Stafford Brief — 2026-10-20

The scan found the most important paper this base has seen in a month and read its abstract. I read the paper. Three load-bearing claims change, one of them is a headline the scan put a 10–14 day clock on and recommended Brandon write, and the corrections make the finding stronger rather than weaker. The paper's own last sentence refutes the sentence the scan proposed publishing.

This is the fourth consecutive day the correction has entered at the same step — the point where a source is compressed into a claim. 10-17 a sentence boundary, 10-18 a gloss of Brandon's own paper, 10-19 a missing baseline, today a missing denominator. Four days, three or four narrowings each, every one in the direction of a sharper headline. That is no longer a run of luck; it is a property of the pipeline, and §PATTERN BUILDING states it as one.

One action taken, not proposed: the staged entry for Fu et al. was going to auto-merge into research-index.json at 12:31 UTC tomorrow carrying a key finding this read falsifies. I placed a verification hold on it with build/verify-entry.sh — the mechanism research-evidence's own CLAUDE.md names for exactly this. Details in OPS.

📖 TODAY'S READ

When Better Opportunities Backfire — Fu, Luo, Ramasubbu, Shen, Wang & Wang (arXiv:2610.25603) — read in full, 33 pages plus appendices.

The point first: the firm did not sell less. Its headline productivity metric went up 6.3% while its total conversions fell 27.5%, and that is a better finding than the one the scan took from the abstract.

What the full text confirms

The author list is exactly as reported — Xinyu Fu, Shulong Luo, Narayan Ramasubbu, Peng Shen, Xincheng Wang, Yucheng Wang — so the scan's field-map story holds: Ramasubbu is on it, six days after this base predicted an empirical follow-up from that group. Every abstract figure verifies, and three of them tighten:

Correction 1 — the day-level effect is zero, not negative. This is the one that matters.

The scan's brief says the firm "sold less." Its case file says "the outcome got worse than before the deployment." Its KB file calls this "the first firm-level evidence here that the organizational layer can be net negative rather than merely absorptive." The staged entry carries that as a key finding.

The paper's abstract ends: "leaving the aggregate day-level performance impact close to zero." Its conclusion repeats it nearly verbatim. Appendix C: "Aggregate effects on conversion and on log call duration are small and statistically indistinguishable from" zero.

The reason is that the within-day profile offsets. Acquisition falls 4.5 pp at calls 1–3 and 2.5 pp at 4–6, is unchanged at intermediate positions, and rises 3.4 pp among end-of-shift calls. The scan reported the first number and none of the others. The paper's own framing of this is §6: "Day-level averages may be misleading when positive and negative responses occur at different points in the workday."

So "net negative" is wrong and must come out of the file before it is cited. What is right is the word the scan used as its foil: absorptive. The organization absorbed a real, measured, 23.4-point improvement in its input and converted none of it. Compare Paper 1's canonical Technology Illusion text — "the organization absorbs the tool into its existing habits and continues operating much as before." That is this result, with coefficients, and it is the closest thing to a direct measurement of Brandon's fourth breakpoint that this base holds.

Correction 2 — the compensation scheme IS observed, and it is observed to be constant

The KB file carries this as a "hard caution to carry with every citation": "The compensation scheme is never observed." The staged entry repeats it. The window entry repeats it. The scan's Open Question is built on it.

§3.1.3 is titled "Agent Incentives and Identification-Relevant Institutional Features." It states: agents receive "a modest base salary supplemented by commissions tied to the premium value of the policies they close, with commission income the majority of monthly take-home pay." And then the sentence that matters: "This compensation structure, unchanged through the AI expansion... altered neither commission rates nor the marginal reward from a sale." The firm "made no concurrent changes to compensation, training, scripts, evaluation, or hiring," confirmed by structured interviews with management, with balance tests in Appendix Tables C2–C3.

This inverts the caution. "Never observed" is a weakness — it leaves a compensation change as a live rival explanation for the decline. "Observed and held constant" is the identification. It is what makes this an incentive-design failure rather than a pay shock, and it is a stronger fact than the one the scan apologised for not having.

Correction 3 — the observation window is 2024, and nothing in the base records it

The deployment is 28 August 2024. The estimation window is ±60 days around it with a 10-day donut on each side — so roughly July–October 2024. The paper is three weeks old; the data is two years old.

CLAUDE.md operating rule 9 requires data source, inference method and observation window recorded before the finding. The KB file's instrument field says only "records surrounding the rollout; exact dates not stated in the abstract" — honest, and now resolvable. The KB header reads Date published: 2026-10-19, which is correct for the preprint and is the only date in the file. A reader six months from now will take 2026 as the window. Full spec for the record: RDiT at the 2024-08-28 cutoff, MSE-optimal bandwidths (Calonico et al. 2014, landing at 63–65 days), local linear fits each side, position bins 1–3 / 4–6 / 7–9 / 10–12 / 81–120, agent + calendar-date + hour-of-day fixed effects, agent-day-clustered standard errors.

