cbrowser 20.2.2 → 21.0.0

package/dist/visual/attention-transport.js

@@ -19,7 +19,9 @@
  * @since v18.28.0
  * @see https://github.com/alexandriashai/cbrowser/issues/159
  */
-import { getPerceptualProfile } from './perceptual-transport.js';
+import { resolvePerceptualProfile } from './perceptual-transport.js';
+import { param } from './model-provenance.js';
+import { tableChecksum, ATTENTION_STATIC_MODIFIER_TYPES, ATTENTION_VALUE_PATTERN_VALUES } from './provenance/attention.js';
 /**
  * Locate where two attention maps differ, as regions rather than cells.
  *
@@ -52,6 +54,8 @@ import { getPerceptualProfile } from './perceptual-transport.js';
  * Pure: no image, no persona. `cellSize` is the cell's size in image pixels.
  */
 export function rankDivergentRegions(salA, salB, rows, cols, cellSize) {
+    const R2 = param("attention.rounding.hundredths");
+    const R3 = param("attention.rounding.thousandths");
     const n = Math.min(salA.length, salB.length, rows * cols);
     let totalA = 0;
     let totalB = 0;
@@ -136,11 +140,11 @@ export function rankDivergentRegions(salA, salB, rows, cols, cellSize) {
             },
             cellCount: count,
             direction: positive ? "personaA" : "personaB",
-            attentionA: Math.round(shareA * 1000) / 1000,
-            attentionB: Math.round(shareB * 1000) / 1000,
-            saliencyA: Math.round((sumA / count) * 100) / 100,
-            saliencyB: Math.round((sumB / count) * 100) / 100,
-            divergence: Math.round((sum / divergence) * 1000) / 1000,
+            attentionA: Math.round(shareA * R3) / R3,
+            attentionB: Math.round(shareB * R3) / R3,
+            saliencyA: Math.round((sumA / count) * R2) / R2,
+            saliencyB: Math.round((sumB / count) * R2) / R2,
+            divergence: Math.round((sum / divergence) * R3) / R3,
             _sum: sum,
             _peak: peak,
             _peakIdx: peakIdx,
@@ -154,25 +158,36 @@ export function rankDivergentRegions(salA, salB, rows, cols, cellSize) {
         regions: regions.map(({ _sum: _s, _peak: _p, _peakIdx: _i, ...r }) => r),
     };
 }
-// ── CIE-Lab Color Space Conversion ──
-function srgbToLab(r, g, b) {
+function labConstants() {
+    const P = (id) => param(`attention.${id}`);
+    return {
+        channelMax: P("srgb.channelMax"), linearKnee: P("srgb.linearKnee"), linearSlope: P("srgb.linearSlope"),
+        gammaOffset: P("srgb.gammaOffset"), gammaScale: P("srgb.gammaScale"), gamma: P("srgb.gamma"),
+        rX: P("xyz.rX"), gX: P("xyz.gX"), bX: P("xyz.bX"),
+        rY: P("xyz.rY"), gY: P("xyz.gY"), bY: P("xyz.bY"),
+        rZ: P("xyz.rZ"), gZ: P("xyz.gZ"), bZ: P("xyz.bZ"),
+        xn: P("white.x"), yn: P("white.y"), zn: P("white.z"),
+        epsilon: P("lab.epsilon"), kappaSlope: P("lab.kappaSlope"),
+        lScale: P("lab.lScale"), lOffset: P("lab.lOffset"), aScale: P("lab.aScale"), bScale: P("lab.bScale"),
+    };
+}
+function srgbToLab(r, g, b, K) {
     // sRGB to linear
     const toLinear = (c) => {
-        c /= 255;
-        return c <= 0.04045 ? c / 12.92 : Math.pow((c + 0.055) / 1.055, 2.4);
+        c /= K.channelMax;
+        return c <= K.linearKnee ? c / K.linearSlope : Math.pow((c + K.gammaOffset) / K.gammaScale, K.gamma);
     };
     const rl = toLinear(r), gl = toLinear(g), bl = toLinear(b);
     // Linear RGB to XYZ (D65)
-    const x = 0.4124564 * rl + 0.3575761 * gl + 0.1804375 * bl;
-    const y = 0.2126729 * rl + 0.7151522 * gl + 0.0721750 * bl;
-    const z = 0.0193339 * rl + 0.1191920 * gl + 0.9503041 * bl;
+    const x = K.rX * rl + K.gX * gl + K.bX * bl;
+    const y = K.rY * rl + K.gY * gl + K.bY * bl;
+    const z = K.rZ * rl + K.gZ * gl + K.bZ * bl;
     // XYZ to Lab (D65 reference)
-    const xn = 0.95047, yn = 1.0, zn = 1.08883;
-    const f = (t) => t > 0.008856 ? Math.cbrt(t) : 7.787 * t + 16 / 116;
-    const fx = f(x / xn), fy = f(y / yn), fz = f(z / zn);
-    const L = 116 * fy - 16;
-    const a = 500 * (fx - fy);
-    const bLab = 200 * (fy - fz);
+    const f = (t) => t > K.epsilon ? Math.cbrt(t) : K.kappaSlope * t + K.lOffset / K.lScale;
+    const fx = f(x / K.xn), fy = f(y / K.yn), fz = f(z / K.zn);
+    const L = K.lScale * fy - K.lOffset;
+    const a = K.aScale * (fx - fy);
+    const bLab = K.bScale * (fy - fz);
     return [L, a, bLab];
 }
 // ── Core: Center-Surround Saliency via W₂ ──
@@ -193,15 +208,27 @@ function srgbToLab(r, g, b) {
  * impractical, and re-deriving it in a separate script is how the Study 1
  * Python port drifted from shipped code. (2026-07-29)
  */
-export async function computeLabSaliency(imagePath, cellSize = 16, surroundRadius = 3) {
+export async function computeLabSaliency(imagePath, cellSize = param("attention.gaze.cellSize"), surroundRadius = param("attention.lab.surroundRadius")) {
+    const K = labConstants();
+    const channels = param("attention.lab.channels");
+    const alpha = channels - 1; // RGBA: alpha is the last byte
+    const alphaCutoff = param("attention.lab.alphaCutoff");
+    const varFloor = param("attention.lab.varianceFloor");
+    // Chrominance boost: a/b channels weighted 3x relative to L (see the loop below).
