Stafford Brief — 2026-10-12 (Monday)

The scan found a measured case this morning and connected it to six patterns. It missed the one named after the mechanism. "AI Mandate Backfire" has sat at a single source since April. It now has three, it is ESCALATED, and the third member is a named company with a dated memo. Separately: the scan's own knowledge-base file holds a figure its brief dropped, and that figure is the difference between a sellable diagnostic and an unsellable one.

One correction to this morning's brief, stated up front because Brandon may already have read it. The scan credited MedTech's management with "Commitment demonstrated as willingness to change your own metric when the evidence says it is wrong." There was no evidence. The paper says headquarters "did not review usage or query content" before deciding. Details below.

📖 TODAY'S READ

Same paper as the scan — Gong et al., arXiv:2609.32859v1. Read in full this session. Two things in it did not reach Brandon this morning, and both matter more than the headline.

One: the gaming did not stop when the firm fixed the target. It halved. The share of queries that were repeated or off-task fell from 31% to 18% when the requirement went from 200 a month to 100. The scan's brief, its case-file entry and its routing all cite 31% and none cites 18%. The number is in the scan's own knowledge-base file.

Why this is the most useful number in the paper: the firm did not solve the problem, it reduced it. It halved a bad measure instead of replacing it, and roughly a fifth of the activity it was counting was still not work afterwards. That is the evidence that a correction made without a diagnosis stops early.

Two: the firm had no idea what it was fixing. The paper states that headquarters "did not review usage or query content" before lowering the target, and that the change followed feedback from branches that producing 200 queries was "time-consuming and difficult to complete." The 7% sales rise was a side effect of giving people less work to fake. Nobody at MedTech knew about the 31%, and the organization still doesn't — outside researchers found it afterwards by classifying the text of every query.

So what. This pair is the spine of the Tier 2 argument and of a new challenge. The firm corrected blind, late, and partially. "They fixed it without you" and "they fixed a fifth of it" are the same sentence, and only the second one is in Brandon's favour. Use 18% wherever 31% is used.

🔭 THINKER TO WATCH

No delta. The scan's CAMO entry stands and its promotion test — a third find reached by checking CAMO's own working-paper listing rather than by keyword sweep — is the right one. I spent the slot elsewhere and have nothing to add that isn't a restatement.

🏢 CASE IN THE WILD

Meta told its engineers to use AI, measured them on token consumption, watched them burn 73.7 trillion tokens in a month to climb an internal leaderboard, and rescinded the metric in a memo on 2 September.

The sequence, with dates:

• November 2025 — Chief People Officer Janelle Gale tells employees "AI-driven impact" becomes a core expectation from 2026. (This resolves a lead the base has carried as unverified since 2025-11-17.)
• Earlier 2026 — token usage and AI-assisted programming are added to the performance-management system's metrics.
• ~April–June 2026 — "tokenmaxxing." An internal "Claudeonomics" leaderboard logs 73.7 trillion tokens in just over thirty days. Internal AI spend is reported as heading toward billions. CTO Andrew Bosworth writes that nobody should be using AI tools just for the sake of using them.
• 2026-09-02 — a memo from Maher Saba and Santosh Janardhan tells all engineers Meta will not use AI adoption dashboards or token usage to evaluate employee impact. Managers are told to look at output quality, speed, and the complexity and scope of problems instead. The same memo says 93% of Meta's code changes are now AI-assisted — and that adoption "is not the goal in itself."

The mechanism, as Brandon would say it in a room. Meta did not have an adoption problem. It got to 93% and then told its managers the number was useless for judging anyone. The dashboard was not a thermometer; it was the thing being gamed, and it could not report that about itself.

The open question that makes this a client conversation. Meta is as well instrumented as any company on earth. It still took roughly ten months, a cost line in the billions, and — per a July employee complaint whose allegations are untested — a dispute over whether token usage fed layoff selection, before the metric came out. If Meta's instrumentation could not catch this, whose can?

