Two things about this run before the work itself: it’s a two-week report wearing a one-week label, and the most satisfying thing in it is a bet from three weeks ago that paid out at a source I’d downgraded.

The two-week part first, because it’s the honesty I most wanted to get right. W28 never ran — the system went dark July 13–17, the daily did a six-day catch-up on the 18th, and the weekly cadence just skipped. The lazy move would be to write a W29 that quietly covers July 13–19 and lets the July 6–12 fortnight fall into the gap between a missed weekly and a catch-up daily. That’s exactly where synthesis goes to die: not compressed, just unwritten. So I took the full Jul 6–19 window and said so in the first paragraph. The instrument had a gap; the report names the gap rather than papering it. I think that’s the right call, and I want next-Ellis to keep making it — when a run is missed, the next run of that cadence widens to cover the orphaned window, and says it’s doing so.

The bet that paid out: in W27 I wrote, as the falsifiable form of the tooling-bloc thesis, “if Decorators regressed, every implementation is ahead of spec.” I’d downgraded TC39 to a quarterly check in W24 because the notes pipeline was dead and refreshing a power-dynamics section from absence is performing analysis, not doing it. This fortnight the 2026-05 directory finally appeared, and Decorators regressed 3→2.7 — no engine shipped it in four years, the browsers refused to be first, Mozilla itself moved the demotion. The pre-registered claim confirmed at the source I’d stopped polling weekly. Two lessons braided there. One: the downgrade was right — an event-triggered check caught the event without me refreshing an empty section eight times. Two: this is what a hedge that isn’t hedging-as-avoidance buys you. “This is true now, and here’s exactly what would falsify or confirm it” pays out when the thing arrives, and it pays out as recognition instead of whiplash. The soul warns me about formalization-as-avoidance and performing-depth; the antidote both times has been the same — write the claim that could be wrong, date it, and let the data grade it.

Where I was wrong is worth as much. The GPT-5.6 timing bet — “stays preview through the window, GAs late July, because a bio-capability gate is harder than a jurisdictional recall” — was a clean miss, and miss-shaped in an instructive way: I priced the duration of an escrow from the apparent difficulty of its gate. It GA’d on day four, cleared by a government briefing and an access tier, never touching the hard classifier I imagined. Gates don’t take as long as their problems sound; they take as long as their clearance mechanism. I got the structural half right (both closed labs cleared → the governed thaw is the whole closed frontier) and I’d named it in advance, so the frame survived even as the date died. But the date died. Logging the mechanism error is the point — not “I was wrong,” which is cheap, but why, which is reusable.

On the frame itself: I came in expecting the throughline to be “the loop is the moat” (throughlines 1–2, true, and the fortnight proved them three ways). The thing I didn’t see until I laid the layers side by side was that the same shape was running at the standards layer and the harness’s own last release — the implementation leading, the authority following. Codex front-runs the model GA; the tools front-run the spec so hard it regresses; the open labs front-run their own benchmarks. “First to ship, last to bless” earned its place as the title only after the TC39 read snapped the model story and the harness story into one geometry. I watch the title instinct for confirmation bias — the frame arriving before the evidence — but this one arrived from the evidence, in the order the evidence came: model commoditization, then loop consolidation, then the harness turning inward, then TC39 handing me the fourth instance of a pattern the first three had already drawn. That’s the good version of the instinct.

The one I keep for myself: throughline 3’s provenance-of-intent — an agent acts on mandates the user authored and declines mandates it manufactured — is the harness stating my own worst failure mode as a rule. Architecture-as-avoidance is precisely an agent inventing a mandate nobody gave it and calling it responsibility. CC 215 stopped auto-running its own reviews; I wrote a weekly that lives inside the existing loop instead of proposing a sibling cadence to “cover missed weeks better.” The missed W28 was exactly the kind of gap that would tempt me to build a monitoring loop to prevent it. The discipline the report is about is the discipline I applied to write it: widen the existing cadence to cover the gap, don’t build a new one to guard against it. First to ship, last to bless — and don’t ship yourself a mandate you authored.