+    const wL = param("attention.lab.weightL");
+    const wA = param("attention.lab.weightA");
+    const wB = param("attention.lab.weightB");
+    const normFloor = param("attention.normFloor");
+    const scale = param("attention.lab.downscale");
+    const fallbackWidth = param("attention.lab.fallbackWidth");
+    const fallbackHeight = param("attention.lab.fallbackHeight");
     const sharpModule = await import('sharp');
     const sharp = sharpModule.default ?? sharpModule;
     // Load and downscale
-    const scale = 0.5;
     const img = sharp(imagePath);
     const metadata = await img.metadata();
-    const width = Math.round((metadata.width || 800) * scale);
-    const height = Math.round((metadata.height || 600) * scale);
+    const width = Math.round((metadata.width || fallbackWidth) * scale);
+    const height = Math.round((metadata.height || fallbackHeight) * scale);
     const { data } = await img
         .resize(width, height)
         .ensureAlpha()
@@ -224,10 +251,10 @@ export async function computeLabSaliency(imagePath, cellSize = 16, surroundRadiu
             const xEnd = Math.min(xStart + cellSize, width);
             for (let y = yStart; y < yEnd; y++) {
                 for (let x = xStart; x < xEnd; x++) {
-                    const idx = (y * width + x) * 4;
-                    if (pixels[idx + 3] < 25)
+                    const idx = (y * width + x) * channels;
+                    if (pixels[idx + alpha] < alphaCutoff)
                         continue; // Skip transparent
-                    const [L, a, b] = srgbToLab(pixels[idx], pixels[idx + 1], pixels[idx + 2]);
+                    const [L, a, b] = srgbToLab(pixels[idx], pixels[idx + 1], pixels[idx + 2], K);
                     sumL += L;
                     sumA += a;
                     sumB += b;
@@ -241,14 +268,14 @@ export async function computeLabSaliency(imagePath, cellSize = 16, surroundRadiu
                 const mL = sumL / count, mA = sumA / count, mB = sumB / count;
                 cellStats.push({
                     meanL: mL, meanA: mA, meanB: mB,
-                    varL: Math.max(0.01, sumL2 / count - mL * mL),
-                    varA: Math.max(0.01, sumA2 / count - mA * mA),
-                    varB: Math.max(0.01, sumB2 / count - mB * mB),
+                    varL: Math.max(varFloor, sumL2 / count - mL * mL),
+                    varA: Math.max(varFloor, sumA2 / count - mA * mA),
+                    varB: Math.max(varFloor, sumB2 / count - mB * mB),
                     count,
                 });
             }
             else {
-                cellStats.push({ meanL: 0, meanA: 0, meanB: 0, varL: 0.01, varA: 0.01, varB: 0.01, count: 0 });
+                cellStats.push({ meanL: 0, meanA: 0, meanB: 0, varL: varFloor, varA: varFloor, varB: varFloor, count: 0 });
             }
         }
     }
@@ -295,14 +322,12 @@ export async function computeLabSaliency(imagePath, cellSize = 16, surroundRadiu
             // Bures-Wasserstein distance for 3D Gaussians (diagonal covariance)
             // BW² = w_L||ΔL||² + w_a||Δa||² + w_b||Δb||² + variance terms
             //
-            // Chrominance boost: a/b channels weighted 3x relative to L.
+            // Chrominance boost: a/b channels weighted 3x relative to L (wL, wA,
+            // wB, read at entry; hand-set, see attention.lab.weightA).
             // This captures the preattentive "pop-out effect" where a uniquely-colored
             // item captures attention even when its luminance is similar to surroundings.
             // Without this, teal text on white scores LOWER than black text on white
             // because teal has less luminance contrast — the opposite of perceptual reality.
-            const wL = 1.0; // luminance weight
-            const wA = 3.0; // chrominance a* weight (red-green axis)
-            const wB = 3.0; // chrominance b* weight (blue-yellow axis)
             const meanDist = wL * (center.meanL - sMeanL) ** 2 +
                 wA * (center.meanA - sMeanA) ** 2 +
                 wB * (center.meanB - sMeanB) ** 2;
@@ -313,12 +338,37 @@ export async function computeLabSaliency(imagePath, cellSize = 16, surroundRadiu
         }
     }
     // Normalize to 0-1
-    const maxSal = Math.max(...Array.from(saliency), 0.001);
+    const maxSal = Math.max(...Array.from(saliency), normFloor);
     for (let i = 0; i < numCells; i++)
         saliency[i] /= maxSal;
     return { cells: saliency, rows, cols, width: width / scale, height: height / scale };
 }
 function computeValueAttentionParams(filter, schwartzValues) {
+    const P = (id) => param(`attention.${id}`);
+    const mode = {
+        centerHeavyCenterBias: P("mode.centerHeavy.centerBias"),
+        motionAttractedExponent: P("mode.motionAttracted.exponent"),
+        textFocusedThreshold: P("mode.textFocused.threshold"),
+        textFocusedBias: P("mode.textFocused.bias"),
+        largeElementsThreshold: P("mode.largeElements.threshold"),
+    };
+    const neutral = P("value.neutral");
+    // Signed coefficients on (value - neutral): positive raises the parameter.
+    const mod = {
+        stimulationExponent: P("valueMod.stimulation.exponent"),
+        securityBias: P("valueMod.security.bias"),
+        securityThreshold: P("valueMod.security.threshold"),
+        achievementExponent: P("valueMod.achievement.exponent"),
+        conformityCenterBias: P("valueMod.conformity.centerBias"),
+        selfDirectionCenterBias: P("valueMod.selfDirection.centerBias"),
+        traditionThreshold: P("valueMod.tradition.threshold"),
+    };
+    const clamp = {
+        exponentMin: P("filter.exponentMin"),
+        exponentMax: P("filter.exponentMax"),
+        biasLimit: P("filter.biasLimit"),
+        thresholdMax: P("filter.thresholdMax"),
+    };
     // Start with the base attention mode
     let exponent = 1.0;
     let bias = 0.0;
@@ -327,45 +377,45 @@ function computeValueAttentionParams(filter, schwartzValues) {
     // Apply categorical attention mode (legacy — still used for accessibility personas)
     switch (filter.attentionMode) {
         case 'center-heavy':
-            centerBias = 0.7;
+            centerBias = mode.centerHeavyCenterBias;
             break;
         case 'motion-attracted':
-            exponent = 0.5; // square root = more things register
+            exponent = mode.motionAttractedExponent; // square root = more things register
             break;
         case 'text-focused':
-            threshold = 0.2; // suppress low-saliency distractors
-            bias = -0.1; // slightly suppress overall
+            threshold = mode.textFocusedThreshold; // suppress low-saliency distractors
+            bias = mode.textFocusedBias; // slightly suppress overall
             break;
         case 'large-elements':
-            threshold = 0.3;
+            threshold = mode.largeElementsThreshold;
             break;
     }
     // Apply value-driven modulation (v18.54.0)
     if (schwartzValues) {
-        const v = (key) => schwartzValues[key] ?? 0.5;
+        const v = (key) => schwartzValues[key] ?? neutral;
         // Stimulation: high → dispersed attention (notices novelty everywhere)
         // Anderson (2013): reward-seeking broadens attentional spotlight
-        exponent -= (v('stimulation') - 0.5) * 0.4; // 0.9 stim → exponent -0.16
+        exponent += (v('stimulation') - neutral) * mod.stimulationExponent; // 0.9 stim → exponent -0.16
         // Security: high → vigilant scanning (boost everything slightly, lower threshold)
         // Pessoa (2009): threat-sensitive individuals show broadened attention