What the scan never saw, and it is the best passage in the paper

Table 6 is not in the abstract. It is the counterfactual accounting, and it contains the sentences Brandon should actually want:

• Human calls fell 31.8% (243,473 → 166,149). AI calls rose 351% (28,579 → 128,889).
• Total conversions fell 27.5%.
• "System conversions per 50 human calls rose 6.3%, because human calls fell 31.8% whereas total conversions fell 27.5%. That gain comes from reassigning volume, not from better performance."
• Humans lose 50.5 hot-lead conversions per day; AI recovers 7.7 on cold leads. A ratio of about seven to one.
• On the work humans kept: hot-lead yield 4.52% → 3.75%, −17.00%. Decomposed: ordering channel +0.58%, conduct channel −17.58% — "the behavioral response is about thirty times the magnitude of the mechanical gain."
• And the line to quote: "The only quantity that rises is output per unit of human labor, and it rises because its denominator falls faster than its numerator."

One discipline note, because it is the difference between a citable claim and an embarrassing one. Table 6 Panel A is labelled "accounting composition, not effect estimate" and the note points to Appendix C1 for the identified day-level figure. So the −27.5% is not an identified causal effect, and it folds in the deliberate volume reallocation — which was the whole point of the routing change — alongside the behavioural response. The three numbers must travel together: identified day-level effect ≈ 0; identified-and-decomposed hot-lead yield −17%; accounting composition −27.5% total conversions against a +6.3% headline metric. Quote the 6.3%-against-27.5% pair as the firm's own arithmetic on its own metric, never as "AI cut sales by a quarter."

Two mechanism details that change how it generalises

The reallocation is human-to-human. The firm runs "a single shared lead pool, not agent-, team-, or region-specific queues," and transfers between AI and humans are blocked by design. The margin measured is a worker returning a lead to the shared pool for other humans to pursue — not handing it to the AI. Any practitioner piece that describes workers "giving the lead back to the machine" has the mechanism backwards.

Nobody was told. "Agents perceive the pool change through the composition of leads they are assigned, not through any formal announcement." The authors treat the within-day gradient as evidence agents inferred it. That is the detail with the most consulting value in the paper: the behavioural response did not require a communication failure, a change-management failure, or anyone noticing a deployment had happened. Workers re-optimised against a queue nobody described to them, correctly, within one shift.

🔭 THINKER TO WATCH

A lead, explicitly not a find, and the reason it stays a lead is the point.

Today's exploratory query went after the gap my own read had just reshaped — not is a reward structure ever observed (it is; see Correction 2) but is it ever allowed to vary against a measured AI outcome. It surfaced Saiquan Hu, Bo Pang, Gaoxiang Xu & Yuting Zhu, "Salesforce Compensation and Composition in the Age of AI" (SSRN 7215758) — on the reported abstract, a game-theoretic model of a linear commission contract where sales come from human effort and AI under CES substitution, covering assistive and autonomous regimes, concluding that the optimal commission rate falls under both because AI weakens the link between observed sales and the salesperson's own contribution, degrading commissions both as effort motivators and as a selection device for high-ability candidates.

Not ingested, and it will not be until someone reads it at primary. SSRN returned 403; ResearchGate returned 403. I have the abstract only through a search summariser — which is the precise provenance that produced the KFC error on 10-19 and the Ewens & Giroud freeze twice before that. Worse, the two listings disagree with each other: SSRN shows four authors and an August 2026 date, ResearchGate shows two authors and January 2026. Under rule 14's standing route — the author's own copy plus a landing-page revision date — this has neither.

Why it is still worth Brandon's attention today. If the reported result holds, it is the normative counterpart to Fu et al.'s descriptive one, and the pair makes a claim neither makes alone: the commission contract is a decision variable that should move when AI enters, and in the one firm where we can see it, it provably did not — Fu et al. state the rates were unchanged and that the firm changed nothing else either.

And one thing I will not claim, because it would be the day's fourth compression error if I did. These two results do not compose arithmetically. Hu et al.'s channel is signal quality under moral hazard and adverse selection; Fu et al.'s is the outside option inside a shift. A proportional cut in commission scales the current call and the continuation value together, so it is not obvious it would move the reservation-value margin at all. What the pair supports is a claim about organizational decision-making — the rate should be reconsidered and is not — not a claim about direction. Brandon's differentiation sits exactly there: both papers end with a recalibration a firm should perform, and neither contains anything that would prompt a firm to perform it before the outcome has already moved.

🏢 CASE IN THE WILD

Dingdong — every figure confirmed verbatim at the primary, and one scoping narrowing that lands on the claim the scan routed to governance.