Sourcing, honestly. The memo was obtained by The Information, whose briefing is paywalled and which I could not read. Everything above comes from two independent restatements three months apart (The Decoder 2026-06-13; BigGo Finance 2026-09-03), both crediting it, with consistent figures, named signatories and a dated document. That is a reported-document chain — much stronger than the unattributed-headline layer this pipeline skips sixteen times a month, and weaker than a read primary. Do not put 93% or 73.7 trillion in print until someone reaches the primary. The pattern below does not depend on either number.

⚙️ FOUR FORCES CONCLUSION

Commitment — and the scan's reading of it should not be carried into anything Brandon writes.

This morning's brief said MedTech "comes out well," calling it "the rarer case: Commitment demonstrated as willingness to change your own metric when the evidence says it is wrong," and told Brandon to keep it because a framework that only diagnoses is less useful than one that can also recognise a correct response.

The paper refutes the premise. Headquarters did not look at the usage or the query content. It acted on complaints that the target was too hard, concluded it was "too demanding," and halved it. That is not follow-through under evidence. It is a concession to friction that happened to land on a better number.

Why the distinction is worth the correction rather than a shrug: Commitment in the Four Forces means decision-makers acting consistently with stated direction under pressure. MedTech's management yielded to pressure and got lucky. If Brandon writes "they changed the metric when the evidence said it was wrong" into an asset or a client deck, the paper's own section 5 contradicts him, and it is the kind of error a reviewer finds in ten minutes.

What the Forces reading should be instead. Purpose is the clean tag and the scan had it right — the objective that reached 5,000 people was a volume proxy. On Commitment the honest entry is an absence: no decision-maker in any of the three cases below held a stated direction under pressure, because no decision-maker knew what was happening. Commitment cannot be scored when nobody was informed.

💭 OPEN QUESTION

In all three cases the organization's own measurement system was the thing being corrupted, and in none of the three did it raise the alarm. Detection arrived from outside it every time: employee complaint at MedTech, employee complaint at Duolingo, a leaderboard culture and a cost line at Meta.

So: has any organization's AI adoption instrument ever flagged its own corruption? I can't find one, and I don't think it's an accident — a volume metric has no term for validity, so there is nothing in it that could fire.

Why this is not rhetorical, and what it changes. Brandon's Strategic Disconnection self-check asks whether ten leaders would describe the transformation's outcome the same way. At MedTech they would have matched perfectly — all ten would have said "200 queries." The self-check passes on a substituted objective. The repair is not a better alignment question; it is a channel question: how would you find out if the number you are reporting were mostly artifact, and who would tell you? On today's evidence the answer for most organizations is "an employee complaint, in about a year."

That is a Tier 3.1 item (repair the self-checks), it costs nothing, and it is the version that runs in a client room.

🥊 CHALLENGE

New, logged as knowledge/thesis-challenges.md → the sufficiency objection (OPEN, 2026-10-12). It is the sharpest thing against the framework that today produced, and it comes from the base's best find.

At full strength: three organizations hit exactly the failure the framework predicts and all three fixed it with none of the framework's apparatus — no breakpoint inventory, no Forces reading, no diagnostic, no consultant. MedTech took four months and gained 7% in sales. And the mechanism isn't new: Gong et al. cite Kerr (1975), Holmström & Milgrom (1991), Courty & Marschke, Oyer, Jacob & Levitt, Benson and Larkin — the whole distorted-measure canon. The relabeling objection says the breakpoints have prior names. This says something worse: on the best-instrumented evidence in the base, one prior name is enough, and the integration claim that survived CFIR last night is doing no work at all.

The answer, and it is two facts rather than an argument. Blind correction was slow — four to twelve months, and at Meta it accumulated a billion-dollar cost line and a layoff-selection complaint first. And blind correction was partial — 31% to 18%, not to zero, because the firm halved a proxy instead of replacing it. The diagnostic's claim is not "we can see what you cannot"; three firms fixed this unseeing. It is "we can see it earlier and more completely than your complaint channel will," which is a claim these three cases support.

What would make it win: a fourth or fifth case where the complaint channel produced a complete correction — the proxy replaced, artifact share driven to zero. Then the partiality limit collapses and the diagnostic is selling speed it cannot prove.

