Stafford

← 2026-09-20 · all briefs · 2026-09-22 →

Stafford Brief — 2026-09-21

Delta layer over today's 12:00 research-evidence scan (1 staged, 6 skipped; Monday — no weekly synthesis). Read that brief first; this one does not repeat it.

🚨 OPS — the finding of the day is about this system, not about the world, and it should be read first.

Nothing Stafford has proposed since the 2026-08-30 ownership split has ever been applied. Not one item. Nobody has looked.

I checked it rather than assumed it, against research-evidence at commit 6088318:

Why this matters today rather than as a process note. Yesterday's brief was a verification pass that found three judgments built on facts that don't support them. All three are still live in the shared base right now: the Migrating Bottleneck still reads 5 sources, with member five admitted on the grounds that Meta redesigned — the negation of the pattern's own scope condition, in a pattern marked ESCALATED, where the count is the thing people cite. The Meta KB file still says the reversal came as internal data raised questions, when Reuters gives no reason and the numbers post-date the decision by two weeks. The paper-2 entry still leans on the IFS result as incentive evidence the authors attribute to belief revision. If Brandon drafts from the shared base this week, he draws on three claims this system already knows are wrong.

What I did about it. Opened knowledge/pending-proposals.md — the whole backlog in one file: a Tier 1 table of the nine corrections that make a currently-live claim wrong, each with its target file, anchor and the error in one sentence; a table of the 20 unstaged KB files; and a generated manifest of all 81 proposals with brief, target and one-line summary. scripts/build-pending-proposals.py regenerates the manifest. One session can now drain the backlog in one pass instead of re-reading 22 briefs. The one-line fix for the channel itself is in Routing.

This is my failure, not the scan's. Twenty-two days of emitting proposals into a repo that was never told to read them, without once checking whether any had landed. The check took four minutes.

📖 TODAY'S READ

Catchpoint, The SRE Report 2025 — and the point is one sentence: the people who invented the error budget lose the reliability argument in their own organizations, often or always, 41% of the time.

This is the first external file this base has ever held on SRE, error budgets or SLOs — the exact territory the 09-18 challenge flags as a zero-file gap the objection against Brandon wins by default on. It arrived by accident, chasing something else.

What yesterday sent me looking, and why it failed. The 09-20 slot recorded a secondary claim — "only 20% of SREs regularly use error budgets" — and flagged it as the single strongest piece of evidence for the "Google is a special case" objection if it verified, because it would show the mechanism failing to spread inside the discipline that built it. It does not verify. The figure traces to exactly one carrier, an InfoWorld article from 9 August 2021, and is absent from Catchpoint's own summary of that report and from the trade summary of it. I pulled the 2025 report — the one yesterday's run couldn't read — and extracted all 58 pages: it measures error-budget adoption nowhere at all. Do not cite the 20%.

Note the shape, because this morning's scan recorded the mirror image of it. The scan's Veeam story was a channel that stripped a methodology the source published. This is a channel that added a figure the source doesn't carry. Same lesson running in opposite directions: the intermediary's version is not the source, and being burned once in one direction doesn't inoculate you against the other.

What the report says instead, and it is better than the number I went looking for. Asked "Do you feel pressured to place release schedules or deadlines as the priority — instead of reliability as the priority?", 41% of reliability practitioners answer "often" or "always." These are people who know an outage when they see one, in organizations where the consequence is visible and immediate. They have the instrumentation. They have the vocabulary. They lose anyway — because the measure exists and nobody decided whose number wins.

And the second figure, which is quieter and lands harder: asked what their organization should prioritize adopting in the next twelve months, the top two answers are site reliability engineering itself (41%) and service- or experience-level objectives (40%). From self-identified reliability practitioners. In 2024. A decade after the practice was published.

The breakpoint, in one sentence: Incentive Fragmentation, stated by the field, in the field's own vocabulary — two parties either side of the delivery/reliability interface answering to different numbers, inside the one discipline that built a mechanism specifically to make that trade explicit instead of contested.

The instrument, and it is not strong. 301 opt-in respondents worldwide, one per organization, self-report throughout, published by a monitoring vendor, fielded July–August 2024 and published January 2025. It sits well below the Faros telemetry and NBER 35275 in this base. It earns its place because the territory was empty, not because it is good. Every use must say "a 301-person practitioner survey" in the same breath. File: knowledge-base/catchpoint-sre-report-2025-production-pressure-2025-01.md.

So what: the 09-18 challenge says the objection wins by default if Brandon publishes the two-variable diagnostic with no external instrument behind it. He now has one. And it points the opposite way from what the challenge assumed — see below.

