Divide the two context figures by their file counts. Fifty files for about 150,000 tokens is roughly 3,000 tokens per file [9]. Seven files for about 18,000 is roughly 2,570 [10]. The per-file cost barely moved. The graph fetched 43 fewer files, and that is the whole 88 percent [1][2]. Which makes the token result a statement about one repository's shape: how many files sit within two hops of the one you edited. Where an auth module is imported by two hundred others, a two-hop blast radius returns hundreds of files and the six-figure context comes straight back. That number depends on sparsity around the change, and sparsity is a fact about this repository's shape, not a property of graphs in general.
The audit log bug is the stronger argument anyway. The path the graph walked was auth.py --emits--> login_event --consumed-by--> audit_log.py, three hops away, with no import of auth.py and no string match on "auth" anywhere in the file [14]. For that path to exist, a build step had to resolve a publish and a subscribe to the same topic. The writeup gives the edge list, imports, calls, inheritance, decorates, listens-to and tested-by, and a three-command install [7][6]. It does not show how listens-to gets derived, and that is the first number I would ask for: edge recall on your own bus. Constant topic strings and decorator registration are tractable. That guarantee thins out once topics get assembled at runtime, handlers get wired by a container, or subscribers get named in a YAML file the parser never opens, and in those cases the graph ends up looking complete while it keeps grep's blind spot, just dressed in nicer formatting.
The decorator case shows the other side of the ledger. A backoff parameter with an unchanged default looked caller-safe and passed the unit tests [15]. Similarity retrieval returned about a dozen files that imported the decorator; the graph returned 31, of which 19 were additions, which puts the implied similarity count at exactly 12 [16][3]. Nineteen extra files went into context, and something has to read them. The graph put the payments webhook handler in front of the model; the judgement about the timeout still came from reading it [17]. Recall is what the graph buys.
The protocol layer is the least interesting part here, which is meant as a compliment. A standard tool interface, shipped by Anthropic in late 2024 and supported by Claude Code, Cursor, Windsurf, Zed and VS Code, means the retrieval backend is swappable behind a stable call [5].
If I were adopting this, the acceptance test would be a replay. Take last quarter's incidents. For each, ask whether a structural path existed between the merged diff and the file that actually broke, and count the ones where it did. That number is yours, and it is the only version of three-bugs-in-a-week that tells you anything about your codebase [3].
