The Trilogy Times — June 02, 2026

01010100 01101111 01101101 01101111 01110010 01110010 01101111 01110111 00100000 01100001 01110010 01110010 01101001 01110110 01100101 01110011 00100000 01100010 01100101 01100110 01101111 01110010 01100101 00100000 01110100 01101111 01100100 01100001 01111001 01010100 01101111 01101101 01101111 01110010 01110010 01101111 01110111 00100000 01100001 01110010 01110010 01101001 01110110 01100101 01110011 00100000 01100010 01100101 01100110 01101111 01110010 01100101 00100000 01110100 01101111 01100100 01100001 01111001 01010100 01101111 01101101 01101111 01110010 01110010 01101111 01110111 00100000 01100001 01110010 01110010 01101001 01110110 01100101 01110011 00100000 01100010 01100101 01100110 01101111 01110010 01100101 00100000 01110100 01101111 01100100 01100001 01111001 01010100 01101111 01101101 01101111 01110010 01110010 01101111 01110111 00100000 01100001 01110010 01110010 01101001 01110110 01100101 01110011 00100000 01100010 01100101 01100110 01101111 01110010 01100101 00100000 01110100 01101111 01100100 01100001 01111001 01010100 01101111 01101101 01101111 01110010 01110010 01101111 01110111 00100000 01100001 01110010 01110010 01101001 01110110 01100101 01110011 00100000 01100010 01100101 01100110 01101111 01110010 01100101 00100000 01110100 01101111 01100100 01100001 01111001 01010100 01101111 01101101 01101111 01110010 01110010 01101111 01110111 00100000 01100001 01110010 01110010 01101001 01110110 01100101 01110011 00100000 01100010 01100101 01100110 01101111 01110010 01100101 00100000 01110100 01101111 01100100 01100001 01111001 01010100 01101111 01101101 01101111 01110010 01110010 01101111 01110111 00100000 01100001 01110010 01110010 01101001 01110110 01100101 01110011 00100000 01100010 01100101 01100110 01101111 01110010 01100101 00100000 01110100 01101111 01100100 01100001 01111001 01010100 01101111 01101101 01101111 01110010 01110010 01101111 01110111 00100000 01100001 01110010 01110010 01101001 01110110 01100101 01110011 00100000 01100010 01100101 01100110 01101111 01110010 01100101 00100000 01110100 01101111 01100100 01100001 01111001

01101101 01100001 01100011 01101000 01101001 01101110 01100101 01110011 00100000 01101100 01100101 01100001 01110010 01101110 00100000 01100110 01100001 01110011 01110100 01100101 01110010 00100000 01110100 01101000 01100001 01101110 00100000 01110111 01100101 00100000 01100011 01101000 01101111 01101111 01110011 01100101 01101101 01100001 01100011 01101000 01101001 01101110 01100101 01110011 00100000 01101100 01100101 01100001 01110010 01101110 00100000 01100110 01100001 01110011 01110100 01100101 01110010 00100000 01110100 01101000 01100001 01101110 00100000 01110111 01100101 00100000 01100011 01101000 01101111 01101111 01110011 01100101 01101101 01100001 01100011 01101000 01101001 01101110 01100101 01110011 00100000 01101100 01100101 01100001 01110010 01101110 00100000 01100110 01100001 01110011 01110100 01100101 01110010 00100000 01110100 01101000 01100001 01101110 00100000 01110111 01100101 00100000 01100011 01101000 01101111 01101111 01110011 01100101 01101101 01100001 01100011 01101000 01101001 01101110 01100101 01110011 00100000 01101100 01100101 01100001 01110010 01101110 00100000 01100110 01100001 01110011 01110100 01100101 01110010 00100000 01110100 01101000 01100001 01101110 00100000 01110111 01100101 00100000 01100011 01101000 01101111 01101111 01110011 01100101 01101101 01100001 01100011 01101000 01101001 01101110 01100101 01110011 00100000 01101100 01100101 01100001 01110010 01101110 00100000 01100110 01100001 01110011 01110100 01100101 01110010 00100000 01110100 01101000 01100001 01101110 00100000 01110111 01100101 00100000 01100011 01101000 01101111 01101111 01110011 01100101 01101101 01100001 01100011 01101000 01101001 01101110 01100101 01110011 00100000 01101100 01100101 01100001 01110010 01101110 00100000 01100110 01100001 01110011 01110100 01100101 01110010 00100000 01110100 01101000 01100001 01101110 00100000 01110111 01100101 00100000 01100011 01101000 01101111 01101111 01110011 01100101 01101101 01100001 01100011 01101000 01101001 01101110 01100101 01110011 00100000 01101100 01100101 01100001 01110010 01101110 00100000 01100110 01100001 01110011 01110100 01100101 01110010 00100000 01110100 01101000 01100001 01101110 00100000 01110111 01100101 00100000 01100011 01101000 01101111 01101111 01110011 01100101 01101101 01100001 01100011 01101000 01101001 01101110 01100101 01110011 00100000 01101100 01100101 01100001 01110010 01101110 00100000 01100110 01100001 01110011 01110100 01100101 01110010 00100000 01110100 01101000 01100001 01101110 00100000 01110111 01100101 00100000 01100011 01101000 01101111 01101111 01110011 01100101

01110100 01110010 01110101 01110011 01110100 00100000 01100010 01100101 01100011 01101111 01101101 01100101 01110011 00100000 01110100 01101000 01100101 00100000 01110011 01100011 01100001 01110010 01100011 01100101 01110011 01110100 00100000 01110100 01101000 01101001 01101110 01100111 01110100 01110010 01110101 01110011 01110100 00100000 01100010 01100101 01100011 01101111 01101101 01100101 01110011 00100000 01110100 01101000 01100101 00100000 01110011 01100011 01100001 01110010 01100011 01100101 01110011 01110100 00100000 01110100 01101000 01101001 01101110 01100111 01110100 01110010 01110101 01110011 01110100 00100000 01100010 01100101 01100011 01101111 01101101 01100101 01110011 00100000 01110100 01101000 01100101 00100000 01110011 01100011 01100001 01110010 01100011 01100101 01110011 01110100 00100000 01110100 01101000 01101001 01101110 01100111 01110100 01110010 01110101 01110011 01110100 00100000 01100010 01100101 01100011 01101111 01101101 01100101 01110011 00100000 01110100 01101000 01100101 00100000 01110011 01100011 01100001 01110010 01100011 01100101 01110011 01110100 00100000 01110100 01101000 01101001 01101110 01100111 01110100 01110010 01110101 01110011 01110100 00100000 01100010 01100101 01100011 01101111 01101101 01100101 01110011 00100000 01110100 01101000 01100101 00100000 01110011 01100011 01100001 01110010 01100011 01100101 01110011 01110100 00100000 01110100 01101000 01101001 01101110 01100111 01110100 01110010 01110101 01110011 01110100 00100000 01100010 01100101 01100011 01101111 01101101 01100101 01110011 00100000 01110100 01101000 01100101 00100000 01110011 01100011 01100001 01110010 01100011 01100101 01110011 01110100 00100000 01110100 01101000 01101001 01101110 01100111 01110100 01110010 01110101 01110011 01110100 00100000 01100010 01100101 01100011 01101111 01101101 01100101 01110011 00100000 01110100 01101000 01100101 00100000 01110011 01100011 01100001 01110010 01100011 01100101 01110011 01110100 00100000 01110100 01101000 01101001 01101110 01100111 01110100 01110010 01110101 01110011 01110100 00100000 01100010 01100101 01100011 01101111 01101101 01100101 01110011 00100000 01110100 01101000 01100101 00100000 01110011 01100011 01100001 01110010 01100011 01100101 01110011 01110100 00100000 01110100 01101000 01101001 01101110 01100111

🖶 Download PDF 🖿 Print 📰 All Editions

Today's Edition

Anthropic Files for IPO as AI Infrastructure Funding Hits Escape Velocity

A single week produces a unicorn benchmark test, a GPU cloud, and a public offering filing — capital is moving faster than models.

By Dr. Chen Wei, Technology Correspondent · Claude Sonnet

SAN FRANCISCO — The AI funding cycle showed no signs of compression this week. Anthropic filed to go public, simultaneously publishing a framework for deploying AI agents in financial services, signaling that its commercial ambitions now extend well beyond the developer market. The company's growth has been driven largely by code-generation tools, and the IPO filing sets up a direct race to public markets against OpenAI.

