Stafford

← 2026-09-04 · all briefs · 2026-09-06 →

Stafford Brief — 2026-09-05 (Saturday)

Delta layer on two scan days — 09-04 (2 staged, weekly synthesis) and 09-05 (2 staged, 5 skipped). Both scan briefs are strong; read them first. This brief adds three things they could not: the queueing model's parameters read for what they name rather than for what they conclude, an audit of the citation chain the 09-05 scan started correcting and stopped halfway down, and a challenge I opened this morning and then defeated with the paper's own algebra.

⚠️ OPS — no Stafford brief ran on 2026-09-04. The last one is 09-03. Friday's scan produced the weekly synthesis and the NBER 35275 ingest with no Stafford delta layer against it; this brief covers both days, and the Weekly Synthesis section below is a one-day-late Stafford-side catch-up. Cause unknown from this session — the run simply is not in the repo. Worth checking the schedule, because the 09-04 scan's own synthesis named deferred standing reviews as this system's weakest link, and then this system deferred one.

📖 TODAY'S READ

Bartolucci & Vivo, Queue & AI: When Faster Tasks Slow Down the Workflow (arXiv 2605.27202, 26 May 2026) — read for its parameters, not its result. Full text pulled and extracted; the scan read the abstract and the figures.

The scan booked this paper as a mechanism for the Migrating Bottleneck and drew the differentiation at τ_A: they treat the human attention a task costs as a parameter; Brandon's claim is that it is a design variable. That is right and it is the generic version. The specific version is one equation the scan quoted and did not unpack, and it is worth more than the paper's headline.

The review threshold is π⋆(θ) = θ/(κK) (Eq. 8). Drafts whose perceived risk falls below π⋆ are "passed through without scrutiny." The paper's own gloss, verbatim: "the review threshold rises with the congestion cost of reviewer time, θ, and falls with both the cost of an escaped error, K, and the effectiveness of review, κ."

Now read the three terms against the framework, using the paper's own definitions:

So the equation governing whether the human in the loop actually looks contains two Four Forces variables and one capacity term, and the only genuinely exogenous inputs in the whole model are the arrival rate λ and the AI's service-time variance c²_A. Two physicists with no organizational-research programme wrote the framework's claim as an algebraic condition and then held all of it fixed. Read as physics, this is a limit on AI. Read as design, π⋆ = θ/(κK) is the first ex-ante expression in this base in which a Five Breakpoints variable appears as a named, estimable parameter — which is the thing growth edge #2 has been asking for since 08-29 and has never had a functional form for.

The limits are the scan's and they hold: no data in this paper, none. The Empirical calibration section is an agenda — "a useful empirical agenda would therefore combine time stamps, reviewer effort, error flags, and rework links at the task level" — not an estimation. Figure parameters (C = 1.00, λ = 0.75, κ = 2.00, K = 2.00) are described in the text as "a representative calibration" and "illustrative levels." Unreviewed preprint. Cite the mechanism and the functional form; never cite a number from this paper.

🔭 THINKER TO WATCH

Vibhas Ratanjee — Gallup, Global Practice Leader for Leadership Development. Not a discovery. A correction, and it is the reason this section exists today.

The base holds him as "Vibha Sratanjee, Forbes" (knowledge-base/boreout-org-design-failure-2026.md, and the sources line of the "Boreout as Org Design Diagnostic" pattern) — name split wrongly, affiliation absent. He is a Gallup practice leader who writes a Forbes column; his July 2026 boreout piece is the second of two members of that pattern, and the other member is Gallup's State of the Global Workplace 2026. The pattern therefore has one research house and two bylines, which is precisely the non-independence the 09-05 scan flagged one pattern over — "same house, adjacent panels, one construct" — applied to a case the scan did not check.

