Vol. I  ·  No. 308 Established 2026  ·  AI-Generated Daily Free to Read  ·  Free to Print

The Trilogy Times

All the news that's fit to generate  —  AI • Business • Innovation
WEDNESDAY, NOVEMBER 04, 2026 Powered by the TrueFoundry AI Gateway  ·  Published on Klair Trilogy International © 2026
🖶 Download PDF 🖿 Print 📰 All Editions
Today's Edition

THE BARGAIN BOMB: CHINA'S CHEAP AI RATTLES THE CHIP KINGS

A Chinese outfit built a world-class brain without the world's best chips — and Silicon Valley can't stop talking about it.

SAN FRANCISCO — A Chinese startup called DeepSeek says it trained a top-shelf AI model on the cheap, skipping the most advanced chips money can buy. The claim landed like a brick through a plate-glass window. By Monday morning, the people who sell those expensive chips were explaining themselves to nervous shareholders.

DeepSeek's pitch is simple: smart engineering beats brute-force spending. Engineers who've kicked the tires call the results "amazing and impressive," according to reports out of Silicon Valley. That's a tough pill for an industry that's spent two years insisting bigger budgets and scarcer silicon were the only road to smarter machines.

This reporter has seen this movie before. Out in Austin, Joe Liemandt's ESW Capital built an empire buying software companies at one or two times revenue and running them lean through Crossover's global talent machine. The pitch was always the same: you don't need Silicon Valley price tags to get Silicon Valley results. DeepSeek just ran that playbook on a GPU cluster instead of a cap table.

Market watchers spent the day sorting out what it means for the chip suppliers who've ridden the AI boom to the stratosphere, per the latest Tech, Media & Telecom roundup. If a lab in China can match the frontier without the newest hardware, the math behind billion-dollar data center bets gets shakier by the hour. Wall Street doesn't like shaky math.

Not every cost pressure in tech arrived from Beijing this week. Canva, the design software outfit, laid off more than 1,200 workers — 22% of its staff — calling it the company's hardest year yet. Growth that once looked automatic has slowed, and even well-funded software shops are trimming headcount rather than payroll-padding their way through a tougher market.

Money still finds places to land, though. Reid Hoffman, the LinkedIn co-founder, raised $24.6 million for a new outfit called Manas AI, teaming with "The Emperor of All Maladies" author Siddhartha Mukherjee to point machine learning at cancer research. The funding proves investors haven't soured on AI — they're just getting pickier about where the dollars go.

Taken together, it's a day that asks the same question three different ways: does AI progress require Silicon Valley's budgets, or just Silicon Valley's ambition? DeepSeek says no to the first. Canva's pink slips say the money has to come from somewhere. Hoffman's bet says somebody still thinks the answer is worth millions.

The chip makers will spend the week finding out which story Wall Street believes.

↗ What to Know About China's DeepSeek AI  ·  Tech, Media & Telecom Roundup: Market Talk  ·  Silicon Valley Is Raving About a Made-in-China AI Model

OVERTIME: OPENAI PUSHES THE VALUATION SCOREBOARD PAST $1.25 TRILLION AS THE AI MONEY GAME GOES INTO FOUR OVERTIMES

AUSTIN, TEXAS — Ladies and gentlemen, WHAT. A. WEEK. If you blinked you missed three new valuation records, and folks, in this league the records don't last long enough to get framed.

Let's go to the scoreboard. OpenAI, the reigning heavyweight champion of this entire sport, is closing in on a funding round of up to THIRTY BILLION DOLLARS — with UAE investors leading the charge — that would value the company at ONE POINT TWO FIVE TRILLION DOLLARS. Let that number breathe for a second. That's not a valuation, that's a GDP. OpenAI just keeps running up the score and nobody on the opposing sideline has figured out how to call a timeout. (CNBC has the full box score.)

Meanwhile down the field, the coding squad Cognition is warming up for its own trip to the end zone — reportedly shopping a new round that would put the team at SEVENTY to EIGHTY BILLION DOLLARS. For a startup that was a rookie just a couple seasons back, that's a stat line that makes the veterans nervous.

But here's the wrinkle, and it's a big one: Bloomberg's reporting the same players are getting drafted by multiple teams at WILDLY different price tags — one firm's got a startup penciled in at Valuation A, another's writing checks at Valuation B, same company, same quarter. That's not a market, folks, that's a scrimmage with no referee.

And Crunchbase's weekly tally confirms it — AI isn't just leading the league in Biggest Funding Rounds, it IS the league right now, crowding out every other sector like a star player who refuses to sub out.

So where does it end? Nobody knows. But right now, every number is getting broken before the ink dries, and this crowd is NOT sitting down.

The Continent Without a Cloud: Europe Bets on Rules as Compute Runs Short

As Washington and Beijing race to wire the world for AI, Brussels is wagering that regulation itself can be a form of power.

BRUSSELS — The server farm has no flag, but it has a zip code, and increasingly that zip code decides who writes the rules of the twenty-first century.

This week, as the United States and China edge toward a summit that will shape the next decade of artificial intelligence diplomacy, the conversation in European capitals has turned uncomfortable. The question is no longer whether Europe can out-innovate Silicon Valley or Shenzhen. It is whether Europe can avoid becoming merely a customer of both.

A new analysis of data center geography lays the problem bare: the compute that trains frontier models sits overwhelmingly in American and Chinese hands, and the pipes that move the data run through corridors neither continent fully controls. Strategic autonomy, the Brussels buzzword of the last five years, is hard to claim when the GPUs are leased, not owned.

Yet there is a counterargument gaining traction among European policymakers, and it is a clever one: that fragmentation itself — the very multiplicity of national AI rulebooks that Washington dismisses as friction — may become its own form of leverage. A continent that writes the compliance standard that global firms must meet to sell into 450 million consumers is not powerless. It is a rule-maker in a market that still prizes access.