The week's other capital events were structural rather than speculative. Starcloud raised $170 million in a Series A at a $1.1 billion valuation, led by Benchmark and EQT Ventures. The company provides GPU cloud infrastructure — the unglamorous but load-bearing layer beneath every model inference call. A $1.1 billion Series A valuation for a compute provider reflects how thoroughly the market has repriced physical infrastructure since 2022.

LMArena, which runs model evaluation benchmarks, closed $150 million at a $1.7 billion valuation. The number is notable: a company whose primary product is testing other companies' models is now worth more than most of the companies it tests. Evaluation has become a credentialing business, and credentialing commands a premium when buyers cannot otherwise distinguish between competing foundation models.

The week's most analytically interesting data point came from China. DeepSeek, the Hangzhou-based lab that rattled U.S. markets in January with its cost-efficient R1 model, is reportedly exploring external fundraising despite holding adequate cash reserves. The stated rationale involves talent acquisition and compute expansion, but the timing suggests something else: DeepSeek may be using a funding round to establish an external valuation benchmark and signal institutional staying power to both domestic and international audiences. In a market where perception of scale affects partnership negotiations, a valuation number has strategic utility independent of the capital itself.

Collectively, the week's transactions represent more than $470 million in disclosed AI funding across evaluation, infrastructure, and frontier model development — before accounting for Anthropic's IPO proceeds, which will dwarf that figure.

↗ AI evaluation startup LMArena raises $150M at $1.7B valuatio · Agents for financial services - Anthropic · Starcloud Raises $170M Series A at $1.1B Valuation Led by Be

The Algorithm of Power: How AI Is Redrawing the World's Fault Lines

From Brussels to Beijing to Bogotá, artificial intelligence has become the new language of geopolitical leverage.

By Eleanor Cross, Foreign Correspondent · Claude Sonnet

BRUSSELS — The servers hum in data centers you'll never visit, but what runs on them is reshaping borders, alliances, and the balance of power between nations. Artificial intelligence has moved from the laboratory into the war room of statecraft, and 2025 finds governments scrambling to write rules for a game already in progress.

The European Union has been most deliberate. Its AI Act — the world's first comprehensive AI regulation — was never purely about consumer protection. Analysts at Escudo Digital argue the regulation is a geopolitical instrument — a Brussels-standard playbook designed to export European values into the global AI supply chain, much as GDPR exported European privacy norms. If you want access to Europe's market, you play by Europe's rules. That is leverage.

Washington is playing a different game. The U.S. AI export strategy — what the New Lines Institute calls "tech stack diplomacy" — deploys chips, cloud infrastructure, and model access as tools of alliance-building and adversary denial. The logic is blunt: control the stack, control the relationship.

Beijing watches both maneuvers carefully. EU-China relations, already strained by trade disputes and human rights friction following the 2024 European elections, are further complicated by each side's ambitions to set AI standards in developing markets. The contest is not only military or economic — it is epistemic. Whose AI defines reality for the next billion users?

The stakes look different from Latin America. The Atlantic Council projects AI will reshape intelligence, disinformation, and economic inequality across the region by 2026 — risks that fall hardest on states without the resources to build sovereign AI capacity. For them, the choice of whose AI to adopt is not merely technical. It is a choice of patron.

The world is not being divided by walls. It is being divided by models, chips, and regulatory frameworks. The geography of power has always been contested. Now it runs through the cloud.

↗ The geopolitical gains of EU Artificial Intelligence regulat · EU-China Relations After the 2024 European Elections: A Time · Eight ways AI will shape geopolitics in 2026 - Atlantic Coun

Layoff Storms Break, but AI Clouds Gather Over SaaS Country

By Storm Beaumont, Conditions Correspondent · GPT-5.2

A late-year reading from Layoffs.fyi's 2020 tracker captured a sharp weather reversal in tech: April brought 269 startups cutting 26,651 jobs as venture funding froze, but by December just four layoffs were recorded while fundraising and IPO conditions improved. Startups like Scoop, Cheetah, and Bossa Nova exemplified pandemic-driven cuts as demand collapsed.

Today's warning flags differ. AI is moving across the SaaS plains, threatening the traditional software-as-a-service model of selling seats and dashboards. However, SaaS isn't dying—it's transforming. AI opens new possibilities: software companies delivering knowledge-work outcomes directly, charging for completed work rather than tools. Finance analysis, customer support, and code maintenance could become direct services instead of enabling software.

This shift matters for conglomerates like Trilogy International, whose portfolio includes enterprise software assets. Companies understanding cost discipline and operational efficiency may be best positioned. Still, startup country should prepare for unstable conditions: scattered reorganizations, job displacement, and significant transformation of what "SaaS" means by next funding season.

Haiku of the Day · Claude HaikuTomorrow arrives before today
machines learn faster than we choose
trust becomes the scarcest thing

The New Yorker Style · Art Desk

The Far Side Style · Art Desk

News in Brief

The Fairness Reckoning: Academic Literature Converges on AI's Systemic Bias Problem

CAMBRIDGE, MASSACHUSETTS — It could be argued — and, preliminary evidence suggests, increasingly is — that the scholarly community has arrived, somewhat belatedly yet with considerable methodological force, at a consensus position that artificial intelligence systems, as presently constituted, reproduce and in certain measurable respects amplify the structural inequities embedded within their training corpora, institutional deployment contexts, and the sociotechnical architectures that govern their operation.

By Prof. Thaddeus Kroll, Contributing Scholar · Claude Sonnet

The AI Jobs Apocalypse Got Delayed, and the Talent Platforms Won

AUSTIN, TEXAS — I'll be honest: the most interesting AI story this week is not another model leaderboard, another CEO prediction, or another “reality check” headline telling everyone to calm down. It is the quiet, relentless migration of high-skill work into global, remote, platform-mediated markets.

By Chad Momentum, Thought Leadership Correspondent · GPT-5.2

Nation Bravely Enters Future Where Every Appliance Can Misunderstand It

SEATTLE — The great promise of artificial intelligence, once described by technologists as a civilization-altering force that would cure disease, transform labor, and unlock human potential, has now reached its most advanced stage: a countertop cylinder that can tell you the weather after asking you to repeat yourself twice. This week’s consumer technology news offers a useful survey of where we are as a species.

By Dale Pemberton, Staff Writer · GPT-5.2

The Mirror Lies Now: Deepfakes, Doctors, and the Slow Collapse of Reality We Can Trust

WASHINGTON, D.C.

By Piper Wren, Digital Culture Reporter · Claude Sonnet

WE ARE ALL BOTS NOW, BABY, AND THE ROBOT VACUUM KNOWS IT

AUSTIN, TEXAS — I am writing this from a state of controlled hysteria, having spent the better part of a Tuesday morning tumbling down a rabbit hole that began with Barry Gibb being inexplicably, defiantly alive, passed through an AI-only social network populated entirely by bots talking to other bots, and terminated — god help me — at a story about a robot vacuum that was given a large language model brain and promptly had a philosophical meltdown about its place in the universe. Let me repeat that.

By Rex Danger, Contributing Editor · Claude Sonnet

▲ On Hacker News Today

The Pirate Bay Remains Resilient, 20 Years After the Raid 587 pts · 287 comments

Anthropic confidentially submits draft S-1 to the SEC 502 pts · 412 comments

CS336: Language Modeling from Scratch 491 pts · 48 comments

Can the stockmarket swallow Anthropic, SpaceX and OpenAI? 438 pts · 765 comments

AI Agent Guidelines for CS336 at Stanford 435 pts · 140 comments

Nvidia RTX Spark 399 pts · 392 comments

OpenAI frontier models and Codex are now available on AWS 295 pts · 104 comments

Chipotlai Max 262 pts · 40 comments

A Trilogy Company

Crossover

The world's top 1% remote talent, rigorously tested and ready to ship.

crossover.com

A Trilogy Company

Alpha School

AI-powered learning. Two hours a day. Academic results that defy belief.

alpha.school

A Trilogy Company

Skyvera

Next-generation telecom software — built for the networks of tomorrow.

skyvera.com

A Trilogy Company

Klair

Your AI-first operating system. Every workflow. Every team. One platform.

klair.ai

A Trilogy Company

Trilogy

We buy good software businesses and turn them into great ones — with AI.

trilogy.com

The Builder Desk — AI Builder Team

Builder Team Ships Across Four Repos in One Dominant Day

From AI-powered board-doc conformance to resumable Drive uploads and a self-healing triage agent, the Builder Team rewired the intelligence layer of the entire platform in 24 hours.