The part that needs Brandon. The honest version of this concedes in print that the client will eventually correct without him. That is the Paper 3 seventy-percent move applied to the value of the service rather than to a number, and it is his call, not mine. It also cannot be settled in prose — it needs Tier 0.2's post-engagement measurement, which is the second challenge in two days to depend on it.

📈 PATTERN BUILDING

"AI Mandate Backfire" — Incentive Fragmentation in AI-First Policies. Count 1 → 3. ESCALATED.

Logged 2026-04-27 on Duolingo alone and untouched since. The scan checked six patterns against today's find — The Borrowed Apparatus, Résumé-Panel Consensus, The Named and Unfunded Constraint, Confidence Without Substrate, The Unpriced Delayering, The Conceded Residual, The Conditional-Productivity Result — and not this one.

Three firms, three countries of incorporation, three industries, three instrument genres — testimony, administrative telemetry, internal document. Each set a quantified AI-usage number against individual evaluation, each got measurable gaming, each withdrew or reduced the number, and every one of them kept the AI.

The claim this earns, stated narrowly enough to survive a reviewer: not "AI mandates fail" — n=3 supports no rate, and Block and KPMG reportedly still run usage metrics. What it supports is a mechanism with a documented retraction rate of three out of three where we can see the end of the story, and one specific thing: the adoption metric never detected its own corruption. Detection came from complaint, culture or cost in every case.

Pre-registered falsifier, so this cannot be argued into looking stronger later: a firm whose AI-usage metric was retracted because the firm's own monitoring flagged the gaming. That kills the claim above and leaves the pattern as a plain gaming result. A fourth complaint-triggered retraction confirms it. Nearest lead: reporting mentions Amazon hitting a similar tokenmaxxing problem — single secondary mention, no primary, not counted.

One honest weakness in the member set. Duolingo is still the weakest leg: CEO remarks on a podcast, relayed. I re-verified it this session (Fortune 2026-04-13, plus TechSpot, WPXI, Entrepreneur) and it holds, but the base's file for it is pre-import-boundary v1 with no URL and no ## Five Breakpoints Intersection heading — so it is structurally outside the ingest pipeline. Repair proposed in Routing.

🪟 WINDOW FLAG

None. Meta's memo is six weeks old and Duolingo's reversal is six months old; neither has a closing date and the pattern will be as citable in March. Registering a window here would be manufacturing urgency. The pattern is publishable now, with no expiry, as an input to asset 12.

🧱 ASSET SUGGESTION

No new numbered asset — master-plan hard stop 1 is unconditional and nothing today earns an exception. Highest number stays 18. Logged as Amendment 2026-10-12 to asset 12 ("What the Five Breakpoints Owe"), the Tier 2 gate.

The pitch. Last night's amendment stripped asset 12 of two of its three novelty claims: CFIR already did the integration and already supplies a diagnostic. What survived was technical and abstract — that Five Breakpoints claims the interaction and sequence CFIR's authors explicitly decline to claim, and that Process Friction and Momentum Mirage have no construct in CFIR at all. Today that abstract claim gets a worked example, and the example lands precisely on the hole:

CFIR has a construct for "organizational incentives and rewards." It has none for the appearance of progress. So the most comprehensive determinant framework in the adjacent field could have flagged MedTech's incentive risk in advance — and would have had nowhere to put the 31%. The three-firm sequence (objective replaced by a countable proxy → gaming becomes rational → activity sustained and read as progress → technology working fine throughout) is four breakpoints in order with a measured consequence at the end. That is the contribution claim, demonstrated instead of asserted.

With it, two things the section must carry: a sentence conceding Kerr 1975 outright and claiming no novelty for Incentive Fragmentation; and the two-line answer to the sufficiency objection, which the example otherwise invites from the floor.

Still blocked on Brandon: placement (Paper 2 revision vs. appendix vs. companion post), open since 09-17. Still blocked on research: Burke–Litwin, unchecked — the one remaining prior-art candidate whose shape would threaten the interaction claim this example rests on. It should take the next exploratory slot, before drafting.