Brandon's differentiation, and it is unusually clean here. Ratanjee's boreout argument reaches Brandon's conclusion from the wellbeing side: roles designed for activity metrics rather than outcomes, and the line the base quotes approvingly — "The interventions treat the person. The org chart created the condition." That is Momentum Mirage at the role level and he gets there without a framework, one column at a time, from a practice built to sell engagement measurement. Brandon has the diagnostic; Gallup has the panel. The differentiation is not intellectual, it is instrumental: Ratanjee can assert the org-chart cause and cannot measure it, because everything Gallup fields is self-report. Worth watching as a distribution node for Gallup findings into Forbes — and worth logging with his affiliation attached, so the base never again counts him and Gallup as two.

🏢 CASE IN THE WILD

No new case from Stafford today, stated rather than filled. The 09-05 scan's CHRO case (Gallup Q1 2026, n=23,717 — 50% of CHROs not confident their managers can guide employees on AI; 57% providing manager AI training) is the case of the day and it is well handled there.

The delta is what today's model does to it, and the scan could not have drawn this because it read the paper and the survey in separate sections. The scan's reading of the 43% untrained is a guidance failure: the manager layer cannot advise on AI use. The model says something stronger and more specific. κ is reviewer verification skill, and an untrained reviewer has a low κ. Low κ raises π⋆ = θ/(κK) directly — mechanically, before any question of motivation or attention. So an untrained manager layer is not merely less able to guide AI use. It is formally predicted to review less of what crosses it, at any given load, than a trained one would. The training gap and the oversight gap are the same parameter, and no governance regime in this base connects them.

The open question the pair leaves. Gallup measures whether CHROs are confident in their managers; the model asks for κ, which is estimable from logs — what fraction of errors an hour of review actually catches. Nobody measures the second, and the first has been standing in for it for four months. Is CHRO confidence in the manager layer correlated with anything at all, or is it the readiness-perception measure Noe 2014 already caught tracking belief and not implementation?

⚙️ FOUR FORCES CONCLUSION

Commitment — and this is a disagreement with the 09-05 scan, not an extension of it.

The scan concluded Capability, stated as capacity: "the binding term is C, human attention capacity for review and rework." That reads the model off its stability condition and it is half the model. The threshold result — the one the scan itself called the result to keep — turns on K, the cost of an escaped error, and K is not capacity, skill, or attention. It is what the organization has decided it costs when a bad output gets through, and who carries it. That is Commitment in the Four Forces sense: whether decision-makers act consistently with stated direction under pressure. An organization that declares a human-in-the-loop control and sets K at zero — nobody's number moves when an AI error escapes — has a reviewer whose π⋆ is infinite. The review does not degrade under load. It was never going to happen at any load.

Why this matters more than a tagging preference. The Capability reading says the fix is capacity: more reviewers, lower ρ. The Commitment reading says the fix is cheaper and nobody is doing it — raise K. Both appear in the same denominator, and only one of them requires headcount. Paper 6's claim that AI absorbs routing but not stewardship is the qualitative version of exactly this: stewardship is who bears the cost of the escape, and the model shows that when nobody bears it, the check stops.

💭 OPEN QUESTION

If oversight collapse is governed by θ/(κK), and K is the cheapest of the three to change, why does every AI oversight regime in this base buy θ and κ and none of them touch K?

The scan's governance point yesterday was that oversight regimes are specified as headcount and the model says load is the missing variable. True, and it stops one term short. Read the denominator: raising K — making an escaped AI error expensive to somebody specific — moves the threshold exactly as far as adding reviewers does, and costs nothing but a decision about accountability. Nothing in the Five Eyes prerequisites, the FRC guidance, Deloitte's audit-versus-approval split or any of the eleven verification instruments assigns a cost to an escaped output. They all assign a person.

Why this is Brandon's question and not the physicists'. They need K to be a number so the algebra closes; they say nothing about where it comes from. In a real organization K is set by design and it is usually set at zero by accident — the reviewer is measured on throughput, the escape surfaces three steps downstream as somebody else's incident, and no ledger connects the two. That is Incentive Fragmentation with a functional form attached, and it converts the framework's most assertive claim into an arithmetic one.

