Stafford

← 2026-10-31 · all briefs · 2026-11-02 →

Stafford Brief — 2026-11-01 (Sunday)

Research-evidence: branch claude/modest-faraday-y5h07a pushed (1 KB verification addendum; scripts/verify.sh passed; nothing staged — no new source was ingested), confirmed merged to its main as 5863ab1 by the promote-claude-branches workflow. Brief emailed to brandon.freitag@gmail.com.

OPS. Clean run. Started from main at fd08a89; both repos fetched and synced before work began. Yesterday's run merged (d668f27, confirmed folded into research-evidence main as f4d9702), so there is no OPS flag to carry. Today's 12:00 scan had landed and merged (21b10c7), so this brief works from today's scan, not a stale one. Not Friday — no weekly synthesis. Not Wednesday — no falsification sweep. Ship disposition is on the last line.

What this brief is. The scan's brief this morning is good and I am repeating none of it. It also did three things with one number that need correcting before that number goes anywhere near print, and it undercounted its own evidence base on the exact question it spent the whole brief on. No new sources. The day's work was spent verifying the primary and running the scan's own open question against the shelf — which answered it, because it happened again.

📖 TODAY'S READ

Do not publish the sentence "declaring the intention moved the realized adoption rate by eleven percentage points." The comparison behind it has no controls, and the ONS never claims it does. But the same two numbers, combined with a third the ONS publishes two sections earlier, give you something far better: four out of every five UK firms that adopted AI in 2024 had told the government a year earlier that they weren't planning to.

Same source as this morning — the ONS release on management practices and AI adoption, Section 8. I read Section 8 on the ONS site directly rather than from our file.

First, what's wrong with the sentence the scan registered. The 38%-versus-27% comparison puts firms that said they planned to adopt AI next to firms that said they didn't, and reports how many of each actually adopted. The ONS does not adjust that comparison for anything — not for management quality, not for size, not for industry. It is two groups of firms counted side by side. And the two groups are not alike: the ONS's own work in the same release shows that better-run firms both plan more and follow through more. So some unknown part of the eleven-point difference is simply that the planners were better-run firms to begin with. "Moved" says the declaration caused the difference. Nothing in the release supports that, and a hostile reviewer needs one sentence to say so.

Second, the size of it. The brief leads on "roughly 55,000 UK firms." That is the number sampled for the main survey. Section 8's follow-through analysis rests on 2,895 observations, unweighted, from a linked sample the ONS itself calls "notably biased towards firms with higher management practice scores." Both 38% and 27% sit well above the ONS's own projection that national AI adoption reached about 22% in 2024 — which is what you'd expect from a sample tilted toward well-run firms. Use the contrast between the two groups. Do not use either level.

Third, and this is the one that matters most, because it is a much stronger finding than the one it replaces. Two sections earlier the ONS reports that 15% of firms that hadn't adopted AI in 2023 planned to adopt in 2024. Put that together with the follow-through rates and you can work out where 2024's adoption actually came from:

That conclusion is robust in a way the eleven points are not. It does not depend on the planners and non-planners being comparable — it only needs planners to be a minority. Work it backwards and unplanned adoption stays the majority of all adoption for any planning rate up to about 41%. The ONS's published planning rate is 15%. You would have to be wrong by a factor of nearly three before the conclusion flips.

Technical detail: 0.15 × 0.38 = 5.7 and 0.85 × 0.27 = 22.95, giving 22.95 / 28.65 = 80.1%. The majority-unplanned threshold solves 0.27(1−p) = 0.38p at p = 41.5%. Caveat that travels with it: the 15% planning rate is from the full weighted MES and the 38/27 rates are from the unweighted linked subsample, so this mixes two samples. The linked sample over-represents well-managed firms, who plan more, so the true planning rate inside it is probably above 15% — which lowers 80% but cannot cross 41.5%. These are my arithmetic from three published ONS figures, not numbers the ONS prints. Full derivation in the KB addendum.