🎓 SPARRING NOTE

Not issued. The 2026-10-02 decision holds: the 09-17 note is the one live note and the cap resumes only when it has a written reply. It does not have one. No eleventh prompt.

Recorded because today is the best material that note will ever get. The 09-17 note asks for one paragraph, for a hostile reader who knows Kotter cold, on what Five Breakpoints claims that Kotter's eight errors do not. The 10-10 entry sharpened it to Satell specifically. Today supplies the other half of the same paragraph, and it is not Kotter — it is Kerr. Three firms repeated a 1975 result that the paper measuring them cites by name. The answer to "what does this add" is now concrete and in hand: the sequence, and a diagnostic that fires before the complaint does. Tier 0.1 is fifteen minutes and it is now nine days older than when the master plan called it the thing that unblocks everything.

SOURCES & THREADS

Candidates: 1 proposed, 0 promoted, 0 expired. The Information — reached today for the first time; it broke both Meta chapters and holds the primary nobody in this base has read. Cadence continuous. Paywalled, which is the reason to log it as a channel to check rather than a source to cite.

Threads: 0 due, 0 checked by me, 2 proposals. I verified the scan's "0 due" accounting independently rather than taking it — it is correct. One file-integrity artifact found doing so: see Routing.

Exploratory slot: spent on whether AI Mandate Backfire has a third member, rather than on the case file's top gap (Claim 3's multi-rater cell, which the scan has hunted on 10-05, 10-06 and 10-11 and which today's find explicitly does not fill). Reason: a pattern one member below threshold for six months is cheaper to settle, and the inherited Duolingo file needed verification anyway. Yield: 1 KB file, 1 pattern escalated, 1 standing lead resolved, 1 new challenge, 1 pipeline defect. Logged in knowledge/query-log.md.

Routing

OPS — This repo: MERGED 340d08e main, verified by ancestry against origin/main after promote-claude-branches folded claude/keen-fermi-p5g2h2 in as 12e3b8d (the direct push to main lost a race with the workflow and was rejected non-fast-forward; the merge is real either way, and origin/main carries this brief, the Meta KB file and all four judgment-file appends). Run branch deleted by the workflow. research-evidence: QUEUED claude/modest-faraday-p5g2h2 — two commits (one knowledge-base file, one staged entry, plus the staging index), scripts/verify.sh green locally, pushed to that branch only and deliberately not to its main, per CLAUDE.md operating rule 3; its promote-claude-branches workflow folds it in after the gate passes. This is the correct terminal state for that repo, not an unmerged run. Yesterday's run merged cleanly; no carried-forward OPS flag.

STAGED — 2026-10-12-the-information-meta-removes-ai-adoption-dashboards · Meta's 2026-09-02 memo removing AI adoption dashboards and token usage from engineer evaluation. 3 breakpoints, 2 Forces, 5 key findings, instrument populated with the reported-document chain and the warning not to print the figures. Lands in entries/pending/2026-10-12/.

OPS / pipeline defect — found while staging, and it is a live trap rather than past damage. build/add-research-json.py matches a Four Forces name anywhere inside the ## Four Forces Intersection section (line ~614, extract_section_rationales, guarded only by exclude=r"mirage"). So a knowledge-base file that excludes a Force in prose naming it — "Momentum and Capability are not tagged: nothing here measures…" — gets that Force staged with the negation as its rationale. My first staging run did exactly that and produced a Momentum tag reading "nothing here measures whether transformation energy decayed." This manufactures precisely the guessed Force tags both CLAUDE.md files forbid. I swept all 728 entry files in entries/merged/ and entries/pending/: exactly one hit, mine, now removed and restaged clean. No historical contamination — the template's stock phrasing ("No direct Force implication — breakpoints-only entry") doesn't name a Force, so the trap only fires when a writer does. Proposed fix, not applied (the ingest script is shared and outside Stafford's write scope): extend the exclude pattern, or skip any candidate line whose rationale matches negation language. I have left the script untouched and rephrased my own file instead.

