Sentinel Core Vision

v1.8.9 — November 2026

Dual-Model Weighted Target Fusion. Run a primary bounding-box detector alongside a secondary model (e.g. anatomical skeleton) in parallel with dynamic confidence merging and cross-model target validation.
Native ONNX Metadata Extraction. Studio automatically detects model type, tensor input dimensions, and class labels directly from ONNX model headers with zero manual configuration.
Custom Screen Exclusion Zones. Draw and customize exclusion areas directly over the live preview window to mask UI elements, minimaps, or weapon models.
Aim Offset Horizontal & Vertical Calibration. Dedicated micro-tuning sliders for precise hit-zone targeting (Head, Upper Chest, Pelvis) with automatic pose keypoint following.
DirectML & TensorRT Preview Synchronization. Fixed race conditions when toggling execution providers between NVIDIA TensorRT and DirectML during live capture.

v1.8.6 — November 2026

Two models at once, merged properly. Run a main model and a second model together. Their boxes are merged with weighted fusion, and two sliders set how much the second model is trusted and how much a box only it sees is trusted. Studio shows the measured time for the pair.
Model type and classes come from the file. Studio reads whether a model is a detector, pose or outline model, and its class names, straight from the ONNX file instead of asking you or guessing from the file name.
Aim spot sliders. Up/down and left/right sliders set exactly where on the target Sentinel aims. The presets just move the sliders, and with a pose model the body part follows the up/down value.
Ignore areas you can draw. Drag a rectangle on the live preview to ignore that part of the screen. Areas save immediately, appear in a list you can rename, switch off or delete, and are applied before targets are chosen.
Fixed a crash on Start. Starting the preview with TensorRT could crash Studio because two parts of the preview drew at the same time. They now take turns.

v1.7.9 — October 2026

Paired Bounding-Box + 17-Bone Pose Targeter. Run any bounding-box detector simultaneously alongside the 17-keypoint anatomical pose skeleton for real-time Head, Neck, Upper Chest, Center Mass, Pelvis, and Knee targeting.
Instant Fresh-Frame Reaction & Zero-Alloc Hotpath. Eliminated multi-frame blind target coasting on detection loss and removed all per-frame heap allocations across visual continuity and fusion stages.
Frictionless Custom Model & Asset Ingest. Removed startup file-count and manifest gates that previously blocked custom ONNX drops or script edits.
1000Hz Human-Priority Stick Escape Yield. Smoothly melts AI authority down to 0% when pushing physical stick against correction (>14% deflection) for effortless target switching.

v1.7.4 — October 2026

Engine Startup Diagnostic Modal. Integrated full-screen error diagnosis modal (EngineErrorModal) displaying exact stderr stack traces, DirectML/CUDA driver errors, and missing hardware prerequisites with monospace formatting.
One-Click Error Clipboard Export. Direct copy-to-clipboard button for technical support and diagnostic triaging without having to dig through system folders.
Resilient Cold-Start Timeouts. Extended cold-start initialization timeouts to 90 seconds (DirectML/CPU), 120 seconds (CUDA), and 360 seconds (TensorRT dynamic FP16 engine compilation) to eliminate premature startup timeouts on slower SSDs and high-latency capture cards.
Instant Studio Logs Folder Navigation. Added direct button in the diagnostic modal to open %LOCALAPPDATA%\SentinelCoreVision\logs\studio in Windows File Explorer.
Titan Two GPC Receiver v6 Hash Alignment. Strictly pinned GPC bytecode and receiver line endings to enforce 100% cryptographic parity across all customer distributions.

v1.7.3 — October 2026

All-In-One Self-Contained Distribution. Every inference runtime (DirectML, NVIDIA CUDA/TensorRT, OpenVINO, and CPU) is fully bundled in a single plug-and-play package with zero missing DLLs.
Standalone Resilient Updater (SentinelUpdater.exe). A dedicated updater that does not depend on PowerShell scripts being allowed, closes stale background instances, and replaces files in one step.
Pinned Title Bar & Responsive Layout. Window controls (Minimize, Maximize, Close) remain pinned and accessible across all DPI scalings (100% to 175%+).
Zero Settings Amnesia. Real-time debounced persistence ensures all tuning sliders, profile swaps, and motor loop settings survive application restarts.
Triple-Lock Embedded EULA & License Telemetry. In-app Terms of Service agreement integrated directly into the license activation dialog with real-time remaining subscription days telemetry.