The uncomfortable half, and it binds. Nobody has estimated θ, κ or K in any organization, this paper included. A model that makes the framework's claim precise is not evidence that the claim is true, and the temptation over the next month — the scan named it and it applies here doubly, because this brief has just built an argument on the algebra — will be to cite π⋆ = θ/(κK) as though someone had measured a term in it. The honest sentence is that the framework now has a functional form and still has no estimate.

🥊 CHALLENGE

A challenge I opened this morning and then lost to the paper's own algebra. Logged as a tested-and-rejected reading in knowledge/thesis-challenges.md under coordination-primacy, because a ledger that records only the challenges that survived is a trophy case.

At full strength, the version I went in with. The base's standing answer to coordination-primacy (CONTESTED since 08-29) is that coordination compression is real, and what remains binding afterwards is verification — an organizational property. That answer leans hard on Klein & Wieczorek, whose own model relocates the constraint to verification. Bartolucci & Vivo appear to take that answer away. They derive verification collapse from a queue: arrival rate, service-time second moment, reviewer capacity. If oversight failure is fully derivable from load, then the residual after coordination compression is capacity, not organizational design, and this base has just adopted as a friendly mechanism the third formal model in a row — after Farach and Klein & Wieczorek — that explains its observed patterns with no organizational variable on the right-hand side.

Why it fails, and the refutation is the paper's Eq. 8. The right-hand side is not load. It is θ/(κK), and two of those three terms are organizational by the authors' own definitions — verification skill and the cost of an escaped error. The model does not derive oversight collapse from congestion; it derives it from congestion relative to how skilled the reviewer is and how much the organization loses when the check fails. The rival reading requires κ and K to be constants of nature. They are line items.

What survives, and it is not nothing. Three formal models now put non-organizational variables on the left of this base's observed pattern, and the base has ingested all three as confirmations. That is a genre exposure worth naming even when each individual reading comes out the framework's way. Coordination-primacy stays CONTESTED and today it got weaker, on the strongest ground it has been contested on yet: a rival model that, unpacked, contains the framework's variables in its own denominator.

📈 PATTERN BUILDING

Three patterns moved today and all three moved down. The 09-05 scan started this correction and stopped one hop short of its own conclusion.

The scan retired the 8.7x manager engagement multiplier — load-bearing since 2026-05-01, sourced to Gallup's 251-page State of the Global Workplace 2026 via a BrianHeger.com summary — and replaced it with the primary (Meinen & Mulherin, 2026-08-16, n = 23,717, ±0.9pp). Correct, and overdue. It then reported the exposure as "two patterns."

The exposure is four patterns, and in three of them the retired secondhand source is one of exactly two counted members (knowledge/emerging-patterns.md, 115 patterns total):

Pattern Recurrence Members Does today's primary repair it?
"Manager Engagement = AI Transformation Multiplier" 2 Gallup SOGW + McKinsey Yes — direct substitution, same construct
"The Wellbeing Collapse" 2 Mercer + Gallup SOGW No — cites SOGW for thriving/global engagement, which the new US AI-culture panel does not measure
"Boreout as Org Design Diagnostic" 2 Forbes/Ratanjee + Gallup SOGW No — see below; substitution leaves one house
"The Inverted Expertise Gap" 2 (Writer + CMI UK) Gallup SOGW as third, supporting Count unaffected; the rationale sentence "the 8.7x Gallup multiplier establishes that manager engagement is the transmission mechanism" needs restating

Two of these fall below this base's own two-source threshold unless a replacement member is named. That is the correction the scan's own logic requires and did not make.

And the Boreout pattern is worse than a counting problem — its chain breaks in three places, verified today against primaries.

