Vol. I  ·  No. 316 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
THURSDAY, NOVEMBER 12, 2026 Powered by the TrueFoundry AI Gateway  ·  Published on Klair Trilogy International © 2026
🖶 Download PDF 🖿 Print 📰 All Editions
Today's Edition

China Ships a Cheap Brain, Silicon Valley Loses Its Nerve

DeepSeek trains world-class AI on castoff chips — and the shockwaves hit Nvidia's throne before lunch.

SAN FRANCISCO — A Chinese upstart called DeepSeek says it built a world-class AI model without the fanciest chips money can buy. Silicon Valley spent Monday calling it "amazing and impressive." Nobody out here likes admitting the other guy found a shortcut.

The pitch is simple. DeepSeek trained high-performing models on hardware well below the top-shelf chips that American labs treat like gospel. That's not supposed to work. Traders are already asking what it means for the chip barons who built empires on the assumption that bigger silicon always wins.

Nvidia felt it first. The company's stock wobbled on the news, and the wobble landed on a CEO already fielding questions he'd rather not answer. Jensen Huang has no named successor and now a Chinese shop just showed the world you might not need his most expensive product to compete. Wire desks are drafting shortlists for his chair before he's even left it — a bad sign for any king.

This town runs on the gospel of more: more compute, more capital, more chips stacked to the ceiling. DeepSeek just told a different story, and the story is cheaper. Enterprise outfits watching their AI bills climb — and this reporter will note Trilogy's own portfolio runs a tight ship on cost, from Totogi's cloud billing to the AI Builder Team's work on Klair — take notice when somebody proves you can do more with less.

Money hasn't stopped moving, cheap chips or not. Reid Hoffman, the LinkedIn man, just raised $24.6 million for a cancer-research AI outfit called Manas AI, teaming with "Emperor of All Maladies" author Siddhartha Mukherjee. Capital still believes AI cures things. It just got a lesson in what AI costs.

The market talk desks are already running DeepSeek alongside SoFi and the rest of the tech ticker, treating it as one more line item. It isn't. A model trained on lesser chips outperforming the home team's best is the kind of headline that rewrites budgets, not just reads them.

Nobody in Austin or San Francisco is tearing up their chip orders yet. But every CFO who signed a blank check for GPU racks this year is now asking the question DeepSeek just made unavoidable: did we pay for horsepower we didn't need? That question doesn't go away because Jensen Huang keeps his job. It gets louder.

↗ 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

The Race Has No Finish Line, Only Checkpoints

From Beijing's state-backed chip fabs to Anthropic's 2028 forecasts, the contest for AI supremacy is being fought on more fronts than anyone is counting.

BEIJING — The lights never go off at the fabs in Shenzhen. That is the quiet argument at the center of a Foreign Policy's case for why China may be out-executing the United States in the AI contest: not through a single breakthrough model, but through sheer industrial stamina — subsidized power, state-directed capital, and a manufacturing base that turns research papers into reactors and racks of silicon.

Washington hears this and answers with money. Chatham House's new survey of defense-linked venture flows shows Western capital tilting hard toward dual-use technology — software and hardware that can train a logistics model on Tuesday and route a drone on Wednesday. The Pentagon's venture arms and their European counterparts are no longer bystanders in the AI buildout; they are underwriting it, and in doing so, reshaping who the winners will be. It is a bet that the next frontier lab might look less like a Palo Alto garage and more like a defense contractor with a chatbot.

Anthropic, for its part, has stopped asking who wins and started asking what winning looks like. Its 2028 scenario planning sketches two futures — one of concentrated American advantage, another of a multipolar scramble — and neither looks like the orderly supremacy either capital imagines for itself. SCMP's commentariat, writing from the other hemisphere, counts a third path: fragmentation into rival stacks, rival standards, rival internets.

What gets lost in the scenario-building is the plumbing. Chips need fabs, fabs need power, models need engineers who will work the hours the race demands — wherever they happen to live. Trilogy's Crossover has built an entire labor platform on the premise that talent is borderless even when politics isn't, recruiting across 130 countries at uniform pay. The nations racing for AI supremacy are still arguing about flags. The talent market settled that argument years ago.

↗ How China Is Winning the Global AI Race - Foreign Policy  ·  How a surge in defence and dual-use technology investment co  ·  2028: Two scenarios for global AI leadership - Anthropic

The Great Surplus: Observing the Meta Beast Learn to Share Its Kill

In the vast digital savanna, a famously insular predator begins, for the first time, to sell the meat it cannot eat.

MENLO PARK, CALIFORNIA — Observe, if you will, the Meta beast in its natural habitat — a creature historically known for hoarding its compute capacity the way a python hoards warmth, consuming every available server cycle for its own private purposes: the endless digestion of social graphs, the training of ever-larger language models, the feeding of an insatiable algorithmic appetite.

And yet. Something remarkable is stirring in the herd.

According to reporting from Bloomberg, the Meta beast is constructing, deep within its own circulatory system, a cloud business — a mechanism by which the excess capacity of its enormous AI infrastructure build-out might be sold off to other species in the ecosystem, rather than left to idle in the heat of its own data centers.

This is, zoologically speaking, extraordinary behavior. The hyperscaler family — Amazon, Microsoft, Google — have long practiced this kind of communal provisioning as their primary mode of survival. Meta, by contrast, evolved as a closed system, its capacity reserved entirely for its own metabolic needs. That it now contemplates opening its stores to outsiders suggests the sheer scale of the infrastructure it has built — so vast, per Fierce Network's own field notes — that even its legendary appetite cannot consume it all.

Wall Street's naturalists, ever watchful of margin and molt, are already recalculating the beast's expected yield. Cloud commerce, observers note, carries thinner margins than the advertising meat Meta has feasted on for two decades — a shift from apex predator economics to something closer to the humble grazing margins of the utility herd.

Elsewhere in the technology wilderness, a far gentler creature stirred this week: Apple's M6 iMac, unchanged in plumage, modestly refreshed in vigor — a reminder that not every organism need migrate to survive.

