mmPixel | 10.43 | 14.15 |
diff --git a/measured-profiles/P1-CHARACTERIZATION.md b/measured-profiles/P1-CHARACTERIZATION.md
@@ -36,16 +36,48 @@ value, but no individual-diode xy or spectral distribution. A neutral,
 primary, secondary, luminance, or ΔE2000 comparison requires each emitter's xy
 and relative radiometric/photometric scale; a luminous-intensity RGB ratio
 alone cannot establish ΔE2000.  Consequently there are **zero admissible
-derived profiles**, no defensible clusters, and no numerical score for
-`TypicalLEDStrip`, `Typical8mmPixel`, or `UncorrectedColor`.  Reporting zero
-error or a five-part illustrative sample would falsely imply full-catalog
-coverage. This is a quantified artifact-completeness result: 0/10 integrated
-parts with admitted extracted spectral/xy data; 0/10 with a measured uncertainty;
-0/10 currently usable for the A1 normalized ΔE calculation. P1 characterization
-remains incomplete pending source extraction and the required numerical report.
-
-When payload access is restored, calculate the three legacy models across
-*every* admissible integrated record: use the profile full-drive white as
+derived profiles** at runtime: 0/10 integrated parts with admitted extracted
+spectral/xy data, 0/10 with a measured uncertainty.
+
+## Legacy-correction accuracy, as a bound
+
+What the datasheets *do* publish -- a dominant-wavelength range and a
+luminous-intensity range per emitter -- is sampled into a grid of profiles
+(each range's ends and midpoint) by `tools/bound_legacy_correction_accuracy.py`,
+and every legacy model is scored across that grid under A1 (relative colorimetry,
+linear-sRGB source, Bradford to the full-drive white, ΔE2000 over a neutral
+ramp, primaries and secondaries). Two assumptions are stated, not taken from
+the PDFs: a Gaussian emitter spectrum with a per-technology FWHM interval, and
+luminous intensity proportional to Y. The generated table is
+[`LEGACY-ACCURACY-BOUND.md`](LEGACY-ACCURACY-BOUND.md)
+(`legacy-correction-accuracy-bound-v1.json`), covering the 9 integrated rows
+with wavelength and intensity data.
+
+Results:
+
+- `TypicalLEDStrip`, `Typical8mmPixel` and `UncorrectedColor` are quantified
+  as ΔE2000 ranges per part; under A1 legacy models miss the datasheet budget
+  (median ≤ 3.0 / p95 ≤ 6.0) at every sampled profile in nearly every
+  part/model cell (see the generated table for the count). The grid is an
+  inner bound on the continuous ranges, not a proof over them. The corrections score worse than
+  `UncorrectedColor` under A1 because they tint the device white, which
+  relative colorimetry penalizes; a supplementary absolute-colorimetry table
+  scores their D65 design intent.
+- The median over the sampled grid alone spans more than the entire 3.0
+  budget in about half the cells. A sampled spread can only understate the
+  continuous one, so datasheet ranges **cannot pin a derived profile to the A1
+  budget**. That is the numerical reason for zero admissible derived profiles,
+  and why accuracy claims wait on P10 measurement.
+- Clustering: per-part ranges overlap at this resolution, so no
+  datasheet-level clustering is defensible; the WS2813 A–D variants differ in
+  intensity only and are not a bin population.
+
+These are datasheet bounds, not measurements, and are kept separate from any
+P10 instrument result.
+
+The bound implements the prescribed method, and P10 measured records reuse
+it unchanged: score the three legacy models across *every* admissible record,
+use the profile full-drive white as
 relative/adaptive white (Y=1, dark surround), test black-safe neutral ramp,
 RGB primaries and CMY secondaries, record neutral chromaticity/Y error and
 ΔE2000, then report count, median, p95 and excluded-part reasons.  Keep