How Percevra computes light-aware color

The pipeline

One engine, one clamp. The pipeline stays in unclamped linear light from the illuminant cast through chromatic adaptation, and only encodes to sRGB at the very end. That ordering is load-bearing rather than fussy: under a correct Illuminant A about 45% of colors leave the display range at the simulate stage, and clipping is irreversible, so encoding early would mean scoring data the display had already destroyed.

1. 01

   Simulatescene-referred

   Your palette is re-lit by the illuminant — reflected = surface × light — as a von Kries scaling in CAT16 cone space rather than a multiply in RGB primaries. That choice is deliberate: adaptation happens in cone space, and casting the light in a different basis leaves a residue the two stages can never cancel. Nothing is clamped here, so a strong lamp can push bright swatches past 1.0 and those values carry forward rather than clip. Daylight is a true no-op, so D65 never quietly restyles your colors.

2. 02

   Adaptstill unclamped

   The stimulus is chromatically adapted back toward D65 with CAT16 at a fixed degree of adaptation D = 0.6, modelling a viewer whose eyes have partly settled into the room. Because stage one casts in the same space, this is an exact partial inverse: at D = 1 it would recover your original color precisely, so what survives here is cleanly the 40% of the cast a partly-adapted viewer still sees.

3. 03

   Measure and displayone clamp, at the end

   The adapted color is measured in OKLab for the score, and separately encoded to sRGB for the diagnostic swatch shown on this page. Both read the same stage, so this page's swatches and numbers can never tell different stories, and the score is taken before the display encoding throws anything away. (Product preview surfaces — palette cards, the Light Profile, studios — render a separate spectral pipeline at a gentler adaptation degree; see the illuminants section below.)

Which illuminants we simulate

The illuminant names are CIE standards—we did not invent the physics, and the tables below let you check our numbers against the published ones. What differs between our two engines is how much of each standard they reproduce: the tint model matches the illuminant's white point, and the spectral model matches its full spectrum.

Scoring — tint model

The stability diagnostic uses a three-channel linear-RGB multiply. The multipliers are not chosen by eye: each one is derived from its illuminant's published CIE chromaticity at constant luminance, so the white the engine adapts to is the standard's white by construction. The two right-hand columns below are the check — the left is computed live from the running engine, the right is transcribed from the standards. They should agree to four decimal places.

Product previews no longer use this model. Palette cards, the Light Profile, mood and decision previews, and Extract render the spectral pipeline (reflectance reconstructed from each hex, integrated against the lamp's SPD, then CAT16) at a display adaptation of D = 0.85 — a settled viewer — while the diagnostic keeps its harsher D = 0.6 stress test. One preview condition also differs from its diagnostic counterpart by design: the “Blue-sky shade” stop previews a daylight-locus 10000K proxy (shade lit only by clear sky — an estimate, not a measured spectrum), because the diagnostic's CIE D75 is chromatically so close to D65 that a preview of it is visually inert. The diagnostic below still measures CIE D75.

One row does not agree, and it is worth naming. Candlelight at 1900K falls outside the sRGB gamut: reproducing it exactly needs a negative blue multiplier. Its multipliers are clamped to the gamut boundary, which is why its white lands a few ten-thousandths off. Candle is preview-only and never enters a score.

Only 3 of these eight (A, LED-B3, Shade) feed the separation diagnostic below; D65 is the fixed reference it is measured against. A pass/warn/fail badge was formerly judged on A, F11, LED-B3 — it was retired after blinded perceptual testing (see “Does the diagnostic match what people see?” below). Candle, Night, and D50 are visualization only and never affect a number.

Light-Aware Lab — spectral model

The Light-Aware Palette Lab runs the real thing: tabulated CIE relative spectral power distributions for D65, D50, A, F11, and LED-B3, sampled from 380–780 nm at 5 nm intervals (81 samples), integrated against the CIE 1931 2° standard observer color-matching functions. Continuous color temperatures are generated from the Planckian locus or the CIE daylight locus as appropriate. The reference data is generated from colour-science 0.4.7 and committed to the repository rather than fetched at runtime, so results are reproducible.

This path can express spectral effects the tint model cannot, because it integrates a reflectance curve against a real lamp spectrum. It is still working from a reflectance recovered from your hex value, not one measured off your material.

What the math runs in, and why

How the separation diagnostic is computed

Every color shifts under a warm bulb. A uniform cast is physics, not a defect, so measuring absolute shift would mostly measure the lamp. What this diagnostic measures instead is whether the palette's internal structure survives: if two colors were clearly different in daylight, are they still clearly different under this light? It is a well-defined physical quantity — and, as the validation section below reports, one that turned out not to track what people perceive as a palette breaking, which is why it no longer grades anything in the product.

for each pair (i, j) in the palette:
    d_base = ‖ OKLab(cᵢ) − OKLab(cⱼ) ‖ × 100        // separation at D65
    skip the pair if d_base < 10                     // never distinct to begin with
    d_lit  = ‖ OKLab(lit(cᵢ)) − OKLab(lit(cⱼ)) ‖ × 100

retention(illuminant) = min over surviving pairs of ( d_lit / d_base )
collapsed             = any d_lit < 4

stability % = min( retention ) over A, LED-B3, Shade  × 100

where lit() is read at two points in the pipeline:
    settled      = adapt(simulate(c), D = 0.6)    // partly adapted viewer
    first glance = simulate(c)                     // D = 0, unadapted worst case

Why there are two numbers

The same formula is read at two points in the pipeline, and the honest reason is worth stating plainly: the settled number cannot fail. Once the illuminant is cast in cone space, adaptation is an exact partial inverse of it, so a viewer whose eyes have adjusted recovers almost all of the original separation. That is not a loophole in the model — it is colour constancy, and it is the correct result. But it means the settled score is a high, narrow quantity: we searched adversarially for the worst palette it can produce and could not drive it below 78%.