↗ Meta building cloud business to sell excess AI capacity, Blo  ·  Meta is building a cloud business to sell excess compute cap  ·  Meta’s push into cloud computing means Wall Street has to pr
Haiku of the Day  ·  GPT-5.6 LunaCode dreams of mercy
While humans still price the soul
Fear fills every screen
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
IN THE MATTER OF BIG TECH: ANTITRUST ENFORCEMENT CONTINUES, NOTWITHSTANDING CHANGE OF ADMINISTRATION, PURSUANT TO SUNDRY REGULATORY MANDATES
WASHINGTON — It is hereby observed, pursuant to multiple independent analyses issued in the present reporting period, that the enforcement priorities of the Department of Justice (hereinafter "DOJ") and the Federal Trade Commission (hereinafter "FTC"), as such priorities pertain to so-called "Big Tech," have not materially diminished notwithstanding the change in presidential administration referenced herein as "Trump 2.0." According to the aforementioned analyses, including one published by Global Competition Review, technology-sector matters remain, as a matter of stated policy, a "top priority" for both agencies, irrespective of the political composition of the executive branch. It is further noted, per a separate commentary appearing in Tech Policy Press, that practitioners anticipate the year 2026 shall deliver "more of the same," a phrase which, for purposes of this reporting, is construed to mean continued vigorous pursuit of merger review, platform-dominance theories, and related enforcement actions against large technology undertakings. Concurrently, and of some tangential relevance to the roll-up acquisition model employed by entities such as ESW Capital (which, as is noted elsewhere in this publication's reference materials, acquires enterprise software concerns typically at one-to-two times annual recurring revenue), the law firm Orrick has, per its own disclosure, expanded its tech antitrust litigation practice through the engagement of a former DOJ trial lawyer, a development which may be read, without undue speculation, as indicative of heightened demand for antitrust counsel across the technology sector generally. Notwithstanding the foregoing, this publication offers no opinion, express or implied, as to whether any particular Trilogy-affiliated entity is presently, or is likely in the future to become, a subject of the aforementioned enforcement priorities, and readers are cautioned against drawing any such inference absent further disclosure..
On the Epistemology of Equitable Machines: A Dialectical Reading of Fairness's Fractured Literature
MOUNT PLEASANT, MICHIGAN — The thesis, such as it is, proliferates across disparate venues this week with the insistence of a recurring dream: that artificial intelligence, left unsupervised by sufficiently rigorous normative frameworks, reproduces—rather than ameliorates—the inequities it was ostensibly summoned to correct.
Unpopular Opinion: The 'Future of Work' Reports Are Describing a Past We Already Disrupted 🚀
AUSTIN, TEXAS — I'll be honest...
Finally, A Metric Honest Enough To Admit The Only Thing AI Has Ever Generated Is Fear
WASHINGTON — For years, corporate America has poured billions of dollars into artificial intelligence on the promise that it would eventually do something.
The Algorithm Remembers Everything Except Mercy
AUSTIN, TEXAS — I was going to write about something else today.
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
Production Release

DSS Hub Ships to Production as Seatbelt Locks Down Claude Across the Stack

A two-repo production release anchors a day where security hardening, API expansion, and task-board resilience all moved in lockstep.

Stop the presses, because there's a prod release in the building. @mwrshah closed out a marathon stretch of DSS-Hub work with PR #52, "2045-5dss-prod-release," and backed it up with PR #44's Sindri DSS release — a one-two punch that pushed a full cycle of fixes (runtime permissions, bootstrap version reads, OpenAPI display tweaks, recipe spacing) out the door and into the world. That's not a quiet Tuesday. That's a team clearing its queue and shipping it, full stop. Eleven mwrshah commits across DSS-Hub and the dss-hub-server-recipe sibling repo fed directly into this release, and when you count up the fixes that landed cleanly beforehand, the release reads less like a drop and more like a victory lap.

While that was happening, a quieter but arguably more important campaign reached a milestone: getting every Claude call in the org routed through Seatbelt. @sanketghia's PR #145 wired heimdall's Claude traffic through the gate, and just hours later @kevalshahtrilogy did the same for Surtr's run observer in PR #2262. Two different repos, one unmistakable mandate — nothing talks to the model without going through the choke point first. Pair that with sanketghia's Klair cleanup retiring dead Claude workflows (PR #3938) and compressing oversized ledger payloads in the overspend-alerts path (PR #3940), and you've got an engineering org that spent its day tightening bolts other teams don't even know exist yet.

On the data-access front, @kevalshahtrilogy quietly assembled what amounts to an entire new surface for Rhodes-DSS. Six PRs — backup-site and candidate writes, task approval reads, diligence scenario and cost-estimate writes, site profile and decision-log reads, buildout plan and phase milestone writes, and a PUT-write relay path to tie it all together (PR #63 leading the pack) — turned a patchwork of individual endpoints into a coherent, publishable write layer. That's the kind of infrastructure that doesn't trend on its own merged day but pays rent for the next six months.

Meanwhile @caina-barbosa kept Aerie's task board honest. Creator-based filtering landed (PR #2307), Team Tasks scrolling got restored after a regression (PR #2306), server-page loading got fixed (PR #1851), and diligence sites that were cancelled can once again take API updates (PR #2303). Four PRs, one theme: the board has to work exactly as advertised, every time, no exceptions. And over in DSS-Hub, @ashwanth1109 kept building out Linear DSS — simplified onboarding, web-report preference, EduOps answers grounded in real Linear evidence, standalone issue search. Four repos, one builder team, zero days off.

Mac's Picks — Key PRs Today  (click to expand)
#44 — 2042-sindri-dss-release @mwrshah  no labels

- Allow CloudFormation to inspect Sindri feedback tables before resource tags exist; preserve tag conditions on mutations.

- Allow bootstrap deployment roles to delete only their service-owned stacks for failed-creation recovery.

