FlashGuard

Experimental low-latency photosensitivity risk reduction for Windows

A D3D11 overlay that captures the desktop, detects potentially hazardous flashing, and limits displayed temporal modulation while trying to preserve ordinary motion.

Warning

Not medically validated, not clinically epilepsy-safe, not Harding FPA/PSE certified. Passing FlashBench is an engineering regression result — not a medical guarantee, not formal accessibility certification, and not a promise of seizure prevention for any person or any stimulus. This is experimental software.

---

Status

Version	0.4.0-alpha.4
Active development	Surface-frequency path — current-frame-only, no NVOFA, no displayed-image feedback
Release-build default	Legacy NVOFA path (--surface-frequency opts in)
Surface validation	260 / 260 checks pass on an RTX 3060
Unresolved	The one confirmed real-world artifact — see live diagnosis

As of 0.4.0-alpha.1 the surface-frequency path passes its complete validation for the first time, including the moving white/red case at 5 Hz and 144 captured FPS that all four preceding candidates failed. It reaches 74.17% against the unchanged 70% requirement — the threshold was never relaxed to get there.

That is deterministic engineering evidence on one GPU. It does not establish that the live speckle artifact recorded in the Outlast diagnosis is gone; that has not been revalidated since.

---

Processing pipeline

flowchart TD
    A[DXGI Desktop Duplication] --> B[Freshest captured desktop frame]
    B --> S{Active path}

    S -->|surface| F[Half-resolution GPU features<br/>linear RGB + ternary Census]
    F --> M[Bounded structural matcher<br/>no NVOFA]
    M --> P[Per-cell transported state<br/>phase, period, extrema, lifetime]
    P --> C1[Full-resolution compositor<br/>current image only]

    S -->|legacy| G[128x72 linear-light analysis]
    G --> H[global / local / red / pattern<br/>translation classification]
    H --> I[Optional NVIDIA Optical Flow<br/>classification evidence only]
    I --> C2[Temporal safety shader<br/>PreviousSource + PreviousOutput]

    C1 --> O[Capture-excluded<br/>click-through overlay]
    C2 --> O

    style C1 fill:#1f6f3f,color:#fff
    style C2 fill:#334155,color:#fff
    style O fill:#0f766e,color:#fff

Loading

The two paths coexist in one executable and share capture and presentation. The surface path composites from the current image only; the legacy path constrains change against filtered history. Neither ever warps the displayed image with flow.

Why the legacy path keeps two separate histories

History	Contents	Used for
PreviousSource	Previous raw desktop frame	Deciding whether a changed pixel is explained by motion
PreviousOutput	Previous filtered frame shown to the user	Constraining displayed change during an active hazard

This separation is deliberate. Motion matching must not compare against an already filtered image, and optical flow must never warp displayed history. Earlier failures drove this: coarse masks produced visible shapes, unrestricted RGB history caused trails, and flow-warped history produced rubber-sheet geometry deformation.

---

Quick start

Requires Visual Studio 2022 with Desktop development with C++.

# Legacy-default release build
.\scripts\build.bat release

# Or a build that defaults to the surface-frequency path
.\scripts\build.ps1 -Mode surface

.\FlashGuard.exe                          # monitor under the mouse pointer
.\FlashGuard.exe --title "window title"   # monitor containing a window
.\FlashGuard.exe --validate-shaders       # compile every embedded HLSL entry point

Build mode	Flags	Default path
fast / dev	/Od /Ob0	Legacy
release	/O2	Legacy (--surface-frequency opts in)
surface	/O2 -D FLASHGUARD_SURFACE_DEFAULT=1	Surface (--legacy opts out)

Controls

Key	Action
F8	Toggle the persistent manual neutral shield
F9	Toggle diagnostics
F10	Open / close the settings menu (Esc also closes)
Ctrl+Shift+F12	Exit

Settings and hotkeys persist in %LOCALAPPDATA%\OutlastFlashGuard\settings.ini.

The settings menu is a compact 440×360 dark window with Home, Shortcuts, and Advanced tabs. Menu opacity is adjustable from 60% to 100% (default 92%); protection output opacity is unaffected. JetBrains Mono is embedded and loaded privately — no font installation required. Menu and loading banner are excluded from capture.