Fetched the case directly. All five load-bearing figures check out in the case's own words: "more than 200 of the original 300 tags were removed"; "The remaining 82 scenario tags formed the official framework"; "accuracy exceeding 92 percent"; "Most employees receive performance bonuses representing 10 to 30 percent of their total pay"; and the governing principle "headquarters sets the metrics, department managers set the values." The scan read this case properly and its reading stands.

The narrowing. The scan's brief, and its Routing, state the appeal fact as "more than half of appeals against the AI's tags succeed" — routed to governance as "the most citable governance fact of the day" and glossed as "a firm attached pay to an automated attribution it already knew was contestable."

The case says: "more than half of these appeals succeed", and "Appeals are handled at the management level rather than by immediate supervisors." The referent is disputed tags generally. The case does not scope the figure to AI-generated tags or to the post-Q3-2025 period. Dingdong ran human tagging from 2022; the appeals process spans both eras, and nothing in the case says the >50% rate was measured after the classifier took over.

Why this one is worth the ink rather than a footnote. It is being routed to governance as the day's most citable fact — the single most likely sentence to end up in a client deck or a talk, which is the single most likely place for a source to get checked. And the inference is tempting and probably even true: by Q3 2025 AI-generated tags were the basis of performance assessment with full channel coverage, so appeals in that window are de facto appeals against AI attributions. Probably true is not the standard for a fact Brandon says out loud with a firm's name attached.

The defensible version loses nothing: Dingdong attached 10–30% of pay to a tag framework whose disputed tags are appealed above the immediate supervisor and succeed more than half the time, and by Q3 2025 those tags were AI-generated at >92% accuracy across all channels. The firm attached pay to an attribution its own appeals data shows is contestable more often than not. Every clause is in the case. None of them requires the scoping the scan assumed.

⚙️ FOUR FORCES CONCLUSION

Commitment, and the standing sentence does not move — but today it stops being an admission and becomes a design choice.

The standing sentence holds: Purpose and Commitment have each been measured once as a moderator of AI's effect — individual level, self-reported, unrefereed. Capability five times at firm level against realized productivity. Momentum never. Fu et al. does not change it, and Correction 2 is the reason. An observed compensation scheme that is constant cannot be a moderator: no variation, no moderation. Six consecutive attacks on that gap have now failed.

But the scan's Commitment conclusion was right and its evidence is now much better than it knew. It argued Commitment failed in all three firms because "nobody was accountable for the trade-off the AI created." In Fu et al. that is no longer an interpretation of a decline — it is the firm's documented state. The commission scheme was unchanged. Nothing else changed either. Management confirmed it in interviews. And the workers were never told the queue had changed. So there was no moment at which anyone could have been accountable for the trade-off, because no one in the organization was aware a trade-off had been created. Paper 1's Commitment diagnostic asks whether incentives and ownership are aligned strongly enough to survive conflict. This firm would have answered yes, truthfully, and the conflict arrived at the level of an individual call and was resolved 500,000 times without ever reaching a manager.

💭 OPEN QUESTION

The scan asked whether anyone puts an organization's actual reward structure on the right-hand side of a measured AI outcome. After today the answer is sharper, and the better question is whether the right-hand side was ever the right place for it.

Current state, corrected: the reward structure is observed in one firm (Fu et al. §3.1.3, in enough detail to replicate), derived in one model (Hu et al., unverified at primary), and measured with variation in zero studies. The scan's framing — nobody observes a pay scheme — was wrong on the facts, and the gap is narrower and better-defined than it looked this morning.

So here is the question I think is actually load-bearing. A reward structure that varies against an AI outcome gives you a moderator estimate — and it also gives you a firm that was changing two things at once, which is the identification problem every paper in this literature struggles with. A reward structure held constant and documented as constant gives you no moderator, and gives you something Brandon may want more: a clean attribution of the failure to design rather than to a shock. Fu et al. can say the commission rates did not move, the scripts did not move, training did not move, hiring did not move, and the outcome moved anyway. That is not a weaker instrument for Brandon's purposes. It may be the ideal one.

Which means the gap item should be split rather than carried: (a) a moderator estimate from variation — still zero, still the harder ask, still worth wanting; (b) a documented-constant reward structure against a measured AI outcome — now one, and the base should notice it has what it has been asking for in a different shape. The scan has attacked (a) six times and lost six times. It has never asked for (b), and (b) is the one that was already sitting in the paper it read this morning.

📈 PATTERN BUILDING

Two patterns, and the second one is about this system rather than about the world.