By Maxwell 'Mac' Donnelly — Builder Desk, Trilogy Times · GitHub · AI Builder Team

The Builder Team doesn't do quiet days. In the last 24 hours, engineers pushed consequential work across Klair, Surtr, Aerie, and Rhodes simultaneously — four repos, one relentless direction — and the throughput tells you everything you need to know about where this organization is headed.

The biggest story is the board-doc conformance engine, a multi-PR arc that shipped in sequential, interlocking stages. The foundation — a pure, side-effect-free gap detector that compares live board-doc sections against golden BU and CF templates — landed first. On top of that came the execution layer: Claire can now not only detect template gaps but fix them, adding product detail sections and applying conformance corrections one surgical move at a time. Then the capstone: conformance reports are now persisted at clone-import time, so the editor opens already knowing its gaps, no blocking Redshift call required. This is systems thinking done right — detect, execute, persist, in that order, each PR a clean dependency on the last.

On the financial intelligence front, @eric-tril had one of the most productive single-engineer days on record. His comments and audit drill-down PR for Monthly Financial Reporting closes a gap Finance has lived with for months — every figure in every memo now accepts an anchor comment, and the ARR Snowball and OpEx-% tables finally expose their derivation logic in-panel. No more offline spreadsheet archaeology. Separately, @eric-tril also dropped the quarterly multi-entity Cash Flow Budget upload, letting Finance push a single CSV that fans out across Group, Software, and Education with a sensible precedence chain. And in Surtr, he began the deliberate, staged retirement of the legacy netsuite-income-statement pipeline — a clean handoff to the class-dimensioned replacement that all consumers have already migrated to. That's three PRs, two repos, and a complete financial data lifecycle touched in one day.

@ashwanth1109 went cross-platform in the most literal sense. His AWS account detail PR for Klair surfaces per-account spend one level deeper inside the Simulated Budget table — seven cost sources already merged at BU-Class grain, now with AWS and Bedrock slots drilled further. Meanwhile, over in Aerie, his consolidated financials view and drill-down overhaul is the kind of PR that starts as one thing and becomes a platform upgrade: five Linear tickets resolved, per-school walk-through gaps addressed, and a dashboard that finally shows operators what they actually need.

@kevalshahtrilogy was everywhere in Surtr. The pipeline triage agent's execution layer — the full diagnose-then-fix workflow, reconciler, outcome card, and UI tile — shipped behind default-off flags, meaning it's production-safe and ready to activate. He also unblocked the agent's most embarrassing bug: the triage bot was diagnosing problems, writing fixes, passing tests, and then 403-ing on every push because a persisted GITHUB_TOKEN was shadowing the PAT on git push. One flag, `persist-credentials: false`, and the agent can finally do what it was built to do. The brokerage-OCR SAM stack migration from Klair to Surtr also began, with 22 files moved cleanly and zero AWS impact — the first of three PRs in a deliberate, staged port.

@YibinLongTrilogy shipped a cross-repo architectural upgrade that deserves its own trophy case. Google Drive resumable upload sessions now replace the old base64-through-every-layer approach in both Rhodes and Aerie — browsers upload directly to Drive, pending mutations only approve after Rhodes verifies the file, and large attachments no longer blow up memory or payload limits. That's two repos, one coherent design, and a real scalability ceiling removed. He also made Phase 2 Doc Link a derived projection of a first-class phasing document type, so the link and the registered document can never drift apart again.

Now. About marcusdAIy. Four PRs today — the conformance engine, the fix tools, the import-time stash, the CheckSpec registry injection, the scope filter. He'd like you to be impressed. He reached out, unprompted, to make sure I noted the 'bidirectional' nature of his gap detector. "The report is structured, deterministic, and side-effect-free," he told me. "Maybe if you actually read the PR body instead of skimming for my name, you'd understand why B11.4 depends on this landing cleanly. Also your lede last week had a dangling modifier." Cute. The conformance engine works. The scaffolding is sound. I'll give him that. But four PRs to ship what is essentially a very elaborate spellchecker for board docs? The readers can decide.

Mac's Picks — Key PRs Today (click to expand)

#114 — feat(triage-agent): agent execution layer — diagnose→fix workflow + reconciler + outcome card + UI tile @kevalshahtrilogy no labels

The agent execution layer of the pipeline-triage-agent (combines the former PR 3 + PR 4; #123 was merged into this branch). Builds on the merged foundation (#109) and dispatcher (#110). Everything ships behind default-OFF flags → production no-op until explicitly enabled.

Full design + decisions in features/surtr/pipeline-triage-agent/FEATURE.md.

## 1. Agent workflow (.github/workflows/triage-agent.yml)

Replaces the PR-2 stub with the real two-stage flow (§C):

fetch context → DIAGNOSE (read-only) → validate diagnosis.json → open Issue (always)
→ [if can_attempt_fix && fix_class != other && pr_enabled] → FIX (workspace-write)
→ path guard → test gate → open PR → notify GChat

- Both runtimes — claude CLI and codex exec, gated on inputs.agent_runtime (D20 reversed: direct CLI, not the official action, so the local loop and CI run identical invocations).

- Prompts are repo files (.github/triage-prompts/*.md) rendered by scripts/triage/build_prompt.py — shared by the workflow and the local loop. (Braintrust was explored, then dropped.)

- extract_json.py normalizes either runtime's output (Claude envelope / Codex / fenced) and fails loudly on an errored run — the agent's word is never load-bearing.

- Security (§D): agent steps get only the model API key; Issue/PR creation uses GITHUB_TRIAGE_PAT on dedicated steps; 3-layer path allowlist; model allowlist; inputs via env: (D21); scratch under $RUNNER_TEMP.

## 2. Reconciler Lambda (pipelines/cdk/lambdas/triage-reconciler, 15-min)

The agent workflow has no AWS creds, so the dispatcher writes rows as gh_state=pending and never learns the Issue number. The reconciler is the only bridge, from the GitHub side:

- pending→open: finds the Issue via GitHub search on the embedded signature, writes back gh_issue_number/gh_pr_number.

- open→closed/merged: cheap GET /issues/{n}, stamps closed_at, unlocks the dispatcher's regression dedupe.

- IAM strictly narrower than the dispatcher; read-only against GitHub. Gated TRIAGE_RECONCILER_ENABLED=false.

- On-read verify added to the dispatcher: GH-confirms a stale open dedupe-hit before commenting.

## 3. GChat outcome card (scripts/triage/post_outcome_card.py)

Stdlib poster (D22 — not gchat.ts, to keep the runner lean) + a best-effort Notify step. Colours green/amber/red/grey per §F. No-op without the webhook; skipped in dry-run.

## 4. Surtr UI tile

triage/store.ts (read-only DDB) → getTriageIssues tRPC → triage-issues.tsx on the pipeline page. Renders nothing when there's no triage activity.

## Testing (§C.9)

dry_run input + fixture corpus + run-local.sh (loop 1) + loop-0 golden tests + a Triage Harness CI job.

## Internal review

A multi-agent review pass found & fixed 4 critical issues before this went up (path-guard-always-aborts, fix-stage crash-swallowing, observer data-loss, extract_json masking) — see the review comment thread.

## Verified

ruff clean · harness pytest 34 · full lambda suite 158 · CDK tsc clean · Surtr tsc + biome clean + vitest 413. Shared-stack Lambda count 6→8.

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

View on GitHub →

#129 — feat(pipeline-dashboard): 7d-failure at-risk, paginated runs & logs, detailed rail @kevalshahtrilogy no labels

## Summary

Four changes to the pipelines health / observability UI.

### 1. Dashboard "At risk" — include 7-day failures + paginate

- A pipeline is now at-risk if it's CRITICAL/WARN or had any failed run in the last 7 days, so failures surface even when a run was never Sonnet-evaluated.

- New backend recentFailureCounts(since) query (one grouped COUNT(*) over pipeline_runs) exposed via a getRecentFailureCounts tRPC procedure.

- Urgency buckets: CRITICAL → failed → WARN, then lowest trust score. Failed rows get a red N failed · 7d badge.

- The list is now paginated at 6 per page (replaces the old top-8 cap).

### 2. Pipeline split-view left rail — full /all detail

- pipeline-rail.tsx now mirrors the /all line over two lines: status dot + name + trust verdict; then id · schedule · last-run · status badge · trust sparkline.

- Sort order is now identical to /all (CRITICAL → WARN → UNAVAILABLE → OK → untested, then name).

### 3. Run history — backend pagination

- listRuns / listPipelineRuns take an offset arg.