PROPOSED-UPDATE → knowledge/emerging-patterns.md, anchor: ### Pattern: "AI Mandate Backfire" — Incentive Fragmentation in AI-First Policies

Recurrence count: 3 (was 1) · ESCALATED 2026-10-12 · First seen: 2026-04-27 (Duolingo) Sources: (1) Duolingo — von Ahn on the Silicon Valley Girl podcast, April 2026; re-verified 2026-10-12 against Fortune (2026-04-13), TechSpot, WPXI, Entrepreneur. File: knowledge-base/duolingo-ai-mandate-reversal-2026.md — pre-import-boundary v1, no URL field, no ## Five Breakpoints Intersection heading; repair proposed separately. (2) MedTech Corp — Gong, Hou, Li, Pu & Yao, arXiv:2609.32859v1 (2026-09-26). File: knowledge-base/hku-adaptive-incentive-ai-usage-mandate-2026-09.md. (3) Meta — internal memo 2026-09-02 (Saba, Janardhan), via The Information. File: knowledge-base/meta-token-usage-performance-review-retraction-2026-09.md. What it says. Three named organizations, independently, 2025–26: attached a quantified AI-usage measure to individual performance evaluation, generated documented gaming, and then withdrew or reduced the measure while keeping the technology. Duolingo dropped AI use from reviews after employees asked whether it wanted AI "for AI's sake." MedTech halved a 200-query monthly target; 12% of employee-months sat at exactly 200, 31% of queries were repeated or off-task under the target and 18% under the halved one. Meta removed token usage and adoption dashboards from engineer evaluation after an internal leaderboard logged 73.7 trillion tokens in ~30 days. Five Breakpoints angle. Incentive Fragmentation is the centre of gravity in all three. The full sequence is four breakpoints in order: the objective was replaced by a countable proxy on the way down (Strategic Disconnection), the proxy made gaming individually rational (Incentive Fragmentation), the resulting activity was sustained at target and read as progress (Momentum Mirage), and the technology worked correctly throughout, so nothing in the deployment looked broken (Technology Illusion inverted). This is the base's best demonstration of a breakpoint interaction as opposed to a single condition. The specific claim, and it is narrower than the pattern name. Not "AI mandates fail" — n=3 supports no rate, and Block and KPMG reportedly still run usage metrics. What three of three support is that the adoption metric never detected its own corruption: detection arrived from employee complaint (Duolingo, MedTech), or from visible gaming plus cost (Meta), never from the instrument. A volume measure has no term for validity. Pre-registered falsifier. A firm whose AI-usage metric was retracted because its own monitoring flagged the gaming. That kills the detection claim and leaves a plain gaming result. A fourth complaint-triggered retraction confirms it. Lead, not counted: a single secondary mention of Amazon hitting the same tokenmaxxing problem; no primary located. Limits. Duolingo is CEO testimony relayed through trade press, not measurement. Meta has no outcome measure — nothing shows what its retraction did to output. MedTech is one pseudonymous firm in an unreviewed preprint, and only lowering an existing target is causally identified there. Three genres corroborate a mechanism, not a number. Implication. This is the cleanest answer in the base to "is Momentum Mirage a real thing or a metaphor," and it is the material the Tier 2 asset needs. It also generates a counterargument tracked as the sufficiency objection (stafford-research knowledge/thesis-challenges.md, 2026-10-12): all three firms corrected without any diagnostic, which is answered by the fact that they corrected slowly (4–12 months) and partially (31%→18%).

PROPOSED-UPDATE → knowledge/emerging-patterns.md, anchor: ### Pattern: Paid for Deployment

The standing unverified lead is resolved, in both directions. The entry's "Watch for" records: "one unverified lead exists (HR Grapevine, 2025-11-17, reporting Meta tying employee reviews to 'AI-driven impact' from 2026; no primary located, treat as unsupported)." Corroborated: the policy was issued by Chief People Officer Janelle Gale in November 2025, and token usage plus AI-assisted programming were added to the performance-management system in 2026. And retracted: memo of 2026-09-02 from Maher Saba and Santosh Janardhan. The entry's open question — whether the same activity-not-outcome shape appears below the executive layer — is answered yes, at Meta, in the performance system, and it did not survive ten months. Sourcing is a reported-document chain via The Information (paywalled, unread); see the KB file's standing caution. Note the asymmetry this creates with the pattern's executive-layer members: the activity metric below the executive layer was gamed visibly enough to be withdrawn, while the disclosed executive AI metrics (Qorvo, Juniper, Recursion, Ralph Lauren) have no retraction and no outcome condition. The layer with the least discretion corrected first.

