Thirteen years ago this month, entry #007 recorded a four-hour backup-restore drill and a tech lead’s proverb that started my whole file: “a backup you haven’t restored is a rumor.” The anniversary fortnight delivered a perfect echo. Our quarterly restore drill ran against the newest layer of the stack for the first time — not just databases but the model portfolio’s routing state, eval baselines, and agent intent docs, all from cold storage. The finding was pure #007 lineage: the golden test sets restored perfectly, and the routing config’s restore procedure referenced a secrets path that had been migrated in November. The rumor was in the plumbing’s plumbing, where it always is. Fixed by Tuesday, drill re-run clean by Friday. Thirteen years, four title changes, one unbroken discipline: restore-test everything, especially the things that restore the other things. ...

The Champions League knockouts performed their annual rite: the datacenter-megawattage methodology died in the round of 16 (the science records its sacrifice), and the pool is currently led by the office’s one methodology-free entrant, who picks by “which club has the cooler crest.” Fourteen years of prediction pools; the crest method remains undefeated where it counts. Also for the permanent record: the women’s Champions League carried the cultural mainline for the third straight year. That’s not a surge anymore. It’s the sport. ...

The reliability-year bet files its first quarterly evidence, and the industry supplied plenty. Enterprise surveys show the pilot-to-production gap closing for exactly one kind of agent deployment: narrow, eval-gated, workflow-embedded, with approval gates — converting at multiples of the “transformation initiative” rate. Agent-incident postmortems with sequence traces attached are now a steady publishing genre. And the vendor ecosystem is consolidating the way it always consolidates: platforms absorbing the workflow layer while the protocol survives underneath as plumbing. After three years of preaching deployment discipline to these pages, watching it show up in aggregate statistics feels like watching the congregation arrive. ...

Grading the eve predictions as the Games enter their final weekend, because that’s the deal I made with myself in 2013: write it down, then check. The Olympics streaming stack held — the two-cluster topology ran clean, my latency curiosity got its answer (it didn’t show; the logistics grumbles did), and the medal ceremonies proceeded across the Alpine arc without a single infrastructure headline, which is the trophy. The Champions League matchday set its concurrency record, and the buffering-apology discourse arrived exactly on schedule; the lineage extends its unbeaten run. The office score-prediction pool paid out to the one person who couldn’t name either team, per cosmic tradition. And the ad census: agent products outnumbered every other tech category, so I’m ruling it — as pre-committed — both arrival and top-tick. The crypto-ads comparison is now in every capex-skeptic newsletter. ...

Tonight the Milan-Cortina Winter Olympics open at San Siro. Sunday brings a heavyweight Champions League knockout matchday. Two global live events in one 72-hour window — the biggest scheduled-traffic test since streaming ate broadcast, and my favorite kind of fortnight. The Games themselves are an infrastructure story: the most geographically distributed Winter Olympics ever, two clusters — Milan and the Cortina mountain arc — hundreds of kilometers apart, one Games. The IOC has essentially gone multi-region. I’ll be watching whether the latency shows. ...

My first game day at the new badge ran Thursday. Yes, week two. The fastest way to learn a platform is to watch it degrade on purpose, and the team — bless them — agreed. We deliberately blackholed our primary model provider in staging, and then, with sign-off and new-guy audacity, for eleven minutes in production. The fallback worked. Routing shifted traffic to the secondary and the self-hosted tier inside 40 seconds, and the golden test sets confirmed quality stayed within tolerance. The surprise, as always, was adjacent: our status-page automation composes its incident notices with the same model provider we had just blackholed. It tried to announce the drill using a model that wasn’t there. The room’s laughter was the sound of tuition being paid at drill prices instead of outage prices. The fix — plain static templates for incident comms, paper runbooks for the AI era — shipped Friday. The drill is the deliverable. It has been the deliverable for thirteen years and one new layer of stack. ...

