Summary

  • Across the four Cloudflare post-incident reviews I read on 21 November 2026 — 21 June 2022, November 2023, 18 November 2025 and 5 December 2025 — the remediation lists add up to 18 items. Not one carries a named owner, a role, a completion date or a target month.
  • The only completion statement in that set is dated 1 May 2026, 164 days after the November 2025 review. It is programme-level, and its one sentence names a second global outage — 5 December 2025, 17 calendar days after the first — among the events the finished work would have avoided.

Eighteen bullets. Zero signatures.

Cloudflare's 5 December 2025 review contains the most useful sentence the company has ever published about its own repairs: “unfortunately, we have not finished deploying them yet”. Seventeen days earlier, the 18 November 2025 review had printed four remediation items and closed that section with “work has already begun”. Between the promise and the admission sits the thing almost no post-mortem prints: a name beside a date.

Do not mistake this for a story about an outage. Read as engineering, these documents are excellent. The November 2025 review names the mechanism down to a component: a change to a database system's permissions caused it to emit duplicate entries into a Bot Management “feature file”, the file doubled in size, the bad file carried more than 200 features, a software limit was hit, and the system panicked. That is what a post-mortem should sound like. It is precise about the bug and vague about the remedy, and the gap is where accountability lives.

The timeline is half-precise, which is its own kind of evidence. The prose says failures began at 11:20 UTC. The table says “11:28 | Impact starts.” Both are printed. Use 11:28 and the first automated test detected the issue at 11:31 — three minutes, a detection record most operators would envy. Use 11:20 and the same page yields a 346-minute impact window instead of 338, an eight-minute divergence the document never reconciles. Main impact was resolved at 14:30, all services at 17:06.

Impact is where the December review is braver than its predecessor. It states that “a subset of customers were impacted, accounting for approximately 28% of all HTTP traffic served by Cloudflare”, narrows the population to customers on the older FL1 proxy with the Cloudflare Managed Ruleset deployed, and notes that the China network was unaffected. The November page, as I read it, prints no percentage and no customer count at all — it distinguishes FL2 customers who saw HTTP 5xx errors from FL customers who saw none but whose traffic all received a bot score of zero.

A reader cannot compare those two impact statements in the same units, and the pages never claim otherwise.

The December mechanism is named even more finely: a WAF request-body buffer raised from 128KB to 1MB to mitigate CVE-2025-55182 in React Server Components, an internal rule-testing tool switched off, the ruleset killswitch applied to a rule with action “execute” for the first time, a nil value, and a Lua exception at modules/init.lua:314. The page gives 08:47 as the start, 08:48 as full impact, 08:50 as the declaration made by automated alerts, 09:11 as the revert, 09:12 as restored — about 25 minutes, exactly as claimed.

What it does not give is a separate identification timestamp: it says only that the issue “was identified shortly after the change was applied”. Treat the declaration time as a declaration time. The impact-to-identification interval is not measurable from that page.

The June 2022 review offers the best small lesson in the set because it is the one page that does the arithmetic for the reader: 19 data centers, 4% of the network, 50% of total requests. It prints 06:27 as the start, 06:32 as the internal declaration, 06:58 as root cause found, 07:42 as the last revert and 08:00 as closed. It then spends 06:58 twice — in the timeline as the moment root cause was understood, in the introduction as the moment the first data center came back online. One minute, two jobs. The November 2023 postmortem does something similar with its start, printing both 11:43 and 11:44 for the same outage, and prints no detection timestamp at all; its mechanism, by contrast, is specific enough to be uncomfortable for a third party — a utility event at 08:50, a vendor that “did not inform Cloudflare that they had failed over to generator power”, observability that could not see the power source change, and batteries that failed after four minutes.

Now count what those four documents promise. November 2025 lists four remediation items. December 2025 lists three. June 2022 lists three categories. November 2023 lists eight. Eighteen in total, across four incidents and three and a half years, and in the text of those pages not one item carries a named owner, a role, a completion date or a target month. What each page carries instead is a forward-looking sentence: “work has already begun”, “We have already started working on the changes outlined above”, “This will have my full attention and the attention of a large portion of our team through the balance of the year”.

Those are commitments to begin. They are not receipts, and a stranger cannot test them.

The only closure in the corpus arrives five and a half months later, and it is worth reading exactly: “Earlier this month, the Cloudflare team finished this work.” Then: “we have now completed the work that would have avoided the November 18, 2025 and December 5, 2025 global outages.” One sentence, programme-level, 164 days after the first review. It does not attach a completion date to any of the four bullets that review printed. And the sentence that certifies the work names the second global outage as one of the things the work would now prevent — seventeen calendar days after the first, on a page that describes the same gap as “two weeks ago”. Three days is nothing as an error. It is everything as a habit: these are documents that can pin a nil value to a line of Lua and cannot pin their own intervals.

That is the delta this piece adds to what BTW has already published on Cloudflare in the directory. The earlier results I read for this object cover status-page entries that never named a product or a cause, postmortems read for what they prove about failover, control-plane recovery and SLA arithmetic, and a pricing and limits group. None of them counted who owns a remediation item; none set a later completion statement against an earlier promise. This one does both.

Bound the finding to what it is. In the five pages I read on 21 November 2026, no remediation item carries an owner, a role, a completion date or a target month, and the only completion statement is the programme-level one. That is a fact about the text of those pages, not about what happens inside the company: no reviewed document shows an item unperformed, and none shows anyone concealing one. The gap is structural and easy to close — three fields beside each bullet (accountable role, target date, closure entry) would turn a promise into a record — which is precisely why its absence across four documents is worth publishing.

What a stranger can do with these pages is narrow and exact. They can verify the mechanism, because Cloudflare names it. They can often verify the detection latency, because the good pages print both endpoints. They cannot verify a repair. The published review tells you what broke and why, and then hands you a list with no name on it and asks you to take the rest on trust.