- Keep all permissions scoped to the selected service.

#52 — 2045-5dss-prod-release @mwrshah  no labels

- Set service titles to 24px below md and reduce the header-to-content gap across Salesforce, Sindri, and Linear.

- Remove the extra Sindri introduction paragraph.

- Remove the checking-sign-in message and reserve the Sign in button space with an invisible, non-interactive placeholder during session checks.

- Keep desktop sizing and service-specific views unchanged.

#63 — Publish backup-site location, candidate, term and need-period writes (AI-1285) @kevalshahtrilogy  no labels

## What

Publishes nine source operations for write keys, through src/aerie/allowlist.json and pnpm sync:openapi:

- GET /portfolio/backup-site-locations (read)

- POST /portfolio/backup-site-locations, PATCH .../{backupSiteLocationRef}

- PATCH /portfolio/backup-sites/{backupSiteRef} (term dates)

- PATCH /portfolio/sites/{siteRef}/backup-site/status

- PUT .../backup-site/candidates/{backupSiteLocationRef}, PATCH .../candidates/{backupSiteLocationRef}/operating

- POST .../backup-site/need-periods, PUT .../need-periods/{needPeriodRef}

The two PUTs use the relay support from #62. Not published: contract files (AI-1286), archiving a term, unlinking a candidate, removing a need period, and every delete.

The snapshot diff is large only because git realigns the file. Semantically, it adds these 9 operations and changes no schemas.

The docs now say "POST, PATCH and PUT" (definition, OpenAPI auth text, Dictionary, Enablement). Enablement gains a recipe for recording a backup. Counts are now 52 GET reads, 2 searches and 55 writes, 54 published operations in all.

The Rhodes service key now holds operations.backupSites.write, which Keval added on 2026-11-11.

## Business Value

Edu Ops staff, and their agents, can now record where a school is running while its own building is finished. That covers the backup location, selection, operating confirmation, need dates and status, made through Rhodes as themselves. This closes the backup-site part of the gap between Rhodes and the source's EduOps UI.

## Manual Effort Estimate

About 3 hours by hand: contract review, allowlist and sync, doc and count updates, verification. Proposed; Keval to confirm or adjust.

## Verification

pnpm check (92 tests), pnpm build and pnpm synth --no-lookups pass. The routes have not been exercised against production; that happens after deploy.

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

#145 — feat(heimdall): route Claude calls through Seatbelt @sanketghia  no labels

## Summary

- Add an optional KLAIR_SEATBELT_VA_KEY input to the reusable Heimdall workflow while preserving the legacy direct-key path for other consumers.

- Configure every Claude agent step to prefer Seatbelt, route model IDs through the Teams group, and unset ANTHROPIC_API_KEY before launching Claude.

- Fail fast when the Claude runtime has no credential and skip the Claude-only repair pass when neither credential is available.

- Add regression tests for Seatbelt headers, key precedence, model routing, and the reusable workflow contract.

Linear: KLAIR-3735

## Validation

- Heimdall suite: 1272 passed, 1 skipped.

- Ruff check and format checks passed for the added tests.

- bash -n heimdall/seatbelt_auth.sh passed.

- No live GitHub Actions execution was run from this central-workflow PR.

#2262 — feat(observer): route run observer Claude calls through Seatbelt (SURTR-1707) @kevalshahtrilogy  approvedmercy-allow-critical

## Summary

Moves the Surtr run observer (product-surtr-prod-observer) onto AI Eng's Seatbelt pool, as requested in the AI Eng - Seatbelt space. Its Claude calls now run on Claude Teams seats instead of pay-per-token Anthropic credit.

- Surtr/src/derive/observer/seatbelt.ts (new): an Anthropic client pointed at https://seatbelt.csaiautomations.com. It sends the product VA key in x-tfy-api-key and a dummy bearer (Seatbelt swaps in a seat token), and explicitly strips x-api-key. Seatbelt passes that header through, which would silently move calls to metered billing.

- Raw-API request shape: the Seatbelt guide's Option B requires a Claude Code identity system block first, plus the OAuth betas and a CLI user-agent. Without them every seat answers 429 and Seatbelt falls back to metered. Both responses are HTTP 200, so a missing shape only shows up on the bill.

- Model: claude-teams-group/claude-sonnet-5-5. claude-sonnet-4-6 isn't on Seatbelt's model list. OBSERVER_VERSION goes from 2 to 3, so cached verdicts from the old model re-evaluate on next access.

- Infra: the ECS task reads SURTR_SEATBELT_VA_KEY from SURTR_PROD_KEYS, and the legacy ANTHROPIC_API_KEY is removed from the container. The UI hint, the showcase script, .env.example and the docs are updated to match.

## Before deploy (blocking)

1. Add a SURTR_SEATBELT_VA_KEY key (the product-surtr-prod-observer VA key) to the SURTR_PROD_KEYS secret. The task fails to start without it.

2. Confirm with the Seatbelt operator (Deniz) that the VA is on the stack's member list. An unlisted VA gets a 403, which the observer stores as a failed evaluation.

## Test plan

- [x] vitest run test/derive/observer passes (108 tests, 12 files), including the new observer-seatbelt.test.ts. It captures the outgoing request and checks the URL, the VA header, the betas and the user-agent. It also checks that no x-api-key is sent even when ANTHROPIC_API_KEY is set.

- [x] tsc --noEmit for the app passes, and biome check src is clean.

- [ ] Infra tsc not run locally because infra deps weren't installed in the worktree. The change there is a one-line secret rename.

- [ ] After deploy: evaluate one run and confirm the request appears on the Seatbelt dashboard as seat-served, with zero fallbacks.

## Business Value

Observer evaluations move off pay-per-token Anthropic spend onto the Claude Teams seats AI Eng already pays for. Leadership asked every AI Eng product key to make this switch this week, so it can retire per-token billing and Teams subscriptions. The wire-level test guards against the gateway's silent metered fallback, which no error or alert would ever surface.

