Date: 2026-03-19 Previous: v2.5.0 release notes Blog post: Isotropic Cortical Sampling
Scrutinizer's default mode now derives its peripheral distortion profile from the FOVI cortical magnification model (Blauch, Alvarez & Konkle 2026). Seven rendering approaches failed before the eighth succeeded: sector geometry drives where and how fast degradation changes — not how pixels change.
The displacement pipeline is now parameterized by cortical sector extent derived from w = log(r + a). Noise frequency scales inversely with sector size; scramble cell size tracks sector extent (capped at 12px). The visual difference from the previous default is subtle — fewer implausible long-range pixel scatters, smoother degradation profile — but the derivation is now principled and traceable to Blauch's formulation.
- 19-test geometry suite validates math against Blauch's Python to 3 decimal places
- Interactive grid comparison (CodePen)
- Cortical manifold visualization
- Grid comparison
V1 displacement components extracted as parameterized GLSL structs (BenderConfig, CutterConfig). Researchers can swap implementations by constructing different configs. Type 1 (Shredder) refactored to delegate — behavior-identical.
DoG band reconstruction artifacts on smooth gradients ("the halo") eliminated via smooth-content detection. When pooledCol ≈ foveaCol, the pipeline snaps to the source color — no Mach bands to amplify. V4 transitions restored to pixel-space.
AI-confabulated paper titles, authors, and article numbers corrected across 8 files using the science agent — a specialized Claude Code subagent that validates citations against CrossRef and BibTeX, detects confabulated titles/authors/DOIs, and flags cross-file inconsistencies. Findings: 12% of BibTeX entries had issues, 1 complete fabrication, 5 partial confabulations. citation-guardian tool built to detect this class of error.
| File | Purpose |
|---|---|
scripts/validate-isotropic-rendering.js |
12-check rendering validation suite |
docs/specs/implemented/mode_graduation.md |
Process spec for promoting modes to default |
docs/drafts/blauch-update-2026-03-19.md |
Draft correspondence re: FOVI adoption |
| Suite | Result |
|---|---|
| Unit (274) | PASS |
| Visual regression | PASS |
| Memory | PASS |
| Performance | PASS (+0.7ms worst p95) |
| Integration | PASS |
| Smoke (7) | PASS |
| Golden (73) | PASS |
| Isotropic rendering (12) | PASS |
| Subband entropy | 4/5 (pre-existing) |
| OCR | Needs recalibration for mode 12 |
- Scrollbar protection bypass: Scrollbar detection now bypasses the full pipeline (not just the memory mask), preventing peripheral distortion from bleeding into native scrollbar rendering.
- Foveal radius menu: Removed separator elements from the foveal radius radio group for cleaner menu layout.
- ArXiv paper: Reframed as "Restricted Focus Viewer," 25% trim, 4 empirical claim corrections, citation audit fixes.
- Science agent (
scripts/science-agent/) added for automated citation validation. - Reading span (Rayner 1998) added to simulation limitations; limitations section reframed.
- Mongrel textures spec rewritten as architectural plan with Tier 3 synthesis path.
- ROADMAP condensed from narrative to table format.
- Default mode changed from 10 (Compute Mongrel) to 12 (FOVI Cortical Grid)
- FOVI blog post (2026-02-28) unpublished pending citation corrections