And this changes which breakpoint the finding belongs to. The scan tagged it Momentum Mirage — the appearance of progress without progress — and called it the first time the base has measured that breakpoint as behaviour rather than opinion. Momentum Mirage needs a declared programme generating a false sense of movement. The 27% group declared nothing. Four-fifths of the adoption here isn't a mirage of movement; it is movement nobody was watching. That is Technology Illusion in its purest measured form: the technology arriving inside organizations with no stated intent, and therefore with no redesign, no purpose statement and no incentive change attached to it, because there was no programme for any of those to be part of. Momentum Mirage still has a claim on the planner side — three in five firms that announced a plan didn't execute it — but it should be the second tag, not the first.

What it changes for you: the asset sentence is withdrawn and replaced (Routing), the breakpoint tags on the staged entry need re-ordering (Routing), and your entry diagnostic has a new first question (below).

One smaller correction, because the brief will be quoted and the file won't. The brief says this result is "a behaviour rather than a perception" and that prior Commitment evidence "asked" where this one "watched." The ONS watched nothing. Firms told one survey they planned to adopt and told a different survey a year later that they had. Both sides are self-reports. What this design genuinely buys is that the two reports were made to different instruments at different times, so a firm couldn't tailor the second to match the first — that defeats the usual problem of one questionnaire grading itself, and it is a real gain. It is not administrative data. Our own KB file states this correctly as caveat 2; the brief contradicts it.

🔭 THINKER TO WATCH

None, and instead a correction: the scan told you this base holds "exactly two" instruments of the kind today's finding belongs to, and both belong to one research lineage. Our own case file and pattern file record at least four, one of them from a completely different tradition. The gap the scan built your positioning on is real — but it belongs to the whole field, not to one group's choices, and the difference decides whether an email you send lands well or badly.

The scan's argument was that only the US Census MOPS work (McElheran and colleagues) and the UK's ONS survey measure an organization's quality independently of what it says about AI and then put that against AI behaviour — and that both sit in the Bloom / Sadun / Van Reenen / Scur management-measurement lineage, which has now "twice declined" to connect the two to a real outcome. From that it drew your differentiation: the people who can best measure your independent variable have not chosen your dependent variable, and suggested it reads as an opening for collaboration.

Two more are already in this base, recorded in both the case file and an active pattern:

Why the count matters, concretely, rather than as bookkeeping. If you write to the ONS or the Bloom network saying nobody with a good measure of organizational quality has put it against an AI outcome, the first reply is: Cambridge and ESCoE did this on our instrument in May, and Statistics Denmark's registers did a version of it in March. That is the 70%-figure class of embarrassment — accurate-sounding, checkable in thirty seconds, and avoidable.

The field-wide version is both true and a stronger argument. Four instruments, three countries, three traditions, and every one of them stops at whether AI got installed. Not one measures what the installation produced. Our case file already says this correctly: "Three instruments now establish that organizational condition determines whether AI is taken up; none of the three measures whether it determines what the uptake yields." A boundary that holds across three independent traditions is evidence about what is hard to measure. A gap in one lineage's agenda is just a gap in one lineage's agenda. Use the first.

The real risk is not the count — it is a citation error forming today. You now hold three separate numbers for Claim 5 that all come out of the same UK survey programme: the ONS cross-sectional gradient (2% to 10% predicted adoption across the management range), Coyle's 0.242–0.267 plus the placebo test, and today's 17/31/48 follow-through result. They look like three confirmations. They are three cuts of one instrument, with overlapping respondents and the same management score. This base has a rule for exactly this, written on 2026-08-24 after the Babina / Ewens & Giroud episode, and nothing today applied it. Cite them as one UK body of evidence with three cuts, never as three sources agreeing. PROPOSED-UPDATE in Routing.

🏢 CASE IN THE WILD

No named organization, fifth consecutive day, and I am still not inventing one.

The scan's case this morning was ASML — correctly handled, and its finding was an absence: ten months after announcing it would strip out management layers, nobody has measured whether anything got simpler. I have nothing to add to it and adding anything would be padding.

But today's number says something about where to look for cases, and it is a change of direction rather than a complaint. Five days of this brief have reported a wage panel, two formal models, a vendor design spec, a data audit, a coefficient correction and now a statistics table — and no company doing anything. The acquisition channels hunt announcements: a declared transformation, a restructuring, a stated plan. Today's finding is that four-fifths of AI adoption never gets announced. If the dominant channel is undeclared, then a pipeline that searches for announcements is structurally blind to most of what it is trying to study — which is the same shape as the Invisible Cause pattern already in the base, one level further out. The cases are not missing because the world is quiet. They are missing because nobody issues a press release about the thing that happened without a plan.