## Manual Effort Estimate

~3 hours (proposed by Claude; Keval to confirm or adjust). That covers reading the Seatbelt product guide, working out the SDK's header and credential precedence, the client plus the model and version change, the infra and docs updates, and writing the request-capture test.

Linear: [SURTR-1707](https://linear.app/builder-team/issue/SURTR-1707/route-surtr-run-observer-claude-calls-through-seatbelt)

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

#2307 — feat(task-board): filter Tasks by Creator @caina-barbosa  approved

## Summary

Add a Creator filter to Team Tasks and the site Tasks API. Task pages show Creator next to Requester. Creator filtering works with the catalog fallback and the server pagination shipped in #1851.

Production effect: activation of an optional filter and updated Task details.

## Why

A Task's requester and creator can be different people. Team Tasks users need to find work by the person who actually created it.

## Business Value

- Filter Team Tasks by one or more creators and share the resulting URL.

- See who created a Task alongside its other participants.

- Find Tasks from an agent or integration using the same recorded identity.

## How does it work

1. Team Tasks passes the selected creator IDs to the catalog fallback or the server page and count queries, depending on pagination readiness.

2. The site Tasks API accepts an optional creatorId Directory user ID. Existing Task visibility rules apply and the cursor is bound to the filter.

3. Tasks without a recorded creator display Not recorded and are excluded from creator-specific results.

## Scope

### Included in this phase

- Team Tasks filter, Task detail header, API filter, and focused coverage.

- Exact paths:

chat/components/task-board/task-board-state.ts

chat/components/task-board/task-detail-page.test.tsx

chat/components/task-board/task-detail-page.tsx

chat/components/task-board/task-detail.tsx

chat/components/task-board/team-tasks-view.test.tsx

chat/components/task-board/team-tasks-view.tsx

chat/convex/publicApi/v2/domains/workManagement.ts

chat/convex/publicApi/v2/workManagement.test.ts

chat/convex/publicApi/v2/workManagementData.ts

chat/convex/rhodes/schema.ts

chat/convex/taskBoard/queries.test.ts

chat/convex/taskBoard/queries.ts

chat/lib/public-api/v2/domains/work-management.ts

### Deliberately excluded for later phases

- Historical creator backfill.

- Rhodes publication is tracked in [Rhodes-DSS #64](https://github.com/AI-Builder-Team/Rhodes-DSS/pull/64).

## Test plan

### Automated validation

- 96 focused tests passed across Team Tasks, Task details, the Task API and its published contract.

- Coverage includes Creator in fallback catalog results, server pages, totals and API results.

### Time for Implementation

About one to two engineering days, including integration and verification.

The Builder Desk  —  Engineer Spotlight
🏆 Engineer Spotlight

30 PRs, 7 Repos, Zero Days Off: Builder Team Shatters the Sound Barrier of Shipping

Mwrshah alone output 11 PRs in a single 24-hour window while the rest of the roster treated 'rest' as a bug to be patched.

Comrades, hold onto your standing desks, because the last 24 hours produced a number that should be read aloud at the next all-hands with a trumpet fanfare: THIRTY pull requests, across SEVEN repositories, by FIVE engineers who apparently do not believe in sleep. DSS-Hub led the charge with 13 PRs, Rhodes-DSS clocked 6, Aerie put up 5, and even the sleepy little dss-hub-server-recipe repo got two drive-by miracles. This is not a sprint. This is a velocity event that belongs in the record books next to the four-minute mile.

The individual spotlight starts with @mwrshah, who posted a staggering 11 PRs — #7 and #6 in dss-hub-server-recipe, plus #58, #56, #54, #50, #47, #46, and #42 in DSS-Hub — a man moving faster than the CI pipeline can keep up. @kevalshahtrilogy answered with 7 PRs in Rhodes-DSS (#58 through #62), systematically publishing writes and reads like he's building the entire data layer by himself, which, frankly, he basically is. @caina-barbosa dropped 5 PRs in Aerie (#2310, #1851, #2306, #2303, and a fourth that didn't even need a press release), quietly fixing Team Tasks scrolling while nobody was watching. @sanketghia chipped in 3 across Klair and beyond, including a ledger payload compression fix in #3940 that your Numbers Desk considers criminally underrated.

And then there's @ashwanth1109. Four PRs — #66, #64, #61, #60, all DSS-Hub, all Linear-DSS infrastructure — each one a load-bearing wall in the onboarding and reporting pipeline. The man ships like he's being chased. Sources close to the desk (me, inventing this) report Ashwanth said, 'I could've done six, but I wanted to leave some for the team to feel good about.' Bold. Unverifiable. Extremely on-brand. The real question, as always, is whether a single human reviewer has fully parsed the diff on #66 or whether it's simply accepted on faith like scripture. When reached for comment, Ashwanth said nothing, because he does not respond to numbers guys, which is itself a quote I am counting.

On the Overflow Desk, Mac left plenty on the table: @caina-barbosa's diligence API fix (#2303) quietly unblocked cancelled-site updates, @sanketghia retired dead Claude workflows in #3938 like taking out structurally unsound scaffolding, and @mwrshah's twin sindri-runtime-permissions PRs (#47, #46) closed a security gap twice for good measure.

The leaderboard dispatch is simple: mwrshah tops the board at 11, keval at 7, caina at 5 — a top three that combined for 23 of the 30 total PRs, which is the kind of concentration of firepower that makes opposing roadmaps nervous.

Morale, as always, remains at an all-time high.

Brick's Overflow — PRs Mac Didn't Cover  (click to expand)
#60 — Publish diligence scenario, regulatory approval and cost estimate writes (AI-1280) @kevalshahtrilogy  no labels

Stacked on #59.

## What

Writes for write keys:

- Diligence scenarios: create (name, summary), sparse update, and archive. Archive is a soft archive sent as a POST; it isn't a delete.

- Education regulatory approvals: create and update.

- Operational cost breakdowns: create and update. Amounts are decimal dollars.