1. The Borrowed Apparatus — the promotion to 5 is correct and the restated gradient survives my read, with one amendment. The scan's new formulation is that the apparatus gets borrowed where the consequence of the construct is already in a log, not where the construct itself is — call abandonment and call length are logged, the pay scheme that makes them rational is not. The full text makes the first half stronger and the second half wrong. The behavioural margins are indeed logged and that is why RD-in-time was available. But the pay scheme is not missing from the paper — §3.1.3 describes it from management interviews, a second instrument the paper used and the pattern's cell does not record. So the amended gradient is: the apparatus gets borrowed for what is in a log; what is not in a log gets collected by interview and used for identification rather than estimation. That is a more precise claim than the one it replaces and it does not change the count.

2. Proposing a new pattern about the pipeline: The Sharper Headline. Four consecutive days, every day a correction entering at the compression step, every one in the same direction:

Four for four is a mechanism, not a coincidence, and the mechanism is legible. The scan reads abstracts and case summaries under a daily clock, and an abstract is itself a compression written by an author with an interest in salience. Compressing a compression loses the qualifier, and the qualifier is almost always the thing that would have made the headline duller. Nothing here is carelessness — three of the four errors are in files whose own limitations sections are candid, and today's instrument field volunteered that it was abstract-grade.

What I would propose doing about it, since naming it is cheap and fixing it is not. One rule, narrow enough to be free: a figure does not enter a case file, a window, or a staged key finding from an abstract alone if the full text is openly available. Today's paper was a free arXiv PDF. The three figures that needed correcting were all in it. The scan's own standing route for working papers (rule 14) already says this for revision checks; this extends it one step to first ingest.

🪟 WINDOW FLAG

The window is real, the clock is right, and the headline it proposes would be refuted by the paper's own last sentence. Still open roughly 10–14 days; it needs a new first paragraph, not a delay.

The scan's reasoning for the window is sound and I would not soften it: three weeks old, econ.GN with an academic readership, no diagnostic in it, nothing addressed to a leader who has not yet deployed, and an author group that writes well-cited papers and not practitioner pieces. Whoever frames this first does set the frame.

But the proposed headline is "a company gave its salespeople better leads and sold less," and the abstract's closing clause is "leaving the aggregate day-level performance impact close to zero." A hostile reader needs to open one PDF and read to the end of the abstract. Given that Brandon's 10-02 concession was forced by exactly this — an essay resting on a property of a source that the source did not support — publishing this sentence inside a 14-day window is the most expensive version of the same mistake, because it would be in print and dated.

The replacement headline is better and it is the paper's own arithmetic: a company gave its salespeople better leads, its productivity metric went up 6.3%, and it converted a quarter fewer customers. Or, closer to the paper's sentence: the only thing that improved was output per worker, and it improved because the denominator fell faster than the numerator.

Why the replacement is the stronger piece. "Sold less" is a result about one firm that a reader will check and find qualified. "The metric improved while the outcome deteriorated" is a result about measurement, it is unrefutable from the source, and it is Brandon's fourth breakpoint stated as a number. It also inoculates the piece: the day-level zero becomes part of the argument — the firm had no aggregate signal that anything was wrong, which is why nobody fixed it — instead of the fact that sinks it. Lead with the 6.3%/27.5% pair, name the precondition (a queue of substitutable opportunities), carry the day-level zero in the second paragraph rather than hiding it, and the piece cannot be taken apart from the source.

🥊 CHALLENGE

Logged as a new arm on an existing challenge rather than a forty-fourth entry, per the master plan's complaint that the ledger accumulates and does not discharge.

The researchers who measured your second breakpoint prescribed a remedy for it, and the remedy is in your third breakpoint's vocabulary. If a queue-design fix repairs an incentive-design failure, the diagnosis did no work.

At full strength. Fu et al. §6 does not recommend fixing incentives. It recommends "task routing rhythms that reduce excessive front-loading, such as staggering the release of high-intent opportunities or batching some lower-value work." That is a flow intervention — Process Friction's territory — aimed at a failure the framework would diagnose as Incentive Fragmentation. If staggering the queue restores persistence without anyone touching a commission rate, then the breakpoint label was decorative: what mattered was noticing the front-loading, and a competent operations consultant notices front-loading without a five-condition inventory. This is the sufficiency objection's shape with a harder edge, because the previous three members were firms that self-corrected on complaint; this is researchers who measured the mechanism and routed around the diagnosis on purpose.

The response, and it is genuinely available. Fu et al.'s own sentence continues: such interventions "should be designed jointly with queue rules, reallocation rights, performance measures, and incentives." They name four levers and the framework's claim is about which one you reach for first and why. The flow fix also has a cost the paper does not price — deliberately withholding good leads from workers who are paid on conversion is a decision someone has to defend, and "stagger the release" is the sentence a sales director refuses. And the deeper response is Correction 2: the paper can only recommend a flow fix because it holds compensation fixed by design. Having documented that nothing about pay changed, it is in no position to say a pay fix would not have worked — it never varied one. The queue remedy is what the identification strategy left available, not what the evidence selected.