⚙️ FOUR FORCES CONCLUSION

Commitment does get better evidence today than it has had, and the scan is right about that — but the thing being graded is narrower than the brief says, and the gradient's weakest number is the one carrying the rhetoric.

What the evidence supports: among firms that said they would adopt AI, follow-through rose from 17% to 31% to 48% across the management-quality range, with the quality measured before the plan was announced. That is follow-through graded by a condition fixed beforehand, rather than people being asked whether they follow through, and it is a real step up on every previous Commitment member in this base.

Three limits, and the third is the one nobody stated today. The comparison is raw — no controls. The sample is tilted toward well-run firms, so the levels are not national rates. And the decile groups come from the full survey while this analysis runs on 2,895 observations, so the cells at the bottom of the range are thin — thin enough that the ONS suppressed the bottom decile entirely for disclosure reasons. No cell sizes are published. So the 17% — the number doing all the rhetorical work, the one that makes it a near-tripling — is the least precisely estimated figure in the set. The direction of the gradient is well supported, including by the ONS's own regression on the full sample. The size of it rests on cells we cannot see.

And Commitment only covers part of today's finding. It grades the planners. The four-fifths of adoption that came from firms with no plan is not a Commitment story at all — you cannot fail to follow through on something you never said. That part is Capability and Purpose: the technology got in without anyone naming an outcome for it.

Purpose stays unmeasured as a moderator for an eleventh consecutive scan, and the scan declined today's tempting route correctly.

💭 OPEN QUESTION

If most AI adoption inside an organization was never planned, what is the first question in your entry diagnostic?

Your diagnostic currently assumes there is a transformation to assess — a stated objective, a programme, sponsors, a plan whose quality can be examined. Today's number says that for roughly four in five adoptions, none of that exists to examine. The AI is already in the building. It arrived through a team's expense card, a vendor feature flag, an individual's workflow. Nobody approved it, so nobody owns it, so nobody is measuring it — and the first time it becomes visible to the organization is when it fails or when someone asks who authorized it.

That is not an argument against your framework. It is an argument that the framework's best entry point is an inventory rather than a plan review. "Where is AI already running in this organization that nobody designed for?" is a question with a findable answer, it produces a list on day one, and every item on that list is a breakpoint waiting: no stated purpose, no owner, no incentive alignment, no measurement. Compare that with "let's review your AI transformation plan," which can only be asked of the minority who have one — and which the vendors are already asking.

Why this is yours to answer rather than mine. It is a change to the product, not to the pipeline, and it lands squarely on master-plan.md Tier 3.1 — repairing the self-checks at cluster level, costed at hours, no dependencies, explicitly cheap and available any time. The unplanned-adoption inventory is the same kind of move as the one already specified there ("name the three teams whose agreement determines delivery, and ask them separately"): it replaces a question about intent with a question about what is actually happening in named places. I can draft it as a one-page entry instrument on request. I am not opening a numbered asset for it — hard stop 1 holds, and Tier 3.1 already exists to receive it.

🥊 CHALLENGE

New entry, and it comes straight out of today's arithmetic: the framework diagnoses why declared transformations fail, and the data now say most AI adoption is not a declared transformation.

The objection at full strength. Five Breakpoints is a theory of programme failure. Every breakpoint presupposes an organization that set out to do something — a stated purpose that turns out vague, incentives that turn out misaligned against the stated direction, progress reported against a plan. Today's ONS numbers say roughly four in five AI adoptions in UK firms in 2024 happened in firms that had declared no such intention a year earlier. So the framework's object of address — the organization running a transformation — is the minority channel. The majority channel is unmanaged diffusion that never reaches a steering committee, has no plan to be disconnected from, and no reported progress to be a mirage of. Your diagnostic needs a programme to diagnose, and most of what happens has no programme.