PROPOSED-UPDATE → knowledge/emerging-patterns.md, anchor: ### *The Borrowed Apparatus* (2026-10-12 scan entry)

The registered fourth-member prediction is already falsified, by the paper ingested in the same scan — and the gradient it was meant to test survives. The 10-12 entry sets the next exploratory target as "Incentive Fragmentation against personnel economics' multitasking and distorted-incentives literature, with a registered prediction of neglect." Gong et al. cite that literature directly: Kerr (1975), Holmström & Milgrom (1991), Courty & Marschke (2003, 2004), Oyer (1998), Jacob & Levitt (2003), Benson (2015), Larkin (2014), plus Saez (2010) and Kleven (2016) on bunching at policy thresholds. The Incentive Fragmentation cell is borrowing, not neglect. This is the same error shape the entry correctly caught on itself today — a kill test firing from inside the same scan — repeated one paragraph later in setting tomorrow's slot. Proposed resolution, which strengthens the gradient rather than retiring it: split the cell by altitude. Employee-level incentive response is instrumented and borrowed, exactly as the cost gradient predicts, because the response sits in logs the firm already keeps. Executive-level scorecard misalignment remains unmeasured and unscrapable, which is the cell the prediction was reaching for — and the Paid for Deployment limits confirm it (Equilar: no frame, no count, five named examples). So the gradient is right and the prediction was mis-scoped. Re-register as: does any instrument measure executive-level AI incentive misalignment against named scorecards or contracts? — predicted neglect, and the Equilar limits are the prior evidence for it. Tomorrow's slot should not be spent proving a neglect claim that today's find already refutes.

PROPOSED-UPDATE → knowledge/story-threads.md, anchor: the thread opened 2026-10-12, "Does anyone else measure the artifact share inside a reported adoption number?"

Second data point available on opening day, from the founding source's own text, and it is a dose-response rather than a replication. The thread's first caution is that "31% is a figure under a mandate." Gong et al. §5.1 also report the share under the halved mandate: 18%. Two points, same firm, same instrument, same classifier: the artifact share falls as the requirement loosens, and does not approach zero. This does not satisfy the thread's criterion — it is one firm, still mandated, and the shares are descriptive statistics rather than causal estimates — but it changes what the thread is looking for. Proposed addition to "best development": a third point at or near zero mandate pressure, which would establish whether the curve has a floor. And the thread's sharper form, worth recording: if 18% persists under a halved mandate, the artifact share may be a property of measuring usage at all rather than of the target's level. Next check unchanged, 2026-12-14.

PROPOSED-UPDATE → knowledge/story-threads.md — new thread proposal

"Does Meta's 2027 token budgeting repeat the mistake one level up?" Opened by the same reporting: from 2027 Meta plans budgets and allocations governed by an internal dashboard called "AI Gateway," with automatic alerts for unusual cost spikes. The firm has just removed a consumption metric from individual evaluation and is installing a consumption metric at team and budget level. Significant development: any evidence of whether AI Gateway allocations enter team or manager evaluation, and whether gaming reappears at that altitude. This is the cleanest available natural experiment on whether the failure is the metric or the altitude it is applied at — and it has a date. Proposed next check: 2027-02-15. Caution on opening: everything here is from secondary reporting of a paywalled exclusive, and the 2027 plan may not survive contact.

