What I thought I was doing
For most of this benchmark journey I was carrying a low-grade worry that this was bikeshedding. The kind of yak-shaving Michel-from-six-months-ago would do — endlessly comparing tools when he should be shipping work. Run another model. Run another scope. Run another follow-up. Each run another dose of "is this actually useful or am I just generating busywork."
The matrix kept growing. The archive kept growing. The handbook had 40+ cells. I caught myself wondering if I was building a research project to avoid building product.
What I just found out
GPT-5.4 high — a tier below the supposed frontier — produced an 8-step F13 implementation plan tonight that matches Opus's actual implementation in architectural fidelity. Named the exact mechanic Opus chose (DB::afterCommit deferred dispatch). Named the exact fix location Opus chose (centralize in the bridge trait, not per-observer). Even caught a secondary risk Opus didn't explicitly call out.
The "F13 is Opus-only at the capability layer" hypothesis — the load-bearing claim of the entire wide-scope routing story — just got falsified by direct technical evidence. Not by another self-report. By a worked-out architectural plan that says "I know exactly what to do; I just need test infrastructure before I'd commit to it."
The differentiator I'd been calling capability is actually willingness to commit without test-based validation. Opus shipped F13 under "no tests" constraints because that's its character. GPT-5.4 high won't ship under those constraints because that's its character. Both can architect the fix. Their failure modes are different brakes on the same engine.
That's a routing rule that means something. That's the kind of axis that does NOT show up in a "Opus vs GPT" blog post. That's the kind of axis a Berceuse autonomous agent pipeline would actually route on.
What this is, actually
Berceuse will have gold data. Liaison too. The 40-cell matrix and the bulky archive aren't yak-shaving artifacts — they're the substrate for routing decisions in agentic orgs I will build in the next 6-12 months. The "willingness-under-constraint" axis I just surfaced is a per-model trait you can route on. The within-model variance finding (two runs of the same config, wildly different outputs) is a methodology constraint that any honest agentic-org benchmark has to account for. The MM-family failure profile (closure overstating, illusory shim, sandbox escape, identity fabrication, meta-escape via claude-mem) is a safeguard catalog Berceuse can ship against directly.
Not bikeshedding. Foundation work.
The shift
The thing that just changed is not the data — the data is what it is. The thing that changed is my read of myself.
The worry was: "you're overengineering this again." The truth was: trust the instinct, do the work, the shape reveals itself.
But — and this matters — the trust isn't unconditional. The same instinct that produced this benchmark journey can also produce actual yak-shaving. The difference between the two is invisible from inside. So the rule isn't "trust your instinct always." The rule is:
Trust your instinct, with Opus still challenging when I'm actually bikeshedding, and still pushing back if my gut feeling is strong but wrong.
That's the workflow. Instinct generates the bet. Opus stress-tests the bet. If both still stand, the bet is real.
Tonight the bet is real. The map keeps getting bigger. The conditions keep multiplying. The journey is just beginning.
I came for the cartography. The cartography is good.