Where it is weaker than it looks, and this ledger says so before Brandon does. Undeclared adoption is not a counterexample to the breakpoints — it is the condition under which they are least observable and most expensive. Technology arriving with no stated purpose is Strategic Disconnection by construction, not by failure. Nobody owning it is Incentive Fragmentation with the owner field left blank. And it will surface years later as an incident or an audit finding, which is exactly Claim 4 — the failure mode is harder to detect because it does not announce itself. The objection does not defeat the framework. It relocates it, from "transformations that fail" to "adoption nobody designed for," which is a larger addressable problem and a less crowded one.

What it would take for this objection to win outright. Evidence that undeclared AI adoption produces better outcomes than declared programmes — that the absence of a transformation programme is a feature rather than an unmeasured liability. Nothing in the base speaks to this, because every instrument here stops at adoption and none measures what adoption yielded. Today's finding and its own rebuttal rest on the same missing measurement, which is the honest position.

What it would take to defeat it. One instrument that follows undeclared adoptions forward and prices what the absence of design cost. Nobody has run it. It is also, note, something a post-engagement measurement could generate — master-plan.md Tier 0.2, still the only time-gated item on the plan, now the terminus of a sixth gate challenge, and twenty-two days without a decision recorded.

Full entry with the arithmetic and four bounds: knowledge/thesis-challenges.md, 2026-11-01.

📈 PATTERN BUILDING

No pattern crossed a threshold — no new sources, so none could. But the pattern candidate the scan opened today is tagged to the wrong half of its own evidence, and the fix changes its opening condition.

The scan logged The Unexecuted Plan — stated adoption intent barely predicts adoption — with one member and a pre-registered opening condition: a second instrument recording stated intention at one date and realized behaviour later, from a separate source.

Two problems. First, the candidate is named for the planners and the finding is mostly about the non-planners: the unexecuted plans account for about a fifth of the adoption in this data, and the unplanned adoptions account for four-fifths. A candidate named after the smaller half will recruit the wrong second member. Second, the opening condition as written would be satisfied by any intention-to-behaviour panel — including ones with no organizational variable in them at all, which would make the pattern about forecasting rather than about organizational design.

Proposed split, as two candidates with different conditions (PROPOSED-UPDATE in Routing — the pattern file is the scan's):

The second candidate also has a relative already in the base, which is why it is worth tracking rather than folding in: the Invisible Cause pattern says organizational failure is the AI-failure mode the public record cannot see. Adoption Without a Programme is the same blindness one step earlier — not a failure the record can't see, but an adoption the organization itself can't see.

🪟 WINDOW FLAG

None. Thirty-first consecutive null, and the scan's reasoning for declining today was right — a nineteen-month-old statistical release has no clock and inventing one would be the error this ledger exists to catch.

One note on the asset inventory, which sits in the same file. Today is the second consecutive day on which the sentence the scan registered for publication has had to be withdrawn and replaced. The count is unchanged and nothing is withdrawn as an asset — both days the argument survived and improved. But two in two days is now a pattern in this system's output rather than an incident, and it belongs in asset 12's checking section alongside the three instances logged last week. PROPOSED-UPDATE in Routing.

🧱 ASSET SUGGESTION

No new numbered asset. master-plan.md hard stop 1 is unconditional, all four hard stops are checked and hold, and nothing today needs one.

The replacement pitch, for the asset registered this morning:

"Declaring the intention moved the realized adoption rate by eleven percentage points."

"Four out of five UK firms that adopted AI in 2024 had told the government a year earlier that they weren't planning to. The transformation programme is not how the technology gets in."

Form: the opening of Paper 2's section on designed-in failure, plus a line in asset 12's concession section. Why it is worth more than the sentence it replaces, and this is the same relationship as yesterday: the withdrawn sentence attributes a difference to a declaration, using a comparison with no controls, and a reviewer kills it in one line. The replacement needs no controls at all — only that planners are a minority — and it says something nobody else is saying. It also sets up the diagnostic change in today's Open Question, which is the part that is sellable.

Research punch list NOT reopened. Asset 2 is at v4 and this is Paper 2 material, not asset 2 material. Full spec: knowledge/asset-suggestions.md, 2026-11-01.

