CHANGELOG.md

Changelog

[2.7.2] - 2026-04-06

Fixed

• Visual memory activating on launch despite Off setting: currentVisualMemory || 20 treated the saved value 0 (Off) as falsy, silently enabling Inhibition of Return mode (limit=20) on every launch. The menu correctly showed "Off" but the memory mask was active, preserving previously-fixated text in the far periphery. Fixed with explicit undefined check.

[2.7.1] - 2026-03-30

Added

• Scanpath replay: ScanpathPlayer (GazeModel drop-in) replays recorded eye movement sequences. COCO-Search18 importer, UEyes duration fix, CLI per-fixation frame capture.
• Gazeplot mode: Visual memory accumulation across fixation sequences with --gazeplot-diagram scanpath overlay on captures.
• Figma plugin v1.5.0: Model fidelity brought to parity with desktop v2.7 — 12-band DoG spatial frequency decomposition, per-band chromatic attenuation (castleCSF), far-periphery decay ramps, rod desaturation. 9 new shader uniforms with biological defaults.

Fixed

• Unified eccentricity master curve: 6 overlapping smoothsteps → 1 C2-continuous curve + power functions, eliminating Mach bands in parafoveal color transitions.
• Luma/chroma split foveal blend (Mullen 1991): Progressive chromatic decay on uniform surfaces, not a cliff.
• Visual memory mask: Suppresses V4 color effects in remembered regions; parafoveal radius (2.5× fovea) for recalled fixation footprint.
• Scanpath coordinate mapping: Correct DPR and image scaling for retina displays.

[2.7.0] - 2026-03-25

Added

• Tier 2.75 WebGPU Compute Pipeline (Mode 14, Default): Laplacian pyramid decomposition (4 bands + DC residual) with per-tile cross-scale magnitude correlation extraction and spectrum-matching synthesis. 15 dispatches per frame via webgpu-pyramid-compute.js. Three WGSL shaders: pyramid-decompose.wgsl (luminance, blur/downsample, band subtraction), pyramid-stats.wgsl (atomic accumulate + finalize → 18 floats/tile), pyramid-synth.wgsl (sine-grating noise + variance matching + bilinear interpolation + multi-band reconstruction). Produces spatially coherent peripheral texture instead of Tier 2.5's colored noise.
• Bilinear Tile Interpolation: Variance, magnitude, and cross-scale correlation bilinearly interpolated between 4 nearest tile centers before scaling noise. Eliminates crosshatch artifact at tile boundaries where adjacent tiles had different content characteristics.
• Isotropic 4-Orientation Noise: Sum of 4 rotated sine gratings (0/45/90/135 degrees) per band with per-pixel phase jitter. Replaces axis-aligned sin(x)*cos(y) which produced diagonal fringing. Noise variance recalibrated to 0.5 for the 4-grating sum.
• Eccentricity-Graded Content Replacement: Synthesis alpha scales with eccentricity — near-fovea (0.15) preserves structure, far periphery (0.8) replaces content. High-variance tiles get more noise disruption than low-variance tiles, producing the crowding mechanism.
• Gaze-Based Stable Seed: Noise seed derived from gaze position instead of pixel coordinates. Eliminates shimmer during fixation.
• Subband Tiling Visualization: 3x2 grid rendering of Laplacian pyramid bands (4 frequency bands + residual + original) with contrast-stretched bandpass for blog use.
• Python Reference Generator: scripts/generate-pyramid-reference.py generates ground-truth pyramid decomposition data via pyrtools for unit test validation.
• Pyramid Decomposition Tests: 30 unit tests validating decomposition against pyrtools reference data.
• Wave 7 Validation Scaffolding: Specs and scripts for 7a (pyramid fidelity), 7b (stats accuracy), 7c (crowding asymmetry). Crowding capture and analysis scripts for Tier 2.5 vs 2.75 comparison.
• Tier 3 TTM Synthesis Plan: Architecture spec (docs/specs/tier3_ttm_synthesis_plan.md) for content replacement within pooling regions.
• Compute Phase in FrameTimer: Timer mark for WebGPU compute dispatch duration.
• Acuity-Gated Saliency: Saliency protection decays with eccentricity (Strasburger et al. 2011, E2=8.0°).
• Eccentricity-Weighted Congestion: Two-scale Feature Congestion — foveal (σ=2.5) + peripheral (σ=5.0).
• Comfort Mode: Toggle (⌘⇧C) adds +1° to clear zone via shader distance offset. Microsaccade-maintained visibility envelope.
• Compute Texture Comparison: capture-compute-texture.js + compare-compute-textures.js for isolated pipeline validation. Wave 7.5: MAD=0.86 between Tier 2.5 (near-black) and Tier 2.75 (structured content).
• Tier 3 Lessons Learned: docs/specs/tier3_lessons_learned.md — failed pure-pooling attempt, fragment shader fixes identified.