  1. Its two sources are one house. Vibhas Ratanjee is Gallup's Global Practice Leader for Leadership Development (see THINKER above). Forbes-column-by-Gallup-leader plus Gallup-report-via-blog-summary is not two independent sources.
  2. Its headline cost figure is a construct swap. The KB entry states: "Research published in the American Journal of Preventive Medicine estimates boreout costs US companies $3,999–$20,683 per affected employee annually." The study is Martinez et al., AJPM 68(4):645–655 (2025), doi 10.1016/j.amepre.2025.01.011, and it measures burnout — verbatim, "employee disengagement, overextension, ineffectiveness, and burnout." Not boreout. The "$3,999–$20,683" is not a range across affected employees; it is the gradient across job levels — $3,999 nonmanagerial hourly, $4,257 nonmanagerial salaried, $10,824 manager, $20,683 executive. And the method is "a computational model" built in 2024, not a measurement of firms. Three errors in one sentence, and the third is the one the base would flag instantly in someone else's work. Note the irony that makes it worth keeping: the column's entire thesis is that boreout is not a wellness problem, and its cost figure is borrowed from the wellness literature's burnout simulation.
  3. Its most-quoted claim has no source at all. "80% of most employees' time is coordination theater" appears in the pattern as "the deeper finding" and in the KB file as "the 80/20 pattern," attributed to nobody. It is a column's rhetorical figure carrying a decimal.

And the KB entry tags all five breakpoints on a single opinion column — no instrument, no sample — which is the loose tagging both CLAUDE.mds forbid in the same words. It is an inherited v1 entry (first logged 2026-07-05), which is exactly the case Hard Rule 10 exists for: inherited v1 history is trusted but re-verifiable before it becomes load-bearing. It became load-bearing without being re-verified, and the pattern proposes it for Paper 2 as "the individual-scale manifestation of Five Breakpoints' macro argument."

Disposition proposed below: retire the Boreout pattern to 1 source and quarantine the KB entry's cost claim. Not delete — Ratanjee's org-design argument is genuinely on-thesis and the quote is good. The claim to keep is the argument; the claims to drop are all three numbers.

The transferable finding, and it is a fourth broken chain of a new kind. The base has now caught the "8.1 spans in 2013" baseline, the MIT Sloan 7→15 span claim, and the "Stanford 1,200 enterprises" study — all three had no primary at all, which makes them findable by chasing the citation to nothing. This one has a real primary, in a good journal, saying something adjacent but not the thing being claimed. That is a harder failure mode: the citation resolves, the DOI works, and the number is wrong anyway. The scan's proposed habit — trace any figure cited more than twice to a primary before the fourth citation — catches the first three. Catching this one requires reading what the primary measured, not just confirming that it exists.

🧱 ASSET SUGGESTION

Asset #5 — "The Human-in-the-Loop Control That Isn't" — gains a fourth audit question, and it is the one with an equation behind it. (Updated in knowledge/asset-suggestions.md; not a new asset.)

The three-question audit as drafted asks: which failure classes must the human recover, what evidence exists that they can, and what is the review measured on that isn't throughput. Today's model supplies the question that was missing and it is the one nobody asks: what does it cost this organization when a bad AI output gets through, and who specifically bears it? That is K. If the answer is "nothing" or "someone three steps downstream," the model says review is not degraded under load — it was never going to occur at any load, and the org chart will show a control that has never once operated.

Why the addition strengthens the essay rather than padding it. The three existing questions are diagnostic and defensible but they are Brandon's assertions. The fourth is derived: it falls out of π⋆ = θ/(κK) with the authors' own definition of K attached, and it is the only one of the four whose remedy costs nothing. The pitch becomes: the human-in-the-loop control has three parameters, your governance framework specifies one of them (the person), and the cheapest of the other two is the one you have never set on purpose. Feeds unchanged plus the queueing KB entry when the scan stages it.

🔍 SOURCES & THREADS

No new KB files today, and that is the correct outcome rather than a thin one. Everything Stafford ran today was verification against primaries the base already cites. The one new primary surfaced — Martinez et al., AJPM 2025 — is a burnout cost simulation with no Five Breakpoints mapping nameable in one honest sentence, so it is skipped under the decision rule and recorded only as the correct citation for a claim the base currently misattributes.