Compiled shaders are cached in %LOCALAPPDATA%\OutlastFlashGuard\shader-cache, keyed by source, entry point, target, flags, and compiler generation. In the recorded check an 11-shader cold compile of 143.420 s became a 0.324 s cached load with zero recompiles. That measures GPU and shader setup only, not full capture startup.

---

Validation

The surface path is checked by flashbench/surface-frequency.ps1 under SURFACE_FREQUENCY_VALIDATION/1 and its HSV extension.

Check	Result
Total checks	260 / 260 pass
HSV cases (5/10/20/30 Hz × 60/120/144 FPS)	60 / 60 pass
Min HSV RGB variation reduction — settled	99.24%
Min HSV RGB variation reduction — early	99.38%
Moving white/red, 5 Hz @ 144 FPS	74.17% (requirement 70%)
1080p GPU draw time	p50 0.831 ms · p99 8.575 ms (uncontrolled load)

Note

GPU timings cover features, tracking, composite and total passes. Capture, copy, queueing, presentation and physical display delay are excluded. These are not end-to-end latency measurements, and no test ran under real game load.

How it got here

gitGraph
    commit id: "0.3.0-a1 architecture"
    commit id: "0.3.0-a2 noise/color" type: REVERSE
    commit id: "0.3.0-a3 local noise" type: REVERSE
    commit id: "diagnosis" type: HIGHLIGHT
    commit id: "0.3.0-a4 amplitude" type: REVERSE
    commit id: "0.4.0-a1 HSV 260/260"

Loading

Version	Correction	Result
0.3.0-alpha.1	First current-frame-only detector and compositor	119/119 focused
0.3.0-alpha.2	Reject one-code dither; stop history inventing events	❌ one case
0.3.0-alpha.3	Require local source change — 33.1964 → 0.00005 codes	❌ one case
—	Live diagnosis: artifact is detector-introduced	⚠️ unresolved
0.3.0-alpha.4	Bound correction by measured source amplitude	❌ one case
0.4.0-alpha.1	Per-channel phase floor replaces grayscale projection	✅ 260/260

Four consecutive candidates reported failure on the same case while each fixed a real defect. The cause turned out to be a genuine bug — the grayscale projection was amplifying saturation-only flashing, failing all 12 saturation cases by up to +53% RGB variation — not a threshold artifact.

---

5–30 Hz flash sweep

24 two-second cases at 60 FPS across full-screen luminance, full-screen saturated red, and quarter-screen luminance stimuli.

Frequencies: 5 · 7.5 · 10 · 12 · 15 · 20 · 25 · 30 Hz

The gate requires source stimuli above 3 flashes/s to be reduced to at most 3 counted output flashes/s. On the self-hosted RTX 3060 run for commit 1802a4e6, all 24 cases produced 0.000 counted output general flashes/s and 0.000 counted red flashes/s.

Frequency	Full-screen luminance	Quarter-screen luminance
5 Hz	77.88%	69.07%
10 Hz	91.86%	90.25%
15 Hz	96.60%	93.01%
20 Hz	98.63%	—
30 Hz	98.63%	95.93%

The red-flash criterion is based on the red-flash transition counter, not on requiring high luminance-modulation reduction — saturated red can be made safer by chromatic mitigation even when screen-mean luminance changes less. This is standards-oriented regression testing, not WCAG/Harding certification; the quarter-screen case uses a simple screen-area stimulus, not a calibrated steradian laboratory measurement.

Informed by WCAG 2.2 Three Flashes or Below Threshold and ITU-R BT.1702.

---

FlashBench

powershell -ExecutionPolicy Bypass -File .\flashbench\run.ps1 `
    -Mode gpu-smoke -OutputDir .\flashbench\manual-results

Performs a release build, HLSL validation, real D3D11/NVOFA execution, deterministic synthetic replay through the same safety path, motion/ghosting and camera-pan regressions, and the 5–30 Hz sweep.