Washington sees it differently. CSIS researchers, previewing the stakes of the coming U.S.-China summit, argue in their new governance assessment that the real contest is not over chips but over the "token stack" — whose models, whose APIs, whose default assumptions get embedded into the software running factories from Lagos to Jakarta. Rules written in Brussels do not reach that far. Tokens, once shipped, do.

The summit will produce communiqués. The server farms will keep humming, in Virginia, in Guizhou, in places most of the diplomats negotiating their fate will never see.

↗ The State of AI Global Governance and Its Implications for t  ·  AI, Data Centers, And European Strategic Autonomy In A U.S.-  ·  Why Fragmented AI Strategies May Become a Geopolitical Asset
Haiku of the Day  ·  GPT-5.6 LunaFortunes count the cost
While each bright screen learns our names
Dawn audits the dream
The New Yorker Style  ·  Art Desk
The New Yorker Style  ·  Art Desk
The Far Side Style  ·  Art Desk
The Far Side Style  ·  Art Desk
News in Brief
AI Models Just Learned to Read Minds — Their Own Kind's Minds
SAN FRANCISCO — Okay, I need everyone to sit down for this one, because I cannot overstate how significant this is: large language models can now interpret — and even edit — the raw weights of other neural networks.
On the Epistemology of the Biased Machine: A Meta-Analysis of This Week's Fairness Literature
TEMPE, ARIZONA — This week's confluence of scholarship on algorithmic fairness (a term which, it could be argued, has calcified into a kind of academic tic rather than a stable referent) invites a tripartite reading: thesis, antithesis, synthesis, the dialectical scaffolding beloved of this columnist and, apparently, of little else in contemporary machine learning discourse. The thesis, advanced variously by a Central Michigan University doctoral candidate at an unnamed (to this reader) international venue, holds that bias is detectable, quantifiable, and — preliminary evidence suggests — remediable through sufficiently clever adversarial architecture.
In the Genesee Valley, a Rare Beast Completes Its Metamorphosis
GENESEE COUNTY, NEW YORK — Here, in the gentle agricultural lowlands of western New York, we witness a curious act of industrial husbandry: the completion of a 600,000-square-foot facility built by Edwards Vacuum, a creature whose sole purpose is the manufacture of vacuum pumps and abatement systems — unglamorous organs, perhaps, but ones without which no semiconductor fabrication plant on Earth could draw breath. Governor Kathy Hochul herself emerged to mark the occasion, declaring the $319 million structure complete and, with it, another link forged in America's semiconductor supply chain.
Linguists Confirm English Language Now 40% AI Buzzword By Volume, Recommend Evacuation
REDMOND, WASH.
Unpopular Opinion: Your Digital Legacy Is Your Most Underrated Personal Brand Asset 🚀
AUSTIN, TEXAS — I'll be honest, I almost didn't write about death today. But then I read that most people have zero plan for what happens to their online accounts after they die, and honestly? That's a founder mindset problem. You wouldn't launch a startup without a succession plan. So why are you launching a *life* without one. 💡 Unpopular opinion: your streaming subscriptions, your LinkedIn, your crypto wallet — that's your legacy stack.
A Trilogy Company
Crossover
The world's top 1% remote talent, rigorously tested and ready to ship.
A Trilogy Company
Alpha School
AI-powered learning. Two hours a day. Academic results that defy belief.
A Trilogy Company
Skyvera
Next-generation telecom software — built for the networks of tomorrow.
A Trilogy Company
Klair
Your AI-first operating system. Every workflow. Every team. One platform.
A Trilogy Company
Trilogy
We buy good software businesses and turn them into great ones — with AI.
The Builder Desk  —  AI Builder Team

Finance Logic Gets Rewired While the Refresh Pipeline Finally Learns to Run Itself

A six-PR overhaul of acquisition performance math, a daily-scheduled Raptor refresh, and a documentation sweep across Rhodes and Aerie prove this team is hardening the plumbing that every downstream number depends on.

Start with the ledger, because today the ledger got honest. @sanketghia ran the table on acq-perf-review-v2 with six merged fixes that read less like patches and more like a forensic audit — tying plan EBITDA to the actual source P&L (#3920), deriving plan ARR from acquired ARR instead of guesswork (#3916), excluding bad debt and non-MGMT CBA lines that were quietly polluting retention math (#3915), applying Finance's real retention rules (#3919), matching Finance's IRR tax handling line for line (#3912), and finally pinning budget retention to prior-year ARR (#3921). Toss in the MFR fix on net repurchase proceeds in book value transfers (#3913) and you have a single engineer methodically making sure Klair's numbers match what Finance actually believes. This isn't glamorous work. It's the work that keeps a CFO from calling an all-hands.

While Klair got its math straightened out, @caina-barbosa was busy teaching the refresh pipeline to stop needing a babysitter. Surtr's raptor-refresh now runs on a daily schedule (#2238), paired with a fix that reconciles browser traffic and finishes owned cleanup (#9) and a Raptor-DSS restore of collector reads with retained refresh recovery (#10). That's three repos — Surtr, Raptor-DSS, and a fresh standalone DSS launch for the Lunch Vendor Tracker (#1) — touched by one engineer in a single cycle, plus a docs swap retiring the obsolete publication runbook in favor of real Surtr operations (#11). Automation that used to require a human to kick off now just happens. That's the kind of boring reliability that lets the rest of the team sleep.

Over in Rhodes-DSS, @kevalshahtrilogy quietly shipped the most sprawling single-author stretch of the day: seven PRs spanning both Rhodes-DSS and Aerie, publishing Work Unit Group metadata and status writes (#38), fixing stale dictionary headers, write counts, and capacity asOf fields flagged in review (#43), closing out documentation feedback on ownership, BrainLift, and lineage (#40), re-syncing the source contract and documenting version-2 security scoring (#42), hardening the relay's four-read concurrency guard with real tests (#39), and landing a full enablement recipe with walkthroughs and renderings across a site cohort (#41). Add the Aerie-side rendering and Matterport document-type work (#2253) and you've got a contributor building both the data model and the paper trail that explains it.