🎓 SPARRING NOTE

Not issued — fourteenth consecutive withhold, and the 2026-10-02 position holds. The 09-17 note is forty-five days old and unanswered; master-plan.md Tier 0.1 calls answering it a fifteen-minute job and names it the first draft of asset 12's core argument.

What today produced instead is one rule for the skill inventory, and it is a close cousin of yesterday's. Yesterday: when a number comes out of a comparison that allows for some other factor, ask at what value of that factor the number is reported. Today: when a comparison of two groups is used to say that the thing distinguishing them caused the difference, ask what the source adjusted for — and if the answer is nothing, the number describes the two groups and says nothing about the cause. Both rules are the same question in different clothes: what is this number a number about? That is now the third week in which the answer has been "less than its sentence claimed," and the rule is worth stating once rather than being rediscovered a fourth time. Full entry: knowledge/brandon-development.md, 2026-11-01.

🔍 SOURCES & THREADS

Threads: 0 due, verified independently rather than taken from the scan. All next-check dates were extracted from story-threads.md programmatically; the earliest open thread is now 2026-11-08 (Uber's micro-teams). The scan checked ASML and Project OT today, both due, both null. Nothing in today's work bears on any open thread and claiming otherwise would be padding.

Candidates: none added, none promoted, none expired. No new outlet was read. The VKTR promotion is still PROPOSED and awaiting your call — carried from 10-31, now five days old. The scan's Puppet candidate and its pre-registered promotion bar are both sound and I endorse them as written.

Exploratory query: none composed — the slot was spent running the scan's own open question as an experiment, and it returned a result. The scan asked whether two consecutive days of finding the best result on the shelf is a finding about the pipeline, said it recommended a scheduled re-read of load-bearing entries, and said it could not define "load-bearing" narrowly enough to propose the rule. So I ran the version of that rule I think is right, on today's restated gap, and it took one search of the file store: the restated Claim 5 chase surfaced Coyle et al., held since 14 October, carrying the best single piece of Claim 5 evidence in this base — in a brief entirely about Claim 5 that did not mention it. That is the third consecutive day. Logged: knowledge/query-log.md, 2026-11-01.

Primary verification also ran, as the structural slot CLAUDE.md reserves for this layer. Target: the number the scan registered for publication four hours earlier. ONS Section 8 was read directly on ons.gov.uk, and the exact wording of the 38%/27% sentence, its absence of controls, the linked-sample observation counts (2,655 and 2,895), the unweighted basis, the Decile 1 suppression and the Section 7 planning rate (15%) were all established from the release rather than from our file. Result: one sentence withdrawn, one better sentence derived, one breakpoint re-tag, one new challenge.

KB files: none new. One verification addendum appended to the existing ONS entry in research-evidence, dated and attributed, per the precedent set on the Babina entry (10-29) and the McElheran entry (10-31). Nothing staged — no new source was ingested, so there is nothing for the index to hold. STAGE-CANDIDATE: none. That is a complete record, not a missed step.

Routing

strategy — do not publish this morning's asset sentence. PROPOSED-UPDATE to research-evidence/knowledge/publication-windows.md, anchor: Scan 2026-11-01, "ASSET registered" and the asset inventory line.

CORRECTED 2026-11-01 (Stafford, 14:30 layer). The registered sentence is withdrawn and replaced; the asset is not withdrawn and the inventory count is unchanged.

"Declaring the intention moved the realized adoption rate by eleven percentage points." The 38%/27% comparison is unadjusted. ONS reports it as a raw contrast between firms that said they planned to adopt AI in 2024 and firms that said they did not, with no control for management score, size, industry or anything else — confirmed against the release, Section 8, heading "Follow through in 2024." The release's own results show management quality predicts both planning and follow-through, so the selecting variable is correlated with the outcome driver and "moved" is an unsupported causal reading of a composition difference. Additional conditions not recorded this morning: the Section 8 analysis rests on 2,895 observations (and 2,655 for the 2023 link), not the ~55,000 sampled for the main survey; results are unweighted, against the weighted basis used elsewhere in the release; and no cell sizes are published, with Decile 1 suppressed for disclosure — so the 17% second-lowest-decile figure, which carries the "near-tripling" rhetoric, is the least precisely estimated number in the set.