-        bias += (v('security') - 0.5) * 0.15;
-        threshold -= (v('security') - 0.5) * 0.1;
+        bias += (v('security') - neutral) * mod.securityBias;
+        threshold += (v('security') - neutral) * mod.securityThreshold;
         // Achievement: high → focused on high-value targets (increase exponent)
         // Concentrate attention on the most salient elements, ignore distractors
-        exponent += (v('achievement') - 0.5) * 0.3;
+        exponent += (v('achievement') - neutral) * mod.achievementExponent;
         // Conformity: high → center-biased (follows expected layout patterns)
         // F-pattern and Z-pattern readers look where convention says to look
-        centerBias += (v('conformity') - 0.5) * 0.3;
+        centerBias += (v('conformity') - neutral) * mod.conformityCenterBias;
         // Self-direction: high → less center-biased (explores periphery)
-        centerBias -= (v('selfDirection') - 0.5) * 0.2;
+        centerBias += (v('selfDirection') - neutral) * mod.selfDirectionCenterBias;
         // Tradition: high → suppresses novel elements (sticks to familiar patterns)
-        threshold += (v('tradition') - 0.5) * 0.1;
+        threshold += (v('tradition') - neutral) * mod.traditionThreshold;
     }
     // Clamp values to sane ranges
-    exponent = Math.max(0.3, Math.min(2.0, exponent));
-    bias = Math.max(-0.3, Math.min(0.3, bias));
+    exponent = Math.max(clamp.exponentMin, Math.min(clamp.exponentMax, exponent));
+    bias = Math.max(-clamp.biasLimit, Math.min(clamp.biasLimit, bias));
     centerBias = Math.max(0, Math.min(1.0, centerBias));
-    threshold = Math.max(0, Math.min(0.5, threshold));
+    threshold = Math.max(0, Math.min(clamp.thresholdMax, threshold));
     return { exponent, bias, centerBias, threshold };
 }
 /**
@@ -377,6 +427,11 @@ function computeValueAttentionParams(filter, schwartzValues) {
  */
 export function applyPersonaFilter(saliency, rows, cols, filter, schwartzValues) {
     const params = computeValueAttentionParams(filter, schwartzValues);
+    const belowThresholdGain = param("attention.filter.belowThresholdGain");
+    const slowProcessingBelow = param("attention.filter.slowProcessingBelow");
+    const speedThresholdBase = param("attention.filter.speedThresholdBase");
+    const speedThresholdSlope = param("attention.filter.speedThresholdSlope");
+    const normFloor = param("attention.normFloor");
     const filtered = new Float64Array(saliency.length);
     const centerRow = rows / 2;
     const centerCol = cols / 2;
@@ -387,7 +442,7 @@ export function applyPersonaFilter(saliency, rows, cols, filter, schwartzValues)
             let s = saliency[idx];
             // Apply threshold floor
             if (s < params.threshold) {
-                s *= 0.2;
+                s *= belowThresholdGain;
             }
             // Apply power curve (exponent controls attention concentration)
             s = Math.pow(Math.max(0, s), params.exponent);
@@ -402,15 +457,15 @@ export function applyPersonaFilter(saliency, rows, cols, filter, schwartzValues)
         }
     }
     // Apply processing speed
-    if (filter.processingSpeed < 0.8) {
-        const speedThreshold = 0.3 + (1 - filter.processingSpeed) * 0.3;
+    if (filter.processingSpeed < slowProcessingBelow) {
+        const speedThreshold = speedThresholdBase + (1 - filter.processingSpeed) * speedThresholdSlope;
         for (let i = 0; i < filtered.length; i++) {
             if (filtered[i] < speedThreshold)
                 filtered[i] *= filter.processingSpeed;
         }
     }
     // Renormalize
-    const max = Math.max(...Array.from(filtered), 0.001);
+    const max = Math.max(...Array.from(filtered), normFloor);
     for (let i = 0; i < filtered.length; i++)
         filtered[i] /= max;
     return filtered;
@@ -421,60 +476,33 @@ export function applyPersonaFilter(saliency, rows, cols, filter, schwartzValues)
  * These represent how inherently attention-grabbing each element type is,
  * independent of visual appearance. A search bar is visually subtle but
  * a power user looks there first.
+ *
+ * The weights (cta 1.0 down to decorative 0.1, price and error 0.9) are
+ * registered as attention.type.*; with the spatial prior in place their
+ * spread contributes nothing measurable on UEyes (see that entry).
  */
-const ELEMENT_TYPE_WEIGHTS = {
-    cta: 1.0, // Calls to action — highest inherent draw
-    heading: 0.8, // Headings — structural anchors
-    form: 0.7, // Form fields — interactive, task-relevant
-    search: 0.7, // Search bars — navigation entry point
-    image: 0.5, // Images — visual anchors
-    navigation: 0.4, // Nav elements — wayfinding
-    link: 0.3, // Regular links
-    content: 0.2, // Body text
-    decorative: 0.1, // Decorative images, spacers
-    price: 0.9, // Prices — high commercial relevance
-    error: 0.9, // Error messages — demand immediate attention
-};
+const ELEMENT_TYPES = [
+    "cta", "heading", "form", "search", "image", "navigation", "link", "content", "decorative", "price", "error",
+];
+function elementTypeWeights() {
+    const w = {};
+    for (const t of ELEMENT_TYPES)
+        w[t] = param(`attention.type.${t}`);
+    return w;
+}
 /**
- * Persona trait → element type weight modifiers.
- * Each persona amplifies or suppresses attention to certain element types
- * based on their cognitive traits and goals.
+ * Persona name → element type weight modifiers, the fallback when no cognitive
+ * traits resolve for the persona (computeTraitModifiers is the normal path).
+ * The table lives in the registry as attention.personaMod.<persona>.<type>.
  */
-const PERSONA_ELEMENT_MODIFIERS = {
-    "first-timer": {
-        cta: 1.2, heading: 1.3, navigation: 1.2, search: 1.0, content: 0.6, decorative: 0.4,
-    },
-    "power-user": {
-        cta: 0.6, heading: 0.5, navigation: 1.4, search: 1.5, content: 0.3, form: 1.2, decorative: 0.1,
-    },
-    "impatient-user": {
-        cta: 1.5, heading: 1.0, navigation: 0.8, search: 1.3, content: 0.2, decorative: 0.1, price: 1.3,
-    },
-    "elderly-user": {
-        cta: 1.0, heading: 1.4, navigation: 1.3, search: 0.8, content: 0.8, form: 0.9, decorative: 0.3,
-    },
-    "cognitive-adhd": {
-        cta: 0.8, heading: 0.7, navigation: 0.5, image: 1.5, decorative: 1.2, content: 0.3, search: 0.6,
-    },
-    "anxious-user": {
-        cta: 0.7, heading: 1.0, price: 1.5, error: 1.5, form: 0.8, decorative: 0.3,
-    },
-    "motor-impairment-tremor": {
-        cta: 1.0, heading: 0.8, navigation: 0.6, form: 0.7, search: 0.9, content: 0.5,
-    },
-    "dyslexic-user": {
-        cta: 1.1, heading: 1.3, image: 1.3, content: 0.4, navigation: 0.9, decorative: 0.5,
-    },
-    "low-vision-magnified": {
-        cta: 1.2, heading: 1.5, navigation: 1.0, content: 0.5, image: 0.8, decorative: 0.2,
-    },
-    "mobile-user": {
-        cta: 1.3, heading: 1.0, search: 1.2, navigation: 0.7, content: 0.4, decorative: 0.2,
-    },
-    "confident-user": {
-        cta: 0.7, heading: 0.5, navigation: 1.3, search: 1.4, form: 1.2, content: 0.3, decorative: 0.1,
-    },
-};
+function staticPersonaModifiers(personaName) {
+    const out = {};
+    if (!Object.hasOwn(ATTENTION_STATIC_MODIFIER_TYPES, personaName))
+        return out;
+    for (const t of ATTENTION_STATIC_MODIFIER_TYPES[personaName])
+        out[t] = param(`attention.personaMod.${personaName}.${t}`);
+    return out;
+}
 /**
  * How many ranked cells to keep as the hotspot pool.
  *
@@ -497,7 +525,7 @@ const PERSONA_ELEMENT_MODIFIERS = {
  * `attentionCompetitors`, a different field, so response size is unchanged.