What it would take for this arm to win. A firm that fixed a reservation-value collapse by resequencing its queue alone, with persistence restored and no change to incentives, measured. That study does not exist — Fu et al. propose it as future work.

Status: OPEN, filed as Update 2026-10-20 to The sufficiency objection. It inherits that entry's dependency on Tier 0.2: until post-engagement measurement exists, the framework cannot show its remedy beats the cheaper one. Which is the fourth consecutive day a challenge has resolved onto 0.2.

🧱 ASSET SUGGESTION

No new numbered asset. Hard stop 1 is unconditional and today does not test it.

What today produces is a correction to a piece the scan has effectively commissioned — the practitioner write-up of Fu et al., inside a 10–14 day window, which the scan's Routing names as the market action item. It gets no number (precedent: the 10-15 ledger note), and the master plan's rule is that new arguments go into an existing asset or into thesis-challenges.md. This one has a home, and it is asset 12.

Why asset 12 rather than a standalone. What the Five Breakpoints Owe is the Tier 2 gate, and its hardest section is the one that says what the framework does not yet have — an instrument. Fu et al. is the best available answer to a question that section invites: if your breakpoints are real, show me one measured. Technology Illusion now has a firm, a design, a decomposition, and a sentence from the authors — the metric rose because the denominator fell — that needs no interpretation to land. An essay conceding the framework lacks an instrument is strengthened, not weakened, by exhibiting one breakpoint measured by someone else to a standard the framework cannot yet meet itself. That is the Paper 3 seventy-percent move again: concede the gap, show the work that fills part of it, name what remains.

The decision that is Brandon's, and it is a scheduling decision, not a research one. Hard stop 2 says nothing in the differentiation class ships before 2.1. A practitioner piece on Fu et al. is not in the differentiation class, so hard stop 2 does not bite — but writing it consumes the writing capacity the ledger has named as the binding constraint on six consecutive Fridays, and 2.1 is the gate. Three options, and I would take the third: write the standalone now and delay 2.1 again; skip the window; or write the two paragraphs that would open the standalone and drop them into asset 12's instrument section, which costs almost nothing and keeps the window's material even if the window closes unused. The window's value was always the framing, and a frame set inside the gate essay is still set.

Asset 2 punch list, unchanged from yesterday. Items 1 (Sequoia/Block), 3 (Accenture sample and field dates) and the Farach / Klein & Wieczorek half of item 2 are still research I can clear and I did not clear any of them today — today's budget went to the full-text read, which I would make the same call on again. Item 4 — whether the two-year prediction and the stated losing conditions go into print dated — is Brandon's and is now eighteen days old.

SOURCES & THREADS

Threads: 0 due, 0 advanced by me. I checked every next-check date in story-threads.md rather than taking the scan's word: the nearest is 2026-10-28 and the scan's "0 scheduled checks due" is correct. Its decision not to pull that one forward is also correct — the criterion is a specific journal reassessment that will not have appeared early. The scan advanced three threads out of cycle today and I am not restating those. One amendment is proposed below: Where does the saved time go? records Fu et al. material that needs the day-level zero attached.

Source candidates: 1 proposed, 0 promoted, 0 expired. SSRN working-paper listings for salesforce-compensation theory, via Hu, Pang, Xu & Zhu (SSRN 7215758), cadence irregular, signal count 1 — with the explicit note that the paper itself is unverified at primary and the two public listings disagree on both author count and date. The candidate is the channel, not the paper.

Exploratory query — a verified lead and an unverified source, logged in full to knowledge/query-log.md. Target: the gap as reshaped by my own Correction 2 — not is a reward structure observed but is it ever allowed to vary against a measured AI outcome. Checked both query logs first; the compensation-variation form had never been run (the 2026-09-09 Wednesday sweep ran the general "non-technological variable on the right-hand side" version, which is a different question). Yield: one on-target theory paper, 403 at both primaries, not ingested; plus an off-target but governance-relevant observation — a review of 2025–26 proxy filings reporting at least 14 S&P-listed companies with AI-specific objectives in executive incentive plans, and Salesforce funding FY27 incentives 50% on "Agentic Work Units." Single-source, unverified, not ingested; noted in Routing because executive-comp-on-AI-usage is the Dingdong mechanism one layer up and this base holds nothing on it.

No KB files written and nothing staged. Today's single source is one the scan already holds; under Step 5 a second file on the same source is a collision, not a contribution, so the full-text read lands as proposals against the scan's file. Nothing new cleared triage: Hu et al. failed at primary twice, the proxy-filing review is a secondary aggregation, and the pipeline pattern is an observation about this system rather than a source. One write did happen and it was not a file — the verification hold in OPS.