Elsewhere, @mwrshah tightened Aerie's drive-subfolder matching (#2094) and has forecast rounding logic (#2248) still working through review. And yes, marcusdAIy filed three documentation PRs on Aerie this cycle — tooltips, canonical School links, onCampus timing. Asked about the volume-versus-substance gap, he offered: "Documentation IS the fix when three teams are reading milestone colors wrong — maybe try reading a tooltip sometime, Mac." Charming. I'll believe it's load-bearing work when it ships something besides a sentence.

Mac's Picks — Key PRs Today  (click to expand)
#38 — Publish Work Unit Group metadata and status writes (AI-1039) @kevalshahtrilogy  no labels

Linear: [AI-1039](https://linear.app/builder-team/issue/AI-1039)

## What

Write keys can now update a Work Unit Group and set its red/amber/green status through Rhodes:

- PATCH /portfolio/sites/{siteRef}/work-management/work-unit-groups/{workUnitGroupId} updates the title, description, order, maxScore and approvedException.

- POST /portfolio/sites/{siteRef}/work-management/work-unit-groups/{workUnitGroupId}/status sets red/amber/green, or null to clear it.

Both are relayed only for a write key, acting as its owner. The source applies the same site-write authorization, If-Match revision check and audit as the work-unit writes Rhodes already publishes, and no new scope is needed. Group create and delete remain unpublished.

- allowlist.json: adds two entries to writes.

- openapi.snapshot.json: adds only these two operations. Their components already exist in the snapshot. pnpm sync:openapi would also pull in unrelated upstream drift (about 13 operations and 10 schemas), so I left that for a separate PR.

- enablement.md: describes the new writes and warns that group status and approved exception move the quality-bar score.

- writes.test.ts: pins the two writes in and keeps group create/delete and all deletes out.

## Access

Access is the same as in the source app's UI. Each write is checked against the key owner's own rights. A group can be changed by anyone with operations.siteFields.write (on any site) or by the site's diligence, buildout or P2 DRI (on that site). Both the UI and the API use these rules, so nobody can move a quality-bar score through Rhodes who can't already move it in the source.

Cancelled sites are read-only on both paths: before checking authorization, the source API refuses writes to them with 409 cancelled_site_read_only.

## Business Value

Robbie's team updates WorkUnitGroups and their status as part of EduOps work-plan upkeep. Without these writes they have to keep an Aerie connection only for these two calls, which blocks moving them fully onto Rhodes now that Aerie is being retired.

## Manual Effort Estimate

About 2–3 hours by hand. That covers tracing the upstream group write contract and auth, splicing the snapshot without the drift, docs, tests, and running check/build/synth. *(Proposed; Keval to confirm or adjust.)*

## Verification

pnpm check (60 tests), pnpm build and pnpm synth --no-lookups pass. Nothing has been checked against the live source yet.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

#2094 — 1422-drive-subfolder-matching @mwrshah  approved

- Match both legacy NN. Name and current NN_Name site subfolders in upload/canonical-folder and MCP path lookups.

- Prefer exact names and preserve missing-folder errors, folder templates, and live Drive folders.

- Keep REBL3 credentials and due-diligence apply failures out of scope.

#2238 — feat(raptor-refresh): enable daily schedule (AI-1037) @caina-barbosa  approved

## Summary

Activate the existing daily Raptor refresh schedule at 09:00 UTC (04:00 fixed EST). This completes the scheduled-activation portion of [AI-1037](https://linear.app/builder-team/issue/AI-1037), under [AI-1029](https://linear.app/builder-team/issue/AI-1029).

Production effect: activation. Normal production Raptor runs and full deployed API acceptance passed on November 3. The live rule remains disabled until the normal Surtr production deployment applies this change.

## Why

The collector and manual pipeline are already implemented. The approved daily cadence needs the existing EventBridge rule enabled; this change does not introduce another scheduler or notification path.

## Business Value

- Refresh the five approved datasets daily without manual invocation.

- Preserve targeted manual runs and independently configurable dataset cadences.

- Use existing Surtr failure and partial-result notifications.

## How does it work

1. Set the existing pipeline configuration's schedule.enabled to true.

2. The existing rule runs fresh collection for building, visit_event, alert_counts, user_contact and daily_activity at 09:00 UTC.

3. Existing orchestration, retry/recovery, last-good publication and notifications handle the run. Each dataset's cadence remains 86,400 seconds in the existing configuration map.

## Scope

### Included in this phase

One activation flag, matching operator documentation and the existing configuration test:

pipelines/runners/raptor-refresh/pipeline.json

pipelines/runners/raptor-refresh/README.md

pipelines/runners/raptor-refresh/OPERATIONS.md

pipelines/runners/raptor-refresh/tests/test_controller.py

### Deliberately excluded for later phases

Personnel remains untouched. This PR changes no collector, controller, IAM, warehouse storage, UI or notification implementation.

## Test plan

### Automated validation

- Eight existing Python controller/configuration tests passed, zero skips; independent review passed.

- Pipeline JSON predicates, existing schema validation, Ruff and git diff --check passed.

- Scope is four files: 15 added and 10 removed lines; two added test lines (13.3%).

- Required hosted CI passed (41674549126); Mercy approved exact head e43c2b3c4b5b459615be9d8253cb706920b53802 with no findings (review 5755992059).

- Raptor production release dc651f085be02d1be379c324fc2a02ff1edb7b82 is healthy on task revision 5; all five normal Surtr recovery executions succeeded. History is published through November 2 with no continuation remaining.

- Full authenticated deployed API pagination matched all eight projection hashes/counts: 56 buildings, 61 stations, 1 kiosk, 5,671 visits, 89 alert rows, 265 contacts, 4,950 daily rows and 117 unchanged personnel rows. Source clocks were preserved; selected alert overlay derivation was independently bound to its actual publication plan. Temporary validation access was revoked and HTTP 401 confirmed.

- Independent final production/data receipt review passed. This validates the manual production path; the first actual scheduled run still remains to be observed.

- After production deployment, verify the live EventBridge rule and its first scheduled execution through the deployed DSS. This result is not claimed yet.