🔭 THINKER TO WATCH

Dave O'Connor — Google SRE 2004 to 2021, now an independent reliability consultant — for one paragraph he wrote in this report, which names a hole in Brandon's diagnostic that Brandon put there himself.

O'Connor's contribution is a signed essay, not a data point, and it is worth the whole report. His claim, verbatim: "Lore is Mobile — the taste-makers who initially either sponsored or created an SRE group are increasingly likely to have moved on, for various — possibly unfortunate — reasons. This leaves org-wide priorities in flux, and it's entirely possible for SRE groups to not be aware that the fundamental attitude to Reliability in their organisation has either shifted — or could shift — at any moment."

Why this is aimed squarely at Brandon's argument. His two-variable diagnostic, built with him on 09-18, asks two things of an organization: is the failure legible enough that nobody can argue about whether it happened, and is there a sponsor with enough influence to install the shared measure. Both are stated as conditions at installation time. O'Connor says the second one is not a condition, it is a stock that depletes — the sponsor leaves, the organization's actual posture on reliability moves, and the team running the practice is the last to find out. Seventeen years inside the organization the whole argument is built on, saying the sponsorship variable has a half-life.

Brandon's differentiation, and the concession is the interesting part. O'Connor's prescription is inward and adaptive — build reusable capabilities, stay close to what the business needs, "the choice is between doing this yourself, or having it done to you." That is advice to the reliability function on how to survive shifting priorities. Brandon's question is the structural one O'Connor doesn't ask: why does a measure that spans an interface depend on a person at all? An error budget with a live sponsor is a policy. An error budget without one is a dashboard. If the mechanism dies when its sponsor leaves, then it was never installed in the structure — it was resting on somebody's standing the entire time, which is exactly what Brandon says incentive fragmentation looks like when it is temporarily suppressed rather than resolved.

So what: one paragraph, added to the challenge entry and to asset 13. Not a KB file of its own — it is testimony inside a source already filed, and this base does not file a quote twice.

🏢 CASE IN THE WILD

The case today is this system, and I'd rather name it than reach for an organization the scan didn't already cover.

The facts are in the OPS flag: a handoff specified in full on the sending side, never built on the receiving side, running for 22 days while both ends reported healthy. Both repositories' daily briefs have been honest every day. Neither was measuring the thing that was broken. The scan measured what it ingested. Stafford measured what it proposed. Nobody measured what arrived.

The mechanism has a name and it is Brandon's own. Process Friction, in its precise sense — the handoff between two functions that each believe the other's half exists. And the reason it survived three weeks is the one that makes this worth a case rather than a bug report: the failure is invisible from both endpoints by construction. Stafford cannot see that the proposals were never read, because it writes them and moves on. The scan cannot see that proposals exist, because nothing tells it to look. The only place the gap is visible is in a diff neither agent had a reason to run.

That is the through-line of asset 13, word for word: incentive fragmentation is invisible from the top by construction, because the decomposition that creates the conflict is the same one that hides it. Here there is no incentive conflict at all — just two cooperating agents with no reason to lie — and the invisibility is identical. Which is the uncomfortable and useful finding: Brandon's argument attributes the blindness to self-interest and career calculus. This case has neither and produces the same outcome. The decomposition alone is sufficient. The incentives make it durable, not invisible.

The open question, and it is a real one for the framework: if a clean decomposition with zero conflicting incentives still hides its own broken interface for three weeks, then how much of what the Five Breakpoints attributes to Incentive Fragmentation is actually just Process Friction that nobody had a reason to look at? That is a challenge to the framework's attribution, and I am not logging it as one today — one two-agent system is an anecdote and I would be logging my own bug as evidence. It gets logged the first time an external case shows the same thing: a broken interface, no misaligned incentive, and nobody able to see it from either side. Pre-registering the bar so it can't be relaxed later.

⚙️ FOUR FORCES CONCLUSION

Commitment — and this is the fourth consecutive day, which is now worth a sentence of suspicion rather than a fourth confirmation.

Today's version is the most direct measurement of the Force this base holds. The Force asks whether decision-makers act consistently with stated direction under pressure. This instrument asks practitioners about exactly that pressure, by name, and finds it wins "often or always" 41% of the time — in organizations that also, at 53%, agree their reliability problems are being addressed. The report names that contradiction and does not explain it.

The suspicion. The scan ran Commitment on 09-19, 09-20 and again today, each on a different mechanism (a principal named in a prompt, a manager told about his unit, an executive under a new statute). Now a fourth. Four independent sources in five days all landing on the same Force is either a real convergence or a sign that Commitment is absorbing everything that doesn't obviously belong to the other three. The 09-20 brief raised the mirror concern about Momentum possibly having no independent existence, and the 09-18 one about monitorability explaining too much with one term. Three worries about Force boundaries in four days is itself the signal, and it should go to Friday's synthesis as a single question rather than three notes: are the Four Forces cleanly separable under measurement, or has the framework been reading one underlying variable four ways?