Changed

• Default Mode: 12 (FOVI Cortical Grid) → 14 (Pyramid Mongrel). Multi-scale pyramid synthesis replaces single-scale oriented noise for peripheral rendering. Previous default accessible via menu.
• Mode-Switch Pipeline Recreation: Switching between Tier 2.5 and Tier 2.75 modes now destroys and recreates the WebGPU compute pipeline with the correct buffer layout. Fixes crash when switching between mode 10/14.
• Cross-Scale Correlation Strength: 0.5 → 0.8 for stronger parent-child band coupling.

Fixed

• Bilinear Stats Interpolation: match_stats shader now interpolates all tile statistics (variance, magnitude, correlation) between nearest tile centers. Eliminates crosshatch artifact at content boundaries.
• Diagonal Fringing: Axis-aligned noise replaced with 4-orientation isotropic sum. Diagonal artifacts at band edges eliminated.
• Eccentricity Blend: Synthesis was applying uniform noise strength regardless of distance from fovea. Now graded from 0.15 (foveal) to 0.8 (far peripheral).

[2.6.1] - 2026-03-22

Fixed

• Scrollbar protection bypass: Scrollbar detection now bypasses the full pipeline, not just the memory mask. Prevents peripheral distortion from bleeding into native scrollbar rendering.
• Foveal radius menu separators: Removed separator elements from foveal radius radio group for cleaner layout.

Changed

• ArXiv paper: Reframed as "Restricted Focus Viewer," 25% trim, 4 empirical claim corrections, citation audit fixes.
• Mongrel textures spec: Rewritten as architectural plan with Tier 3 TTM synthesis path.
• ROADMAP: Condensed from narrative to table format.

Added

• Science agent: scripts/science-agent/ for automated citation validation against source papers.
• Reading span (Rayner 1998): Added to simulation limitations documentation.

[2.6.0] - 2026-03-19

Blog post: Isotropic Cortical Sampling | Release notes

Added

• FOVI Cortical Grid (Mode 12, Default): Isotropic cortical sampling from Blauch, Alvarez & Konkle (2026, arxiv:2602.03766) is now the default mode. Sector geometry derived from w = log(r + a) parameterizes the displacement pipeline — noise frequency scales inversely with sector extent, scramble cell size tracks sector extent (capped at 16px, floor 8px). Sector geometry drives transition rate, not rendering mechanism. 19-test geometry suite validates math against Blauch's Python to 3 decimal places.
• Bender/Cutter Extraction: V1 displacement components extracted as parameterized GLSL structs (BenderConfig, CutterConfig) with applyBender() and applyCutter() functions. Researchers can swap implementations by constructing different configs. Type 1 (Shredder) refactored to use extracted functions — behavior-identical.
• Smooth Content Detection: DoG band reconstruction artifacts on gradients (the "halo") eliminated by detecting when pooledCol ≈ foveaCol and snapping pooledCol to foveaCol. On structured content (text, edges), the detection is a no-op.
• Isotropic Rendering Validation: scripts/validate-isotropic-rendering.js — 12-check suite covering angular isotropy, readability destruction, texture preservation, dark mode scatter, and mode comparison. 24 comparison captures across 4 pages × 3 modes × 2 fixations.
• Mode Graduation Spec: docs/specs/mode_graduation.md — process for promoting modes to default (test baselines, toggles, docs).
• ArXiv Figure: FOVI Geometry: Print-quality 2×3 composite figure (docs/arxiv-paper/figures/fig-fovi-geometry.html) showing Gaussian → Polar → Isotropic progression with pooled-content visualization. Self-contained HTML with PNG export.
• Interactive Grid Demo: Side-by-side visualization of uniform, polar, and isotropic cortical sampling grids (scrutinizer-www/demos/isotropic-grid-demo.html). Aspect-ratio coloring, pooled-content mode with selectable stimuli (checkerboard, text, radial, noise). Interactive controls for rings, spokes, field extent.

Changed

• Default Mode: 10 (Compute Mongrel) → 12 (FOVI Cortical Grid). The sector-parameterized displacement produces fewer implausible long-range pixel scatters with a smoother degradation profile. Previous default accessible via menu.
• V4 Halo Fix: Restored pixel-space transitions for blur blend, contrast ramp, bypass transition, and fovea protection. corticalStrength-based transitions were too wide — created visible Mach bands on smooth gradients. Widened magnocellular contrast ramp from 4.5px to 22.5px to match blur blend zone.
• Isotropic Cutter Cell Size: Floor raised from 4px to 8px, cap from 12px to 16px. At 4px the Cutter grid fragmented characters at the fovea, dropping OCR recognition to 56% (below the 70% threshold). At 8px, foveal recognition is 84% while far-peripheral degradation remains at 52% — passing all four OCR validation criteria.