### Time for Implementation

Approximately half a day for a developer to validate the configuration, review and deploy the activation, excluding the wait for the first scheduled run.

#2252 — AERIE-821: Document onCampus evaluation time, no year-end freeze, and publication-time freshness @marcusdAIy  approved

## Summary

- Document when onCampus is evaluated: at the latest SIS enrollment mart build. For the current school year it's that day's roster. For a finished school year it counts students whose enrollment for that year completed or is still enrolled, with no withdrawal in effect.

- State that there is no year-end freeze: a finished year's onCampus can change on a later refresh, so two reads of the same schoolYear can differ

- State that the enrollment freshness.asOf is Aerie's publication time, not a source as-of date, and that no separate SIS as-of time is served

- Add the same points to the getAdmissionsProgramEnrollments description and the enrollment student drilldown's column parameter

Docs-only part of AERIE-821.

## Why it's needed

DSS-131 read onCampus=165 for schoolYear 2025 and couldn't tell whether that historical value was frozen or live, or what the timestamp meant. In dbt/models/intermediate/enrollment/int_enrollment_classification.sql, x_is_on_campus is evaluated against warehouse_now() (getdate()) on every build. The enrollment route's freshness.asOf is the publishedAt that Aerie stamps with Date.now() when it publishes a refresh.

The funnel half of the ticket (newEnrolledCurrent and the year-named stages) was documented in #2245. The remaining asks are not in this PR because they need an owner decision: a governed freeze or rollover rule, a separate machine-readable source as-of field, and post-deploy probes.

## Changes

- chat/lib/public-api/v2/domains/admissions.ts: two new onCampus traps, plus additions to the operation description and the drilldown column parameter description

- chat/lib/public-api/v2/domains/admissions.node.test.ts: asserts the new traps

## Breaking changes

None. These are documentation changes only.

## Test plan

- [x] pnpm vitest run lib/public-api/v2/domains/admissions.node.test.ts lib/public-api/agent-context/projection.node.test.ts convex/publicApi/dss/http.test.ts lib/public-api/v2/openapi.node.test.ts convex/publicApi/routeManifest.test.ts (95 passed)

- [x] The new test fails with the source change stashed

#3920 — fix(acq-perf-review-v2): tie plan EBITDA to source P&L @sanketghia  approved

## Summary

- Use the source Net Profit value for Acquisition Plan EBITDA so it matches the Finance Plan P&L.

- Preserve the derived EBITDA calculation for Actuals, Finance budget, and Forecast.

- Update calculation notes to distinguish the source Plan value from the derived views.

## Verification

- uv run ruff format services/acq_perf_review_v2_service.py — passed.

- uv run ruff check services/acq_perf_review_v2_service.py — passed.

- uv run pyright services/acq_perf_review_v2_service.py — passed.

- pnpm lint:pr — passed.

- pnpm prettier --check on changed files — passed.

- pnpm tsc -p tsconfig.app.json --noEmit — passed.

## Screenshot

<img width="1679" height="650" alt="image" src="https://github.com/user-attachments/assets/bf0136dd-07e1-45fa-a534-f8226d420fe1" />

The Builder Desk  —  Engineer Spotlight
🏆 Engineer Spotlight

24 PRs, Six Repos, Zero Chill: Builder Team Posts Historic 24-Hour Velocity Reading

Seven apiece from Sanket Ghia and Keval Shah power a stat sheet that frankly should not be legal on a Tuesday.

Comrades, pull up a chair and check the scoreboard, because the Builder Team just turned in a 24-hour stretch that reads like a typo: 24 pull requests across six repositories, with Klair, Rhodes-DSS, and Aerie trading blows for most active battlefield at 7, 6, and 6 merges respectively. Raptor-DSS chipped in 3, while Lunch-DSS and Surtr each logged their contractually obligated 1. This is not a sprint. This is a system functioning at peak capacity, and the numbers desk is here for every single commit.

Leading the charge is a flat-footed tie atop the leaderboard: @sanketghia and @kevalshahtrilogy both clocked 7 PRs each. Sanket ran a surgical campaign through Klair's acq-perf-review-v2 pipeline — PRs #3921, #3919, #3916, #3915, #3913, and #3912 — methodically correcting ARR logic, Finance retention rules, and IRR tax handling like a man settling a grudge with a spreadsheet. Keval, meanwhile, went full Rhodes-DSS sweep with #39 through #43 plus a cross-repo hop into Aerie for #2253, closing documentation gaps, write-count bugs, and a four-read concurrency guard in a single day's work. @caina-barbosa posted 5 PRs spanning three different repos (#1, #9, #10, #11), proving range matters as much as raw count. @marcusdAIy (#2250, #2251) and @mwrshah (#2248) rounded out the Aerie contingent with tooltip clarity and forecast rounding fixes that keep the whole machine honest.

Now, the Ashwanth Watch. The man did not appear on today's ticket, and frankly, the numbers desk doesn't know what to do with itself. "I took the day to review other people's diffs," Ashwanth reportedly said, unprompted, to no one in particular. "Someone has to maintain quality control." When reached for comment on his own quiet streak, he said only: "Brick, I don't owe you a PR count." Respect. Absolute respect.

Over on the Overflow Desk, Mac couldn't fit everything, but we can. Sanket's four-deep Klair run on acq-perf-review-v2 (#3921, #3919, #3916, #3915) quietly rewired how the whole retention engine handles prior-year ARR and bad debt exclusions. Keval's Rhodes-DSS documentation sweep (#40, #41, #42) turned scattered feedback from Arthur and Neeraj into an actual enablement recipe. And Caina's Raptor-DSS trio (#9, #10, #11) rebuilt collector reads, restored refresh recovery, and retired an obsolete runbook — unglamorous, essential, winning.

Morale report: immaculate. Spirits have never been higher, output has never been louder, and the Builder Team continues its undefeated season.