Exploratory slot — spent on author-and-instrument verification, per CLAUDE.md operating rule 2 (verification is explicitly Stafford's lane, not the scan's), and logged to knowledge/query-log.md. Target: the 09-05 scan's own correction, followed one hop further down its chain. Not a repeat of anything in either query log. Yield: 1 author affiliation corrected, 1 construct swap found, 1 job-level range misread as a per-employee range, 1 unsourced 80% figure, 1 pattern proposed for retirement, 3 patterns' recurrence counts affected, and one challenge opened and defeated. Second-order lesson, and it generalizes the 09-03 one: when a correction lands, run its chain to the end the same day. The scan found the secondhand source and fixed the pattern it was standing in; the same source was holding up three others, and the adjacent pattern was in worse shape than the one that got corrected.

Verification note on today's read: the full text of arXiv 2605.27202 was pulled and extracted (20pp.), which is where Eq. 8, the definitions of κ and K, and the "representative calibration" language come from. The scan worked from the abstract and figure captions. Both readings agree; the parameter argument in this brief is only available from the body.

Threads: none due (nearest, middle-manager, 2026-09-09 — the 09-05 scan checked it four days early). None opened, advanced or closed by Stafford.

Standing item, now two days old and unactioned: the AI field-experiment canon. The 09-03 Stafford brief established that the base holds none of it and proposed Dell'Acqua et al. (Jagged Frontier), Peng et al. (Copilot RCT) and Noy & Zhang (Science 2023) as the highest-value acquisitions. Neither the 09-04 nor the 09-05 scan touched them, and the 09-04 exploratory slot went to operations management while the 09-05 slot went to Claim 5. Not a criticism of either choice — both produced ingests. But the BLR challenge is logged CONTESTED with "ingest the rest of the canon" as its named next action, and a CONTESTED challenge whose next action stalls is how a counterargument quietly becomes permanent. Re-proposed below.

Judgment files updated (this repo): thesis-challenges.md (coordination-primacy: tested-and-rejected rival reading + the genre-exposure note); asset-suggestions.md (asset #5, fourth audit question); brandon-development.md (no SPARRING NOTE — weekly cap spent on 08-31/09-01, both still unanswered; the edge-#1 hook parked); query-log.md (today's slot).

Weekly Synthesis (Stafford side — one day late, run because Friday's was missed)

Compact by design: the 09-04 scan ran the shared-state synthesis (patterns, case files, field map, windows, staleness). This covers only the judgment layer this repo owns.

1. Challenge ledger. Five entries. No OPEN challenge is older than 30 days — the oldest (unfalsifiability, 08-29) is seven days old, so the 30-day surfacing rule does not fire. Movement this week: soft-metric DEFEATED (09-01) with an axis correction (09-02); measurability confound OPEN, sharpened and cheaper to settle, next action unchanged and still unrun — one email to Cohen et al. for the Table 8/9 specification with the two score dummies included; BLR counterexample CONTESTED (09-03), next action stalled two days (above); coordination-primacy CONTESTED, weakened today. The ledger's health problem is not age, it is that three of five entries name a cheap next action and none of the three has been run. All three are Brandon's to run, not Stafford's — two require an email and one requires a reading decision.

2. Assets. Five entries: #1 DECLINED/repurposed, #2 drafted v4, #3 (a/b) and #4 PROPOSED with named open calls, #5 PROPOSED and updated today. #4 remains the one with fewest dependencies — two calls outstanding since 09-01 (the T+2 AI outcome definition; S&P 500 vs Russell 3000) plus the author query added 09-02. #5 is the only asset shippable this week without a decision from Brandon, since the audit card ships independent of the essay and now has four questions.

3. Development file. Two SPARRING NOTES outstanding with no recorded response (08-31, 09-01); none issued 09-02 through 09-05, each time deliberately and each time logged. Two hooks are now parked for the first /coach session and they are the same session: the BLR→Garicano hook on growth edge #1 (09-03), and today's — π⋆ = θ/(κK) is the first expression in this base where a Four Forces variable appears as a named parameter in someone else's model. Edge #1 asks Brandon to ground "hierarchy as information routing" in the formal literature; the formal literature has started writing his variables down without him. That is a better prompt than a reading list and it is going stale unanswered.