Fixed

• Isotropic Mode Was Inert (Ship-Blocking): v1_distortion_type: 5 (cortical_isotropic) was silently overwritten to type 1 (shatter) every frame by a legacy mongrelMode override on peripheral.frag line 1806. The guard condition exempted types 2, 3, and 4 but not type 5 — so mode 12's sector geometry was computed but never reached the shader's isotropic code path. All captures, demos, and validation prior to this fix were running shatter distortion with isotropic uniforms ignored. Fixed by adding config.v1_distortion_type != 5 to the exemption list.
• Arxiv Citation Accuracy (Science Agent): Corrected FOVI paper title to "A biologically-inspired foveated interface for deep vision models" across 6 files using the science agent for CrossRef/BibTeX validation. Fixed author order (Alvarez before Konkle) where reversed. AI-confabulated title variants ("Foveation of inputs...", "Foveated vision in neural networks") replaced with actual paper title.
• Missing Uniform Declaration: u_num_cortical_rings was passed from JS but not declared in the shader.
• OCR Baseline Stale: Re-frozen OCR baseline at 1x DPR (1920×944) after mode 12 graduation changed capture geometry. Previous baseline was frozen at 2x DPR (3840×1888), causing foveal radius mismatch.

Validation

• OCR Profile (Mode 12): Fovea 84% | Parafovea 60% | Near 62% | Far 52%. Monotonically declining. Overall 57% recognition (mode 0 baseline: 40%). Far-peripheral degradation passes ≤55% threshold. Foveal preservation passes ≥70% threshold.
• OCR Profile (Mode 0, reference): Fovea 100% | Parafovea 67% | Near 44% | Far 31%. Overall 40%.

[2.5.0] - 2026-03-16

Added

• 12-Band DoG Decomposition: Extended from 8 half-octave bands (LOD 0.0–4.0) to 12 bands (LOD 0.0–6.0), covering 5.66–0.088 cpd. The four new low-frequency bands capture panel backgrounds, hero images, and page-level color regions — scales where peripheral color perception is strongest.
• Swatch-Aware Chromatic Preservation: mip[12] Oklab chrominance magnitude distinguishes large uniform color regions from mixed/text content. Swatches retain up to 30% more color at the same eccentricity. Boost applies to frequency bands only (not DC residual) to prevent color halos.
• Color Search Validation (Wave 1): Three-tier validation suite against published psychophysics. 24px dots at 5 eccentricity rings (2.5°–12.4°). Tier 1: 9/9 (monotonic decrease, BY ≥ 1.5× RG). Tier 2: 2/3 (Bowers ratio pass, green tracking pass, rendered-vs-model 5/20). Tier 3: 3/3 (Hansen correlation r=1.000, BY > RG every ring).
• Smoke Test Pipeline: npm run capture-smoke — 6-shot sanity check across 3 Electron batches. Covers basic render, mode switch, saliency debug, scroll, off-center fixation.
• Capture Infrastructure: Manifest-based skip logic (capture-manifest.js), batch orchestrator (capture-runner.js). Golden captures group shots by URL and reuse Electron instances.
• New Test Suites: mip-fidelity.test.js (DoG vs rect sampling), isotropic-sectors.test.js (cortical grid geometry), stimulus-domain.test.js (spectral mismatch, gamut, crowding). 258 tests across 11 suites (was 138).
• Reference Pages: color-spectrum-v2.html (improved band layout), chroma-uniform.html (uniform chromaticity stimulus), grid-comparison.html (MIP vs cortical grid), ocr-text-grid.html (readability testing).
• Mode 12 Infrastructure: Isotropic cortical sampling mode scaffolding (grid geometry not yet active in rendering).

Changed

• RG Decay Calibration: rg_decay 0.072 → 0.085. BY/RG sensitivity ratio at ring 5 (12.4°) now 2.59 vs Bowers et al. (2025) 2.72 at 15° (5% off, was 21%). Single-exponential cannot capture biphasic RG decay at far eccentricities — documented limitation.
• Base Desaturation Ramp: lgn_ramp_end_mult 2.0 → 6.0 for mode 0. Extends smoothstep from 1°–2° to 1°–6°, eliminating color cliff at parafoveal boundary. When per-band chromatic pooling is active, base ramp reduced to 40% strength to prevent stacking.
• Base Desat Stacking Fix: Per-band attenuation and base desaturation were multiplying, over-desaturating the periphery. Base ramp now scales to 40% when per-band is active.