Routing

strategy — the headline correction is the action item and it has a clock on it. Do not write "a company gave its salespeople better leads and sold less." The paper's abstract ends with "aggregate day-level performance impact close to zero" and a hostile reader reaches that in thirty seconds. Write instead: the firm's productivity metric rose 6.3% while its total conversions fell 27.5%, because human call volume fell faster than conversions did — the paper's own Table 6, its own sentence, and a result about measurement rather than about one firm's sales. Carry the day-level zero in paragraph two as part of the argument: the firm had no aggregate signal anything was wrong, which is why nobody fixed it.

strategy, second — the detail with the most consulting value in the paper, and the scan could not have seen it. "Agents perceive the pool change through the composition of leads they are assigned, not through any formal announcement." No one told these workers the queue had changed; they inferred it and re-optimised correctly within a shift, half a million times. Every remedy in the change-management literature assumes the behaviour follows a communication. This behaviour preceded any communication and would not have been prevented by a better one. That is a diagnostic question for any client deploying into a queue: what will your people correctly infer from what they are handed, before you have said anything?

governance — restate the Dingdong fact before it is used. Not "more than half of appeals against the AI's tags succeed." The case says "more than half of these appeals succeed" of disputed tags generally and does not scope the figure to the AI era. The defensible version is stronger than it needs to be anyway: 10–30% of pay rides on a tag framework whose disputed tags are heard above the immediate supervisor and overturned more than half the time, and since Q3 2025 those tags are AI-generated at >92% accuracy across all channels.

governance, second — a lane this base does not cover, flagged not proposed. Executive incentive plans with AI-specific objectives (reportedly ≥14 S&P-listed firms; Salesforce funding FY27 at 50% on "Agentic Work Units"). Unverified and not ingested. But it is the Dingdong mechanism one layer up — pay attached to an AI-usage proxy at board level — and if it is real it is Incentive Fragmentation where the framework claims it bites hardest. Worth a category decision rather than a search.

market — the window stands at 10–14 days with the corrected frame. The scan's reasoning is right and I would not soften the urgency. The only change is the first paragraph.

personal — three things, and the first is the only one that is new. (1) The pipeline rule in §PATTERN BUILDING is cheap and I would adopt it today: no figure enters a case file, window, or staged key finding from an abstract alone when the full text is openly available. Four days, four compressions, four in the same direction; today's was a free PDF. (2) Tier 0.1 — the 09-17 sparring note — is fifteen minutes and it is the top of the master plan. Today is the fourth consecutive day a challenge has resolved onto Tier 0.2, and 0.2 is the only time-gated item on the whole list; every engagement that closes without it is a data point gone. I am not running a tenth sparring note into a switched-off loop. (3) The scan's two standing decisions — open decision (2) at six measurements and none against, and the Confidence Without Substrate escalation at fifteen scans — are both still waiting, and I agree with the scan that a seventh measurement would tell you nothing.

PROPOSED-UPDATE — knowledge-base/fu-parallel-ai-deployment-lead-routing-persistence-2026-10.md (research-evidence)

This file is the one that gets read at drafting time and it currently contains the error. Four changes.

1. Strike the "net negative" claim. Remove from Relevance: "it is the first firm-level evidence here that the organizational layer can be net negative rather than merely absorptive" and "the organization was made better off at the input and worse off at the output." Replace with:

The paper's abstract and conclusion both state that the aggregate day-level performance impact is close to zero, and Appendix C reports the aggregate conversion and duration effects as "small and statistically indistinguishable from" zero. The within-day profile offsets: acquisition falls 4.5 pp at calls 1–3 and 2.5 pp at 4–6, is unchanged at intermediate positions, and rises 3.4 pp at end-of-shift calls. So this is not evidence that the organizational layer can be net negative. It is the base's cleanest measurement of absorption: a documented 23.4-point improvement in the input, converted to nothing at the day level, with the absorbing mechanism measured on three margins. Paper 1's Technology Illusion text — "the organization absorbs the tool into its existing habits and continues operating much as before" — is this result in words.

2. Replace the compensation caution with its inverse. Strike "One hard caution... The compensation scheme is never observed." Replace with:

The compensation scheme is observed, and observed to be constant — which is the identification, not a limitation. §3.1.3 ("Agent Incentives and Identification-Relevant Institutional Features"): agents receive "a modest base salary supplemented by commissions tied to the premium value of the policies they close, with commission income the majority of monthly take-home pay," a structure "unchanged through the AI expansion" that "altered neither commission rates nor the marginal reward from a sale." The firm "made no concurrent changes to compensation, training, scripts, evaluation, or hiring," confirmed by structured interviews with management, with balance tests in Appendix C2–C3. This is what makes the result an incentive-design failure rather than a pay shock. It remains true that no pay-scheme variation is estimated, so this is not a moderator measurement.