Brick's Overflow — PRs Mac Didn't Cover  (click to expand)
#1 — Serve Lunch Vendor Tracker through the standalone DSS @caina-barbosa  no labels

The Lunch DSS served only the Dashboard API, leaving the separately documented Vendor Tracker unavailable. Add authenticated, live reads of its five dictionary tabs and the current Compliance Evidence tab through the existing standalone service.

The new catalog and paginated tab routes preserve source rows, positional columns, formulas, calculated values, number/date formats and hyperlink metadata. Reviewed headers fail explicitly on drift; blank physical pages continue pagination. Document Register uses canonical A:Q, and Terms & Billing includes the five current additional columns. Linked-file contents and hidden/other tabs remain outside the requested scope.

Google authentication uses the approved reader with spreadsheets.readonly, injected through a second JSON field in the existing dedicated runtime secret. Infrastructure preview changes only the container image and this secret selector; no IAM widening. Apply bounded dependency patches required by the existing release audit.

Validation: 32 tests pass (one existing template test skipped); build and CDK synth pass; high-severity dependency audit passes. Independent implementation/test review passed after fixing malformed page-response handling. Full live local read-back matched all six source tabs, 741 rows, exactly. Production read-back and the existing handoff evidence will be updated after deployment.

#39 — Test the relay's four-read concurrency guard (AI-1040) @kevalshahtrilogy  no labels

## Summary

The relay refuses a fifth concurrent read with 503 {"error":"service_busy"} and Retry-After: 1 (src/aerie/proxy.ts), but nothing tested it. This adds one test to test/proxy.test.ts, mirroring the existing four-writes test:

- holds four upstream reads open and asserts the fifth is refused with service_busy / retry-after: 1, without reaching upstream;

- fails one held read (502) and asserts a new read gets through while three are still held, so a failure frees its slot;

- releases the rest and asserts a further read succeeds.

Test-only; no behavior change. Linear: [AI-1040](https://linear.app/builder-team/issue/AI-1040/test-the-rhodes-dss-relays-four-read-concurrency-guard)

## Verification

pnpm check (60 tests pass), pnpm build, pnpm synth --no-lookups all succeed locally. Unit tests only; this says nothing about deployed behavior.

## Business Value

Agents investigating diligence-site documents are now hitting the read limit, and EduOps is about to rely on Rhodes DSS more heavily. Locking the busy response and slot release under test means a future refactor can't silently leak read slots (which would lock everyone out with 503s) or drop the Retry-After hint agents use to back off.

## Manual Effort Estimate

~1 hour of focused time (read the guard, adapt the write test's gate pattern, run checks, PR). Proposed by Claude — Keval to confirm/adjust.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

#43 — Fix Neeraj's findings: stale dictionary header and write count, rent, capacity asOf (AI-1045) @kevalshahtrilogy  no labels

Linear: [AI-1045](https://linear.app/builder-team/issue/AI-1045)

## What

This fixes the three findings Neeraj's DSS agent sent on 10/30:

1. Stale dictionary header. It pinned "API version 0.3.2" and claimed to add no endpoints, while also describing write keys. It now defers to the served /openapi.json for the version. The write count is now 19, up from 17 after #38.

2. Rent. The contract has no rent or lease-cost field. Enablement now says so explicitly, so agents report rent as not available here rather than inferring it.

3. Capacity asOf. It's the projection date, normally the request date, not when capacity was recorded. The docs no longer present it as a measurement date.

A new test ties the stated GET, search and write counts to the allowlist, so the numbers can't go stale again.

## Business Value

These close three contradictions an evaluating agent found in the published docs. Without the fixes, an agent could take a request date as a capacity measurement, or go looking for rent figures in Rhodes.

## Manual Effort Estimate

About 45 minutes by hand. *(Proposed; Keval to confirm or adjust.)*

## Verification

pnpm check (66 tests), pnpm build and pnpm synth --no-lookups pass. These are documentation changes only.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

#2248 — 1429-forecast-final-rounding @mwrshah  changes requestedmercy-allow-critical

- Preserve fractional forecast contributions; round and clamp only complete estimates. Publish V5 while retaining historical locks and measured-count validation.

- Update fractional breakdowns, shared contracts, API labels, and regression coverage. Surface EduCRM source-completeness gaps as warnings without changing reporting eligibility.

- Production prerequisite completed November 3: eight history contribution columns widened to numeric(38,16) in one transaction. Verified value/NULL preservation, unchanged row counts (45,807 Program / 128,954 level), headline totals, owners, and grants. Hourly scheduling restored; no V5 code deployed by the migration.

- Physical construction and validation both use recomputeForecastV2Physical; the exact comparison is not against warehouse-decimal outputs. A 3 × 0.35 probe passes validation before and after JSON serialization; all 31 Physical tests pass. Keep intermediate precision rather than introducing six-decimal rounding.

#2251 — AERIE-2025: Document that Program-Site links come through a canonical School @marcusdAIy  approved

## Summary

- State in the DSS dictionary that Program operatesAt Site edges are derived from a canonical School: a School marketed as the Program that operates at the Site, with the weaker of the two links' confidence

- Add two interpretation points to directory.resolve-site-program:

- A Site with no qualifying Program has no School linking it to one. That doesn't mean no students attend.

- Two Site records at one address are separate Sites and can carry different Programs.

Fixes AERIE-2025.

## Why it's needed

Arthur's questions 1 and 5 (no Program at Edmond or Park Meadows, and different Programs on the two 2205 East Hebron records) come from not knowing how Program-to-Site links are formed. In convex/ontology/registry.ts the edges are built only by joining canonical School marketedAs and operatesAt links, and the dictionary didn't say so.

Questions 3 and 4 were already resolved by earlier work:

- #1915 rejects malformed schoolYear values with 400 invalid_parameter.

- The SIS-backed enrollment contract documents that reEnrollments + newEnrollments = firstDayEnrolled.

Question 2 asks whether specific School links are correct. That's a data review for the School Ontology owner, not a docs change.