Fixed

• Analyzer Sample Offset: analyze-color-search.js was sampling at RINGS[r] + halfBand (outer edge of band) instead of RINGS[r] (band center). At outer rings the sample landed in the gap between bands, causing spurious monotonicity failures.
• Foveal Boundary LOD: textureGrad callers now use V1-distorted UV derivatives (dFdx(distortedUV)), fixing over-blur radially and under-blur tangentially at the foveal boundary.
• Validation Eccentricity Label: Bowers comparison now correctly labeled as ring 5 (12.44°), not "~15°".

[2.4.0] - 2026-03-13

Added

• Reading Span (Rayner 1998): Asymmetric foveal envelope shifts protection zone in the reading direction during horizontal pursuit over text content. Three gates (speed, horizontality, text structure) prevent false activation. Fovea center shifts by up to radius × 0.7. Directional velocity tracking (vx/vy) in GazeModel. Menu toggle: Vision Model → Reading Span (Rayner). Per-mode config: ON for presentation/default, OFF for research modes.
• Wave 6: COCO-Periph Validation Scaffolding: System-level peripheral encoding validation against Harrington et al. 2024. Download, capture, analyze, validate scripts. Published psychophysics data. Validation adversary agent spec. npm run wave6 runs the full pipeline.
• Citation Export Fields: foveaDeg and pxPerDeg metadata in PNG captures for angular calibration documentation.
• Reading Span Capture Script: 4-scenario trajectory capture (scripts/capture-reading-span.js) with mid-sweep screenshot timing. Tests static, LTR, RTL, and disabled conditions.
• Isotropic Cortical Sampling Spec: Design document for isotropic log-polar sampling.

Changed

• Fovea Degree Correction: fovea_deg 2.0→1.0 (1° radius = 2° diameter). Default fovealRadius 90→45px. Settings migration via _foveaDegMigrated flag halves existing user radius on first launch. All shader px_per_deg derivations updated.
• Saccadic Blindness Default: Now ON by default (was OFF).
• Menu Labels: Foveal radius options now show px/° units. "Double Vision" renamed to "Drunken Reading".

Fixed

• Mid-sweep capture timing: Trajectory-based test captures now screenshot during active motion, not after velocity decay.

[2.3.0] - 2026-03-11

Added

• WebGPU Compute Mongrel Pipeline (Tier 2.5): Real-time metamer texture synthesis via two-pass WebGPU compute. Pass 1: tile-based Oklab statistics extraction (luminance, chrominance variance). Pass 2: oriented noise synthesis from extracted moments. ~900 lines of pipeline code, under 0.3ms on integrated GPU. Graceful fallback to Tier 1 when WebGPU is unavailable.
• Ratio Reconstruction: Dual-LOD structure map sampling in processLGN(). Sharp LOD 0 for type/density, blurred LOD 4 for rhythm via R/G ratio (stable across content boundaries). Soft density gate feeds blurred density into the crowding sigmoid so V1 strength tapers smoothly at DOM edges instead of cliff-edging. Structure map texture upgraded to LINEAR_MIPMAP_LINEAR with generateMipmap() on upload.
• Fovea Radius Granularity: Three new options around the 90px sweet spot — Tight (70px), Relaxed (110px), Wide (130px) — with separator-grouped menu layout.
• Oklab Variance Metrics: Luminance and chrominance variance computation in compare-crowding-modes.js for quantitative mongrel quality comparison.
• Seeded Flanker RNG: Deterministic mulberry32 PRNG for crowding captures — identical flanker placement across runs for reproducible A/B comparison.
• Auto-Fallback Safety Harness: 60-frame rolling window frame budget monitor with 30fps floor. Automatically disables WebGPU compute path and falls back to fragment shader pipeline if sustained frame drops detected.

Files

• renderer/webgpu-crowding-compute.js — pipeline manager (device, buffers, bind groups, dispatch)
• renderer/shaders/crowding-stats.wgsl — pass 1: tile statistics extraction
• renderer/shaders/crowding-synth.wgsl — pass 2: oriented noise synthesis
• renderer/webgpu-probe.js — WebGPU capability detection
• renderer/webgpu-safety.js — frame budget monitor and auto-fallback