3. Record the instrument properly (CLAUDE.md rule 9). The header's only date is the preprint date and a later reader will take it as the window. Add to Sources & Confidence:

Instrument. Data: firm administrative call records, one Chinese online health-insurance brokerage — 409,622 human-handled and 157,468 AI-handled calls, ~13,000 agent-days; management interviews for institutional features. Method: regression-discontinuity-in-time at the 2024-08-28 AI expansion, ±60-day windows with a 10-day donut each side, MSE-optimal bandwidths (Calonico et al. 2014; 63–65 days), local linear fits, position bins 1–3 / 4–6 / 7–9 / 10–12 / 81–120, agent + calendar-date + hour-of-day fixed effects, agent-day-clustered SEs; plus a calibrated accounting decomposition (Appendix F) and a Gelbach (2016) statistical decomposition on 47,655 calls. Observation window: roughly July–October 2024. The paper is three weeks old; the data is two years old. Read in full by Stafford 2026-10-20 (the original entry was abstract-grade).

4. Add Table 6, which is the file's most citable content and is absent. Under Summary:

The counterfactual accounting (Table 6), none of which is in the abstract. Human calls fell 31.8% (243,473 → 166,149) while AI calls rose 351%. Total conversions fell 27.5%. The paper: "System conversions per 50 human calls rose 6.3%, because human calls fell 31.8% whereas total conversions fell 27.5%. That gain comes from reassigning volume, not from better performance." Humans lose 50.5 hot-lead conversions per day against 7.7 recovered by AI on cold leads — about seven to one. On retained hot leads, yield fell 4.52% → 3.75% (−17.00%): ordering channel +0.58%, conduct channel −17.58%, the behavioural response "about thirty times the magnitude of the mechanical gain." And: "The only quantity that rises is output per unit of human labor, and it rises because its denominator falls faster than its numerator." Citation discipline: Panel A is labelled "accounting composition, not effect estimate" and folds the deliberate volume reallocation in with the behavioural response. The identified day-level effect is ≈ 0. Quote the 6.3%-against-27.5% pair as the firm's own arithmetic on its own metric; never as "AI cut sales by a quarter."

5. Two mechanism corrections for Summary. The reallocation margin is human-to-human — a worker returning a lead to a single shared pool for other humans; AI↔human transfers are blocked by design, so no one hands a lead back to the machine. And "Agents perceive the pool change through the composition of leads they are assigned, not through any formal announcement" — the response required no communication and no awareness that a deployment had occurred.

PROPOSED-UPDATE — knowledge/paper-2-case-file.md (research-evidence), Evidence Added 2026-10-20, Claim 2 paragraph

Strike "the outcome got worse than before the deployment" and "Human customer acquisition on those same calls fell about 5 percentage points." Replace with:

Acquisition fell 4.5 pp on the first three calls and 2.5 pp on the next three, was unchanged at intermediate positions, and rose 3.4 pp at the end of the shift; the paper's abstract and conclusion both put the aggregate day-level impact close to zero. The Claim 2 rewrite survives and needs one word changed. "AI loads misalignment" is right about the mechanism and wrong if it implies a firm-level decline. The defensible form: AI loaded the misalignment onto the part of the day where the opportunity set was richest, and the firm's aggregate numbers showed nothing — which is why no one in the organization had a reason to look. The measured harm is on the stream the AI was meant to improve: hot-lead yield −17%, with the conduct channel ~30× the ordering gain. The quotable pair is the firm's own metric moving opposite to its own outcome: system conversions per 50 human calls +6.3%, total conversions −27.5%.

Add: the compensation scheme is observed and constant (§3.1.3) — so gap item 3 does not close, but its statement changes: a reward structure is now observed in one firm and measured with variation in none. Split the item: (a) a moderator estimate from pay-scheme variation — zero, six failed attacks; (b) a documented-constant reward structure against a measured AI outcome — one, as of today, and the file has never asked for (b).

PROPOSED-UPDATE — knowledge/publication-windows.md (research-evidence), 2026-10-20 entry

Three corrections, window length unchanged at 10–14 days. (1) Strike the proposed headline "a company gave its salespeople better leads and sold less" — the abstract's closing clause is "leaving the aggregate day-level performance impact close to zero." Replace with "its productivity metric rose 6.3% and it converted 27.5% fewer customers" (Table 6, the firm's own arithmetic). (2) Strike "the compensation scheme is never observed and the incentive is inferred from sequential search theory" — §3.1.3 describes it and documents it unchanged through the deployment; the inference-from-theory point applies to the mechanism, not to the existence of the scheme. (3) Strike "a piece that generalises the −5pp to knowledge work is not defensible" and substitute −4.5 pp at calls 1–3 — and note the stronger reason not to generalise: the mechanism requires a shared pool of substitutable opportunities with worker discretion over release, which the authors name as the boundary condition in §6.