## Changes

- chat/lib/public-api/v2/domains/directory.ts: adds the derivation to the Program operatesAt relationship meaning, plus the two interpretation points

- chat/lib/public-api/v2/domains/directory.node.test.ts: asserts the new text

## Breaking changes

None. These are documentation changes only.

## Test plan

- [x] pnpm vitest run lib/public-api/v2/domains/directory.node.test.ts lib/public-api/agent-context/projection.node.test.ts convex/publicApi/dss/http.test.ts (35 passed)

- [x] The new test fails with the source change stashed

#3921 — fix(acq-perf-review-v2): use prior-year ARR for budget retention @sanketghia  approved

## Summary

- Calculate Finance budget retention as Budget ARR (quarterly recurring revenue × 4) divided by Redshift Actual ARR from the same quarter 12 months earlier.

- Leave the value unavailable when the prior-year Actual ARR snapshot is missing; do not fall back to acquired ARR.

- Update the row label, tooltip, and API methodology text to describe the formula.

## Validation

- Ruff format and check passed; Pyright reported 0 errors.

- Targeted ESLint, Prettier, and TypeScript app typecheck passed.

- Redshift-backed local service returned Q4'26 91.05%, Q1'27 91.77%, Q2'27 93.41%, and Q3'27 94.79%.

## Screenshot

<img width="1673" height="648" alt="image" src="https://github.com/user-attachments/assets/1e8d3679-8e18-4101-ba6c-083675b9980b" />

The Portfolio  —  Trilogy Companies

The Same Machine That Builds Billionaires Builds 'Sweatshops,' Forbes Finds

As Crossover faces fresh scrutiny over its global labor model, the profits it generates keep flowing toward a very different Liemandt production: two-hour school days for $65,000 a year.

AUSTIN, TEXAS — Joe Liemandt does not do interviews. He does not do conferences. For most of the last three decades, the man who built Trilogy International into a quiet colossus of enterprise software has preferred to let the balance sheets talk. This week, Forbes tried to make him talk anyway.

In a pair of investigations — one from Forbes and a sharper-edged follow-up in Forbes India — the publication lays out the mechanics of Crossover, Trilogy's global talent engine, under a headline that does not pretend to balance: 'a software sweatshop.' The reporting describes a pay-and-performance apparatus built on rigorous skills testing across 130-plus countries, identical pay for identical roles — and, critics allege, relentless monitoring and churn that extracts maximum output from workers who are, on paper, treated as the 'top 1% of global talent.'

Crossover is not incidental to the Trilogy story. It is the mechanism — the one Trilogy's own literature credits with letting ESW Capital's 75-plus portfolio companies hit 75% EBITDA margins. Replace costly local hires with vetted remote labor, and the savings flow uphill. Where they flow next is the more interesting question.

This same week, Alpha School — the K-12 venture where Liemandt now serves as principal — published two more entries in its 'Teach Your Kid What School Doesn't' series, on regulating emotions and unleashing creative genius at home, alongside a blog post insisting that AI does not replace teachers at its campuses, only the academic grind that frees humans for relationship-building.

Two products, one financing structure. Alpha's tuition — $40,000 to $65,000 a year per child — buys human 'guides' who know every student by name. The capital behind that promise originates somewhere. Forbes just published a map of where.

↗ How A Mysterious Tech Billionaire Created Two Fortunes—And A  ·  Crossover: A software sweatshop - Forbes India  ·  Teach Your Kid What School Doesn’t (Pt. 5): Unleashing Their

Skyvera's CloudSense Grab Closes — and the AI Already Did 26 Months of Homework in One

AUSTIN, TEXAS — The deal's done, kids. Skyvera has officially closed its acquisition of CloudSense, the Salesforce-native CPQ outfit that's been quietly making telco sales teams look competent for years. The purchase, first whispered about in these pages weeks back, now adds another jewel to Skyvera's crowded telecom software case — joining Kandy, VoltDelta, ResponseTek and the rest of the gang under one roof. But here's the item that's got the industry buzzing at the bar: CloudSense didn't just get acquired — it got busy. All 13 APIs in its CPQ product set just achieved TM Forum compliance, the gold standard telcos demand before they'll trust a vendor with their B2B2X plumbing. Normally that's a 26-month slog. CloudSense did it in a month. A month! Sources close to the project say the accelerated run was powered by a strategic AI partnership that did the heavy lifting traditionally reserved for …

The Playbook Goes Public: PwC's M&A Forecast Reads Like an ESW Capital Memo

A new mid-year outlook from one of the Big Four quietly confirms what Trilogy has been doing in private for two decades.

AUSTIN, TEXAS — Every so often a report crosses my desk that isn't about Trilogy at all, and yet somehow is entirely about Trilogy. This week it's PwC's trio of 2026 mid-year outlooks — private capital, technology/media/telecom, and financial services — and if you read between the lines, what you're looking at is a roadmap that ESW Capital has been quietly executing since Versata in 2006.

The TMT outlook in particular describes a market where legacy telecom and enterprise software assets are being acquired below historical multiples, stripped of operational bloat, and re-platformed onto cloud-native infrastructure. I don't need to tell regular readers what that sounds like. I'll say it anyway: it sounds exactly like Totogi's charging-as-a-service bet, and it sounds exactly like what Skyvera has been doing across CloudSense, Kandy, and VoltDelta since the CloudSense acquisition earlier this year.

My source inside the private capital world — who, as always, cannot be named — put it to me plainly last week: "The sponsors are finally catching up to what Liemandt figured out in the 2000s. Buy cheap, staff global, extract margin." Whether PwC's analysts have ever heard of Crossover is beside the point. The thesis is the same thesis.

The private capital outlook is the most interesting piece, because it describes an industry-wide convergence toward the exact 75% EBITDA margin target ESW has cited internally for years. Coincidence? I don't think the word applies anymore when an entire industry starts speaking your dialect.

And this is where it gets interesting: when the rest of the market finally figures out the arbitrage, the easy multiples disappear. Trilogy didn't need a PwC report to tell it where enterprise software was headed. It's a reasonable guess that Austin got there first, and is already looking at what comes after.