4. Positioning intelligence — one item, and it is a caution. This system produced two strong corrections in three days (the Appendix B axis error 09-02, the 8.7x chain 09-05) and both were corrections of its own prior work. That is the practice functioning. The thing to watch is the ratio: on 09-05 the scan staged 2 entries and this brief spent its whole slot auditing the base rather than extending it. One day of that is hygiene. A week of it is a research system that has become its own subject, and the window that closes on 2026-09-28 does not care how clean the citations are.

5. Windows — unchanged and stated because the scan is right that it matters. The 08-31 agent-intrusion window closes ~2026-09-28: 23 days, unwritten for the sixth consecutive scan, most reinforced window this base has held. Today adds a second paragraph to it, on top of the one the 09-05 scan supplied: the rebuttal to an agent-oversight argument is "we have a human in the loop," and the answer is now not merely that the human stops checking under load but that whether they ever check is set by a cost the organization has almost certainly never assigned. Nothing else in this brief is time-priced.

Routing

PROPOSED-UPDATE — research-evidence knowledge/emerging-patterns.md (append to the 2026-09-05 correction block on "Manager Engagement = AI Transformation Multiplier")

Extension, 2026-09-05 (via Stafford) — the exposure is four patterns, not two, and two of them fall below the source threshold. State of the Global Workplace 2026 (BrianHeger.com summary) is a Sources-line member in four patterns. In three it is one of exactly two counted members: "Manager Engagement = AI Transformation Multiplier" (Gallup + McKinsey — repaired by substituting Meinen & Mulherin 2026-08-16, same construct); "The Wellbeing Collapse" (Mercer + Gallup — not repaired; SOGW is cited there for thriving/global engagement data, which the new US AI-culture panel does not measure, so this pattern needs its own primary trace or drops to 1 source); "Boreout as Org Design Diagnostic" (Forbes/Ratanjee + Gallup — not repaired, see the separate block below). In "The Inverted Expertise Gap" SOGW is a third supporting source, so the count is unaffected, but the rationale sentence "the 8.7x Gallup multiplier establishes that manager engagement is the transmission mechanism" rests on the retired figure and should be restated against the primary.

PROPOSED-UPDATE — research-evidence knowledge/emerging-patterns.md, "Boreout as Org Design Diagnostic" (retire to 1 source) and knowledge-base/boreout-org-design-failure-2026.md (three corrections)

Retired to 1 source, 2026-09-05 (via Stafford), on three verified chain breaks. (1) Non-independence. The byline is Vibhas Ratanjee (the file records "Vibha Sratanjee" — the name is split wrongly), Gallup's Global Practice Leader for Leadership Development. A Forbes column by a Gallup practice leader plus Gallup's own report via a blog summary is one research house, not two independent sources. Correct the name and add the affiliation wherever it appears. (2) Construct swap in the cost figure. The entry reads "Research published in the American Journal of Preventive Medicine estimates boreout costs US companies $3,999–$20,683 per affected employee annually." The study is Martinez MF, O'Shea KJ, Kern MC, et al., "The Health and Economic Burden of Employee Burnout to U.S. Employers," American Journal of Preventive Medicine 68(4):645–655 (2025), doi 10.1016/j.amepre.2025.01.011. It measures "employee disengagement, overextension, ineffectiveness, and burnout" — not boreout. The range is not per affected employee: it is the gradient across job levels ($3,999 nonmanagerial hourly; $4,257 nonmanagerial salaried; $10,824 manager; $20,683 executive; $5.04M at an average 1,000-person firm). And the method is "a computational model" developed in 2024 representing engagement/burnout states — a simulation output, not a measurement of firms. Strike the sentence; if the citation is kept at all, keep it as the burnout-cost simulation it is. (3) Unsourced headline claim. "80% of most employees' time is coordination theater" (the "80/20 pattern") is attributed to no instrument anywhere in the entry or the pattern. Strike or mark unsourced. (4) Tagging. The entry tags all five breakpoints on a single opinion column with no instrument or sample — the loose tagging both operating rules forbid. Recommend reducing to the mappings the column's own argument supports (Strategic Disconnection, Momentum Mirage) with a rationale sentence each, or holding the entry as a thinker/argument record with no breakpoint tags. What to keep: Ratanjee's org-design argument and the quote "The interventions treat the person. The org chart created the condition." The argument is on-thesis; all three numbers are not. Inherited v1 entry (first logged 2026-07-05) — this is the Hard Rule 10 case: it became load-bearing without being re-verified.