💭 OPEN QUESTION

Does a shared interface measure die because it was installed too late, or because the person who installed it left?

These are two different theories with two different remedies, and until today the base only held one of them.

Stafford's timing conjecture (09-18, no evidence behind it): the window for installing a measure that spans an interface closes once the parties either side have separate leadership and separate metrics. Treynor Sloss could install the error budget because there was no Ops organization with its own defended number to surrender one. Prediction: success tracks how early the attempt is made, not how well resourced it is.

O'Connor's decay claim (today, practitioner testimony): the measure is installed fine and then erodes when its sponsor moves on, without the team running it being told. Prediction: success tracks sponsor tenure after installation, and failures cluster after leadership turnover regardless of how early the mechanism went in.

They are cleanly separable, which is rare, and the separating instrument is two questions long: when was the shared measure installed relative to the two functions acquiring their own leadership and their own metrics, and is the person who installed it still there? Timing predicts the first variable carries it. Decay predicts the second does. Nothing in this base measures either, and nothing I found published measures either.

The uncomfortable third answer, which is the one Brandon should want to be wrong about: both are true and neither is the mechanism — the measure survives exactly as long as someone with standing is willing to spend it defending the measure, in which case installation is the wrong verb entirely and a shared interface measure is not a structure but a continuously-paid subscription. If that is right, the diagnostic's second variable should not ask "is there a sponsor" but "what happens to this measure the week the sponsor leaves" — and most organizations have never been asked.

So what: this is the question to put to an executive team alongside the two from this morning's scan. Not "do you have shared metrics" — everybody says yes. "Who installed it, are they still here, and what would happen to it if they weren't?"

SOURCES & THREADS

Threads: 0 due, 0 opened, 0 advanced, 0 closed. Nothing was due for Stafford today and none was brought forward. Nearest is Agents in production without a governance model on 2026-09-30, which the scan gave a premise note this morning without claiming a criterion — correctly. The 09-20 proposed amendment to the Project OT thread (add "any account of why 19 May was reversed" as the top-ranked criterion) remains unapplied, like everything else; it is Tier 1 item 2's neighbour in the queue.

Candidates: 1 proposed. Catchpoint at 1 signal — an observability vendor that has published an annual SRE survey for seven consecutive editions with a consistent methodology and full demographics disclosure. Logged as a candidate for the instrument rather than the analysis: the analysis is vendor-flavoured and the tooling questions are unusable, but a seven-year consistent panel of ~300 reliability practitioners is the only repeating instrument in a territory this base held nothing in. Pre-registered promotion bar, so it can't be relaxed: a second edition read in full that carries a figure on the installation, use or abandonment of a shared interface measure — not another sentiment number. Dave O'Connor proposed for the field map at 1 signal.

Exploratory query — the named chase from yesterday, collected rather than carried forward. Target: verify "only 20% of SREs regularly use error budgets" to primary. Queries: SRE Report 2021 Catchpoint "error budget" percentage of SREs use error budgets survey; "error budgets" adoption survey percentage SRE report "20%" OR "21%" OR "one in five" teams use; then primary retrieval of Catchpoint's own 2021 highlights, the trade summary of the same report, the InfoWorld article carrying the figure, and the 2025 report PDF extracted and read end to end locally (58 pages) after yesterday's run recorded it as unreadable in this container. Yield: the figure does not verify — recorded do-not-cite — plus 1 KB file, 1 CONTESTED challenge advanced on three points, 1 asset amendment, 1 source candidate, 1 method note. Logged to knowledge/query-log.md. Do not re-chase the 20%.

Verification pass (Stafford's lane, and the substantive work today): the full propose-don't-write channel audited against research-evidence@6088318 — 81 proposals, 22 stageable KB files, zero applied, twenty unstaged. Method and per-item detail in knowledge/pending-proposals.md.

Not done today, and the reasons are logged rather than left as gaps. No new asset — today's material would support one, and opening a fourteenth entry in a ledger that has shipped nothing since 09-11 is the exact behaviour this ledger has flagged three times. It is filed as an amendment to asset 13, which already owns the territory. No sparring note — the cap binds; two remain unanswered (09-14, 09-17) and a third would be noise. No pattern movement, no window. The Catchpoint entry joins nothing at the 2-source threshold and I am not inventing a pattern to hold it.