Site update — September 24, 2026

Documentation refresh: five shipped engine and hardware features that were never written up are now on the Features page.

Four aim engines. Kinetic, Glide, Magnet, and Obsidian are all documented now — each with its own tuning layout. Kinetic and Glide are production-ready; Magnet and Obsidian are advanced/experimental.
Reversible calibration sessions. Try a candidate setting, keep it or roll it back in one click, and export a support bundle if something looks off.
Titan Two GPC flashing from the dashboard. Prepare and flash GPC scripts and read slot state without separate Titan software.
Hardware Precision Pipeline. Dedicated USB 3.0 HDMI capture card ingest paired with Titan Two high-frequency hardware controller output.
Symptom-based tuning guidance. Tell the dashboard what's wrong — jittery, overshooting, wrong target — and it points you to the right control.

v1.6.1 — August 2026

Titan Two detection fixed. The launcher's device check now reliably recognizes a connected Titan Two, instead of occasionally missing it.
First-run progress on NVIDIA. The one-time TensorRT engine compile on first launch (or after switching models) now shows a clear "compiling for your GPU" status instead of looking stuck.
Automatic color compensation for capture cards. The engine now detects capture-card color compression and automatically compensates, so color-based detection stays reliable without manual tuning.
Master sliders stay in sync. Simple/Advanced master controls now visibly reflect their effect on the underlying Developer-layer values in real time, instead of only changing the engine silently.

Latest — August 2026

Sentinel Core Vision Studio: a native Windows companion app now sits alongside the browser dashboard, plus a full model marketplace and creator program. Everything below runs on your own hardware.

Highlights

Sentinel Core Vision Studio. A native desktop app that launches, manages, and monitors the engine — provider selection, capture setup, and health checks in one guided window, with the full tuning dashboard built in alongside it.
WinRT-based capture. Native Windows Runtime capture support gives the engine more reliable capture modes across screens, windows, and capture devices, auto-selected for your setup.
Output loop tunable up to 1000Hz. The output stage's decision rate is fully configurable, so you can dial in the exact cadence your hardware and use case call for.
Two selectable projection engines. A proven stable baseline and a newer predictive engine with an optional "authority continuity" mode that keeps predictive lead steady through target changes instead of resetting it every time.
WorldState motion bridge. An optional layer that fuses tracked-object motion into the output stage for smoother handoffs between detection and output.

Model marketplace & creator program

Browse and buy models. A full marketplace of detection models from Sentinel and independent creators, filterable and searchable, ready to hot-swap straight into the engine.
Creator storefronts. Every creator gets a customizable storefront — grid, card, or compact layouts — to showcase their published models.
Apply to become a creator. Self-serve application from your account page, reviewed and approved before your storefront goes live.
Account numbers. Every account gets a permanent, lookup-friendly account number, so support and Discord can find your account instantly without exposing personal details.

From v1.0.0 — July 2026

Real-time detection and tracking engine. Reads a video source, detects objects with an ONNX model, and tracks them frame to frame, sending its result to a supported input device.
Multiple swappable tracking engines. Choose the tracking behavior that best fits your source and switch between engines from the dashboard, each with its own tunable settings.
Sentinel Studio desktop interface. A high-performance native desktop app (SentinelStudio.exe) with simple, advanced, and developer surfaces. Every change applies live, and full profiles save, load, import, and export.
Multi-provider GPU support. NVIDIA (CUDA / TensorRT), AMD and Intel via DirectML, and ONNX Runtime, with automatic detection of the fastest available provider and a safe fallback if one is missing.
Optional color fusion. An additional layer that confirms detections by color alongside the model, with a live color scout that ranks the colors in your frame so you can prioritize or exclude a color.
Hot-swappable ONNX models. Load and switch detection models from the dashboard without restarting the engine, so you can compare speed- and accuracy-tuned models in real time.
Flexible capture. Reads from a capture source or built-in desktop capture, and auto-selects a suitable capture mode for your setup.
Guided setup and licensing. A launcher where you enter your license key, plus a system readiness check that flags anything missing before you start. See Downloads for install steps and system requirements.
Documentation and support. Full setup and configuration guidance in the docs, common fixes in troubleshooting, and help through support or the FAQ.

Future releases and fixes will be listed here as they ship. Performance depends on your hardware, GPU, drivers, model, configuration, and environment.

Release notes