bi-01 — Analysis
What this module is actually training
Not knowledge of Chromium. Convergence of search. The measurable output is: given a browser behaviour you have never investigated, how many queries until you are reading the right code, and do you notice when you are in the wrong subsystem?
Required invariants of a good navigation
- The prediction is made before the search. Naming the expected subsystem first is what turns a search into a calibration measurement. A search without a prediction teaches nothing about your model.
- Zero hits is information, not failure. It means one of: generated code, wrong subsystem, wrong spelling of a term of art, or the feature is flag-gated. Broadening to a soup of common words discards that information.
- Callers and callees come from cross-references, never from reading. Gate questions 3 and 4 have a mechanical answer; inferring them from nearby code is how wrong call graphs get written down and repeated.
- The query is recorded, not the path. Paths rot — this repository has two dated instances
(
ng_prefixes,TraceWrapperMember). Queries transfer across subsystems and across years.
Failure modes
| Failure | Symptom | Correction |
|---|---|---|
| Concept instead of token | "layout" returns tens of thousands of hits | convert to symbol: or a spec phrase |
| Wrong layer | two hours in //content for Blink internals | answer "which process?" before searching |
| Assuming dead code | "this is never called" | check generated code, .mojom, feature flags |
| Depth-first descent | four hours in //base/containers | keep a deferred list; stop when the answer stops changing |
| Trusting a remembered path | confidently wrong | check the date in the verification log |
What would falsify the approach
If spec-phrase search failed on spec-defined subsystems, Technique 1 would be worthless. It does not, and the reason is structural: Blink quotes specification step text in comments, which makes the specification a search index over the implementation. That property is why the parsing lab is tractable on day one with no build.
The technique degrades precisely where there is no spec — compositor, scheduler, GPU. There the
naming grammar (Technique 3) and the trace-event bridge (bi-12) take over. Knowing which
technique applies to which subsystem is the actual skill.
Evidence this module produces
- a query log with failures included
- a prediction-accuracy count out of 8
- one candidate failing WPT recorded for
bi-14 - the three techniques that worked best for you, which differ by person more than people expect