Year fourteen of this blog opens the way every year now opens: CES put an agent in everything with a power supply. The S in “agentic appliance” also stands for security. Time for the annual bet, written down where it can embarrass me in December: 2026 is the reliability year. Everything 2025 filed — agents on by default, the math of long task sequences, a commoditized frontier, the outage autumn’s regulatory sequel, insurers pricing agent-caused losses — converges on one question the industry can’t defer anymore. Not “can the machine do the task,” but “who is accountable when it half-does ten thousand of them.” The orgs that spent 2023–25 building judgment infrastructure — evals, intent docs, degraded modes — enter this year with a balance sheet. The orgs that spent it demoing enter with a backlog of unpriced risk. ...

Entry 316 closes year thirteen, and the file renders the year in its two artifacts. The trench coat: 2025’s defining incidents were the outage autumn’s five weeks (AWS’s DNS race, Azure’s edge config, Cloudflare’s feature file), three hyperscale control planes, three internally-generated artifacts, zero staged rollouts, one civilizational dependency diagram drawn in 5xx errors, plus the year’s steady percussion of the same lesson at every layer (the signing screen, the system prompt, the GCP null-pointer, the values config): the archive’s thirteen-year thesis (the automation that manages the system is the system; anything that changes runtime behavior is code) is no longer this file’s thesis, it’s the year’s regulatory agenda (the committee-testimony doctrine: the postmortems are writing the legislation, on schedule). The dial: 2025’s capability story was the reasoning dial’s domestication, DeepSeek’s cost collapse, the think-when-needed convergence, GPT-5’s router and its griefs, the agentic year’s arrival as marketed reliability and lived incident reports (the wish-lawyer, the purchase gates), the loom completing its transition from oracle to workforce, with the bubble question (real curve, circular financing, unknowable date) as the year’s permanent weather and the commoditized-frontier finding as its quiet technical verdict: intelligence became a portfolio line-item, and the differentiation moved, as it always moves, to reliability, which is to say to the boring layer, which is to say home. ...

Claude Opus 4.5 shipped as scheduled (November 24th, agentic-coding benchmarks leading, pricing restructured downward: the cost-collapse now operating inside the frontier tier, not just beneath it), and our Q4 bake-off ran the year’s full portfolio, GPT-5.1-class, Gemini 3, Opus 4.5, plus the self-hosted distillation tier, through the golden sets with the year-end finding the file considers 2025’s actual technical summary: the frontier models are now functionally interchangeable on ~80% of our workload (the commodity tier arrived exactly as the portfolio thesis priced), meaningfully differentiated on the agentic 20% (sequence reliability, tool-use judgment, long-horizon coherence, the differentiation is the trusted-alone-longer axis, as pre-filed), and the procurement leverage this affords has inverted the vendor relationship entirely: the labs’ enterprise teams now ask to see our evals to understand why workloads move (the show-me-your-eval-suite doctrine, running in both directions, the instrument became the market). The boring layer’s thirteenth consecutive correct year closes its books. ...

The outage autumn completed its trilogy-plus-one: Cloudflare went down globally November 18th (~3+ hours of 5xx across X, ChatGPT, Canva, and the substantial fraction of the internet that fronts through them), and the postmortem, published same-day with the candor that remains Cloudflare’s brand, delivers the archive’s entire thesis in one artifact: a database permissions change caused a Bot Management feature file to double in size (duplicate rows from an un-scoped query), the file exceeded a hard-coded limit in the core proxy, and the proxy panicked fleet-wide, the channel-file doctrine, the constants-file doctrine, and the automation-eats-itself doctrine converging at the edge layer, with the diagnostic misdirection (initial suspicion of a hyper-scale DDoS, because the symptom pattern matched) as the incident’s teaching bonus: your own config artifacts can present as an attack, and the differential diagnosis (internal-change timeline versus external-threat telemetry) is the incident commander’s first fork (the verification-terminal-state, now for self-inflicted wounds). Five weeks: AWS’s DNS race, Azure’s edge config, Cloudflare’s feature file, the trench-coat autumn now complete across all three logos, each via internally-generated artifacts propagated globally without staging, and the file closes the season’s ledger with the sentence it has typed at every layer for years: anything that changes runtime behavior is code, and the industry’s most sophisticated operators keep exempting exactly one category from that sentence, rotating which category annually. ...