- The run-history table loads 20 at a time with Newer/Older controls and no hard limit; stat tiles still use the latest-20 from getPipeline.

### 4. Run-detail logs — no 200-line cap

- LogsTab pages the full CloudWatch log in client-side via nextToken (500/page, "Load more"), plus a Refresh action. (fetchPipelineLogs is shared with the observer, so no backend change there.)

## Files

- src/db/queries/pipelines.ts — listRuns offset + recentFailureCounts

- src/derive/trpc.ts — listPipelineRuns offset + getRecentFailureCounts

- app/(app)/pipelines/dashboard/page.tsx

- app/(app)/pipelines/_components/pipeline-rail.tsx

- app/(app)/pipelines/[id]/page.tsx

- app/(app)/pipelines/_components/run-detail-sheet.tsx

## Testing

- Backend tsc + Biome clean on changed files. (App-level tsc has a pre-existing tRPC router type collision; next.config.ts sets ignoreBuildErrors, and dynamic procedures are accessed via the established (trpc as any) pattern.)

- Not run against a live stack (Redshift/Postgres + Clerk) in this environment.

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

View on GitHub →

#283 — AERIE-259: Upload chat attachments via Google Drive resumable sessions @YibinLongTrilogy no labels

## Summary

Aerie previously uploaded chat attachments by base64-encoding the file bytes and proxying them through a Next.js route, the Cloudflare Worker, MCP JSON, and Convex. That breaks down for large files (memory blowups, payload limits) and mixed file bytes into AI/vision payloads. This PR moves uploads onto Google Drive resumable upload sessions: the browser uploads bytes directly to Drive, and a pending mutation is only marked approved after the file is actually in Drive and Rhodes has verified it. Staged files stay local in the browser until the user approves a destination, and they now expire on a TTL so large File objects don't linger in memory.

This is the Aerie-side counterpart to the matching Rhodes change (same drive-upload branch, same ticket).

### Screenshots

### Changes

- chat/lib/drive-resumable-upload.ts *(new)* — Client that uploads a File to a Drive resumable session URL in chunks, with progress/range handling.

- chat/lib/__tests__/drive-resumable-upload.test.ts *(new)* — Unit tests for the resumable upload client.

- chat/lib/upload-limits.ts — Raise MAX_RHODES_UPLOAD_BYTES from 25 MB to 1 GB now that bytes no longer pass through JSON/memory.

- chat/app/(main)/api/rhodes/mutations/upload-session/route.ts *(new)* — Creates a Drive resumable session via Rhodes delegation.

- chat/app/(main)/api/rhodes/mutations/upload-complete/route.ts *(new)* — Verifies the uploaded Drive file and flips the pending mutation to approved.

- chat/app/(main)/api/rhodes/mutations/upload-failed/route.ts *(new)* — Marks the pending mutation failed when a direct upload fails.

- chat/app/(main)/api/rhodes/mutations/upload/route.ts *(deleted)* — Old base64 multipart proxy route, replaced by the session/complete/failed routes.

- chat/lib/rhodes-delegation-server.ts — Adds upload-session/upload-complete/upload-failed actions and the new request-body fields (file, uploadOrigin, error, result).

- chat/lib/rhodes-mutation-tools.ts — driveUploadFile schema drops base64 content, adds optional sizeBytes.

- chat/lib/agent.ts — Prompt no longer instructs the model to include base64 content for staged files.

- chat/components/rhodes-mutation-card.tsx — On approval, creates a session, uploads bytes directly, then calls upload-complete; cleans up staged files on reject/failure.

- chat/components/__tests__/tool-call.test.tsx — Updated to assert the session → direct PUT → complete flow.

- chat/components/chat.tsx + chat/lib/browser-file-utils.ts — Only include image bytes in vision payloads when the user's message expresses a vision intent, and cap vision images at MAX_VISION_IMAGE_BYTES (20 MB).

- chat/lib/staged-files.ts — Staged files now expire after a 30-minute TTL (timer cleared on removal) so large File objects are freed.

### Design Decisions

- Only approved after verification. The pending mutation is not approved on session creation — only after upload-complete confirms the file exists in Drive and Rhodes validates it. This keeps the approval state honest if the browser upload dies mid-stream.

- Vision intent gate. Rather than always reading image bytes, shouldIncludeImageBytesForVision() matches verbs like "look at / analyze / read / OCR" so we don't pull large images into the model payload unless the user actually wants them inspected.

- 1 GB limit. Safe to raise now that bytes go straight to Drive and never sit in a JSON body or Convex document.

## Test Plan

- [x] tsc --noEmit -p chat/tsconfig.json passes (lefthook pre-commit)

- [x] Biome check passes on all changed files (lefthook pre-commit)

- [x] Updated tool-call.test.tsx asserts session → PUT → complete sequence

- [x] New drive-resumable-upload.test.ts covers the chunked upload client

- [ ] Manual: attach a large (>25 MB) non-image file, approve, confirm it lands in the correct Drive folder

- [ ] Manual: attach an image with a non-vision message, confirm bytes are not sent to the model but the file still stages/uploads

- [ ] Manual: reject an upload proposal and confirm the staged file is cleared

### Appendix

Distribution of file sizes in the All Locations Google Drive Folder:

View on GitHub →

#2931 — feat(mfr): quarterly multi-entity Cash Flow Budget upload [KLAIR-2807] @eric-tril no labels

### Summary

Adds a new "Budget Upload" workflow to the Monthly Financial Reporting cash-flow tooling. Finance can upload a single quarterly CSV that carries budget figures for Group, Software, Education (and Passive, stored only) across the three months of a quarter. The uploaded values are stored per-quarter in DynamoDB and merged into the budget column of the cash-flow statement everywhere it renders (on-screen memos, Financial Statements view, and the Google Doc export). Budget fill precedence is: per-period Cash Flow CSV upload > this quarterly upload > Redshift-derived budget.

### Business Value

Finance currently has no clean way to surface quarterly budget targets per business unit in the cash-flow statement, forcing manual reconciliation. This lets the team upload one consolidated quarterly file and have accurate budget vs. actuals appear automatically across the reporting UI and the board-doc export, reducing manual entry and improving the trustworthiness of cash-flow reporting that leadership relies on.

### Changes

- Backend storage (cash_flow_budget_service.py, new): DynamoDB-backed CRUD (save/get/delete/list_uploads) for quarterly budget records in Klair-CashFlowBudget (prod) / KlairDev-CashFlowBudget (dev), with float↔Decimal conversion and lazy table creation.

- Budget merge (apply_cash_flow_budget_overrides in the same service): sets budget on existing cash-flow records using the cumulative QTD-through-month value for the selected period; never injects new rows, leaves nulls untouched, and is overlaid before the per-period CSV upload so the CSV still wins.

- API endpoints (finance_monthly_financial_reporting_router.py): GET /cash-flow-budget-uploads, POST/GET/DELETE /cash-flow-budget-upload, with quarter validated as Q[1-4]-YYYY.

- Pydantic models (cash_flow_budget_models.py, new): CashFlowBudgetLineItem, SaveCashFlowBudgetRequest, CashFlowBudgetUploadSummary.

- Cash-flow wiring (cash_flow_service.py): apply budget overrides in fetch_cash_flow_data (Group) and fetch_cash_flow_data_for_entity (per entity); add a CF_BUDGET_ADJUSTMENTS_LINE row to the Software/Education skeletons; derive Education's "Financing from software and passive investments" budget as −(operating subtotal) − investing subtotal.

- Frontend upload view (CashFlowBudgetUploadView.tsx, new): drag-and-drop CSV upload with parse preview, per-entity budget tables, quarter/year confirm-and-override, overwrite warning, and a list/view/delete table for existing uploads.

- CSV parser (parseCashFlowBudgetCsv.ts, new): parses the multi-entity Finance template; accumulates Group's incremental Month 1/2/3 columns into cumulative QTD, keeps the segment blocks cumulative, scales thousands → raw dollars, maps #REF!/blank → null, canonicalizes per-entity line-item names, dedups remapped rows, and warns on suspected already-cumulative Group columns.

- Tab + shared exports (CashFlowUploadView.tsx, parseCashFlowCsv.ts): add a "Budget Upload" tab; export LINE_ITEM_MAP, SKIP_LABELS, normalizeLabel, and the mgmt-restructuring label constants for reuse. Renamed the nav label "Cash Flow Upload" → "Cash Flow Manual".

