SIGNAL·WATCHER DEV

The masked renderer strings are worthless as GPU identity — Chrome hardcodes "WebKit WebGL", Firefox sanitizes, and under privacy.resistFingerprinting the debug extension returns null — but the SHAPE of the masking identifies the engine and the privacy configuration, which is why both halves of the pair are shown. getContext('webgl') returning null is meaningful in its own right: Chrome 137+ desktop removed the automatic SwiftShader fallback, so null now means no usable GPU rather than a slow software path. Adapter limits are the hardware's; a bare requestDevice() would report spec defaults (8192) instead, so no device is requested here and none is reported. isFallbackAdapter is documented as always false on users' devices — probed for completeness, non-functional as a signal. requestAdapterInfo() is never called: it was removed in Chrome 135 and its absence serves as a version rung instead.

UA-specific memory needs crossOriginIsolated and a button press (it is async and gated)

The display period is measured, not assumed: 90 Hz and 120 Hz panels exist and scoring them against a hardcoded 16.67 ms invents jank that is not there. Percentiles are reported rather than a mean because a 250 ms stall inside an otherwise good second still averages ~55 fps and looks healthy. Read this alongside visibilityState — a backgrounded tab is throttled to about 1 Hz and would otherwise read as catastrophic jank.

origin trial in M125, desktop only — absence here is expected on Android and in-car. The three stages are reported separately because presence is not function: the constructor is undefined in headless Chrome and a function in headed Chrome, and even headed it delivered no sample in 12 s on macOS. A green stage 1 with a silent stage 3 is a real and common outcome.

context values fill in after "Audio route probe" — no AudioContext is built on load, so nothing steals audio focus

Two names exist for the same quantities and they differ by 1000×. Legacy playoutStats (fallbackFramesDuration/Events, totalFramesDuration) is in MILLISECONDS; shipped playbackStats (underrunDuration/Events, totalDuration) is in SECONDS. Both are probed and normalised here — durations to seconds, latencies to ms. Counters are cumulative, so this card diffs successive reads rather than reporting an instant. Availability is contested upstream (chromestatus lists it behind a flag at milestone 146; an independent test found it live with no flags on 150). This page makes no support claim: it reports only which name exists here, whether reading threw, and whether the numbers moved.

Portable fallback for browsers with no playback statistics: the RATE at which currentTime advances against the wall clock, measured over a sliding window. Healthy sits at 0.999–1.003; a starved graph reads ~0.33. The absolute offset is NOT drift — it is a constant ≈31 ms of outputLatency, shown separately and never scored. The first window is discarded (startup reads ~0.94) and two consecutive bad windows are required before starvation is called.

Brackets the engine from what it can DO, then cross-checks the UA string. A disagreement is itself the finding: it means the UA is reduced, frozen or spoofed. The bound is a floor only — no upper limit is claimed, because a rung being absent can also mean the feature was removed.

Each row isolates one engine. AudioContext.setSinkId present with HTMLMediaElement.setSinkId absent identifies Chrome Android with no UA involved. On iOS every browser is WebKit, so audioSession present alongside a Chrome UA means iOS Chrome, not Blink. permissions.query({name:'speaker-selection'}) throws a TypeError in Chromium instead of returning a status — throw-vs-return is recorded, because that difference is itself a discriminator.

Because getAutoplayPolicy is Firefox-only, the universal oracle is the state of a freshly constructed AudioContext: running = autoplay allowed, suspended = blocked. It is captured on the first context this page builds (never on load) and timestamped, because in Chromium it is not stable per-origin — the Media Engagement Index can flip it between visits.

"addressable channels" is what maxChannelCount reports — a channel budget, not a speaker count. No web API exposes the real speaker layout, so a value of 6 does not mean six speakers exist. On decodingInfo, only the supported column carries information for audio-only configurations: every supported audio codec reports powerEfficient: true in Chromium, and smooth is defined against video frame pacing. Both are shown but neither is meaningful here.

any "present" here = potential in-browser route to a CAN/OBD dongle without phone

usageDetails is a Chromium extension to the standard estimate and breaks the total down per storage system — its presence alone is an engine signal, and the bucket NAMES it returns move between Chromium versions.

The extension COUNT is more discriminating than any single limit: it moves with driver, GPU family, ANGLE backend and Chromium version together. Precision formats separate mobile GPUs that lack true high-float in fragment shaders from desktop ones that do.

Every value here is the ADAPTER's — the hardware ceiling. No requestDevice() is called anywhere on this page: a device reports spec defaults rather than hardware, and on Chrome 140+ requesting one consumes the adapter. GPUSupportedLimits exposes its values on the prototype, not as own properties, so the limits are read from a fixed name list rather than enumerated — a name missing from that list is a limit this page does not ask for, not a limit the adapter lacks.

All three are promptless and silent — the audio one renders through an OfflineAudioContext, which produces no sound and takes no audio focus. FNV-1a, inline, no dependencies. A hash that changes between two reloads on the same device is the finding: it means the browser is randomising the surface (Brave's farbling, privacy.resistFingerprinting), which is far more informative than the hash itself.

queryLocalFonts() is never called — it prompts. Detection is the measureText width method: a family is present if its rendered width differs from all three generic fallbacks. document.fonts.check() is probed too but is NOT used as the detector: in Chromium it answers true for families that do not exist, so what it reports is its own behaviour rather than the font set. Whether it discriminates is recorded as its own row. The SET of families is one of the strongest OS discriminators reachable without a permission.

The January/July offset pair reveals the DST RULE rather than the current offset, which identifies the timezone far more tightly than the zone name alone — and disagrees loudly when the zone name is spoofed but the underlying ICU data is not. The supportedValuesOf counts track the bundled ICU version, which moves with the browser build.

Presence of crypto.subtle says nothing about which algorithms work, so the keys are actually generated — behind a button, because RSA-OAEP alone can cost hundreds of milliseconds. Ed25519 and X25519 are the sharp rungs: they shipped unflagged in Chromium 137, long after the rest, so a working Ed25519 dates the build on its own.

No STUN server is configured, so this works fully offline and reveals only host candidates — which is the point. Whether those hosts arrive as raw private IPs or as .local mDNS hostnames is a per-engine, per-version policy and a strong discriminator. getCapabilities() needs no peer connection and no permission at all; the codec list and its clock rates vary by platform build far more than the codec matrix above suggests.

enumerateDevices() itself never prompts. Empty labels with a non-zero device count is the un-granted state; populated labels mean camera or microphone permission has already been given to this origin — so the label column is read as a permission probe, not as device names. getSupportedConstraints() varies widely across platform builds and is one of the cheapest breadth wins available.

Zero voices is a real and informative result, not a failure — it is what an embedded or in-car build with no TTS engine installed reports, and it means every speech feature in the app above is dead on that device. The list populates asynchronously in most engines, so voiceschanged is listened for and the card updates itself; a count that stays at zero after the event has fired is a much stronger claim than one read once on load.

Math.tan(-1e300) is not a stunt: the result comes from the platform's libm and differs between V8 on glibc, V8 on Android's bionic, JavaScriptCore and SpiderMonkey, so it separates engines that every string-based check reports as identical. The observed timer granularity is measured by spinning performance.now() until it moves — 5 µs means crossOriginIsolated, 100 µs is the standard Spectre-mitigated clamp, and coarser values mean an additional privacy mode is active. navigator.plugins is a fixed fake set in Chromium rather than a real plugin list, which is exactly why its length and contents identify the engine.