[2.2.0] - 2026-03-09

Added

• Oriented DoG Bands (Phase 1-3): Orientation-selective band attenuation in peripheral.frag. Phase 1: oblique effect — cardinal edges get M-scaling cutoffs pushed ~50% further (Appelle 1972). Phase 2: 4-channel V1 simple cell energy decomposition (H/V/D45/D135) replacing cos(2θ) (Hubel & Wiesel 1962). Phase 3: radial-tangential anisotropy — tangential edges +30%, radial -15% (Toet & Levi 1992). 3 uniforms: u_dog_oriented, u_dog_orient_bias, u_dog_radial_bias. 4-tap MIP-1 gradient with BGRA-corrected luminance, gradient magnitude gate. Enabled on modes 0 and 1.
• Orientation Diagnostics: Debug level 4 (4-channel energy: R=H, G=V, B=diagonal) and debug level 5 (band weights + orientation tint with fovea blend). Menu: Simulation → Utility → Orientation Diagnostics.
• Keyboard Shortcuts: Direct keyboard access to visualization modes via Simulation menu accelerators.
• Oriented DoG Capture Script: A/B capture comparing oriented vs isotropic DoG output across reference pages. TEST_DOG_ORIENTED and TEST_DOG_ORIENT_BIAS env vars for parametric testing.
• Validation Report Format: Claim/Basis/Result structure with pass/partial/fail badge pills, replacing Published/Validation/Result.
• Validation Regression Tests: 23 Jest tests encoding psychophysical validation findings (Waves 1–3: chromatic decay, spatial frequency ordering, density-gated crowding) as automated guards against modes.json parameter regressions. Loads parameters from source of truth and validates against Bowers 2025 published data.

Changed

• Mongrel textures spec: Updated for v2.1 with timestamps added to all specs.
• Roadmap: Spec index table with status triage added. Linguistic priming spec refreshed to v3, roadmap condensed.

Fixed

• Debug 5 fovea circle: White fovea circle in orientation diagnostics mode corrected.
• Toolbar clipping: Fixed toolbar clipping in visualization mode keyboard shortcut additions.
• Scroll-to-top in captures: Capture scripts now scroll to page top before capture, eliminating scroll-position variation.
• Arxiv table overflow: Fixed table overflow in two-column layout.

Docs

• README: DOM-aware rendering rationale, LGN/V1/V4 pipeline table with deep links, structure map as DOM analysis, validation case study, psychophysical validation section with published data table.
• Arxiv paper: Gaussian comparison conditions added to saliency and color-search analysis. Communications review fixes.
• Release notes (v2.1): Undefined technical terms defined inline. Blog post links, published data links, grad student project cross-links added.

[2.1.0] - 2026-03-08

Added

• 8 Half-Octave DoG Bands: peripheral.frag upgraded from 4 octave-spaced to 8 half-octave bands. 9 MIP levels (LOD 0.0–4.0), √2 frequency spacing, smoother blur gradient. Half-integer LODs use hardware trilinear interpolation natively.
• Five-Wave Psychophysical Validation: Automated test suite against published data — chromatic decay (Hansen 2009), spatial frequency (Rovamo 1979), crowding geometry (Bouma 1970), saliency protection (Itti & Koch 2001), mixed-density UI (Halverson & Hornof 2011). 5 capture scripts, 5 analysis scripts, 25+ golden captures.
• 15 Experimental Stimulus Pages: Open-source HTML psychophysical stimuli (color-search, spatial-acuity, crowding variants, saliency-popout, face-test, halverson-mixed-density). Menu: Go → Reference Pages → Experimental Stimulus.
• Capture Infrastructure: TEST_LOAD_TIMEOUT for heavy external pages, capture-appendix-baselines.js, appendix figures (fig-a1, fig-a2).
• Band Weight Analysis: scripts/analyze-dog-bands.js — pure math analysis of DoG band weights across eccentricity for both linear and CMF paths.

Fixed

• Toolbar clipping: getSize() → getContentSize() for child view bounds. Title bar height was eating into toolbar, cropping URL bar and eye toggle.
• Polar sector R:T ratio: peripheral.frag spoke count computed from biased ring width, cancelling 2:1 radial elongation. Fixed to use unbiased width.
• V1 far-peripheral growth: Growth factor tuned 0.5→1.5 via capture-analyze loop against crowding reference pages.
• Composite Rovamo correlation: Replaced per-band Spearman with frequency-weighted composite metric (r=0.600).

Docs

• Arxiv paper: Walton 2021 contradiction fixed, WebGPU tiered roadmap, band equations updated to half-octave.
• Developer guide, primer, foveated vision model, oriented DoG spec all updated for 8-band architecture.

[2.0.0] - 2026-03-07

Added

• Minecraft Mode (Mode 4): CMF-driven block quantization (4-64px) with per-channel Oklab neighbor averaging. Fovea-relative grid, RG boundaries merge while YV stays sharp.
• Minecraft Eyeball (Mode 8): Polar variant — wedge-shaped sectors sized by CMF, radially elongated ~2:1 approximating TTM-predicted pooling geometry. Same Oklab chromatic decay.
• Blueprint Mode (Mode 3): ARIA-typed wireframe from live DOM. Structure map alpha encodes role IDs (0–12), shader renders color-coded bounding boxes. Fovea shows original content, periphery transitions to blueprint wireframe.
• Density-Gated Crowding: Sigmoid gate on structure map density modulates V1 distortion strength. Dense content gets full crowding; isolated elements spared (0.3× floor). Threshold=0.6, steepness=20.0.
• Eccentricity Scaling: u_ecc_scaling uniform (default 0.75, Brown et al. 2023) modulates pooling growth rate across all CMF-enabled modes.
• Toggle Effects On/Off: Simulation → Behavior → Toggle Effects (Cmd+E). Provides menu access to the toolbar eye button toggle.