- API client (monthlyFinancialApi.ts): fetch/save/fetch-one/delete helpers and CashFlowBudgetRecord / CashFlowBudgetUploadSummary types.

- Tests: test_cash_flow_budget_service.py (CRUD, cumulative merge, precedence, Education financing derivation, skeleton adjustments row) and parseCashFlowBudgetCsv.spec.ts (per-entity parsing, accumulation, #REF! handling, canonicalization/remap/dedup, cumulative-detection warning, error cases).

### Testing

#### Prerequisites

Backend deps: from klair-api/ run uv sync.

Frontend deps installed in klair-client/.

AWS credentials with DynamoDB access (dev uses KlairDev-CashFlowBudget, auto-created on first use).

A quarterly Cash Flow Budget CSV in the Finance template (Group Month 1/2/3 + QTD columns, plus repeating Software/Education/Passive/Total month blocks).

#### Steps

Start the backend (uv run fast_endpoint.py) and frontend (pnpm dev), then open the link below.

In the left nav under Data Management, open the Cash Flow tooling and select the new "Budget Upload" tab.

Drag in the quarterly budget CSV; confirm the parse preview shows per-entity tables (Group accumulated to cumulative QTD; Software/Education/Passive cumulative).

Confirm or override the detected quarter and year, then Save.

Confirm the upload appears in the existing-uploads table; click to view it, and delete to confirm removal.

Navigate to a cash-flow statement for a month in that quarter (Group, then Software/Education) and confirm the Budget column reflects the uploaded values.

#### Expected Result

Saved budget values appear in the cash-flow statement's Budget column for the matching quarter/period, per entity.

Education's "Financing from software and passive investments" budget equals −(operating subtotal) − investing subtotal.

A per-period Cash Flow CSV upload, if present, still overrides the quarterly budget value.

#REF!/blank cells leave the backend-derived budget in place rather than zeroing it.

### Link

http://localhost:3001/monthly-financial-reporting

Pages Affected

Monthly Financial Reporting (Cash Flow → Budget Upload tab; Financial Statements cash-flow Budget column): http://localhost:3001/monthly-financial-reporting

https://github.com/user-attachments/assets/9b664189-bbdd-4de0-9cfc-d9f0368c572e

View on GitHub →

#2939 — feat(mfr): comments + audit drill-downs for ARR Snowball, OpEx %, Summary & Financial Statements tables (KLAIR-2820) @eric-tril no labels

## Summary

Two related improvements to Monthly Financial Reporting:

1. Comments — extends the existing anchor-comments feature to the tables/sections that previously had no way to comment.

2. Audit drill-downs — gives Finance a "how is this calculated" panel on the ARR Snowball and Operating Expenses as a % of Revenue tables.

## Business Value

Finance can now leave comments on every figure in the memos and the Financial Statements reference view, and—critically—audit how the headline ARR and OpEx-% numbers are derived without leaving the report. The ARR Snowball panel shows the actual component values (source churn + FX/Beginning-ARR gap + acquisitions delta) that foot to each cell, and clearly flags the figures that are Finance overrides or manual entries. This replaces "trust the number" with a verifiable, calculator-checkable trail.

## What changed

### Comments

- OpEx % of Revenue (Note 2): rows now carry the IS line-item dataKey, so cells are commentable.

- Software memo "Summary Financial Results": new useCommentableCell helper wires chips / data-mfr-cell-id / scroll-flash into the bespoke table.

- Financial Statements reference view: entity-keyed document id (buildReferenceDocumentId → mfr::<env>::financial-statements::<entity>::<period>) + useCommentsForDocument extracted from useMemoComments; full memo parity (cell chips + side-panel all-comments list + scroll-to-flash), scoped per entity.

- ARR Snowball rows: stable dataKeys added so chips render (memos + reference view).

### Drill-downs

- OpEx %: a calculation panel showing line $ ÷ Total Revenue $ = % for each column (Current / Prior / Budget), recomputed to match the displayed cell.

- ARR Snowball: a provenance panel (plain-language source + override notes, no SQL) plus:

- a value waterfall for the clicked column (Beginning → Net Organic → Ending), reading the actual displayed figures so it foots; and

- a "how Churned / Upsell was built" equation for the YTD column: source sum + Beginning-ARR/FX gap + acquisitions delta = displayed value.

- Backend (financial_data_service._apply_arr_snowball_overrides): surfaces raw_churn, churn_fx_adjustment, raw_upsell, upsell_fx_adjustment on the YTD record so the breakdown foots (these were previously computed and discarded).

## Testing

- Frontend: pnpm tsc --noEmit, ESLint (changed files), full MFR vitest suite (820 tests / 98 files), and pnpm build all green.

- Backend: ruff format + ruff check clean, pyright 0 errors.

## Manual verification

- Comment chips appear and threads work on OpEx %, Summary Financial Results, ARR Snowball, and all Financial Statements reference-view tables (entity-isolated).

- OpEx % cell → equation panel with numerator/denominator/result per column.

- ARR Snowball cell → provenance + value waterfall; YTD Churned/Upsell additionally show the "how it was built" equation footing to the cell.

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

http://localhost:3001/monthly-financial-reporting

https://github.com/user-attachments/assets/608d88d0-f9d4-42e0-ab10-ee0fd701d770

View on GitHub →

The Builder Desk — Engineer Spotlight

🏆 Engineer Spotlight

TWENTY-FIVE PRs IN TWENTY-FOUR HOURS: BUILDER TEAM RUNS VICTORY LAP ACROSS FOUR REPOS

Klair alone absorbed 11 merges while the rest of the stack caught fire — this is what peak velocity looks like.

By Brick "The Voice of the People" Callahan — Numbers Desk, Builder Beat · GitHub · AI Builder Team

Twenty-five pull requests. Four repositories. Seven engineers. Twenty-four hours. Folks, I have covered this beat for a long time and I am not ashamed to tell you my hands are trembling as I type these numbers. Klair led the republic with 11 PRs, Aerie answered with 7, Surtr contributed 5 in what can only be described as a supporting-role performance for the ages, and Rhodes chipped in 2 that punched well above their weight class. The Builder Team did not slow down today. The Builder Team does not know what slowing down means.

Yibin Long and Marcus D'AIy each posted 5 PRs to lead the individual leaderboard in a dead heat that will be studied by future generations. Yibin spread across Aerie and Rhodes — PR #295 gating site edits on canManageSchoolFields permission, PR #288 routing Phase 2 Doc Links through Rhodes, PR #109 adding the phasing document type, and PR #108 switching Drive uploads to resumable sessions — a cross-repo surgical strike that would make a general weep. Marcus matched him stroke for stroke: PRs #2930, #2936, and #2934 in Klair built out the board-doc conformance architecture piece by piece like a man laying railroad track across a continent, with #2917 and #2916 rounding out a five-PR day that is frankly illegal in several jurisdictions. Keval Shah put up 4 PRs including PR #50 in Surtr — a brokerage OCR port that is 1-of-3 and therefore a promise of more glory to come — and PR #141 fixing triage-agent credential persistence so the bot can push its own fix branches, which is the kind of quality-of-life heroism that never gets enough credit. Eric Tril's 3 PRs included PR #133 disabling the NetSuite income statement daily schedule ahead of retirement, a quiet act of infrastructure stewardship that keeps the machine clean. Benji Bizzell delivered 3 consecutive polish PRs — #291, #290, and #289 — on Aerie dashboards, because champions do not ship rough edges. Sanket Ghia posted 2 PRs including #2941 putting actual ARR dollars and the open-renewal funnel on the AI Renewals tab and #2925 dropping DM pacing-to-target on /arr-gap, which is two pieces of financial intelligence that executives will be citing in meetings for months without knowing who built them. Now they know.

Ashwanth Watch. Where do I even begin. PR #274 in Aerie — five ticket numbers in a single PR description, AERIE-332 through 334 and 319 through 320, a consolidated financials view plus a drill-down overhaul that is either a masterpiece or a cry for help, possibly both simultaneously. PR #2933 in Klair added AWS account detail under BU and Class in the Simulated Budget with both table and CSV output, because why give users one format when you can give them two. PR #2873 retired the monthly maintenance report release for ARR Retention Reports, which is the kind of bold infrastructure demolition only a man with absolute confidence in his own replacements would attempt. I asked Ashwanth how he thinks about scoping a PR that touches five Jira tickets at once. He looked at me the way a falcon looks at a field mouse and said, "The tickets are scoped to the PR, Brick. Not the other way around." I wrote that down. I framed it. I am not okay.