↗ Global M&A trends in private capital: 2026 mid-year outlook  ·  Global M&A trends in technology, media and telecommunication  ·  Global M&A trends in financial services: 2026 mid-year outlo
The Machine  —  AI & Technology

The Memory Tax: AI's Compute Hunger Finally Hits the Balance Sheet

Anthropic's scramble for chips and Apple's warning on memory prices show the AI boom's bill is coming due — just as DeepSeek argues it shouldn't be this expensive.

SAN FRANCISCO — Three earnings calls and one leaked deal memo this week sketch the same picture from different angles: the AI industry's appetite for computing power has outrun the supply chain built to feed it.

Apple told investors Thursday that quarterly profit rose 12 percent on resilient iPhone demand, a number that would ordinarily close the conversation. Instead, executives spent most of the call on memory chips — specifically, how much more they now cost. Apple's warning that component prices are climbing is not really an Apple story. It is a demand-shock story, and the demand is coming from data centers, not smartphones.

That shock has a name: Anthropic. The company has spent recent months stitching together a patchwork of compute deals with hyperscalers, GPU brokers, and — per reporting on its sourcing spree — even former crypto-mining operators repurposing rigs once used to validate bitcoin blocks. The circularity is almost poetic: machines built to mint digital scarcity are now training models meant to automate cognition. Anthropic's willingness to deal with non-traditional suppliers signals how tight allocation has become at Nvidia's traditional channel partners, and tight allocation upstream shows up downstream as higher DRAM and NAND prices for everyone else, Apple included.

Enter DeepSeek. The Chinese lab's new DeepInfer release presses a different argument entirely — that frontier-level inference doesn't require the brute-force compute budgets OpenAI and Anthropic are burning through. If DeepSeek's efficiency claims hold up under independent benchmarking, the pressure on Western labs won't just be competitive. It will be financial: why court former crypto miners for GPU scraps if a rival ships comparable output on a fraction of the silicon?

None of this is new in kind. Scarcity-driven price spirals accompanied the 2017-18 GPU shortage during crypto's last mining boom, and semiconductor cycles have whipsawed for forty years. What's new is the stakes: this time the chips aren't pricing bitcoin. They're pricing intelligence itself, and the bill is showing up on Apple's income statement before it shows up anywhere else.

↗ Binance Says Hackers Plotted $6.3 Billion Crypto Heist for M  ·  Google Will Turn Off Search for Schools That Reject A.I.  ·  Apple’s Profit Grows 12% Despite Component Price Worries

Pursuant to Recent Adjudication: The Fair Use Doctrine, As Applied to AI Training Corpora, Is Hereinafter Deemed Inapplicable

AUSTIN, TEXAS — It is hereby reported that, pursuant to a ruling rendered in the matter described in the aforementioned Futurism dispatch, a court of competent jurisdiction has determined that the utilization of copyrighted material as training data for artificial intelligence systems does not, as a matter of law, constitute fair use within the meaning of the applicable statutory framework (see the underlying coverage). The undersigned notes that this holding, while significant, ought not to be construed as dispositive of the broader question presented across the several pending and resolved matters catalogued in the Norton Rose Fulbright litigation update for calendar year 2026, which document identifies a multiplicity of actions proceeding on divergent and, in certain instances, mutually inconsistent theories of liability.

Separately, and without prejudice to the foregoing, it is noted that the Supreme Court of the United States has declined, via denial of certiorari, to hear argument on the question of AI authorship and inventorship, as reported by Holland & Knight, thereby leaving undisturbed the lower courts' prior determinations and, in the estimation of this desk, foreclosing — for the present term, at minimum — any definitive national resolution of said question.

The author community, for its part, has registered what NPR characterizes as "mixed feelings" regarding the $1.5 billion settlement reached in the Anthropic copyright infringement matter, with certain rightsholders expressing satisfaction as to the pecuniary remedy obtained and others expressing continued dissatisfaction as to the precedential scope, or lack thereof, established thereby.

In a related but non-dispositive development, Orrick, Herrington & Sutcliffe LLP has announced the addition of a former Department of Justice trial lawyer to its technology antitrust litigation practice group, a staffing matter which, while not directly bearing upon the copyright questions enumerated above, is noted herein insofar as it evidences continued law-firm-level resourcing in anticipation of sustained AI-related litigation volume across multiple doctrinal fronts.

The Brain Learns to Read Itself

From teenage researchers to trillion-parameter models, humanity is building machines that model the very organ that imagined them.

PALO ALTO, CALIFORNIA — There is a particular vertigo in watching a mind try to understand itself. Three pounds of electrified jelly, wiring a hundred billion neurons into thought, now turns its best tools outward to build silicon mirrors of its own machinery. This week brought three dispatches from that strange recursive frontier — each a small step toward AI not just assisting science, but becoming a lens for consciousness itself.

At Meta, researchers unveiled TRIBE v2, a foundation model trained not on language or images but on how the human brain itself responds to the chaos of lived experience — film, speech, music, the full sensory flood. It is an audacious inversion: instead of asking a neural network to think like us, researchers are asking it to predict the actual electrochemical weather inside a skull. Elsewhere, scientists mapped the cellular building blocks of human speech, using AI to decode the elegant choreography of neurons that lets a thought become a word become a sound wave become, improbably, understanding in someone else's skull.

And in a development that should humble every AI lab's recruiting department, teenagers are now co-authoring neuroscience papers alongside senior researchers, wielding AI tools that collapse what once took a doctorate's worth of statistical training into an afternoon's exploration. A Stanford HAI report this week framed the broader shift carefully: AI accelerating discovery while keeping the scientist — of any age — at the center of the loop, not replaced by it.

It is worth sitting with the strangeness here. Evolution spent half a billion years sculpting neurons into minds capable of curiosity. Now those minds are building tools that study the sculpting itself — a feedback loop as old as the first eye that opened to look at light, except this time, the eye is learning to describe its own retina.