  */
 function hotspotPoolSize(totalCells) {
-    return Math.min(4000, Math.max(200, Math.round(totalCells * 0.05)));
+    return Math.min(param("attention.hotspotPool.max"), Math.max(param("attention.hotspotPool.min"), Math.round(totalCells * param("attention.hotspotPool.share"))));
 }
 /**
  * Compute trait-driven element modifiers from cognitive traits.
@@ -505,15 +533,18 @@ function hotspotPoolSize(totalCells) {
  * ANY persona's trait values, not just the 11 hardcoded ones.
  */
 function computeTraitModifiers(traits) {
-    const t = (key) => traits[key] ?? 0.5;
+    const P = (id) => param(`attention.traitMod.${id}`);
+    const missing = P("missingTrait");
+    const reading = { threshold: P("reading.threshold"), base: P("reading.base"), slope: P("reading.slope") };
+    const t = (key) => traits[key] ?? missing;
     return {
         // Low patience → fixate on CTAs and search, skip content
-        cta: 1.0 + (1.0 - t("patience")) * 0.5,
-        search: 0.7 + (1.0 - t("patience")) * 0.6,
-        content: 0.2 + t("patience") * 0.4,
+        cta: P("cta.base") + (1.0 - t("patience")) * P("cta.slope"),
+        search: P("search.base") + (1.0 - t("patience")) * P("search.slope"),
+        content: P("content.base") + t("patience") * P("content.slope"),
         // High curiosity → explore headings, images, more elements
-        heading: 0.8 + t("curiosity") * 0.4,
-        image: 0.5 + t("curiosity") * 0.5,
+        heading: P("heading.base") + t("curiosity") * P("heading.slope"),
+        image: P("image.base") + t("curiosity") * P("image.slope"),
         // Navigation weight is deliberately FAMILIARITY-INDEPENDENT.
         //
         // This read `0.4 + (1.0 - t("siteFamiliarity")) * 0.6`, which made the
@@ -533,32 +564,36 @@ function computeTraitModifiers(traits) {
         //
         // 0.7 is the previous formula evaluated at familiarity 0.5, so a persona
         // with a neutral (or absent) value sees no change at all.
-        navigation: 0.7,
+        navigation: P("navigation"),
         // High readingTendency → content draws attention
         // (overrides patience for reading-heavy personas)
-        ...(t("readingTendency") > 0.6 ? { content: 0.4 + t("readingTendency") * 0.4 } : {}),
+        ...(t("readingTendency") > reading.threshold ? { content: reading.base + t("readingTendency") * reading.slope } : {}),
         // Low selfEfficacy → error messages, help text draw attention
-        error: 0.5 + (1.0 - t("selfEfficacy")) * 0.5,
+        error: P("error.base") + (1.0 - t("selfEfficacy")) * P("error.slope"),
         // High riskTolerance → prices don't scare, CTAs are natural
-        price: 0.5 + (1.0 - t("riskTolerance")) * 0.5,
+        price: P("price.base") + (1.0 - t("riskTolerance")) * P("price.slope"),
         // Decorative: ADHD-like attention (low metacognitivePlanning = distracted by visuals)
-        decorative: 0.1 + (1.0 - t("metacognitivePlanning")) * 0.4,
-        form: 0.5 + t("proceduralFluency") * 0.3,
+        decorative: P("decorative.base") + (1.0 - t("metacognitivePlanning")) * P("decorative.slope"),
+        form: P("form.base") + t("proceduralFluency") * P("form.slope"),
     };
 }
+/** Goal words that never count as keywords. */
+const GOAL_STOP_WORDS = [
+    "the", "a", "an", "to", "for", "of", "in", "on", "at", "is", "it", "my", "i", "me", "and", "or", "how", "do", "can",
+];
 /**
  * Compute goal relevance boost for an element.
  * If the element's text matches the user's goal, massively boost its weight.
  * This is the "task-driven attention" component (Yarbus 1967).
  */
-function computeGoalRelevance(elementText, goal) {
+function computeGoalRelevance(elementText, goal, minKeywordLength, gain) {
     if (!goal || !elementText)
         return 0;
     const goalLower = goal.toLowerCase();
     const textLower = elementText.toLowerCase();
     // Extract goal keywords (remove stop words)
-    const stopWords = new Set(["the", "a", "an", "to", "for", "of", "in", "on", "at", "is", "it", "my", "i", "me", "and", "or", "how", "do", "can"]);
-    const goalWords = goalLower.split(/\s+/).filter(w => w.length > 2 && !stopWords.has(w));
+    const stopWords = new Set(GOAL_STOP_WORDS);
+    const goalWords = goalLower.split(/\s+/).filter(w => w.length >= minKeywordLength && !stopWords.has(w));
     if (goalWords.length === 0)
         return 0;
     // Score based on keyword matches
@@ -571,9 +606,41 @@ function computeGoalRelevance(elementText, goal) {
         return 0;
     // Partial match: some keywords found
     const matchRatio = matches / goalWords.length;
-    // 2.0x boost for full match, proportional for partial
-    return matchRatio * 2.0;
+    // 2.0x boost for full match (attention.relevance.gain), proportional for partial
+    return matchRatio * gain;
 }
+/**
+ * The element weights the semantic layer paints with for this persona: base
+ * weight per type, the persona's modifier per type (from its cognitive traits,
+ * else the static table, as buildSemanticMap chooses), the weight of an unknown
+ * type, and the goal-keyword constants.
+ *
+ * Exported for the in-page journey heatmap (ATTENTION_OVERLAY_SCRIPT in
+ * cognitive/index.ts), which draws inside the browser and cannot call this
+ * module. It received a hand-written copy of these tables that had drifted:
+ * its navigation weight still read siteFamiliarity, dropped here on
+ * 2026-08-05. It now receives this. (ISC-48.1, 2026-10-22;
+ * tests/attention-overlay-config.test.ts checks it against buildSemanticMap.)
+ */
+export function semanticWeightTable(personaName, cognitiveTraits) {
+    return {
+        typeWeights: elementTypeWeights(),
+        personaMods: cognitiveTraits ? computeTraitModifiers(cognitiveTraits) : staticPersonaModifiers(personaName),
+        typeFallback: param("attention.type.fallback"),
+        relevanceGain: param("attention.relevance.gain"),
+        minKeywordLength: param("attention.goal.minKeywordLength"),
+        stopWords: [...GOAL_STOP_WORDS],
+    };
+}
+/** Text cues for Schwartz-value relevance, in the order their boosts apply. */
+const VALUE_PATTERN_REGEX = {
+    trust: /secur|trust|verif|guarant|privacy|encrypt|safe|protect/i,
+    social: /review|rating|star|trusted.?by|\d+[kK]?\+?\s*users|testimonial/i,
+    novelty: /\bnew\b|beta|launch|updat|just.?added|latest|early.?access/i,
+    metrics: /\d+%|\d+x\s|roi|faster|improv|performance|\$\d/i,
+    urgency: /limited|only\s*\d|hurry|expire|last.?chance|act.?now/i,
+    community: /community|join|forum|together|open.?source|contrib/i,
+};
 /**
  * Build a semantic attention map from DOM elements.
