diff --git a/conditional-tracing-explainer.md b/conditional-tracing-explainer.md
new file mode 100644
index 0000000..584f93e
--- /dev/null
+++ b/conditional-tracing-explainer.md
@@ -0,0 +1,84 @@
+# Explainer: conditional tracing
+
+# Overview
+Recently, there were two discussions that seem separate but might end up being related to the same use case.
+
+## More granular information in Long Animation Frames & event timing
+Context: w3c/long-animation-frames#3
+
+A script entry point is often not granular enough, e.g. it can be a React callback that hides all the underlying slow functions,
+and also creates faux attribution: the wrapper function is "blamed" for slowness when it just passes the execution to an underlying function.
+A similar issue exists for event timing, where the provided timing attributes don't always capture the whole story, and don't always coincide with a more attributable long animation frame.
+
+## Annotated user timing in devtools
+Context: w3c/user-timing#111. 
+
+Chromium devtools already provides a way to annotate user timing entries by adding details to the user timing `detail` attribute.
+Apart from the desire to make those web-exposed annotations somewhat interoperable/standard (kind of like the `console` APIs), a question was raised about the efficiency of user timing entries for the purpose of tracing at high granularity.
+Creating user timing entries (mark/measure) has several issues:
+- Those functions receive a dictionary with each call, causing a GC load
+- User timing entries are buffered in an infinite array by default.
+- User timing entries have one global namespace, making it difficult for separate libraries to use them without stepping on each other's toes.
+
+# Proposal: conditional tracing using a `PerformanceTrack`
+
+This proposal is to have a new type of object, that represents a "track" - a namespace of sorts for low-overhead user measurements.
+A track is like a mini performance timeline, with the following characteristics:
+- Entries are not buffered in the regular performance timeline
+- Relevant entry types (`long-animation-frame`, `event`) can buffer and report track entries that are reported during the lifetime of their corresponding entries
+- Apart from the entry name, all the other annotations are per-track rather than per-entry, to avoid overhead.
+- All devtools-specific annotations are done using the `console` object.
+
+## Shape:
+
+```js
+const track = new PerformanceTrack("framework");
+
+// This buffers new entries for not-yet-created LoAF/event observers.
+track.startConditionalBuffering({entryType: "long-animation-frame"});
+
+// At some point, add a mark. This mark does not appear in the performance timeline
+track.mark(name);
+
+// At some other point, add another mark
+track.mark(name);
+
+// Combines mark->mark as a single entry with duration. This measure does not appear in the performance timeline.
+track.measure(name);
+
+// Adds devtools-specific information about a given mark/measure
+console.describe(trace.mark(name), details);
+
+// Returns a function that calls inner_function and measures it duration.
+// Addeded to LoAF scripts as an entry point
+track.bind(inner_function, name, [additional_args]);
+
+const observer = new PerformanceObserver(entries => {
+    for (const loaf : entries.getEntriesByType("long-animation-frame") {
+      // This will have all the marks & measures from the attached tracks
+      for (const trackEntry of loaf.trackEntries) {
+        // report or use the track entry for attribution
+      }
+      
+      // This will have now also include functions captured with track.bind
+      loaf.scripts
+    }
+});
+observer.observe({entryType: "long-animation-frame"});
+
+// This adds "tracks" to the LoAF entries observed by `observer`.
+observer.attachTrack(track);
+
+// Gives devtools additional information about this track
+console.describeTrack(track, { color: "blue" });
+```
+
+## Some notes:
+We use an object rather than something like `performance.trace` for the following reasons:
+- It allows adding detailed annotations without having to create a dictionary during performance critical execution.
+- It creates a natural namespace: both the reporter and the observer have to access the same object, rather than rely on (leaky) strings
+- From an ergonomic perspective, adding a global `performance.*` function might be confusing as the result doesn't end up in the performance timeline directly.
+- Using the `bind` function doesn't guarantee attribution, because it doesn't include microtasks (unlike script entry points). This needs to be handled with care when using it for "blame"
+- It might look confusing to mix web-facing and `console` APIs in the same function, however it's intentional as devtools is an "augmentation" here.
+
+