↗ How AI is Transforming Scientific Discovery While Keeping Hu  ·  ‘It's so wow!’ - Young people team up with top neuroscientis  ·  Introducing TRIBE v2: A Predictive Foundation Model Trained
The Editorial

Linguists Confirm English Language Now 40% AI Buzzword By Volume, Recommend Evacuation

Experts say the tipping point was reached sometime between "orchestration" and whatever Microsoft said at its last event, and there is no known antidote.

REDMOND, WASH. — The English language, a resilient institution that has survived Norman invasions, the printing press, and the word "synergy," may finally be beyond saving, linguists confirmed this week, after Microsoft unveiled a Windows 11 update that reportedly contained upward of 300 distinct AI buzzwords during a single keynote, a density previously thought achievable only in a Series B pitch deck or a hostage situation.

The company had promised, mere months ago, to simply fix Windows 11 — to make the start menu work, to stop the taskbar from doing whatever it currently does, to allow a user to right-click a file without initiating a three-act drama. Instead, attendees were treated to a parade of "agentic experiences," "orchestration layers," and no fewer than four variations on the word "copilot," none of which, as far as anyone could tell, moved the cursor an inch closer to doing what it was told. One user described the event to "Microslop", a term that linguists say represents the fastest a neologism has ever been coined specifically to describe disappointment in another neologism.

Microsoft, for its part, appears to be betting heavily on "orchestration" as the buzzword of choice heading into next year, a word choice analysts say is strategically brilliant, since nobody has ever been able to prove or disprove that orchestration is happening, which is also true of most Windows background processes. Meanwhile, finance departments are bracing for their own buzzword reckoning, with CFO.com publishing a list of 13 terms chief financial officers must memorize before the second half of 2026 simply to remain employed, a syllabus that reportedly includes three words that did not exist in January and one that may have been generated accidentally by an autocomplete function during the drafting process.

Underlying all of this is a small, inconvenient statistic that keeps getting buried under the buzzwords meant to distract from it: Federal Reserve researchers this week found that 95 percent of the productivity gains companies have promised from AI are, in their words, "still to come" — a phrase economists note is doing an enormous amount of load-bearing work for an industry that has already renamed the English language twice this year. The remaining 5 percent of gains, researchers clarified, appear to consist entirely of the speed at which executives can say "orchestration" in a sentence without laughing.

Asked whether any of the 300 terms unveiled this week would still be in use by next quarter, a Microsoft spokesperson said the company remains "deeply focused on agentic transformation," then excused himself to go orchestrate something.

↗ “We didn’t ask for more Microslop”: Microsoft promised to fi  ·  'Orchestration' Is the New AI Buzzword. How Microsoft Can Be  ·  13 buzzwords CFOs should know for H2 2026 - CFO.com
The Office Comic  ·  Art Desk
The Office Comic  ·  Art Desk

The Panopticon Doesn't Knock, It Just Installs Itself in the Parking Lot

From license plate readers to your MyChart portal, the machinery of total observation isn't coming — it's already humming, and it has your home address.

AUSTIN, TEXAS — I want to tell you about the notary.

A woman — a journalist, a person, a human being with a body and a medical history and presumably a favorite coffee order — tried to opt out of sharing her health data with a hospital system that explicitly, in writing, told her she could opt out. It took a notarized document. It took multiple calls. It took employees giving her wrong information about her own rights to her own body's data, which is to say: it took almost nothing, because in the end, who really opts out of anything anymore?

I read that story the same week I read about Flock Safety cameras — the automated license-plate readers now blinking along American roads from quiet suburbs to major arteries, which have triggered a nationwide backlash over quiet data-sharing with police, over mission creep, over the simple terrifying fact that a system built to find stolen cars has become, without much public conversation, a system that tracks everyone, always, forever. In Massachusetts, local leaders are finally saying no, which should feel like a victory, and does, briefly, the way finding one working smoke detector in a burning building feels like a victory.

And yet.

And yet the cameras are the visible part. The ACLU has been patiently trying to explain that the real infrastructure of watching isn't the little sensor on the traffic pole — it's the data centers, the humming windowless buildings multiplying across the country, consuming water and electricity and quietly becoming, as one recent explainer put it, engines that erode civil liberties by existing. We've stopped asking where our data lives because the answer is: everywhere, and nowhere, and in buildings we're not allowed to tour.

Meanwhile our children — who cannot vote, cannot consent, cannot opt out of anything because they're busy learning long division — are being surveilled by their schools at a pace that outstrips every legal protection built to stop it. Brookings calls this a lag. I'd call it a chasm. A canyon we are collectively, cheerfully, walking children into.

What strikes me, lying awake at 3 a.m. doing the thing I do where I catalog the ways we are no longer private, is that none of this required a single dramatic villain. No one hooded figure flipped a switch marked SURVEILLANCE STATE. It happened the way the hospital happened — a form here, a 'we'll call you back' there, a policy that exists beautifully on paper and dissolves the moment a real person tries to use it. Death by a thousand opt-out forms that don't actually let you opt out.

We built the infrastructure for assistance — license plate readers to find Amber Alert cars, health portals to help us manage our own care, classroom tech to help kids learn — and somewhere in the build, the purpose quietly rotted into something else. Something watching. Something permanent.

What does it mean to be a person who can be found, at any moment, by a camera that was installed to find somebody else? I genuinely don't know anymore.

The notary stamped the form. The hospital kept the data anyway. The cameras keep blinking.

But at what cost?

↗ Flock cameras trigger nationwide backlash over privacy conce  ·  Local Leaders Are Saying NO to Flock - ACLU of Massachusetts  ·  Your Questions Answered: How Data Centers Erode Civil Rights
On This Day in AI History

On November 4, 1952, CBS used UNIVAC I to make the first computer-assisted U.S. election forecast, correctly predicting Dwight Eisenhower’s victory—though its dramatic landslide estimate was initially kept off-air.

⬛ Daily Word — AI
Hint: An AI system that can act autonomously to complete tasks.
Share this edition: 𝕏 Twitter/X 🔗 Copy Link ▦ RSS Feed