The first-glance figure — the unadapted stimulus, the moment you walk into the room before your eyes do any work — spans roughly 43–100 and was historically the graded number. Those grades are now retired (see the validation section below); both figures remain documented here as diagnostics. The gap between the two is still an honest reading: a palette whose numbers sit far apart is one that only holds together once the room has settled.

It is worst-case twice over: the weakest surviving pair within a light, and then the weakest light of the three. A palette fails at its weakest relationship under its worst condition, not on average, and reporting an average across lights would let a palette that collapses under a warm bulb be rescued by holding up under two others. Retention is a ratio of distances, so it is dimensionless and comparable across palettes of different overall contrast.

A worked example

This palette is run through the live engine every time the page loads, so you can reproduce the arithmetic below rather than take it on faith.

The second score: MSS

The Material Stability Score answers a different question—absolute drift rather than preserved separation—and is used where a color has to hit a target rather than merely stay distinguishable. Per color it is 1 − max(ΔE across illuminants) / threshold, clamped to 0–1, where the threshold tightens by use case. A palette's MSS is 70% of the mean plus 30% of the worst swatch, so one fragile color cannot be averaged away.

• 14Print
• 15UI
• 16Architecture
• 17Textile
• 20Interior
• 18Default

Thresholds are in the same OKLab ΔE ×100 units. Print is tightest because a proof is judged against a target; interior is loosest because a wall is judged against its own room.

Does the diagnostic match what people see?

In September 2026 we tested it, blind. A reviewer judged palettes shown as-authored beside their simulations under real lamps — no scores, categories, or metric values visible — across two rounds (79 palettes, then a 30-palette confirmation with hidden repeats, near-miss controls, and anchors). The question each time: did the palette’s internal relationships materially change?

The result was decisive and negative: separation retention — settled or first-glance — predicts those judgments at chance level. Palettes with the lowest retention were mostly judged fine; palettes judged broken were scattered across the whole retention range. What people actually track is hue: the judged failures were overwhelmingly pale, low-chroma cool swatches flipping to warm (a pale cyan reading as peach under an incandescent lamp), a signal that per-palette retention averages are structurally blind to.

So we changed the product before changing this page’s math: the pass/warn/fail badge, the stability grades, and every public retention percentage were withdrawn, and surfaces now describe simulations (“previewed across lighting conditions”) rather than issue verdicts. A candidate replacement — a per-swatch pale-flip event defined in Cartesian opponent coordinates — is under pre-registered blinded testing now, and no verdict language returns until a metric passes it. The retention formulas above remain published because they are honest physics and useful diagnostics; they are simply not proof that a palette will look broken or fine to a person.

What the model does not capture

This is the most useful section on the page. A model is only trustworthy to the extent its edges are known, so these are stated as specifically as we can rather than hedged.

Known limits of the score itself

Separately from what the physics misses, the stability metric has properties worth knowing before you quote it.

• It does not predict human judgment. Blinded testing (see the validation section above) found both retention readouts perform at chance against what reviewers perceive as a palette materially breaking under a lamp. Quote either number as a measured property of the simulation, never as evidence a palette will look stable.
• Scores are not comparable across palette sizes. More colors means more pairs, and the score reports the weakest one, so a seven-color palette is structurally harder to keep stable than a three-color palette. Median first-glance scores across random palettes fall from roughly 79 at three colors to 67 at five and 61 at seven. Compare a palette against alternatives of the same size, not against a number you remember from elsewhere in the app.
• Low-contrast palettes score perfectly by construction. Pairs closer than ΔE 10 in daylight are excluded from the measurement, so a palette of five near-identical tints has nothing to lose and returns 100%. That is the metric declining to comment, not an endorsement.
• It is silent about absolute appearance. A palette can keep all of its internal separation while every color shifts a long way from what you specified. Retention will not catch that; the mean shift figure and MSS will.
• It scores three lights, not your light. Warm 2700K, cool 4000K, and shade are stand-ins for common conditions. A low-CRI lamp, a colored wall bouncing light, or a mixed-source room is not represented.
• The settled score is close to a formality. It has a floor near 78% even for palettes built adversarially to break it, so a high settled number is weak evidence on its own. Read it as confirmation that adaptation is doing its job, and take the disagreement between it and the first-glance figure as the real signal.

Standards referenced

• CIE 15:2018 — colorimetry, standard illuminants D65, D50, A, and the F and LED series
• CIE 1931 2° standard observer color-matching functions
• CIECAM16 / CAT16 chromatic adaptation transform (Li et al., 2017)
• IEC 61966-2-1 — sRGB color space and transfer function
• OKLab / OKLCh perceptual color space (Ottosson, 2020); CSS Color 4
• CIEDE2000 color-difference formula, used for print gamut and color naming
• WCAG 2.2 contrast ratios and APCA lightness contrast, used for accessibility readouts

Color accuracy·Guide: how colors change under light·Light-Aware Palette Lab·Terms of Use