  *
@@ -607,20 +674,24 @@ relevanceScores,
  */
 typeWeights) {
     const map = new Float64Array(rows * cols);
-    const TYPE_W = typeWeights ?? ELEMENT_TYPE_WEIGHTS;
+    const TYPE_W = typeWeights ?? elementTypeWeights();
+    const typeFallback = param("attention.type.fallback");
+    const relevanceGain = param("attention.relevance.gain");
+    const minKeywordLength = param("attention.goal.minKeywordLength");
+    const neutral = param("attention.value.neutral");
+    const edgeEpsilon = param("attention.grid.edgeEpsilon");
+    const normPercentile = param("attention.normPercentile");
+    const normFloor = param("attention.normFloor");
     // Use trait-driven modifiers if traits available, else fall back to static table
     const modifiers = cognitiveTraits
         ? computeTraitModifiers(cognitiveTraits)
-        : (PERSONA_ELEMENT_MODIFIERS[personaName] || {});
-    // Value-driven semantic patterns (imported from attention-quality)
-    const VALUE_PATTERNS = {
-        trust: { pattern: /secur|trust|verif|guarant|privacy|encrypt|safe|protect/i, values: { security: 1.0, conformity: 0.5 } },
-        social: { pattern: /review|rating|star|trusted.?by|\d+[kK]?\+?\s*users|testimonial/i, values: { conformity: 1.0, security: 0.6 } },
-        novelty: { pattern: /\bnew\b|beta|launch|updat|just.?added|latest|early.?access/i, values: { stimulation: 1.0, selfDirection: 0.5 } },
-        metrics: { pattern: /\d+%|\d+x\s|roi|faster|improv|performance|\$\d/i, values: { achievement: 1.0, power: 0.4 } },
-        urgency: { pattern: /limited|only\s*\d|hurry|expire|last.?chance|act.?now/i, values: { stimulation: 0.5, achievement: 0.4 } },
-        community: { pattern: /community|join|forum|together|open.?source|contrib/i, values: { benevolence: 0.8, universalism: 0.6 } },
-    };
+        : staticPersonaModifiers(personaName);
+    // Value-driven semantic patterns (imported from attention-quality). The
+    // weights are attention.valuePattern.<pattern>.<value>, summed in this order.
+    const VALUE_PATTERNS = Object.entries(VALUE_PATTERN_REGEX).map(([name, pattern]) => ({
+        pattern,
+        values: ATTENTION_VALUE_PATTERN_VALUES[name].map((v) => [v, param(`attention.valuePattern.${name}.${v}`)]),
+    }));
     for (const [elementIndex, el] of elements.entries()) {
         // Determine element type
         let elType = el.type || "content";
@@ -633,7 +704,7 @@ typeWeights) {
         else if (el.isDecorative)
             elType = "decorative";
         // Layer 1: Base weight by element type
-        let weight = TYPE_W[elType] ?? 0.2;
+        let weight = TYPE_W[elType] ?? typeFallback;
         // Layer 2: Persona modifier (trait-driven or static lookup)
         const modifier = modifiers[elType] ?? 1.0;
         weight *= modifier;
@@ -646,22 +717,22 @@ typeWeights) {
         // is a floor to degrade to, never the intended path.
         const judged = relevanceScores?.[elementIndex];
         if (judged !== undefined && judged > 0) {
-            weight *= (1.0 + judged * 2.0); // same 3x ceiling as the keyword path
+            weight *= (1.0 + judged * relevanceGain); // same 3x ceiling as the keyword path
         }
         else if (goal && el.text) {
-            const goalBoost = computeGoalRelevance(el.text, goal);
+            const goalBoost = computeGoalRelevance(el.text, goal, minKeywordLength, relevanceGain);
             if (goalBoost > 0) {
                 weight *= (1.0 + goalBoost);
             }
         }
         // Layer 4: Schwartz value relevance — elements matching persona values get boost
         if (schwartzValues && el.text) {
-            for (const [, pattern] of Object.entries(VALUE_PATTERNS)) {
+            for (const pattern of VALUE_PATTERNS) {
                 if (pattern.pattern.test(el.text)) {
                     let valueBoost = 0;
-                    for (const [valueName, valueWeight] of Object.entries(pattern.values)) {
-                        const personaValue = schwartzValues[valueName] ?? 0.5;
-                        valueBoost += (personaValue - 0.5) * valueWeight;
+                    for (const [valueName, valueWeight] of pattern.values) {
+                        const personaValue = schwartzValues[valueName] ?? neutral;
+                        valueBoost += (personaValue - neutral) * valueWeight;
                     }
                     weight *= (1.0 + Math.max(0, valueBoost));
                 }
@@ -691,9 +762,9 @@ typeWeights) {
             const x0 = box.x, x1 = box.x + box.width;
             const y0 = box.y, y1 = box.y + box.height;
             const startCol = Math.max(0, Math.floor(x0 / pxPerCol));
-            const endCol = Math.min(cols - 1, Math.floor((x1 - 1e-6) / pxPerCol));
+            const endCol = Math.min(cols - 1, Math.floor((x1 - edgeEpsilon) / pxPerCol));
             const startRow = Math.max(0, Math.floor(y0 / pxPerRow));
-            const endRow = Math.min(rows - 1, Math.floor((y1 - 1e-6) / pxPerRow));
+            const endRow = Math.min(rows - 1, Math.floor((y1 - edgeEpsilon) / pxPerRow));
             for (let r = startRow; r <= endRow; r++) {
                 const cellTop = r * pxPerRow;
                 const overlapY = Math.max(0, Math.min(y1, cellTop + pxPerRow) - Math.max(y0, cellTop));
@@ -726,9 +797,9 @@ typeWeights) {
     // exceptional cells clip, which is standard practice for rendering saliency.
     const sorted = Array.from(map).filter((v) => v > 0).sort((a, b) => a - b);
     const p95 = sorted.length > 0
-        ? sorted[Math.min(sorted.length - 1, Math.floor(sorted.length * 0.95))]
-        : 0.001;
-    const denom = Math.max(p95, 0.001);
+        ? sorted[Math.min(sorted.length - 1, Math.floor(sorted.length * normPercentile))]
+        : normFloor;
+    const denom = Math.max(p95, normFloor);
     for (let i = 0; i < map.length; i++)
         map[i] = Math.min(1, map[i] / denom);
     return map;
@@ -769,13 +840,14 @@ typeWeights) {
  * attention for a specific persona. 0.05 is the best available estimate, not a
  * validated constant for the goal-directed case.
  */
-export function blendSaliencyMaps(visual, semantic, visualWeight = 0.05, semanticWeight = 0.95) {
+export function blendSaliencyMaps(visual, semantic, visualWeight = param("attention.blend.legacyVisualWeight"), semanticWeight = param("attention.blend.legacySemanticWeight")) {
+    const normFloor = param("attention.normFloor");
     const blended = new Float64Array(visual.length);
     for (let i = 0; i < visual.length; i++) {
         blended[i] = visual[i] * visualWeight + semantic[i] * semanticWeight;
     }
     // Renormalize
-    const max = Math.max(...Array.from(blended), 0.001);
+    const max = Math.max(...Array.from(blended), normFloor);
     for (let i = 0; i < blended.length; i++)
         blended[i] /= max;
     return blended;
@@ -810,8 +882,6 @@ export function blendSaliencyMaps(visual, semantic, visualWeight = 0.05, semanti
  *
  * Pass `spatialPrior: 0` to disable it for any of those cases.