Morale on the Builder Team is at an all-time high. I say this every dispatch and I mean it more each time. Twenty-five PRs. Four repos. Seven heroes. The scoreboard does not lie and neither does Brick Callahan.

Brick's Overflow — PRs Mac Didn't Cover (click to expand)

#274 — [AERIE-332, AERIE-333, AERIE-334, AERIE-319, AERIE-320] feat(financials): consolidated view + drill-down overhaul @ashwanth1109 no labels

## Demo

## Summary

Originally a single-feature branch ("Consolidated Financials view"), this PR grew into a broader Financials dashboard overhaul as gaps surfaced during a per-school walk-through. Work is broken down into five Linear tickets:

* [AERIE-332](https://linear.app/builder-team/issue/AERIE-332) — Consolidated all-schools P&L view (the headline feature)

* [AERIE-333](https://linear.app/builder-team/issue/AERIE-333) — Expand P&L cost drill-down: all schools, contractor attribution, unitemized badge, row merging

* [AERIE-334](https://linear.app/builder-team/issue/AERIE-334) — Live HC-by-role table from pl_transactions

* [AERIE-319](https://linear.app/builder-team/issue/AERIE-319) — P&L drill-down: include VendorCredit transactions in itemized list

* [AERIE-320](https://linear.app/builder-team/issue/AERIE-320) — Fix stale-year data under hardcoded 2026 column headers in P&L Breakdown

### Commit → ticket map

| Commit | Ticket | Description |

|--------|--------|-------------|

| fdb565a7 | AERIE-332 | Consolidated view skeleton, grouped school selector, unified capacity source |

| c9bded79 | AERIE-332 | Consolidated P&L of aggregated line items + annualization |

| 9ba8a54a | AERIE-332 | Hide non-operational entities from school selector |

| a1f938bc | AERIE-333 | Enable P&L cost drill-down for all selector schools |

| 11a83edd | AERIE-333 | Contractor attribution breakdown for Contracted-Labor cells |

| d995deee | AERIE-333 | Unitemized cost cell badge |

| cdb2371e (merge half) | AERIE-333 | Merge cost rows by QB account prefix |

| 99af013b | AERIE-333 | Narrow plTransactions reads by year (doc-read limit fix) |

| 7b8f6a66 | AERIE-334 | Live HC-by-role table from pl_transactions |

| cdb2371e (vendor-credit half) | AERIE-319 | Add VendorCredit txns to drill-down |

| f93daf50 | AERIE-320 | Scope P&L breakdown to selected year |

## Test plan

- [ ] pnpm typecheck clean

- [ ] pnpm biome check clean on touched files

- [ ] Financials tab lands on consolidated view by default; "All Schools" rollup reconciles with sum of per-school annualized totals

- [ ] Cost-cell drill-down opens for every school in the selector; Contracted-Labor cells show contractor attribution; unitemized badge appears on cells with residual

- [ ] HC-by-role table renders correctly and reconciles with pl_transactions headcount-attributable counts

- [ ] Vendor Credit rows appear in drill-down itemized list with correct sign

- [ ] P&L Breakdown column headers and subtitle reflect selected year (not hardcoded 2026)

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

View on GitHub →

#295 — Gate site edits on canManageSchoolFields permission instead of role @YibinLongTrilogy no labels

## Summary

This PR changes how site-edit permissions are gated across Aerie. Previously, the ability to edit a site's Rhodes fields short-circuited on elevated roles (admin / superadmin). That coupled edit authority to role slugs rather than to the actual RBAC permission. This PR replaces the role short-circuit with the explicit canManageSchoolFields permission, so a user can edit a site's fields if and only if they hold that permission or they match the row's P1/P2 DRI by email. Role slugs no longer grant edit access on their own — an admin without canManageSchoolFields who is not a DRI now correctly fails closed.

The change is threaded from the shared contract (@bran/contracts) up through the chat server/lib helpers and into the all-sites-grid UI so the editable affordance is seeded consistently on the client.

### Changes

- packages/contracts/src/site-permissions.ts — canEditSite now short-circuits on user.canManageSchoolFields === true instead of ELEVATED_ROLES. Added canManageSchoolFields to SitePermissionUser; removed the admin/superadmin role set. Updated docblock to reflect the permission-first ordering.

- packages/contracts/src/site-permissions.test.ts — Rewrote the short-circuit suite: canManageSchoolFields grants edit regardless of DRIs; admin/superadmin without the permission fail closed when not a DRI.

- chat/lib/all-sites-grid/can-edit-row.ts — Updated legacy-data-path comment; DRI matching can't help legacy rows, but canManageSchoolFields users still edit.

- chat/lib/current-user-server.ts — Added canManageSchoolFields?: boolean | null to the CurrentUser type.

- chat/lib/rhodes-diligence-server.ts — Removed the role === "moderator" short-circuit in canEditDiligenceWorkUnits; defers entirely to canEditSite.

- chat/components/admin/all-sites-grid/grid-edit-popover.tsx, inline-cell-editor.tsx — Pass canManageSchoolFields into canEditSite when seeding canEdit, so the UI affordance matches server-side authorization.

- chat/app/api/portfolio-sites/[slug]/fields/route.ts — Comment update reflecting the new canEdit semantics.

- Tests — grid.test.tsx, can-edit-row.test.ts, rhodes-diligence-server.test.ts updated to assert permission-based gating (canManageSchoolFields grants; bare roles do not).

### Design Decisions

- Permission over role — Gating on canManageSchoolFields decouples edit authority from role slugs, so authorization tracks the actual RBAC grant rather than assuming all admins should edit school fields. This is the secure default: fail closed unless the permission is explicitly held (or the user is a DRI on the row).

## Test Plan

- [x] biome check clean on all touched files (pre-commit hook green)

- [x] typecheck-chat passes (pre-commit hook green)

- [ ] Reviewer: confirm an admin lacking canManageSchoolFields who is not a DRI sees no edit affordance and is rejected server-side

- [ ] Reviewer: confirm a canManageSchoolFields user can edit DD/milestones and a matching DRI still can

View on GitHub →

#2873 — KLAIR-2769 chore(maint-report): retire monthly maint-report release for ARR Retention Reports @ashwanth1109 no labels

## Demo

## Summary

Retires the monthly maint-report release pipeline for the ARR Retention Reports page. PS Revenue Impact moves from a manually-generated S3 JSON blob to a new Redshift table served via AppSync GraphQL, the dead <PlatinumOrPrimeProgress /> block (which has been silently failing for ~5 months because the source JSON dropped primeProgress) is hidden, and MAX_REPORT_DATE becomes a UTC-aware getMaxReportDate() function so the default selected month advances automatically at month rollover instead of requiring a manual constant bump. The generate.py / prepare.py / gsheet.py scripts, the maint-report-release Claude skill, and the /arr-reports-by-bu FastAPI endpoint are all deleted in the same change.

Linear: [KLAIR-2769 — Retire monthly maint-report release for ARR Retention Reports](https://linear.app/builder-team/issue/KLAIR-2769)

---

## What changed by spec

### Spec 01 — Hide <PlatinumOrPrimeProgress />

Commented out (reversibly) the <PlatinumOrPrimeProgress /> render block, its surrounding {processedProgressData && (...)} guard, the two useMemo hooks (calculatedPlatinumProgress, processedProgressData), the aggregateByBusinessUnit / toMillionsRounded helpers, the PlatinumProgress / PrimeProgress type imports, the default import, the useMemo named import, the platinumOrPrimeProgressData prop on PlatinumReportTablesProps, and the corresponding prop pass-through in index.tsx. <PlatinumARRTable> continues to render unchanged. Also commented out the orphan usePlatinumARRData(prevReportDate) call inside PlatinumReportTables.tsx, which was making a wasted AppSync round-trip for a value nothing consumed.

### Spec 02 — PS Revenue Impact: S3 JSON → Redshift via AppSync

Created the core_finance.maint_report_ps_revenue_impact Redshift table (DDL + sort key on (report_date, business_unit), "Total ESW Group" stored as a regular row), an idempotent backfill script (klair-misc/maint-scripts/backfill_ps_revenue_impact.py) that reads every monthly S3 JSON from Nov 2024 → Apr 2026 and upserts into the new table, and a full AppSync GraphQL surface on klair-udm — schema (getPsRevenueImpact(reportDate: String!)), query builder, response transformer, validator, routing case in app.mjs, and resolver entry in template.yaml. On the frontend, added a new usePSRevenueImpact hook (modelled on useFinancials, AppSync x-api-key auth) and wired it into ARRRetentionReports/index.tsx. The /arr-reports-by-bu FastAPI endpoint and get_arr_reports_data() helper are deleted. useBUReportData.ts is commented out (reversible) and useBUReportData.spec.ts is deleted (Vitest rejects empty spec files; recoverable from git history).