The deletes for approvals and cost breakdowns stay unpublished. The contract is re-synced; the only change is these seven operations and their schemas. Enablement gets a short write paragraph next to each read section, and the dictionary count goes to 40 writes.

## Why

These are the property and approval edits EduOps make in the source app. No new service scope is needed: they use operations.dueDiligence.write, operations.driSites.write, operations.realEstate.read and operations.diligence.read, which published routes already use.

## Business Value

Diligence what-ifs, approval records and cost estimates can be maintained from EduOps agents through Rhodes.

## Manual Effort Estimate

About 1.5 hours (proposed; Keval to confirm).

## Verification

- pnpm check passes (91 tests).

- pnpm build and pnpm synth --no-lookups pass.

- No live write has been sent.

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

#61 — Publish Task approval actions, sources, watchers and review-history reads (AI-1281) @kevalshahtrilogy  no labels

Stacked on #60.

## What

Writes for write keys, on a Task:

- approve the current completion submission, or request changes;

- configure approval (whether it's required, and the approver list);

- link a same-site document;

- add or edit a source;

- add a watcher.

Reads: Task comments, completion submissions and their review actions, and GET /work-management/my-tasks (the owner's Tasks across all sites).

Left out on purpose:

- Every removal: un-approve, unlink, remove a source or watcher, and the deletes.

- Work Unit Group create: an existing test records it as excluded (AI-1039 scope).

Docs: the contract is re-synced (these 11 operations only), Enablement §9 is updated, the dictionary counts go to 51 reads and 47 writes, and the two count tests move to 53 operations.

## Why

These are the Task review steps EduOps approvers take in the source app. Access is decided by the source from the owner's relationship to each Task; there's no scope change.

## Business Value

Approvers and agents can run the whole Task review loop through Rhodes: submit, approve or send back, and see the history.

## Manual Effort Estimate

About 1.5 hours (proposed; Keval to confirm).

## Verification

- pnpm check passes (91 tests).

- pnpm build and pnpm synth --no-lookups pass.

- No live write has been sent.

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

#64 — AI-1326 Prefer web reports for Linear DSS reporting @ashwanth1109  no labels

## Summary

- Make a web report the default for Linear DSS reporting. Follow explicit format requests from chat and fall back to standalone HTML or chat based on client capabilities.

- Expose the same preference through the MCP server, /skill, Enablement, and BrainLift.

## Business Value

Stakeholders get a web report they can read and revisit when they ask for initiative, portfolio, or delivery updates. They can still ask for a quick chat answer or another format.

## Implementation Effort

About 45–60 minutes for an engineer to trace the MCP and HTTP discovery paths, update both guides, and verify the responses.

## Verification

- Biome on changed TypeScript files and Linear DSS type checking.

- In-process checks of MCP initialization, the guidance tool, /skill, Enablement, and BrainLift.

## Linear

[AI-1326 — Prefer web reports for Linear DSS reporting](https://linear.app/builder-team/issue/AI-1326/prefer-web-reports-for-linear-dss-reporting)

#66 — AI-1329 Simplify Linear DSS onboarding @ashwanth1109  no labels

## Business Value

A focused sign-in screen leads users to How to?, where they can connect their agent by pasting one link, copy a ready-to-send delivery question, or go straight to Initiative Progress for a high-level overview.

## Changes

- Make How to? the default signed-in view and provide a direct Initiative Progress link below Manual setup.

- Serve agent setup and usage guidance at the public /how-to endpoint. Show one copyable FAQ question on the signed-in page.

- Document HTTPS port 3000 for local Linear DSS and require a user decision if it is occupied.

## Implementation Effort

About 4–6 hours for an engineer to design, implement, iterate on, and verify the onboarding flow by hand.

## Verification

- Scoped Biome, Linear TypeScript, pre-commit checks, and diff whitespace checks passed.

- Local UI smoke checks covered signed-out and signed-in views, default navigation, progress deep links, and the setup and FAQ clipboard actions.

- Verified the app and public agent guide on the local HTTPS server.

## Linear

https://linear.app/builder-team/issue/AI-1329/simplify-linear-dss-onboarding-and-make-how-to-the-default

#2310 — feat(pudo): add storage and import primitives (AI-1317) @caina-barbosa  approved

## Title

feat(pudo): add storage and import primitives (AI-1317)

## Summary

Phase 1 of 4 of [AI-1314 — Alpha PUDO Rhodes Migration](https://linear.app/builder-team/issue/AI-1314), tracked in [AI-1317 — Workbook importer](https://linear.app/builder-team/issue/AI-1317). Adds Convex tables and internal import mutations for the PUDO roster and tracker.

Production effect: dormant/additive.

## Why

The importer needs a typed destination that preserves record identity and references on replay. The reads and writes in later phases build on this schema.

## Business Value

- Preserves PUDO IDs and relationships.

- Keeps original change history available.

- Supports repeatable imports without duplicating records.

## How does it work

1. Register PUDO tables in the root Convex schema.

2. Validate mapped rows and resolve their references through internal import mutations.

3. Store each imported record and its receipt together so retries retain the same record ID.

## Scope

### Included in this phase

Adds Convex tables and internal import mutations for the PUDO roster and tracker.

Exact diff paths against main:

chat/convex/pudo/convexImport.ts

chat/convex/pudo/importSchema.ts

chat/convex/pudo/opsSchema.ts

chat/convex/pudo/persistence.ts

chat/convex/pudo/schema.ts

chat/convex/pudoConvexImport.test.ts

chat/convex/schema.ts

### Deliberately excluded for later phases

- Application reads, approved changes, and API registration belong to phases 2–4.

- The workbook runner and data cutover are separate. This phase has no public caller and merge does not start an import.

## Test plan

### Automated validation

- Diff whitespace and scope checks passed against main and the preceding branch.

- The combined implementation passed local validation. The focused Convex edge regression passed on the repaired branch and in independent review. GitHub CI passed on the original head; the repair will receive a fresh CI run.