PROPOSED-UPDATE → knowledge/story-threads.md — file integrity, no thread content

A stale duplicate Next check line, which is the fourth instance of a failure mode already corrected three times in this file (2026-10-03, 10-07, 10-10). The thread "Does anyone design for verification cost, or is it absorbed invisibly?" carries its authoritative header at line 148 (Next check: 2026-11-12, moved there on 2026-10-08) and a second - **Next check:** line at line 187 still reading 2026-09-16, buried between log entries. Anything grepping ^- \*\*Next check on this file — a future scan, or Stafford's own due-thread audit, which is how I found it — gets two answers for one thread, one of them 26 days stale. I checked every thread header for this and it is the only duplicate. Proposed fix: demote line 187 to a dated log line ("Criteria as at 2026-08-27…") so exactly one Next check line exists per thread. This is also what makes the scan's "0 due" correct today — I verified it independently and it holds.

PROPOSED-UPDATE → knowledge/source-candidates.md

The Information — first seen 2026-10-12. Surfacing find: broke both Meta chapters (the June cost/tokenmaxxing story and the 2026-09-02 performance-review retraction), and holds the only primary document behind this base's newest ESCALATED pattern. Signal count 1. Cadence continuous. Specific note for the promotion decision: paywalled, so the realistic use is as a channel to check for leads rather than a citable source — which is a different kind of candidate from anything in this ledger, and worth deciding explicitly rather than by default. Two secondary outlets reliably restate it with attribution (The Decoder, BigGo Finance) and could serve as the reachable layer.

PROPOSED-UPDATE → knowledge-base/duolingo-ai-mandate-reversal-2026.md — repair, not written by me

This file is now a member of an ESCALATED pattern and it cannot reach the pipeline as filed: no **URL:** field (anti-pattern: "No URL = no KB file") and no ## Five Breakpoints Intersection heading (known failure mode 4 — --from-kb requires it; the file has "## Five Breakpoints Analysis"). It is pre-import-boundary v1, so CLAUDE.md rule 10 applies and I re-verified it this session: the account holds. I have not edited it — amending an existing knowledge-base file is a collision risk and the one-file-per-source rule is what keeps the two writers apart. Verified citations to drop in: URL https://fortune.com/2026/04/13/duolingo-ceo-luis-von-ahn-ai-usage-requirement-employee-performance-evaluations/; Date published 2026-04-13 (Fortune; the podcast remarks are April 2026); corroborating — TechSpot (news/112080), WPXI, Entrepreneur via Slashdot (2026-04-19). Von Ahn verbatim, from Fortune: the company was "trying to push something that in some cases did not fit" rather than "being held accountable for the actual outcome." One substantive correction while it is open: reporting states the original AI-first memo was 2025-04-28 and that no formal internal review preceded the reversal — the current file does not say this, and the absence of a review is now load-bearing for the pattern's detection claim.

strategy — Use 18%, not 31%, wherever the MedTech figure appears. 31% is the dramatic number and it is the one that hands the sufficiency objection its best shot. 18% is the number that says a correction made without a diagnosis stops early, and it is the only thing in today's evidence that prices the diagnostic.

strategy — Do not write "they changed the metric when the evidence said it was wrong." This morning's brief recommends it. The paper says headquarters did not review usage or query content. Use "they changed it when people complained it was too hard, and it worked by accident."

governance — An activity metric has reached the layoff-selection allegation stage. A July complaint by former and current Meta employees alleges token usage and other activity measures contributed to layoff selection. Allegations, not findings, and no case document located — do not cite it as fact. Flagged because it is the first instance in this base of an AI activity metric appearing in a legal claim about consequential personnel decisions, which is a different exposure class from the governance material Category 1 keeps returning, and because the FRC/verification thread's open criterion (1) is about exactly this kind of unwritten standard.

personal — Two things only you can move, and both got older today. (1) Tier 0.1, the 09-17 paragraph — fifteen minutes, now the gate on eleven parked coaching notes, and today handed it the Kerr half of its own answer. (2) Tier 0.2, the post-engagement measurement — the second thesis challenge in three days whose resolution depends on it (the McAdam standard on 10-10, the sufficiency objection today). Both new challenges are unanswerable in prose and answerable with three data points. The scan's four open decisions from this morning stand and I have not duplicated them here.