Changed

• Chromatic decay recalibrated: RG decay 0.059→0.072, YV decay 0.004→0.014 (Bowers et al. 2025 suprathreshold measurements).
• Mode count: 9→10 modes (4 research, 4 presentation, 1 experimental).

Fixed

• Toolbar URL overflow: Long URLs now truncate with ellipsis instead of pushing the eye toggle off-screen.
• Toolbar icon alignment: Eye toggle and version label vertically centered in 40px toolbar.

Docs

• Arxiv paper synced with shipped features; new Open Questions section.
• 6 stale doc status lines updated (Planned→Shipped).

[1.9.0] - 2026-03-04

Added

• Per-Channel Chromatic Pooling (castleCSF): Replaces uniform chrominance reduction with biologically accurate per-channel decay in DoG band reconstruction. RG (red-green) attenuates ~2.5× faster than achromatic; YV (blue-yellow) persists into far periphery. Both channels now have per-band frequency-dependent attenuation — large colored regions preserve hue further than small ones for both RG (k_ef=0.003) and YV (k_ef=0.008). Suprathreshold correction (u_supra_exponent = 0.5) converts castleCSF detection thresholds to appearance (Jiang, Shooner & Mullen 2022). 6 uniforms. Enabled on modes 0, 1, 9.
• Chromatic pooling on Mode 9 (Congestion-Gated Pooling): Mode 9 now inherits full per-channel RG/YV chromatic pooling from Mode 0.
• Saliency vs Congestion Split View: Side-by-side heatmap rendering — saliency (indigo-white, "What pops out?") and congestion (blue-yellow-red, "How cluttered?") with labeled divider. Menu: Simulation → Utility → Congestion Report.
• scrutinizer-audit CLI + MCP Server: Headless analysis pipeline (Playwright) reusing congestion-core.js scoring. CLI: positional URLs, sitemap parsing, desktop+mobile viewports, JSON/HTML output, CI fail-above gate. MCP: analyze_url, analyze_urls, compare_pages tools.
• Crowding Diagnostics: Two reference pages (crowding.html, crowding-stimulus.html), simulation limitations doc, density-gated crowding spec.
• Saccadic Blindness: Foveal/parafoveal regions shrink proportionally to mouse velocity (smoothstep(4, 10, velocity)), simulating saccadic suppression. Menu toggle, off by default. Both shaders (peripheral.frag, peripheral.frag).
• Golden captures: Chromatic pooling on/off variants for color-spectrum (modes 0, 1) and dashboard pages.
• CMF MIP mapping unit tests: Validate logarithmic CMF→MIP level conversion.

Fixed

• Chromatic eccentricity decoupling: Per-band RG/YV decay now uses true gaze eccentricity (visual_ecc) instead of V1 distortion-coupled eccentricity. A pixel at 15° was previously seen as ~0.6° by the chromatic path — castleCSF decay was near-zero. Spatial band weights (w0-w3) still use coupledEccentricity.
• Suprathreshold chromatic correction: castleCSF k_e values are detection thresholds, not appearance. Added power-law compression (exponent 0.5) — reds at 10° retain ~51% instead of ~26%.
• Red Kill Switch gating: Guarded with u_chromatic_pooling < 0.5 || u_dog_enabled < 0.5 — per-band RG decay at correct eccentricity handles red suppression. Base desaturation always runs (complementary, not alternative).
• Chromatic pooling override: _chromaticPoolingOverride flag pattern prevents updateConfigFromMode() from clobbering manual toggle on every frame.
• CMF→MIP mapping: Corrected to ln(r+a) - ln(a) with cortical_max normalization (commit 5be3b2b).

Removed

• Mode 8 (Gaussian Desaturation): Removed. castleCSF per-channel pooling (v1.9.0) supersedes the Gaussian vs smoothstep desaturation comparison. The mode was also broken — missing chromatic_pooling in config caused triple desaturation stacking on saturated content. The Gaussian functional form (exp(-r/σ)) is wrong for RG decay: Bowers, Gegenfurtner & Goettker (2025) showed RG attenuation is biphasic (steep to ~15°, then shallower), not pure exponential. Gaussian color decay remains in Mode 6 (CMF standalone) where it serves a different purpose.