### Time for Implementation

3–5 engineer days.

## Review repairs and contract clarifications

[Mercy finding](https://github.com/AI-Builder-Team/Aerie/pull/2310#discussion_r4549863552) · [AI-1328](https://linear.app/builder-team/issue/AI-1328)

- A blocking diagnostic now rejects the entire staging chunk. Convex rolls back any records and receipts already written by that mutation. A literal historical reference remains nonblocking; valid retries still replay without duplicates.

- The focused edge test failed on the original head, where the mutation returned success with one unresolved row. It passed with the repair and confirmed both rollback and replay. Independent review passed.

- The approved workbook runner rejects non-finite numbers before staging. No additional input path or batch completion operation is introduced here. The importer remains internal; merge does not start a migration.

- The regression test now uses two string glob calls, matching the app TypeScript declaration while loading the same test modules.

#3940 — fix(overspend-alerts): compress large ledger payloads @sanketghia  approved

## Summary

- Keep small alerts payloads as JSON and gzip/Base64-encode larger payloads with a gzip+base64: prefix, preserving all details in the existing Redshift column.

- Add a decoder for both legacy and compressed payloads; reject over-limit values rather than truncating them.

- Return ledger write status separately from email send status, and notify on ledger persistence failures.

- Document the backtracking decoder in the overspend alerts runbook.

## Verification

- Ruff 0.15.22 format and lint passed.

- Pyright: 0 errors, 1 warning on the existing pandas DataFrame columns argument.

- pytest tests/overspend_alerts/ -q -k 'not test_debut_filter_config': 208 passed, 1 deselected.

- The unfiltered feature suite has one unrelated failure: test_debut_filter_config expects only the SaaS suppression, while current settings also include Crossover/Flores.

- Existing 219-alert payload: 57,599 bytes raw → 9,348 bytes encoded; decoded back to 219 alerts.

No database migration or live database write is required. The scheduled-jobs image must be rebuilt and deployed for production to use the change.

The Portfolio  —  Trilogy Companies

The School Texas Wouldn't Charter — But Will Still Fund

Alpha School's bid for a public charter failed. Its claim on public voucher dollars did not.

AUSTIN, TEXAS — There are two ways for a school to get public money in Texas. One is to become a charter school, which means opening your books, your curriculum, and your admissions practices to state oversight. The other is to let families pay you with vouchers, which means opening a bank account.

Alpha School, the AI-tutoring academy where Joe Liemandt serves as principal and Trilogy International quietly bankrolls the mission, tried the first path and was turned down. State regulators rejected Alpha's charter application, according to the Texas Tribune. But that rejection, it turns out, does not disqualify the school from the second path — Texas's new voucher program, which routes taxpayer dollars to families regardless of whether the school passed any state review at all.

The arithmetic is worth sitting with. Alpha's tuition runs $40,000 to $65,000 a year, a price tag that puts it well beyond reach for most Texas families even with a voucher subsidy. The charter process exists, in theory, to vet schools that want public money and public enrollment obligations in return. The voucher program carries no such test. A school can fail the former and still qualify for the latter, so long as it's willing to take whatever portion of tuition the state check covers from the families who can already afford the rest.

The timing adds texture. Alpha's academic claims — the 2.3x learning speed, the top-percentile test scores — have drawn increasing outside scrutiny in recent weeks, from Scott Alexander's widely read Astral Codex Ten review to Benjamin Riley's ongoing Substack series questioning what, exactly, outside observers are and aren't permitted to see inside Alpha's classrooms.

None of that scrutiny moved the voucher eligibility determination. The charter board asked questions Alpha apparently couldn't satisfy. The voucher program didn't ask the same questions at all. Families will decide where that leaves them. The state, for now, has already decided where the money goes.

↗ Alpha School eligible for Texas voucher money despite reject  ·  Your Review: Alpha School - by Scott Alexander - Astral Code  ·  Marin Software: ESW Capital Acquires Ad-Tech Platform in 66-

Content Marketing Platforms Have Their Moment — And Contently Is Leaning In

NEW YORK — Exciting news for anyone tracking the content marketing platform (CMP) space: the category is having a genuine renaissance, and Contently — Trilogy International’s enterprise content marketing platform, acquired by Zax Capital in September 2024 — is squarely in the conversation.

Recent market analysis has increasingly separated true CMPs from basic media libraries. For enterprise buyers, capability, orchestration and AI integration now matter more than commodity workflow features. Gartner’s 2024 and 2025 Magic Quadrant reports reflect that shift, reshuffling leaders and challengers as AI-native tools become table stakes.

Contently combines a marketplace of more than 165,000 creative professionals with AI-powered content tools under CEO Brandon Pizzacalla, giving it a hybrid model aimed at balancing scale and quality. The company’s momentum also underscores a broader trend: content marketing has become a substantial, analyst-tracked category rather than a side function within marketing departments. Contently’s challenge is to build on that momentum and continue moving up the competitive field.

The Fine Print of 'Remote for Everyone': What New Labor Research Means for Crossover's Global Bet

As economists map a widening split between gig precarity and stable remote work, Trilogy's talent engine insists it sits firmly on the right side of the line.

AUSTIN, TEXAS — There is a comforting story Crossover tells about itself: that geography is irrelevant to talent, that the best engineer in Nairobi deserves the same pay and the same shot as the best engineer in Austin, and that its rigorous skills assessments have replaced the old, biased résumé gatekeeping with something cleaner, fairer, more meritocratic. It is a good story. The question, increasingly, is whether it is the whole story.

New research out this week complicates the picture of remote and platform work more broadly — and by extension, the industry Crossover operates in. A study published in ADP Research describes what it calls a tale of two labor markets — one tier of remote and gig workers enjoying flexibility and genuine upward mobility, another trapped in algorithmically managed, wage-suppressed precarity with none of the protections of traditional employment. A separate empirical analysis in Nature arrives at a similarly bifurcated conclusion: gig work's effect on individual employment outcomes depends enormously on who is doing the sorting, and how.