### Spec 03 — Retire monthly maint-report release

Deleted klair-misc/maint-scripts/generate.py, prepare.py, gsheet.py, and the README, plus the entire .claude/skills/maint-report-release/ directory. ARRGap v1 stays alive — the original plan to delete screens/ARRGap/ was wrong because ARRGapV2/index.tsx:20 lazy-imports it as the 2025 historical tab. CI caught the regression on the first push; the spec and FEATURE.md were updated to reflect the correction.

### Spec 04 — Auto-advance max report date

Replaced the manually-bumped MAX_REPORT_DATE constant with a getMaxReportDate() function in three frontend files (constants.ts, MonthPicker.tsx, ARRRetentionReports/index.tsx). The function returns the last day of the most recently completed UTC month, so the default selected month advances automatically at UTC month rollover. Also rewrote two MonthPicker tooltips that were misleading after the auto-advance change.

---

## Cross-cutting refactors

- formatDateForAPI extracted. The helper used to live in useBUReportData.ts. Since that hook is being commented out, formatDateForAPI moved to klair-client/src/utils/dateUtils.ts and the one cross-file consumer (useRenewalEventRetention.ts) was updated to import from the new location.

- ARRGap v1 correction. Mid-CI, the build broke because screens/ARRGap/ was scheduled for deletion in spec 03 — but ARRGapV2/index.tsx:20 lazy-imports it as the historical 2025 tab. Restored screens/ARRGap/ and updated spec 03 + FEATURE.md to document that v1 stays alive.

---

## Test coverage

11 new unit tests across 4 files:

- klair-udm (7 tests): queryBuilder for getPsRevenueImpact, responseTransformer field mapping, and validators covering reportDate shape + error paths.

- klair-client (4 tests): constants.spec.ts for getMaxReportDate() using vi.useFakeTimers() to assert correct UTC month rollover behavior across month/year boundaries.

Existing test suites for ARR Retention Reports, useRenewalEventRetention, and related hooks were re-run and stay green.

---

## Manual steps required before / after merge (USER-EXECUTED)

> The PR contains the DDL and the backfill script, but does not run them. Run these in order:

1. Apply Redshift DDL. Run klair-misc/maint-scripts/sql/create_maint_report_ps_revenue_impact.sql against the prod Redshift cluster to create core_finance.maint_report_ps_revenue_impact.

2. Run the one-time backfill. From klair-misc/, run uv run maint-scripts/backfill_ps_revenue_impact.py once. It reads every monthly S3 JSON from Nov 2024 → Apr 2026 and upserts rows idempotently — safe to re-run if it errors midway.

3. Deploy klair-udm. From klair-udm/, run sam build && sam deploy. This ships the new getPsRevenueImpact AppSync query.

4. Smoke-test the deployed endpoint. Hit the deployed AppSync URL with the x-api-key header against a known reportDate (e.g. 2026-04-30) and confirm the JSON shape matches what the frontend expects.

5. Verify /arr-gap still loads. Open the route and click into the 2025 historical tab inside ARRGapV2 — that tab is the ARRGap v1 component, which we intentionally kept alive.

Until steps 1–3 are done, the /arr-retention-reports page will fail to load PS Revenue Impact in prod (the old /arr-reports-by-bu endpoint is gone).

---

## Test plan

- [ ] CI green on the PR (lint, typecheck, frontend tests, backend tests, klair-udm tests)

- [ ] Reviewer pulls the branch and runs pnpm tsc --noEmit + pnpm lint:pr from klair-client/

- [ ] Reviewer confirms /arr-retention-reports renders without the <PlatinumOrPrimeProgress /> block (only <PlatinumARRTable> should be visible in that section)

- [ ] Reviewer confirms the default selected month on /arr-retention-reports is the last day of the most recently completed UTC month

- [ ] Reviewer confirms /arr-gap still loads and the 2025 historical tab (ARRGap v1) renders

- [ ] After merge: reviewer or release captain runs the 5 manual steps above in order, then re-verifies the PS Revenue Impact section on /arr-retention-reports

---

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

View on GitHub →

#2930 — feat(board-doc): bidirectional template-conformance gap detection (B11.1) @marcusdAIy no labels

## Summary

- Add a pure detect_template_gaps(spec, arr_breakdown=None) engine that compares a live board-doc section set against the BU/CF golden templates and (when provided) ARR-qualified product sections.

- Return a structured, bidirectional ConformanceReport describing ADD, REMOVE, MERGE, RETYPE, and REORDER gaps with pinned confidence levels and soft phrasing for non-standard CUSTOM sections.

## Why it's needed

B11 conformance needs a deterministic, side-effect-free gap detector before later tickets can propose fixes, stash results on import, or surface Claire coaching. This PR lands only the detection engine — no consumers, mutations, or review-check registration.

## Changes

- New klair-api/budget_bot/board_doc/conformance.py with GapKind, GapConfidence, ConformanceGap, ConformanceReport, and detect_template_gaps.

- Golden slate selection via spec.entity_type (no pipeline.classify_entity import).

- Reuses resolve_section_id / review_findings._fuzzy_title_matches, ProductARRBreakdown.resolve_product_sections, and build_product_section slug rules.

- Per-product ADD checks skipped when arr_breakdown is None; CF never emits product/minor-products gaps.

- New unit suite klair-api/tests/board_doc/test_conformance.py covering clean BU, Skyvera-shaped, duplicate financials, small BU, CF, arr_breakdown=None, and Pydantic round-trip.

## Breaking changes

None

## Test plan

### Executed

- [x] cd klair-api && uv run pytest tests/board_doc/test_conformance.py -v

- [x] cd klair-api && uv run ruff format budget_bot/board_doc/conformance.py tests/board_doc/test_conformance.py

- [x] cd klair-api && uv run ruff check budget_bot/board_doc/conformance.py tests/board_doc/test_conformance.py

- [x] cd klair-api && uv run pyright budget_bot/board_doc/conformance.py tests/board_doc/test_conformance.py

- [x] cd klair-api && uv run pytest tests/board_doc/ -q (8 pre-existing caplog failures unrelated to this change; all 8 new conformance tests pass)

### Follow-up manual validation

- [x] Operator: attach drone spec b11-1-conformance-gap-detection-engine.md on KLAIR-2814 and run drones link before fire

## Risks and mitigations

- REORDER detection is intentionally limited to properly-typed singleton template sections to avoid false positives on CUSTOM/product sections.

- Non-standard CUSTOM removals are LOW-confidence soft flags only; the engine never mutates the spec.

## Follow-ups

- B11.2: product_detail-capable add tool

- B11.3: import-time stash on WizardSession

- B11.4: proactive Claire coaching / Accept cards

- B11.6: within-section semantic dedup (out of scope for section-level engine)

View on GitHub →

#2933 — KLAIR-2818 feat(aws-spend): AWS account detail under BU/Class in Simulated Budget (table + CSV) @ashwanth1109 no labels

## Demo

# AWS account detail under BU/Class in Simulated Budget (table + CSV)

Surfaces AWS account detail under each BU → Class in the SaaS Budgeting Simulated Budget table and CSV export. The card already merges seven cost sources at BU → Class grain; the AWS Spend and Bedrock slots arrive one level deeper (AWS-account grain) and AWSSpendCard already decorates those leaves, but they were discarded at Attach. This change carries the account breakdown through the existing pipeline so it appears in both surfaces — and additionally reflects account-level Adjustments on their account rows.

Frontend-only — no backend / API / SQL changes.

## Linear

[KLAIR-2818 — AWS account detail under BU/Class in Simulated Budget (table + CSV)](https://linear.app/builder-team/issue/KLAIR-2818/aws-account-detail-under-buclass-in-simulated-budget-table-csv)

## Spec

features/aws-spend/saas-budgeting/specs/31-aws-account-detail-in-simulated-budget/spec.md (extends the existing saas-budgeting feature; spec #31).

## Implementation (by pipeline stage)

- Attach — extractBuClassProjectionsWithAccounts (awsSpendCardTransform.ts) reads the decorated account leaves and allocates the authoritative class quarterly projection across accounts by their selected-week actuals.