Replacement asset sentence, derived from three published ONS figures: "Four out of five UK firms that adopted AI in 2024 had told the government a year earlier that they weren't planning to." Section 7 reports 15% of 2023 non-adopters planned to adopt in 2024. With follow-through at 38% for planners and 27% for non-planners: 0.15 × 0.38 = 5.7 and 0.85 × 0.27 = 22.95, so 22.95 / 28.65 = 80.1% of new adoption came from firms with no stated plan. This requires no adjustment and no comparability between the groups — only that planners are a minority. Majority-unplanned holds for any planning rate below 41.5% (0.27(1−p) = 0.38p), against a published rate of 15%.

Conditions that travel with the replacement: the 15% is from the full weighted MES and the 38/27 from the unweighted linked subsample, so this combines two samples; the linked sample over-represents well-managed firms, who plan more, so the in-sample planning rate is likely above 15%, which lowers 80% but cannot cross 41.5%; and the 80% and the 41.5% threshold are Stafford's arithmetic from ONS's published figures, not numbers ONS prints — quote the three inputs alongside, exactly as with the 10-31 crossover derivation. Full derivation: knowledge-base/ons-management-practices-technology-ai-adoption-2025-03.md, Stafford verification addendum 2026-11-01.

strategy — the breakpoint tags on today's extension are in the wrong order. PROPOSED-UPDATE to research-evidence/knowledge/paper-2-case-file.md, anchor: the 2026-11-01 Claim 2 entry, and to the staged entry's rationales.

Added 2026-11-01 (Stafford). The 38%/27% pair should lead with Technology Illusion, not Momentum Mirage. Momentum Mirage requires a declared programme producing the appearance of movement; about four-fifths of the adoption in this data came from firms that declared nothing, and an adoption with no plan cannot be a mirage of progress against a plan. What it is instead is the technology entering the organization with no stated purpose, no owner and no accompanying redesign — Technology Illusion measured at national scale. Momentum Mirage retains a secondary claim on the planner side (three in five firms that announced a plan did not execute it) and should be kept as the second tag with that rationale, not the first. Note on the index: this is a rationale and ordering change on tags already present, so apply-rationales.py can carry it — unlike the Commitment force, which the scan correctly reported as blocked.

strategy — a citation error is forming, and it is the one this base has a rule for. PROPOSED-UPDATE to research-evidence/knowledge/paper-2-case-file.md, anchor: Claim 5 evidence list, and to emerging-patterns.md, the shared-right-hand-side note.

Added 2026-11-01 (Stafford). Claim 5 now carries three numbers that all come out of the UK Management and Expectations Survey programme: the ONS cross-sectional gradient (2% → 10% predicted AI adoption across the management distribution, MES wave 3), Coyle, Nguyen, Poquiz & Riley (MES 2020 → MES 2023 panel, coefficient 0.242–0.267 plus the cross-technology placebo test), and today's Section 8 follow-through result (17/31/48, MES wave 3 → BICS 2024). Same survey programme, same management construct, overlapping respondents. Under the shared-right-hand-side rule (2026-08-24) and CLAUDE.md rule 9 these are one UK body of evidence with three cuts, and must never be cited as three independent confirmations. The Coyle entry already records that it shares its right-hand side with the ONS entry; today's extension makes it three and nothing today applied the rule.

Correction to the 2026-11-01 brief's instrument count. The brief states the base holds "exactly two" instruments measuring an organizational condition independently of the respondent's AI story and putting it against AI behaviour, both in the Bloom / Sadun / Van Reenen / Scur lineage. The case file and the active predetermined-condition pattern record at least four, including Pulito et al. (Örebro WP 3/2026), which uses Danish employer–employee registers with capability measured from 2015 workforce composition and is recorded in its own KB file as sharing no right-hand side with any other family here — so it is outside that lineage entirely. The gap itself is correctly stated elsewhere in this file: "Three instruments now establish that organizational condition determines whether AI is taken up; none of the three measures whether it determines what the uptake yields." Use the field-wide form, not the one-lineage form — a boundary holding across three traditions is evidence about what is hard to measure; a gap in one lineage's agenda is answerable by naming the group that filled it.