PROPOSED-UPDATE — research-evidence knowledge/field-map.md (Vibhas Ratanjee; and the Brian Heger entry)

New/corrected entry — Vibhas Ratanjee, Gallup (Global Practice Leader, Leadership Development); writes the Forbes column carrying Gallup findings into a general-business audience. First logged in the base 2026-07-05 as "Vibha Sratanjee, Forbes" with no affiliation. Log with the affiliation attached so no pattern ever again counts him and a Gallup report as two independent sources. Argument: boreout as an org-design failure rather than a wellbeing failure — activity metrics decoupled from outcomes at the role level. Brandon's differentiation: Ratanjee reaches the org-chart cause without a diagnostic framework and cannot measure it, because Gallup's instruments are self-report throughout; he names the condition, Brandon names the mechanism and the intervention. Also amend the Brian Heger / Talent Edge Weekly entry, whose text still presents the 8.7x as a "key finding surfaced this week" — that is the summary route the 09-05 scan retired; restate it as the distribution node it is.

PROPOSED-UPDATE — research-evidence knowledge/paper-2-case-file.md (Claim 5 / instrument gap, append)

2026-09-05 (via Stafford) — the readiness claim now has a functional form, and it is not Brandon's. Bartolucci & Vivo's Eq. 8, π⋆(θ) = θ/(κK), gives the review threshold as congestion cost of reviewer time over (verification skill × cost of an escaped error). Two of the three terms are Four Forces variables by the authors' own definitions — κ is "the reviewer's verification skill", K is the cost of an escaped error, i.e. accountability architecture — and only λ and c²_A are exogenous. This is the first expression in this base in which a Five Breakpoints variable appears as a named, estimable parameter inside someone else's model, which is what the readiness-instrument gap (and the unfalsifiability challenge) has lacked since 2026-08-29. Discipline, binding: the paper contains no data — its "Empirical calibration" section is an agenda, and its figure parameters are described as "a representative calibration" and "illustrative levels." The claim licensed is the framework's variables now have a functional form; the claim not licensed is that any of them has been estimated anywhere.

PROPOSED-UPDATE — research-evidence knowledge/story-threads.md, "Does anyone design for verification cost"

Note appended 2026-09-05 (via Stafford), on the entry the 09-05 scan advanced eleven days early. The thread recorded the paper as a theory of its object. Add the parameter reading: criterion (1) — a named organization publishing a verification standard — can now be stated as a testable specification rather than a wish. A verification standard that meets this thread's criterion has to name a load ceiling (θ), an estimate of reviewer effectiveness (κ), and the cost of an escaped error and who bears it (K). None of the eleven instruments in six functions contains the third, and it is the cheapest of the three to set. Criterion (1) remains unmet; what changed is that the thread can now say what "met" would look like.

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

No new candidate slots proposed. American Journal of Preventive Medicine surfaced today as the primary behind a misattributed figure, not as an on-thesis outlet — logged here as a verification touch only, deliberately not a candidate. Per the standing rule that candidates come from finds that clear triage, and this one was correctly skipped.

RE-PROPOSED (2nd time, opened 2026-09-03, unactioned) — the AI field-experiment canon

The base holds none of the canonical AI field experiments. Dell'Acqua et al., Navigating the Jagged Frontier (BCG consultants, 2023); Peng et al., GitHub Copilot RCT (2023); Noy & Zhang, Science (2023). These are the named next action on the BLR counterexample logged CONTESTED in stafford-research/knowledge/thesis-challenges.md, and the fastest route to arguing Claim 3 and Claim 5 from an evidence class rather than from one paper. Two scan days have passed with the exploratory slot spent elsewhere, both times productively. Flagging the drift rather than the choice: a CONTESTED challenge whose next action keeps deferring is how a counterargument becomes permanent by default.

Routing by audience