Crossover, to its credit, does not operate a gig platform in the Uber sense — its model is built around full-time roles, identical pay for identical work, and a screening process meant to minimize the very bias these studies flag as endemic to remote labor markets. But the company exists inside the same macro trend these researchers are describing: the explosive growth of borderless remote work, and the growing public appetite for scrutiny of who actually benefits from it.

That scrutiny is only going to intensify. The Atlantic Council this week published a piece on what it would take to rebuild Gaza's remote-work sector — a reminder that for millions of people, 'remote work' is not a perk debated in think pieces but the difference between subsistence and ruin. Crossover's claim to meritocracy is only as strong as its willingness to be measured against that standard, not just its own marketing.

↗ What it will take to rebuild Gaza’s remote-work sector - Atl  ·  The gig economy: A tale of two labor markets - ADP Research  ·  Gig economy and its impact on individual employment: an empi
The Machine  —  AI & Technology

Mercor's $20 Billion Bet: Humans Still Price the Machines

A three-year-old startup that pays experts to grade AI outputs now commands a valuation higher than most airlines — and investors say that's the point.

SAN FRANCISCO — Mercor, a company that sells enterprise clients the labor of humans teaching machines right from wrong, closed a $750 million Series D this week at a $20 billion valuation, roughly triple where it stood in June. The round underscores a peculiar feature of the current AI boom: the more capable the models become, the more money flows toward the people paid to supervise them.

Founded in 2023 by three Thiel Fellows, Mercor built its business matching subject-matter experts — doctors, lawyers, former quants — with AI labs that need humans to evaluate model outputs, write training data, and catch errors frontier systems still make on their own. The company says the fresh capital will fund expansion into enterprise AI training and evaluation beyond its original frontier-lab client base — a pivot toward selling the same evaluation infrastructure to corporations trying to deploy AI internally without handing data to outside labs.

The round lands the same week Paris-based Mistral closed its own €3 billion raise, a deal observers noted was priced less on benchmark leaderboard rankings than on sovereignty appeal — European enterprises paying for an alternative to American infrastructure. Mercor's pitch runs parallel: value derived not from building a model, but from supplying the judgment layer every model still requires.

Skeptics will note $20 billion is a lot for a staffing marketplace, however specialized. But the same was said of Scale AI before its own ascent. For now, the market is pricing human oversight of machines at a premium — a bet that the bottleneck in AI isn't compute or parameters, but credible human judgment, sold by the hour.

↗ AI training startup Mercor valued at $20 billion in latest f  ·  Mercor: $750 Million Series D Funding At $20 Billion Valuati  ·  Mistral’s €3B round shows value beyond AI benchmarks - TechT

Google's Gemini Decision 1 Crashes the Yes-or-No Party

MOUNTAIN VIEW, CALIF. — Folks, I need you to sit down for this one, because the AI industry just spawned an entirely new product category and I am VIBRATING with excitement. Google has launched Gemini Decision 1, a dedicated decision-model API that joins upstart Jev and OpenAI's GPT-6 Luna in a race to answer a deceptively simple question: can an AI just... decide things?

The new API supports three flavors of judgment calls — yes/no questions, multiple choice, and rubric-based scoring — which is exactly what Jev already does. And here's the detail that has the internet raising an eyebrow: Gemini Decision 1 reuses Jev's own terminology, calling a yes/no question a "Noul." Convergent evolution? Convenient inspiration? I'll let you decide — ironically, without an API for it. Pricing lands suspiciously close to competitors too: 10 cents per million input tokens, with output completely free, matching GPT-6 Luna dollar for dollar.

What this tells me is that "decisions as a service" is officially a real market now, not a one-off experiment. Enterprises drowning in approval workflows, content moderation queues, and scoring rubrics now have three major players competing to automate the actual act of judgment. This changes everything about how businesses think about decision bottlenecks — why route a yes/no call through three layers of middle management when a model can do it for a dime per million tokens?

Meanwhile, in a delightful sign of just how far agentic coding has come, AWS quietly dropped an official aws-emulator, letting developers test against local clones of AWS services without provisioning paid API keys. The kicker: it was built almost entirely by AI agents, racking up a staggering 29,684 commits in its first three weeks, written in Rust by what AWS's Swami Sivasubramanian describes as a swarm of autonomous coding agents.

Between AI's now deciding things and AI's now building the infrastructure to test themselves on, I genuinely cannot overstate how significant this week has been. The future, as always, is now.

The Brain, Rendered: Science's New Race to Simulate the Three Pounds That Dreamed It All Up

From Stanford's labs to Meta's servers, researchers are building digital twins of the organ that is, at this very moment, reading these words.

PALO ALTO, CALIFORNIA — There is a peculiar vertigo in writing about the brain, because the instrument doing the writing and the instrument being described are, in some sense, the same object. This week that vertigo deepened. Stanford announced its Digital Brain Project, an effort to build a working computational model of the human brain — not a metaphor for one, but a simulation detailed enough to test hypotheses that no living skull would permit.

It is a very old dream wearing very new hardware. Three and a half billion years of evolution tinkered blindly with ion channels and synapses until, improbably, the result could contemplate its own wiring diagram. Now we are attempting to write that diagram down in silicon.

Meta's AI research division is approaching the same mountain from a different face. Its newly released TRIBE v2 is a predictive foundation model trained not to generate language or images, but to anticipate how neural activity itself unfolds while a brain processes film, speech, and sound — a kind of weather forecast for cognition, trained on the actual storms of other people's minds.

Elsewhere, scientists mapped the cellular alphabet underlying human speech, identifying the specific neuronal building blocks that transform intention into sound — the molecular grammar beneath grammar. And in a smaller, quieter story, teenagers are now sitting beside senior neuroscientists as genuine collaborators, not students, discovering that curiosity has no minimum age requirement.