Changed

• Mode 4: Cyberpunk → Minecraft (Block Pooling): Blocks sized to CMF MIP level (4-64px), fovea-relative grid, channel-independent neighbor color averaging in Oklab. Demonstrates customizing the baseline simulation — same CMF math, visible as block geometry.
• Rename fovi_* → cmf_*: Config keys, shader uniforms, and JS variables renamed from fovi_ prefix to cmf_. "FOVI" is the name of a specific model; the feature is cortical magnification function (CMF) — not FOVI-specific.
• CMF log path default for Modes 0 and 1: Research modes now use logarithmic MIP scaling via CMF instead of legacy linear normalizedEcc * 2.5. Mode 7 retains cmf_enabled: false as frozen comparison baseline.
• Doc terminology scrub: Replaced "desaturation" with "chromatic pooling" / "chrominance reduction" across 8 docs where describing biology. Peripheral color is pooled (mean chromaticity preserved), not lost.
• Hansen et al. 2009 citation correction: Paper measures detection thresholds only, not suprathreshold appearance. Separated from Jiang et al. 2022 citation. PDF archived in docs/research/.

[1.8.0] - 2026-03-03

Added

• Feature Congestion Pipeline (Rosenholtz 2007): Simplified Oklab-space Feature Congestion — local variance across L, |a|, |b| channels with fixed σ=2.5. Dual-worker architecture: saliency (256px real-time) + congestion (1024px on-demand). Scoring: sqrt(congestion_p90 × 0.7 + edgeDensity_p90 × 0.3) × 100. Validated against Python reference (Spearman ρ = 0.93 at 768px).
• ComplexityHUD: Interactive draggable overlay with Score / Stats / Spatial tabs. Congestion heatmap on TEXTURE4 (blue → yellow → red). Scroll/navigation-aware — hides on scroll, restores when fresh results arrive.
• Mode 9: Congestion-Gated Pooling: High-congestion regions get up to 2× MIP pooling boost via coupledEccentricity × (1 + congestion). Auto-starts high-res congestion worker on mode selection; recomputes on scroll/navigation. Tests Rosenholtz (2012) prediction that clutter and crowding share the same summary-statistic computation. Tagged experimental category.
• Validation pipeline: Three-script architecture (validate-congestion.py, extract-congestion.js, compare-congestion.js) for Spearman rank correlation against Rosenholtz reference.

Fixed

• Parafoveal blur band: Replaced flat eccentricityScale = 0.15 across the entire parafovea with a smoothstep ramp from 0.0 (inner) to 0.15 (outer boundary). Inner parafovea stays sharp — word-length cues survive per Rayner (1998). Widened fovea-to-pooled MIP blend from fovea_radius × 0.1 to × 0.5 to eliminate the abrupt soft step.

Changed

• Linear M-scaling: Replaced geometric 2x band cutoffs (0.3, 0.6, 1.2, 2.4 × E2) with Rovamo & Virsu (1979) formula. New cutoffs: E2 × (2^k − 1) giving 1, 3, 7, 15 × E2. Coarse structure persists further into periphery.
• E2 recalibration: High-Key 0.5→0.15, Biological 0.4→0.12 to preserve band-0 onset under linear cutoffs.
• Approximate Laplacian pyramid: Qualified "Laplacian pyramid" terminology across all docs, arxiv paper, and shader comments. Hardware MIP uses box/bilinear, not Gaussian (Burt & Adelson 1983). Added citation to references.bib.
• Output clamping: Final color clamped to [0,1] to prevent negative-going band artifacts.
• Golden captures: Removed 348MB bulk captures from tracking, gitignored. Curated mode-comparison captures in docs/golden/.
• .claude config: Checked in settings, skills, and agent memory.

[1.7.1] - 2026-02-28

Changed

• Simulation Menu Reorganization: Restructured the Simulation menu to foreground behavioral simulation over rendering presentation.
  • Behavior submenu is now first — contains Visual Memory, Enable Structure Map, Enable Saliency Modulation (the cognitive processes being modeled).
  • Experimental simulation models (FOVI, Legacy v1.6, Gaussian Desaturation) moved from Utility to Behavior — these are alternative pipelines for testing, not rendering styles.
  • Utility submenu replaces the old "Aesthetic Mode" nesting — contains the 6 rendering/presentation modes plus Show Structure Map / Show Saliency Map debug views.
  • Content Analysis submenu removed — its items redistributed to Behavior (Enable toggles) and Utility (Show toggles).
  • Foveal, Peripheral, and Visual Overlay submenus unchanged.
  • Aesthetic mode radio buttons now sync across Behavior and Utility submenus via menu rebuild on selection.
  • No shader or modes.json changes — purely a menu organization change.