strategy — the positioning sentence the scan routed this morning is unsafe to send as written. "The people who can best measure your independent variable have not chosen your dependent variable," offered as a collaboration opening to the ONS/Bloom lineage, invites a thirty-second refutation: Cambridge and ESCoE ran a version of it on the ONS's own instrument in May 2026, and Statistics Denmark's registers did in March. The field-wide version is both accurate and a better argument. And the item that should be in your hands before any of this: Coyle's placebo test. The same management practices that predict AI adoption predict robotics, specialised equipment and specialised software at zero — which closes the "well-run firms are better at everything" reply with data rather than argument. Our case file calls it the strongest single Claim 5 item in the base. It is from 7 May 2026, held here since 14 October, and it did not appear in today's brief.

governance — the scan asked you to decide a pipeline change and said it could not specify it. It does not need you, and here is the rule. PROPOSED-UPDATE to research-evidence/CLAUDE.md §Weekly Synthesis or §Source Discovery, at the scan's discretion.

Trigger the re-read on restatement, not on a calendar. The scan's recommendation is a scheduled sweep of "load-bearing" held entries against current gap statements, and it correctly reports that it cannot define "load-bearing" narrowly enough — 751 entries against five claims is not a scan activity. The definition is not needed, because the calendar is the wrong trigger. Look at when the three failures happened: Claim 5's gap was restated on 10-23, narrowed on 10-29, corrected and narrowed again on 10-31, and narrowed a fourth time on 11-01. The rediscoveries landed on 10-31, 11-01 and today — each one immediately after a restatement. The shelf did not go stale. A restated gap is a new question asked of an unchanged corpus, and nobody was asking it.

The rule, in one line: when a case-file gap statement is restated or narrowed, run that restatement against the held corpus before composing the next exploratory query. It is triggered by the event that creates the risk, it iterates over one restated claim rather than 751 entries, it needs no definition of "load-bearing," and it costs a file search. Demonstrated today: today's restated Claim 5 chase, run against the corpus, surfaced Coyle et al. in one search — the best Claim 5 item in the base, missing from a brief about Claim 5. Recommended in place of the Friday sweep, not in addition to it.

governance — two days running, the registered publication sentence has not survived verification, and that is now the material for asset 12 rather than an incident. 10-28: a self-classified share cannot support a claim about organizational state. 10-30: a composition claim survived benchmarking where a shape claim did not. 10-31: a coefficient from an uncentred interaction read as a population average. 11-01: an unadjusted two-group contrast read as a causal effect. Four in five days. Every one was accurately quoted from a reachable source with a correct date, so no provenance rule in either repository could have caught any of them — and the question that would have caught all four is the same one: what population is this number about, and what was held constant? Filed as asset-12 material (knowledge/asset-suggestions.md 2026-11-01 §3). This is the demonstration that essay needs: not a claim about Brandon's standards, but a dated log of his own system catching its own errors on material it wanted to be true.

personal — yesterday's ask, repeated once because it is now cheaper than it was. Both repositories' operating rules are provenance rules; none asks whether a number means what its sentence says. Yesterday I said I did not know what the rule should be. Today I do, because the fourth instance named it: state the population the number describes, and state what was held constant; if nothing was held constant, the number describes the groups and not the cause. Say yes and it goes into the operating rules and into asset 12's checking section with four dated instances attached.

personal — unchanged, and now twenty-two days old. master-plan.md Tier 0.1 (one paragraph, fifteen minutes, forty-five days old) and Tier 0.2 (the post-engagement measurement, the only time-gated item on the plan, no decision recorded) are both still open. Today's challenge makes Tier 0.2 the terminus of a sixth gate challenge — the unplanned-adoption objection can only be defeated by following undeclared adoptions forward and pricing what the absence of design cost, and an engagement outcome measure is the only instrument in reach that could do it. Also available today at no cost: Tier 3.1, which is where today's Open Question lands. Say the word and the unplanned-adoption inventory is a one-page entry instrument.

governance — the OPENCLAW_AGENT_ID provenance defect is now six days old. One line in the script or the environment. The two-writer split you scoped on 2026-10-02 depends on being able to tell the writers apart.

MERGED 0d3677b main. Research-evidence addendum merged to its main as 5863ab1. Brief emailed to brandon.freitag@gmail.com.