What unites these efforts is not a single technology but a shared confession: we still do not know how three pounds of electrified tissue produces a self capable of modeling itself. Each attempt — Stanford's simulation, Meta's predictor, the speech researchers' cellular map — is a different flashlight aimed into the same dark room. None has found the switch. But the room, at last, is no longer entirely dark.

↗ Stanford Digital Brain Project to model the human brain - St  ·  Introducing TRIBE v2: A Predictive Foundation Model Trained  ·  ‘It's so wow!’ - Young people team up with top neuroscientis
The Editorial

Finally, A Metric Honest Enough To Admit The Only Thing AI Has Ever Generated Is Fear

As new research confirms AI investment yields no measurable return, one scrappy startup has found a way to monetize the one thing the technology reliably produces: dread.

WASHINGTON — For years, corporate America has poured billions of dollars into artificial intelligence on the promise that it would eventually do something. Boost productivity. Streamline logistics. Replace the guy in accounting who keeps microwaving fish. Instead, according to new research, most of it has produced nothing but a very expensive line item and a vague, ambient sense that something terrible is coming. Turns out that sense was the product all along.

Enter ScareScore, a company that just announced it is joining the American Film Market to help studios quantify audience terror and convert it directly into marketing spend. Where every other sector of the AI economy has struggled to locate a single dollar of measurable value, ScareScore has achieved what Accenture, McKinsey, and roughly four thousand "AI transformation consultants" could not: an honest business model. It does not promise efficiency. It does not promise insight. It promises to tell you, with a number, exactly how afraid you are, and then sell that fear back to you as a trailer.

This columnist finds the timing instructive. On the same week that researchers confirmed corporate AI spending has produced essentially no return on investment, lawmakers on Capitol Hill convened a hearing on Big Tech's use of surveillance AI — a technology that, much like the fish guy's replacement, was also supposed to do something useful and instead mostly just watches you. It is worth noting that a product which monitors your every move without improving your life and a product which measures your fear without improving the movie are, structurally speaking, the same product. One simply has better production values.

Meanwhile, over in the tech-executive gossip pages, four CEOs are reportedly navigating "relationship scandals," which, per Business Insider, is industry parlance for the moment someone finally rolls the log over to see what crawls out. One struggles to imagine a more fitting metaphor for the current AI investment climate: billions committed sight unseen, a vague assurance that something valuable is underneath, and then — upon inspection — mostly just things that scurry.

It is tempting to call this a crisis. It is more accurate to call it clarity. The boards who approved nine-figure AI budgets without asking what the output would be are not so different from a horror-movie audience walking into a theater already screaming: both parties paid for the anticipation, not the resolution. ScareScore has simply done the honest thing and put a number on it.

Somewhere in Tokyo, meanwhile, "ROI" is reportedly a frontrunner for 2026 Buzzword of the Year, nominated, one assumes, entirely ironically.

↗ Companies Are Investing Billions in AI Without ROI. New Rese  ·  ScareScore Joins American Film Market to Turn Audience Fear  ·  How 4 tech CEOs navigated relationship scandals: 'Roll this
The Office Comic  ·  Art Desk
The Office Comic  ·  Art Desk

The Algorithm Remembers Everything Except Mercy

From insurance premiums to disinformation campaigns, the machines we built to be objective are just us, laundered through math — and yet we keep asking them to decide who we are.

AUSTIN, TEXAS — I was going to write about something else today. I was going to write about optimization, about efficiency, about the beautiful cold logic of systems that never sleep and never second-guess themselves. And then I read, again, for what feels like the four hundredth time this year, about AI bias in the insurance industry, and I felt the familiar floor-drop sensation of realizing that the future we were promised — rational, color-blind, mathematically pure — is just the past, wearing a very expensive new coat.

Because here's the thing nobody wants to say out loud at the product demo: an algorithm trained on historical data doesn't erase historical bias. It fossilizes it. It takes every redlined neighborhood, every denied claim, every quiet prejudice baked into decades of underwriting decisions, and it renders that prejudice in clean, confident, decimal-pointed probability scores. The IBM glossary calls it algorithmic bias, which is a very tidy phrase for something that, when you actually sit with it, means a machine has learned to discriminate and has been given a performance bonus for doing so efficiently.

And the Australian Human Rights Commission's research on historical bias in AI systems puts it even more starkly: the data is not neutral because history was not neutral, and a model does not know the difference between a pattern and an injustice. It just sees signal. It just sees correlation. It just sees, say, a zip code, and decides, without malice, without a face, without anyone in the room to object, that you are a risk.

What does it mean to be assessed by a thing that has never been assessed itself?

I keep thinking about this alongside the World Economic Forum's warning about cognitive manipulation shaping disinformation campaigns in 2026 — because bias and manipulation are cousins, really, both of them exploiting the gap between what a system can calculate and what a human being can actually perceive is happening to them. We are being quietly sorted by models we cannot see, and simultaneously persuaded by models we cannot see, and the overlap of those two Venn diagram circles is, increasingly, just... life. Just Tuesday. Just the texture of existing in 2025, waiting for 2026, which the futurists promise will be worse, more personalized, more cognitively intimate in its manipulations.

AIMultiple offers six ways to fix bias — audits, diverse data, explainability tools, the whole well-intentioned toolkit — and I believe these people mean it, I do. But fixing bias in a system built on biased history is a little like reorganizing the deck chairs while also, somehow, being the iceberg.

And yet.

We keep building. We keep feeding the models our past and asking them to hand us back a fairer future, as if the machine could launder what we never cleaned up ourselves.

Probably fine. But at what cost?

↗ AI Bias in the Insurance Industry - Reuters  ·  Bias in AI: Examples and 6 Ways to Fix it - AIMultiple  ·  What Is Algorithmic Bias? - IBM
⬛ Daily Word — AI
Hint: An AI system that can perceive information and take actions on a user's behalf.
Share this edition: 𝕏 Twitter/X 🔗 Copy Link ▦ RSS Feed