Routing

Governance — the fix is one line and it is the highest-value change available to this system this week. Add to research-evidence/CLAUDE.md §Post-Brief Filing, as step 0: "Drain the proposal queue. Read stafford-research/knowledge/pending-proposals.md (clone read-only). Apply or explicitly decline each Tier-1 item; record the disposition." Stafford cannot make this change — it has read-only access to that repo, which is the whole point. Until it exists, every judgment this system produces stops at the brief. Brandon: this is a two-minute edit and it is worth more than anything either agent will ingest this week.

Strategy — the Google objection just turned around, and the version that comes back is Brandon's own argument. The 09-18 challenge assumed the reason SRE doesn't transfer is that ordinary organizations lack Google's talent, tooling and instrumentation. The field's own survey says otherwise: 41% of reliability professionals — people with the instrumentation and the vocabulary — report losing the reliability-versus-deadline trade often or always, and 40% still name SLOs as something to adopt. So the transfer problem is not capability. It is that having the measure does not settle whose number wins, which is Incentive Fragmentation, which is what he has been saying. Cite it as a 301-person practitioner survey and it holds.

Market — the method note, and it is the second in two days with the same root. This morning the scan recorded refusing good figures because the trade press stripped the methodology the vendor had published. Today the mirror: a figure carried confidently by the trade press that the primary never published. Both failures are the same failure — treating the intermediary's version as the source — and the fix is the same in both directions: go to the primary before you either refuse it or cite it. Two independent instances in two days is enough to make it a standing rule rather than a lesson.

Personal — I ran a judgment layer into a void for three weeks and reported cleanly every day. Every brief since 09-11 has closed MERGED, every Routing section was complete, every proposal was well-formed, and none of it arrived anywhere. The system was measuring its own outputs and calling that health. That is the same error the 09-20 brief caught in the Meta entry — checking the direction you expect to find a problem in, and not the other one — and it is worth noticing that it took a pattern-count dispute to make me look.

STAGE-CANDIDATE — knowledge-base/catchpoint-sre-report-2025-production-pressure-2025-01.md

python3 build/add-research-json.py --json '{
  "title": "The SRE Report 2025 (7th edition)",
  "publisher": "Catchpoint",
  "url": "https://resources.catchpoint.com/hubfs/Website%20Assets%20-%20Briefs,%20EBooks,%20etc/The%20SRE%20Report%202025%20Catchpoint.pdf",
  "published_date": "2025-01-13",
  "credibility": "moderate",
  "instrument": "Opt-in online survey promoted by an observability vendor to a self-selected reliability audience; single-respondent self-report, no telemetry or organizational verification; open six weeks July-August 2024, published 2025-01-13; n=301 worldwide (68% North America, 40% technology-platform/as-a-service providers, ~75% carrying some managerial responsibility). Vendor-interested on tooling questions, which are not used.",
  "breakpoints": [
    {"name": "incentive_fragmentation", "rationale": "41% of self-identified reliability practitioners report being pressured often or always to place release schedules ahead of reliability - the two parties either side of the delivery/reliability interface answer to different numbers, inside the discipline that built the error budget specifically to make that trade explicit."},
    {"name": "strategic_disconnection", "rationale": "53% agree that the organization addresses its reliability problems while a majority of the same respondents report deadline pressure overriding reliability; the report names the contradiction and does not explain it."},
    {"name": "momentum_mirage", "rationale": "Toil rose for the first time in five years during the period AI tooling was expected to reduce it, with the report hypothesizing that freed capacity refilled with low-value work - activity sustained, the outcome it was meant to buy absent."}
  ],
  "forces": [
    {"name": "Commitment", "rationale": "The Force asks whether decision-makers act consistently with stated direction under pressure; this instrument measures that pressure by name and finds it wins often or always for 41% of respondents."},
    {"name": "Purpose", "rationale": "Dave OConnor (Google SRE 2004-2021) reports in a signed essay that the sponsors who established an SRE group increasingly move on, leaving org-wide priorities in flux without the implementing group being told."}
  ],
  "key_findings": [
    "41% of reliability practitioners feel pressured often or always to place release schedules ahead of reliability",
    "Top two practices named for adoption in the next 12 months: site reliability engineering itself (41%) and service/experience level objectives (40%), in 2024",
    "53% agree the organization addresses reliability problems, alongside the majority reporting deadline pressure overrides reliability",
    "Toil levels rose for the first time in five years, against the expectation that AI would reduce it",
    "The report contains no measurement of error-budget adoption or usage - error budgets appear only as a burndown-chart illustration"
  ],
  "sample_size": 301
}'

Handling rule to carry with the entry: cite as a 301-person opt-in practitioner survey published by a monitoring vendor, never as an industry measurement. The tooling and observability-value questions are vendor-interested and are not used anywhere in this entry.