  */
-const WEB_SPATIAL_PRIOR_ROWS = 12;
-const WEB_SPATIAL_PRIOR_COLS = 16;
 const WEB_SPATIAL_PRIOR = [
     0.3438, 0.4761, 0.6428, 0.8103, 0.9375, 1.0000, 0.9920, 0.9220, 0.8109, 0.6831, 0.5571, 0.4409, 0.3344, 0.2406, 0.1686, 0.1216,
     0.3225, 0.4491, 0.6086, 0.7704, 0.8973, 0.9657, 0.9669, 0.9055, 0.8004, 0.6767, 0.5533, 0.4380, 0.3308, 0.2359, 0.1636, 0.1172,
@@ -826,13 +896,63 @@ const WEB_SPATIAL_PRIOR = [
     0.1355, 0.1682, 0.2067, 0.2426, 0.2686, 0.2818, 0.2811, 0.2677, 0.2465, 0.2236, 0.2021, 0.1813, 0.1578, 0.1310, 0.1061, 0.0877,
     0.1150, 0.1385, 0.1658, 0.1908, 0.2086, 0.2173, 0.2160, 0.2056, 0.1899, 0.1736, 0.1587, 0.1442, 0.1278, 0.1094, 0.0923, 0.0795,
 ];
-/** Fitted operating point, all five CV folds unanimous. See `applyGazeCalibration`. */
-export const GAZE_CALIBRATION = { visualWeight: 0.2, sigmaCells: 2.5, spatialPrior: 1.0 };
+/**
+ * The fitted table, checked against its registered checksum on every use.
+ *
+ * The 192 cells are registered as ONE fitted entry (attention.spatialPrior.checksum,
+ * with its shape as .rows and .cols) rather than one per cell; provenance/attention.ts
+ * says why. Reading the checksum through param() and refusing a table that no
+ * longer hashes to it is what ties this literal to the registry: a cell edited
+ * here without re-registering stops the model instead of shipping unrecorded.
+ */
+function spatialPrior() {
+    const rows = param("attention.spatialPrior.rows");
+    const cols = param("attention.spatialPrior.cols");
+    const checksum = param("attention.spatialPrior.checksum");
+    if (WEB_SPATIAL_PRIOR.length !== rows * cols || WEB_SPATIAL_PRIOR_CHECKSUM !== checksum) {
+        throw new Error(`WEB_SPATIAL_PRIOR does not match its registration (${WEB_SPATIAL_PRIOR.length} cells, checksum ${WEB_SPATIAL_PRIOR_CHECKSUM}; ` +
+            `registered ${rows}x${cols}, checksum ${checksum}). Re-fit or re-register it in provenance/attention.ts.`);
+    }
+    return { table: WEB_SPATIAL_PRIOR, rows, cols };
+}
+const WEB_SPATIAL_PRIOR_CHECKSUM = tableChecksum(WEB_SPATIAL_PRIOR);
+/**
+ * Fitted operating point, all five CV folds unanimous. See `applyGazeCalibration`.
+ *
+ * A live view of the registry (attention.gaze.*): each field reads its entry
+ * through param() when read, so this object cannot disagree with the values the
+ * model runs on.
+ */
+export const GAZE_CALIBRATION = Object.freeze({
+    get visualWeight() { return param("attention.gaze.visualWeight"); },
+    get sigmaCells() { return param("attention.gaze.sigmaCells"); },
+    get spatialPrior() { return param("attention.gaze.spatialPrior"); },
+});
+/**
+ * The cell size the calibration was fitted at: research/study1 runs every
+ * stimulus at CELL = 16 (run-optimize.ts, run-semantic.ts). computeLabSaliency
+ * halves the image before celling, so one fitted cell spans 32 capture px and
+ * `sigmaCells` 2.5 is 80 capture px. Registered as attention.gaze.cellSize; this
+ * export is that entry's value, read once at load.
+ */
+export const GAZE_CALIBRATION_CELL_SIZE = param("attention.gaze.cellSize");
+/**
+ * `GAZE_CALIBRATION.sigmaCells` rescaled to a grid of `cellSize`, so the
+ * smoothing keeps its fitted width in pixels.
+ *
+ * The blur runs in cells, and until 2026-10-21 every caller passed 2.5 cells
+ * whatever its grid: attention_analysis (cellSize 4) smoothed at a quarter of the
+ * fitted width, journey_heatmap_gif (8) at half. Exactly 2.5 at 16, so the
+ * Study 1 numbers do not move.
+ */
+export function gazeSigmaCells(cellSize) {
+    return param("attention.gaze.sigmaCells") * (param("attention.gaze.cellSize") / cellSize);
+}
 /** Separable Gaussian blur over the cell grid. */
 export function smoothSaliency(map, rows, cols, sigma) {
     if (sigma <= 0)
         return map;
-    const rad = Math.ceil(sigma * 3), inv = 1 / (2 * sigma * sigma), k = [];
+    const rad = Math.ceil(sigma * param("attention.smooth.radiusSigmas")), inv = 1 / (2 * sigma * sigma), k = [];
     for (let d = -rad; d <= rad; d++)
         k.push(Math.exp(-d * d * inv));
     const tmp = new Float64Array(map.length), out = new Float64Array(map.length);
@@ -877,16 +997,21 @@ export function smoothSaliency(map, rows, cols, sigma) {
  * 0.0005. Human inter-observer ceiling on the same data is 0.8112, so this
  * closes about 52% of the reachable gap. All five folds selected the same
  * values.
+ *
+ * `sigma` is in cells and its default is right only on the fitted grid
+ * (GAZE_CALIBRATION_CELL_SIZE); on any other grid pass `gazeSigmaCells(cellSize)`.
+ * The spatial prior needs no rescaling: it is looked up by normalised position.
  */
-export function applyGazeCalibration(map, rows, cols, sigma = GAZE_CALIBRATION.sigmaCells, priorStrength = GAZE_CALIBRATION.spatialPrior) {
+export function applyGazeCalibration(map, rows, cols, sigma = param("attention.gaze.sigmaCells"), priorStrength = param("attention.gaze.spatialPrior")) {
     let m = smoothSaliency(map, rows, cols, sigma);
     if (priorStrength > 0) {
+        const prior = spatialPrior();
         const out = new Float64Array(m.length);
         for (let r = 0; r < rows; r++) {
-            const pr = Math.min(WEB_SPATIAL_PRIOR_ROWS - 1, Math.floor((r / rows) * WEB_SPATIAL_PRIOR_ROWS));
+            const pr = Math.min(prior.rows - 1, Math.floor((r / rows) * prior.rows));
             for (let c = 0; c < cols; c++) {
-                const pc = Math.min(WEB_SPATIAL_PRIOR_COLS - 1, Math.floor((c / cols) * WEB_SPATIAL_PRIOR_COLS));
-                const p = WEB_SPATIAL_PRIOR[pr * WEB_SPATIAL_PRIOR_COLS + pc];
+                const pc = Math.min(prior.cols - 1, Math.floor((c / cols) * prior.cols));
+                const p = prior.table[pr * prior.cols + pc];
                 out[r * cols + c] = m[r * cols + c] * (1 - priorStrength + priorStrength * p);
             }
         }
@@ -895,6 +1020,17 @@ export function applyGazeCalibration(map, rows, cols, sigma = GAZE_CALIBRATION.s
     return m;
 }
 // ── Public API ──
+/** Metric definitions and display rounding shared by both analysis paths. */
+function readMetricParams() {
+    return {
+        entropyEpsilon: param("attention.metric.entropyEpsilon"),
+        concentrationShare: param("attention.metric.concentrationShare"),
+        competitors: param("attention.report.competitors"),
+        r2: param("attention.rounding.hundredths"),
+        r3: param("attention.rounding.thousandths"),
+        r4: param("attention.rounding.tenThousandths"),
+    };
+}
 /**
  * Compute persona-specific attention analysis for a screenshot.