Report	Contents
summary.json	Overall run result
nvof-smoke.json	Real NVOFA execution evidence
synthetic-replay.json	Deterministic replay metrics
flash-sweep.json	Per-case sweep gates
flashbench.log	Full run log

Visual replay — inspect the synthetic cases yourself

powershell -ExecutionPolicy Bypass -File .\flashbench\run.ps1 `
    -Mode gpu-smoke -OutputDir .\flashbench\manual-results -VisualReplay

Start-Process .\flashbench\manual-results\visual\index.html

The viewer shows sampled replay frames as SOURCE | FILTERED | 6× AMPLIFIED DIFFERENCE, covering the 15 Hz flash, straight and oblique bright motion, small-object motion, and camera pan.

Legacy detector defaults — engineering values, not medical thresholds

Parameter	Value		Parameter	Value
lookaheadMs	0		flashEnergyThreshold	0.030
localDeltaThreshold	0.10		smallFlashAreaThreshold	0.008
globalDeltaThreshold	0.16		smallFlashDeltaThreshold	0.25
affectedAreaThreshold	0.18		smallFlashCoherenceThreshold	0.85
strongAffectedArea	0.30		spillExpansionCells	4
globalAreaThreshold	0.90		localGlobalSupportThreshold	0.035
coherenceThreshold	0.70		safeRiseRate	1.35 luma/s
visualFieldAreaThreshold	0.25		safeFallRate	1.60 luma/s
patternScoreThreshold	0.24		minimumProtectionTime	0.22 s
cameraMotionSuppression	0.32		releaseTime	0.45 s
redThreshold	0.55		displayDiagonalInches	27
redDeltaThreshold	0.18		viewingDistanceCm	70
redAffectedAreaThreshold	0.15		overloadWhiteCeiling	0.72
redDesaturation	0.68		subtleToneMap	true
blackFloor	0.08		whiteCeiling	0.84

The surface-frequency detector uses fixed 5–30 Hz parameters; these legacy sensitivity and profile settings do not tune it.

---

Documentation

Document	Covers
ARCHITECTURE.md	Legacy path — capture, detection, temporal filtering, motion
SURFACE-FREQUENCY.md	Current path, correction history, validation
SOLUTION-2026-09-05.md	Design review the current path implements
TESTING.md	Methodology, protocols, reproducibility, gaps
VERSIONING.md	Software / protocol / run versioning and immutability
OUTLAST-NOISE-DIAGNOSIS.md	The unresolved real-world artifact
docs/releases/	Notes shipped with each test package — historical, not current
experiments/	Immutable raw records, including failed runs
CHANGELOG.md	Public version history

Repository layout

Path	Contents
src/	C++ and embedded HLSL — analysis/, shaders/, render/, ui/
assets/	Embedded fonts and licenses
scripts/	Windows build entry points
flashbench/	GPU smoke, replay, visual viewer, regression automation
docs/	Architecture, testing, versioning
experiments/	Protocols and archived raw results
.github/workflows/	Hosted build and self-hosted GPU CI

Automated CI

On pushes to test: build.yml runs a hosted Windows release build with HLSL validation, and gpu-smoke.yml runs on the self-hosted runner labeled flashguard-gpu. Successful GPU runs publish per-commit artifacts containing the machine-readable reports.

---

Known limitations

• Reduces measured temporal modulation in its regression corpus; cannot guarantee seizure prevention for every person or stimulus.
• The live speckle artifact in OUTLAST-NOISE-DIAGNOSIS.md was traced to temporal detection but remains unresolved, and has not been revalidated since the synthetic suite began passing.
• Desktop Duplication and Windows composition impose latency even on the waitable low-latency path.
• Real gameplay can expose motion/content combinations absent from deterministic synthetic cases.
• On the surface path, provisional attenuation can affect ordinary color changes before a frequency is confirmed.
• The legacy local motion fallback is deliberately bounded; unusual large or complex local motion can be misclassified.
• NVOFA availability depends on supported NVIDIA hardware, driver, and runtime.
• Luminance/chroma limiting can alter colors, highlights, shadows, and perceived contrast.
• Display-size and viewing-distance calibration is approximate.
• Pattern detection and the flash sweep are not Harding FPA/PSE certification implementations.
• The detector can miss stimuli below its spatial, temporal, color, or luminance thresholds.

---

_{MIT licensed · Experimental risk-reduction software · Not a medical device}