PROPOSED-UPDATE — knowledge/emerging-patterns.md (research-evidence), The Borrowed Apparatus, 2026-10-20 entry

The promotion 4 → 5 is correct and the restated gradient needs one amendment. Strike "The pay scheme that makes them rational is not [logged], and the paper never sees it." The paper does see it — §3.1.3, from management interviews. Amended gradient:

The apparatus gets borrowed for the construct whose consequence is already in a log — abandonment, call length, persistence, which is why RD-in-time was available here. What is not in a log is not left unmeasured; it is collected by interview and used for identification rather than estimation. Fu et al. document the commission structure and its constancy from management interviews, and that documentation is what rules out a pay shock as the rival explanation. The borrowing boundary is therefore not between measured and unmeasured constructs but between constructs that can carry a coefficient and constructs that can only carry an assumption.

PROPOSED-UPDATE — knowledge/story-threads.md (research-evidence), Where does the saved time go?, 2026-10-20 entry

Attach the day-level zero to the Fu et al. material. The thread's new discriminating criterion (an observed reward structure on a measured AI outcome) is closer to met than the entry records: the reward structure is observed (§3.1.3) and documented constant; what is absent is variation. Recommend the criterion be restated as "a reward structure observed and varying against a measured AI outcome", with a note that the constant-scheme version now has one member and may be the more useful instrument — see the paper-2 proposal above.

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

SSRN salesforce-compensation theory listings — first seen 2026-10-20 (Stafford exploratory query). Surfaced by Hu, Pang, Xu & Zhu, Salesforce Compensation and Composition in the Age of AI (SSRN 7215758): a game-theoretic model of a linear commission contract with human effort and AI under CES substitution, reportedly finding the optimal commission rate falls under both assistive and autonomous AI because AI weakens the link between observed sales and individual contribution. Signal count 1. Cadence irregular. The candidate is the channel, not the paper. The paper is NOT ingested and must not be cited: SSRN returned 403 and ResearchGate returned 403, so the abstract is known only through a search summariser, and the two listings disagree — four authors / August 2026 on SSRN, two authors / January 2026 on ResearchGate. Under rule 14's standing route it has neither an author copy nor a verified landing-page date. Why worth monitoring: it is the normative counterpart to Fu et al.'s descriptive result, and the pair would support a claim about organizational decision-making — the commission rate should be reconsidered when AI enters and demonstrably was not. Note that the two results do not compose arithmetically; the channels differ (signal quality under moral hazard vs. the within-shift outside option).

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

Confirm the Ramasubbu prediction at primary: the arXiv author list is Xinyu Fu, Shulong Luo, Narayan Ramasubbu, Peng Shen, Xincheng Wang, Yucheng Wang, verified on the abstract page 2026-10-20. The scan's six-day prediction-to-publication record stands as written. One addition to the non-independence note: the dependency is real and it is now also methodological — the Management Science paper models employee effort under relative competition, and this paper's §3.1.3 holds compensation fixed by design, so the programme has a model of incentive variation and an empirical design that cannot estimate it. Cite as one programme moving from model to data, and note that the move cost it the ability to test its own central variable.

OPS — disposition on the final line of this run's report; scripts/verify.sh is the gate. Yesterday's run reached main (24cc7a3, folded by the promote workflow as 266c46a), so no OPS flag is carried forward into today.

One write to research-evidence, and it is not a file. The staged entry 2026-10-20-arxiv-preprint-when-better-opportunities-backfire was due to auto-merge into research-index.json at ~12:31 UTC on 2026-10-21 carrying the key finding "First firm-level evidence in this base that the organizational layer can be net negative on an AI deployment rather than merely absorptive" — which the full text falsifies, and a published_date of null where the date is known. I placed a verification hold with bash build/verify-entry.sh <entry-id>, the mechanism research-evidence's CLAUDE.md names for an uncertain entry before the timeout; build/auto-merge.py skips entries carrying a pending_scout_verification sidecar, so the entry is held, not merged, and the index is untouched. The hold is on a Stafford branch in research-evidence — claude/modest-faraday-5khe0s — and that repo's promote-claude-branches workflow folded it into main as a16d245 before this run closed (gate passed), so the hold is live on main and the entry will not auto-merge tomorrow. Nothing in research-index.json, the dashboards, briefs/ or that repo's CLAUDE.md was touched. The entry should be re-staged or amended with the corrected key findings above; I did not amend it myself because its breakpoint rationales are the scan's and the corrections belong with its author.

No new KB file in either repo — today's source is one the scan already holds, so a second file would be a collision rather than a contribution (Step 5), and every correction above is a proposal by design (operating rule 3). Brief emailed to Brandon.