  *
@@ -902,7 +1038,8 @@ export function applyGazeCalibration(map, rows, cols, sigma = GAZE_CALIBRATION.s
  * 1. Visual saliency — W₂ on CIE-Lab distributions (what POPS visually)
  * 2. Semantic attention — DOM element classification × persona priorities (what MATTERS)
  *
- * Blended at 5% visual / 95% semantic. The direction is the one the literature
+ * Blended at 20% visual / 80% semantic (GAZE_CALIBRATION, refitted with the spatial
+ * prior on 2026-08-04; 5% was the optimum before the prior existed). The direction is the one the literature
  * predicts — task-driven attention dominates bottom-up saliency for goal-oriented
  * users (Yarbus 1967, Henderson 2003, Tatler et al. 2011) — but the number is no
  * longer read off those papers. It was swept against 495 UEyes web stimuli and
@@ -921,9 +1058,11 @@ export function applyGazeCalibration(map, rows, cols, sigma = GAZE_CALIBRATION.s
  * are interchangeable. Typical runtime: 5-30ms vs 1500-3000ms for the full
  * version.
  */
-export async function analyzeAttentionFromDOM(viewportWidth, viewportHeight, domElements, personaName, cellSize = 16, schwartzValues, goal, cognitiveTraits, relevanceScores) {
+export async function analyzeAttentionFromDOM(viewportWidth, viewportHeight, domElements, personaName, cellSize = param("attention.gaze.cellSize"), schwartzValues, goal, cognitiveTraits, relevanceScores, accessibilityTraits) {
     const startTime = performance.now();
-    const profile = getPerceptualProfile(personaName);
+    const M = readMetricParams();
+    const resolvedProfile = resolvePerceptualProfile(personaName, accessibilityTraits);
+    const profile = resolvedProfile.profile;
     // Look up Schwartz values from built-in profiles if caller didn't pass any
     let values = schwartzValues;
     if (!values) {
@@ -966,9 +1105,17 @@ export async function analyzeAttentionFromDOM(viewportWidth, viewportHeight, dom
     //
     // This path exists so callers are INTERCHANGEABLE with analyzeAttention (its
     // own docstring says so). Calibrating one and not the other would silently
-    // make the fast path and the full path disagree about where a persona looks --
-    // the live cognitive-journey overlay drawing a different map from the audit.
-    const filtered = applyGazeCalibration(applyPersonaFilter(semanticMap, rows, cols, profile.visualFilter, values), rows, cols);
+    // make the fast path and the full path disagree about where a persona looks.
+    //
+    // Sigma keeps the fitted width in PIXELS (80, at device pixel ratio 1). This
+    // grid is not halved, so a cell here spans cellSize px, which is the cell
+    // analyzeAttention gets from cellSize x downscale. gazeSigmaCells(cellSize)
+    // alone smoothed this path at 40 px, half the fitted width, at every
+    // cellSize (ISC-47, fixed 2026-10-22; tests/attention-dom-gaze-width.test.ts).
+    // The live cognitive-journey overlay does NOT run this function: it draws
+    // its own in-page copy (ATTENTION_OVERLAY_SCRIPT in cognitive/index.ts), so
+    // no shipped overlay moved with this fix.
+    const filtered = applyGazeCalibration(applyPersonaFilter(semanticMap, rows, cols, profile.visualFilter, values), rows, cols, gazeSigmaCells(cellSize * param("attention.lab.downscale")));
     // Hotspots, entropy, concentration — same shape as analyzeAttention
     const indexed = Array.from(filtered).map((s, i) => ({ s, i }));
     indexed.sort((a, b) => b.s - a.s);
@@ -984,15 +1131,16 @@ export async function analyzeAttentionFromDOM(viewportWidth, viewportHeight, dom
     for (let i = 0; i < filtered.length; i++) {
         const p = filtered[i] / totalSaliency;
         if (p > 0)
-            entropy -= p * Math.log2(p + 1e-10);
+            entropy -= p * Math.log2(p + M.entropyEpsilon);
     }
     entropy = entropy / Math.log2(filtered.length || 1);
     const sorted = Array.from(filtered).sort((a, b) => b - a);
-    const top20Count = Math.ceil(filtered.length * 0.2);
+    const top20Count = Math.ceil(filtered.length * M.concentrationShare);
     const top20Sum = sorted.slice(0, top20Count).reduce((a, b) => a + b, 0);
     const concentration = top20Sum / totalSaliency;
     return {
         persona: personaName,
+        perceptualProfile: { resolution: resolvedProfile.resolution, name: resolvedProfile.resolvedProfile },
         saliencyMap: {
             cells: filtered,
             rows, cols,
@@ -1001,19 +1149,23 @@ export async function analyzeAttentionFromDOM(viewportWidth, viewportHeight, dom
             hotspots,
         },
         alignmentScore: 1, // No base to compare against
-        entropy: Math.round(entropy * 1000) / 1000,
-        concentration: Math.round(concentration * 1000) / 1000,
-        attentionCompetitors: hotspots.slice(0, 5).map((h) => ({
+        entropy: Math.round(entropy * M.r3) / M.r3,
+        concentration: Math.round(concentration * M.r3) / M.r3,
+        attentionCompetitors: hotspots.slice(0, M.competitors).map((h) => ({
             region: `(${h.x}, ${h.y})`,
-            saliency: Math.round(h.saliency * 100) / 100,
+            saliency: Math.round(h.saliency * M.r2) / M.r2,
         })),
         transportCost: 0,
         computeTimeMs: Math.round(performance.now() - startTime),
     };
 }
-export async function analyzeAttention(screenshotPath, personaName, cellSize = 16, schwartzValues, domElements, goal, cognitiveTraits, relevanceScores) {
+export async function analyzeAttention(screenshotPath, personaName, cellSize = param("attention.gaze.cellSize"), schwartzValues, domElements, goal, cognitiveTraits, relevanceScores, accessibilityTraits) {
     const startTime = performance.now();
-    const profile = getPerceptualProfile(personaName);
+    const M = readMetricParams();
+    const downscale = param("attention.lab.downscale");
+    const visualWeight = param("attention.gaze.visualWeight");
+    const resolvedProfile = resolvePerceptualProfile(personaName, accessibilityTraits);
+    const profile = resolvedProfile.profile;
     // v18.54.0: If no values passed, try to look them up from the built-in value profiles
     let values = schwartzValues;
     if (!values) {
@@ -1056,13 +1208,13 @@ export async function analyzeAttention(screenshotPath, personaName, cellSize = 1
         // blend to take more from the visual channel. Re-fitted WITH the prior and
         // smoothing in place -- all five CV folds selected 0.2. Carrying 0.05 over
         // would have been importing half of an optimum.
-        combinedSaliency = blendSaliencyMaps(baseSaliency.cells, semanticMap, GAZE_CALIBRATION.visualWeight, 1 - GAZE_CALIBRATION.visualWeight);
+        combinedSaliency = blendSaliencyMaps(baseSaliency.cells, semanticMap, visualWeight, 1 - visualWeight);
     }
     else {
         // No DOM data — fall back to visual-only.