[1.7.0] - 2026-02-28

Added

• FOVI Cortical Magnification Function: Added FOVI's CMF (Blauch, Alvarez & Konkle, 2026, arxiv:2602.03766) as a switchable MIP-sampling mode. This implements one component of FOVI — the spatial resolution falloff — not the full kNN sensor manifold or neural network pipeline.
  • New uniforms: u_cmf_enabled, u_cmf_a, u_cmf_color_sigma.
  • When enabled: logarithmic MIP scaling via CMF (log2((r_deg + a) / a)) and CMF-derived DoG band cutoffs.
  • Modes 0 & 1 retain legacy pipeline (cmf_enabled: false) — FOVI's CMF models cortical magnification only, not the full peripheral detail loss (crowding, RF growth, contrast sensitivity). Existing parameters already approximate the combined perceptual effect.
• Mode 6: FOVI (Cortical Magnification): FOVI's CMF as a standalone MIP-sampling mode, orthogonal to Scrutinizer pipeline.
  • Single CMF-driven MIP blur (no DoG band decomposition).
  • Gaussian color decay (V4 style 6) is Scrutinizer's own addition for perceptual comparison — not from FOVI, which is purely spatial.
  • No LGN gating or V1 distortion — clean comparison baseline.
• Mode 7: Legacy v1.6 (Comparison): Frozen v1.6 pipeline snapshot for A/B dogfooding.
  • Linear MIP scaling, hand-tuned DoG cutoffs, cmf_enabled: false.
  • Identical to pre-v1.7 mode 0 behavior.
• Mode 8: Gaussian Desaturation (Experimental): Controlled experiment isolating desaturation curve shape.
  • Identical pipeline to mode 0 (same DoG, LGN gating, V1 distortion, MIP).
  • Replaces smoothstep ramp with Gaussian exponential decay: 1 - exp(-r_deg / sigma).
  • Single variable changed for perceptual matching study: which curve better models the rod-cone transition?
  • Sigma (4.0 default) maps to effective cone density falloff distance in degrees.

Changed

• modes.json metadata version bumped to 2.0.0 (new cmf_* pipeline params).

[1.6.0] - 2026-02-28

Changed

• De-Monolith Refactor: Extracted three domain modules from the 969-line scrutinizer.js monolith:
  • gaze-model.js (166 lines) — oculomotor system proxy (velocity tracking, fixation detection, saccadic suppression)
  • visual-memory.js (254 lines) — visuospatial working memory (fixation buffer, mask rendering, decay)
  • content-analysis.js (356 lines) — pre-cortical feature extraction (structure map scanning, saliency, DOM observation)
  • scrutinizer.js is now a thin Pipeline Orchestrator (535 lines) with backward-compatible property proxies.
• Unit Tests: Added 138 unit tests for pure-function modules (oklab-utils: 73, gestalt-processor: 41, color-saliency-map: 24).

Added

• DoG Peripheral Reconstruction: New biologically-inspired peripheral rendering mode replacing simple MIP pooling with Difference-of-Gaussians band decomposition.
  • Decomposes hardware MIP chain into 4 approximate Laplacian pyramid bands (box/bilinear, not true Gaussian) with M-scaling rolloff per band.
  • Preserves low-frequency structure (layout, buttons, large text) while filtering high-frequency detail (serifs, fine textures).
  • Gated by dog_enabled uniform — legacy MIP pooling preserved when disabled.
  • New uniforms: u_dog_enabled, u_dog_e2 (M-scaling half-resolution eccentricity), u_dog_sharpness (band rolloff sharpness).
  • Enabled by default in High-Key and Biological modes; other modes unchanged.
  • Near-zero additional cost: reuses existing hardware MIP chain from gl.generateMipmap().

[1.5.0] - 2026-01-30

Added

• Mobile Emulation: New "Mobile Emulation" submenu in View menu.
  • Simulates iPhone viewport (390x844), scale factor (3.0), and User Agent.
  • Automatically resizes and locks window to phone dimensions.
  • Restores previous window size and desktop mode when disabled.
• Touch Simulation (Alpha): Added support for synthesizing touch events.
  • Hold Option (Alt) + Click while in Mobile Emulation mode to trigger touchStart sequence instead of mouse events.
  • Note: This is an experimental feature to unblock testing of touch-only interactions.
• Responsive Toolbar: Redesigned toolbar URL input for better usability on narrow (mobile) screens.
  • Replaced inline text input with a clickable trigger button.
  • Added dedicated URL entry dialog window.

Changed

• Window Management: Adjusted window bounds saving logic to ignore mobile emulation resizing, preserving user's desktop window preference.
• Toolbar: Updated toolbar layout to prevent overflow artifacts in small windows.