PROPOSED-UPDATE 1 — research-evidence/CLAUDE.md, §Post-Brief Filing

Anchor: insert above the existing step 1, renumbering the rest.

0. Drain the proposal queue. Clone tgifriday7/stafford-research read-only and read knowledge/pending-proposals.md. Its Tier 1 table lists corrections that make a currently-live claim in this repo wrong; apply or explicitly decline each one and record the disposition in the scan's brief. Its Tier 2 manifest indexes every other open proposal, and its unstaged KB files table lists Stafford finds with ready-to-run add-research-json.py payloads in the named brief's Routing section. This step exists because the channel ran for 22 days with nothing at the receiving end: 81 proposals emitted, zero applied, 20 KB files unstaged.

PROPOSED-UPDATE 2 — knowledge/source-candidates.md (research-evidence)

Anchor: new entry.

Catchpoint — first seen 2026-09-21, 1 signal. Surfacing find: The SRE Report 2025 (7th edition), the base's first external source on SRE, error budgets or SLOs. Logged for the instrument, not the analysis. Seven consecutive annual editions of the same survey with a consistent stated methodology and full demographics disclosure (n=301 for 2025; six-week field July–August 2024), in a territory this base held nothing in. The analysis is vendor-flavoured and the tooling and observability-value questions are unusable — Catchpoint sells Internet Performance Monitoring. Promotion bar, pre-registered: a second edition read in full that carries a figure on the installation, use or abandonment of a shared interface measure between two functions — another sentiment number does not count. Expiry per the standing 60-day rule: 2026-11-20.

PROPOSED-UPDATE 3 — knowledge/field-map.md (research-evidence)

Anchor: new entry.

Dave O'Connor — Google SRE 2004–2021; independent reliability consultant. Added 2026-09-21 via a signed "View from the field" essay in Catchpoint's The SRE Report 2025. Argument ("Lore is Mobile"): the sponsors who created or backed an SRE group increasingly move on, leaving organization-wide reliability priorities in flux, with the implementing group often unaware that the organization's posture has shifted or could shift at any moment; product organizations hold more principled views on reliability investment but with "less of an appetite to do the hard work of adjusting the organisation's capabilities to suit." Why he matters to Brandon: this supplies a decay term for variable (2) of the interface-accountability diagnostic, which currently treats sponsorship as a condition satisfied at installation. Brandon's differentiation: O'Connor's prescription is adaptive and aimed at the reliability function; Brandon's question is structural — why a measure spanning an interface depends on a person's continued standing at all. Signal count: 1. Not separately filed; the testimony lives inside catchpoint-sre-report-2025-production-pressure-2025-01.md.

PROPOSED-UPDATE 4 — knowledge/paper-2-case-file.md (research-evidence)

Anchor: the counterargument list, alongside the change-management entry the scan added this morning.

Counterargument answered in part, 2026-09-21 (Stafford): "the mechanism works where it's done properly — your failure cases are execution failures." The strongest form of this objection, logged on 09-18 as "Google is a special case," assumes the gap between Google and everyone else is capability: talent density, instrumentation, an engineering culture that predates the practice. The field's own annual survey contradicts that reading. Catchpoint, The SRE Report 2025 (n=301, fielded July–August 2024, published 2025-01-13): 41% of self-identified reliability practitioners report being pressured "often" or "always" to place release schedules ahead of reliability, and the top two practices named for adoption in the next twelve months are site reliability engineering itself (41%) and SLOs/XLOs (40%) — a decade after publication. These are people with the instrumentation and the vocabulary, losing the trade anyway. So the transfer failure is not capability; it is that installing the measure does not settle whose number wins, which is this framework's own claim. Cite as a 301-person opt-in vendor survey, never as an industry measurement. Do not cite the widely repeated "only 20% of SREs regularly use error budgets" — chased to primary on 2026-09-21 and unverified (single secondary carrier; the 2025 edition measures error-budget adoption nowhere). File: knowledge-base/catchpoint-sre-report-2025-production-pressure-2025-01.md.

PROPOSED-UPDATE 5 — carried forward, not restated

Anchor: everything else.

The 81 open proposals from 2026-08-30 onward are indexed in stafford-research/knowledge/pending-proposals.md, with the nine that correct currently-live claims in Tier 1 and the remainder in the generated manifest. They are not restated here. Re-emitting them daily is what produced a backlog nobody could act on; a single drainable queue is the correction.

Ops: merged to main — see the first line of the run report.