         //
         // This fallback changes what the heatmap MEANS, and said so nowhere. With
-        // DOM elements the map is 35% bottom-up saliency + 65% semantic (element
+        // DOM elements the map is 20% bottom-up saliency + 80% semantic (element
         // type x persona priority) — a model of attention. Without them it is the
         // bottom-up layer alone: centre-surround colour contrast in CIE-Lab. Those
         // are different quantities under one name, and two of the three production
@@ -1077,16 +1229,17 @@ export async function analyzeAttention(screenshotPath, personaName, cellSize = 1
     }
     const attentionModel = domElements && domElements.length > 0 ? "visual+semantic" : "visual-only";
     // Apply persona perceptual filter + value-driven attention modulation, then
-    // foveal smoothing and the fitted spatial prior (see applyGazeCalibration).
-    const filtered = applyGazeCalibration(applyPersonaFilter(combinedSaliency, baseSaliency.rows, baseSaliency.cols, profile.visualFilter, values), baseSaliency.rows, baseSaliency.cols);
+    // foveal smoothing and the fitted spatial prior (see applyGazeCalibration),
+    // the smoothing scaled to this grid so its width in pixels is the fitted one.
+    const filtered = applyGazeCalibration(applyPersonaFilter(combinedSaliency, baseSaliency.rows, baseSaliency.cols, profile.visualFilter, values), baseSaliency.rows, baseSaliency.cols, gazeSigmaCells(cellSize));
     // Build saliency map with hotspots
     const indexed = Array.from(filtered).map((s, i) => ({ s, i }));
     indexed.sort((a, b) => b.s - a.s);
     const hotspots = indexed.slice(0, hotspotPoolSize(indexed.length)).map(({ s, i }) => ({
         row: Math.floor(i / baseSaliency.cols),
         col: i % baseSaliency.cols,
-        x: (i % baseSaliency.cols) * cellSize * 2, // Scale back to original
-        y: Math.floor(i / baseSaliency.cols) * cellSize * 2,
+        x: (i % baseSaliency.cols) * cellSize / downscale, // Scale back to original
+        y: Math.floor(i / baseSaliency.cols) * cellSize / downscale,
         saliency: s,
     }));
     const saliencyMap = {
@@ -1094,7 +1247,7 @@ export async function analyzeAttention(screenshotPath, personaName, cellSize = 1
         /**
          * Which model produced `cells`.
          *
-         * "visual+semantic" — 35% centre-surround CIE-Lab saliency blended with 65%
+         * "visual+semantic" — 20% centre-surround CIE-Lab saliency blended with 80%
          * DOM-semantic attention (element type x persona priority).
          * "visual-only" — the bottom-up layer alone, i.e. local colour-contrast
          * pop-out. Useful, but NOT a model of where a person looks.
@@ -1104,7 +1257,7 @@ export async function analyzeAttention(screenshotPath, personaName, cellSize = 1
         cols: baseSaliency.cols,
         width: baseSaliency.width,
         height: baseSaliency.height,
-        cellSize: cellSize * 2, // Scaled
+        cellSize: cellSize / downscale, // Scaled
         hotspots,
     };
     // Compute attention metrics
@@ -1114,13 +1267,13 @@ export async function analyzeAttention(screenshotPath, personaName, cellSize = 1
     for (let i = 0; i < filtered.length; i++) {
         const p = filtered[i] / totalSaliency;
         if (p > 0)
-            entropy -= p * Math.log2(p + 1e-10);
+            entropy -= p * Math.log2(p + M.entropyEpsilon);
     }
     const maxEntropy = Math.log2(filtered.length);
     entropy = entropy / maxEntropy; // Normalize to 0-1
     // Concentration: fraction of total saliency in top 20% of cells
     const sorted = Array.from(filtered).sort((a, b) => b - a);
-    const top20Count = Math.ceil(filtered.length * 0.2);
+    const top20Count = Math.ceil(filtered.length * M.concentrationShare);
     const top20Sum = sorted.slice(0, top20Count).reduce((a, b) => a + b, 0);
     const concentration = top20Sum / totalSaliency;
     // Alignment: how well does the persona's saliency match the base (unfiltered) saliency?
@@ -1142,18 +1295,19 @@ export async function analyzeAttention(screenshotPath, personaName, cellSize = 1
         transportCost += Math.abs(baseSaliency.cells[i] / (totalSaliency || 1) - filtered[i] / (totalSaliency || 1));
     }
     // Attention competitors: group hotspots into regions
-    const attentionCompetitors = hotspots.slice(0, 5).map((h, i) => ({
+    const attentionCompetitors = hotspots.slice(0, M.competitors).map((h, i) => ({
         region: `(${h.x}, ${h.y})`,
-        saliency: Math.round(h.saliency * 100) / 100,
+        saliency: Math.round(h.saliency * M.r2) / M.r2,
     }));
     return {
         persona: personaName,
+        perceptualProfile: { resolution: resolvedProfile.resolution, name: resolvedProfile.resolvedProfile },
         saliencyMap,
-        alignmentScore: Math.round(alignmentScore * 1000) / 1000,
-        entropy: Math.round(entropy * 1000) / 1000,
-        concentration: Math.round(concentration * 1000) / 1000,
+        alignmentScore: Math.round(alignmentScore * M.r3) / M.r3,
+        entropy: Math.round(entropy * M.r3) / M.r3,
+        concentration: Math.round(concentration * M.r3) / M.r3,
         attentionCompetitors,
-        transportCost: Math.round(transportCost * 10000) / 10000,
+        transportCost: Math.round(transportCost * M.r4) / M.r4,
         computeTimeMs: Math.round(performance.now() - startTime),
     };
 }
@@ -1162,11 +1316,17 @@ export async function analyzeAttention(screenshotPath, personaName, cellSize = 1
  * Returns the L1 divergence between their normalized saliency maps (described
  * elsewhere as Wasserstein; it is not, see attentionDivergence) and the
  * connected regions where attention differs most.
+ *
+ * `accessibilityTraitsA` / `B` are each persona's own accessibility traits, so
+ * a persona without a built-in perceptual profile is filtered through one
+ * synthesized from them, as analyzeAttention does when given them (ISC-48).
  */
-export async function compareAttention(screenshotPath, personaA, personaB, cellSize = 16, domElements, goal) {
+export async function compareAttention(screenshotPath, personaA, personaB, cellSize = param("attention.gaze.cellSize"), domElements, goal, accessibilityTraitsA, accessibilityTraitsB) {
+    const r4 = param("attention.rounding.tenThousandths");
+    const regionCount = param("attention.report.divergentRegions");
     const [analysisA, analysisB] = await Promise.all([
-        analyzeAttention(screenshotPath, personaA, cellSize, undefined, domElements, goal),
-        analyzeAttention(screenshotPath, personaB, cellSize, undefined, domElements, goal),
+        analyzeAttention(screenshotPath, personaA, cellSize, undefined, domElements, goal, undefined, undefined, accessibilityTraitsA),
+        analyzeAttention(screenshotPath, personaB, cellSize, undefined, domElements, goal, undefined, undefined, accessibilityTraitsB),
     ]);
     // One pass computes the aggregate AND the regions from the same normalized
     // maps, so the regions are by construction the contributors to the number
@@ -1179,8 +1339,8 @@ export async function compareAttention(screenshotPath, personaA, personaB, cellS
     return {
         personaA: analysisA,
         personaB: analysisB,
-        attentionDivergence: Math.round(divergence * 10000) / 10000,
-        divergentRegions: regions.slice(0, 10),
+        attentionDivergence: Math.round(divergence * r4) / r4,
+        divergentRegions: regions.slice(0, regionCount),
     };
 }
 //# sourceMappingURL=attention-transport.js.map