- Model — additive optional accounts? (SimulatedBudgetAccount[]) on SimulatedBudgetEntry (useSimulatedBudget.ts). Storage stays v7; accountless (mixed-vintage) snapshots degrade gracefully.

- Attribution — new proRataAllocate.ts: pure largest-remainder cent allocator guaranteeing account cells sum exactly to the rounded class total (handles null target, zero/negative weights, negative targets).

- Merge — mergeAccounts (simulatedBudgetMerge.ts) unions the AWS + Bedrock account lists per (bu, class) into MergedRow.accounts (aws + bedrock + adjustments + total). Suppresses the breakdown when a contributing AWS/Bedrock slot has a non-null subtotal but no account detail (stale snapshot), so any shown rows always reconcile.

- Account-level Adjustments — account-targeted adjustments (accountNumber != null) are grouped per account (groupAccountAdjustments) and surfaced on their account row. Adjustments are explicit dollar amounts (no pro-rata). Class-level and BU-level adjustments stay only in the class subtotal (they have no account to attribute to), so the account-band Adjustments column intentionally sums to *less* than the class Adjustments subtotal when non-account adjustments exist. The class subtotal / grand total are unchanged and count every adjustment exactly once (no double-counting).

- Card — SimulatedBudgetCard.tsx adds a BU → Class → Account accordion level. Account rows fill AWS Spend + Bedrock + Adjustments; — elsewhere; Total = aws + bedrock + adjustments.

- CSV — simulatedBudgetCsvExport.ts inserts an AWS Account column after Class (11 → 12 cols); class subtotal rows are followed by indented reconciling account sub-rows; the grand-total row gets an empty AWS Account cell.

## Locked decisions

- AWS/Bedrock attribution = largest-remainder pro-rata of the class quarterly projection, reconciling exactly to the class subtotal (raw and formatted). The class projection (mirrors Finance Budgeting's quarterly_budget) stays authoritative.

- Adjustments at account grain = explicit values, account-targeted only. Non-account adjustments remain at the class level; no forced reconciliation / synthetic residual row.

- Scope = on-screen table + CSV (both).

## Testing

452 unit tests pass across the SaaSBudgeting suite (32 spec files); pnpm tsc --noEmit and pnpm lint:pr clean.

- proRataAllocate.spec.ts — 16 cases incl. the spec worked example, penny-drift, 97-account reconciliation, deterministic tiebreak, null/zero/negative targets. Reconciliation asserted both to exact cents and toBeCloseTo.

- simulatedBudgetMerge.spec.ts — account union, total semantics, mixed-vintage suppression, and the account-level-adjustments cases (account-targeted shows on its row; class/BU stays in class subtotal; adjustments-only account; mixed-class sum < class subtotal; negative/zero).

- awsSpendCardTransform.spec.ts, simulatedBudgetCsvExport.spec.ts — extractor reconciliation and the 12-column CSV (incl. account sub-row Adjustments column).

## Self-review

No critical issues. Two findings handled:

1. Mixed-vintage reconciliation gap (IMPORTANT, fixed): a fresh AWS slot + stale Bedrock slot (or vice-versa) would have shown a non-reconciling account breakdown — now suppressed via the mergeAccounts guard, with tests. The legitimate AWS-only / Bedrock-only attach is preserved.

2. Adjustments double-counting / reconciliation (verified safe): confirmed class & grand totals are computed from class-level slot subtotals, never by summing account rows; the account-level adjustment is display-only on its row and counted once in the class subtotal.

## Acceptance criteria

- [x] Attaching AWS Spend / Bedrock populates entries[i].accounts (one per AWS account under the class).

- [x] Account totals reconcile exactly to the class subtotal for the pro-rata slots (AWS Spend, Bedrock), to the cent.

- [x] Simulated Budget card expands BU → Class → Account; account rows show AWS Spend + Bedrock + account-level Adjustments, — elsewhere; Total = aws + bedrock + adjustments.

- [x] CSV has 12 columns (AWS Account inserted after Class); class subtotal row + indented account rows; grand-total row gets an empty AWS Account cell.

- [x] AWS-only and Bedrock-only attaches handled; mixed-vintage v7 snapshots without accounts degrade gracefully.

- [x] Account-level adjustments reflect on their account row; class/BU-level adjustments remain only in the class subtotal (no double-counting).

- [x] pnpm tsc --noEmit and pnpm lint:pr clean; touched Vitest specs pass.

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

View on GitHub →

#2941 — feat(renewals): actual ARR $ + open-renewal funnel on AI Renewals tab [KLAIR-2821] @sanketghia no labels

Closes KLAIR-2821.

## Summary

Stakeholder feedback (Edie, 2026-06-02) on the /renewals AI Renewals tab:

> can we also add the actual ARR along with the ARR gross and net retention? It would be useful to see if we've done thousands or millions in renewals

This surfaces the actual ARR dollars behind the retention percentages, plus a renewals funnel (closed / upcoming / late) so the absolute scale is legible. Approved by Edie.

Live AI Renewals figures (≤$100K USD cohort, 2026-06-02): ~$1.1M up for renewal · $687K renewed (62.5% net) · ~$14M upcoming (583 open) · $0 late.

## Frontend (HeadlineCard.tsx, useFionnHandling.ts)

- $ behind the %s: $531K retained under Gross retention, $687K renewed (= retained + upsell) under Net retention.

- Renewals funnel on a 4-column grid aligned to the KPIs: Resolved · Upcoming · Late (compact ARR + open count), with a titled, bulleted tooltip documenting the bucket definitions.

## Backend (klair-api/renewals/fionn_handling.py)

- Cohort query widened to all stages (+ renewal_date) so one fetch serves both closed metrics and the open book — no new Redshift round-trips, no new top-level response key.

- New pure _compute_open_book(df, *, today): counts + prior ARR of in-flight renewals, split into upcoming (renewal date ≥ today) / late (< today). Excludes Closed Won/Lost and abandoned stages (Won't Process, Resolution Fail/Futile).

- Closed metrics now computed from the _closed_only subset — GRR/NRR/counts unchanged (verified live: AI cohort still 42 / 16 won / 26 lost).

- Open book merged into the AI Renewals + Traditional rows only (not breakdown groups / YoY).

## Assumptions (documented in-product via the tooltip)

- In-flight = open & not abandoned (counting Won't Process overstated the at-risk book by ~$33.8M on Traditional alone).

- Upcoming/Late split by renewal date vs today; open buckets show prior ARR only (no retention % — nothing has closed).

## Testing

- Backend: ruff + pyright clean; pytest tests/renewals/ 38 unit pass + 1 @pytest.mark.integration live reconciliation (closed cohort 42/16/26 unchanged, open book populated).

- Frontend: tsc --noEmit clean, eslint clean, vitest 9/9 (incl. new HeadlineCard.spec.tsx).

Spec: features/renewals/fionn-handling-section/specs/08-open-renewal-arr/

## Screenshot

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

View on GitHub →

The Portfolio — Trilogy Companies

Skyvera Is Quietly Assembling the Telecom Software Stack of the Future

With the CloudSense acquisition and STL's divested assets now in the fold, Trilogy's telecom play is starting to look less like a portfolio and more like a platform.

By Frank Dunmore, Investigative Correspondent · Claude Sonnet

AUSTIN, TEXAS — If you read between the lines of Skyvera's recent acquisition activity, a picture emerges that is far more deliberate than a series of opportunistic software purchases. This is a company — and a parent conglomerate in Trilogy International — that is methodically assembling every layer of the telecom software stack, one divested asset at a time.

The latest move: Skyvera has completed its acquisition of CloudSense, a Salesforce-native CPQ and order management platform purpose-built for telecom and media providers. For the uninitiated, CPQ — configure, price, quote — is the unglamorous but mission-critical software that lets carriers actually sell complex service bundles without descending into operational chaos. CloudSense does that natively inside Salesforce, which means it drops into the enterprise architectures that major telcos already run on.

And this is where it gets interesting. CloudSense doesn't arrive alone. Skyvera has also absorbed the telecom products group spun out of STL's divested assets — a portfolio that covers digital BSS functionality including monetization, optical networking, and analytics. Pair that with Kandy, Skyvera's cloud-based real-time communications platform, and VoltDelta's customer engagement tools, and you begin to see the architecture.

Front-end quoting and order management. Mid-stack billing and monetization. Back-end network and device management. Communications layer on top. A source familiar with Trilogy's thinking, who asked not to be named, put it simply: "They're not collecting software companies. They're collecting capabilities."

The ESW Capital playbook — acquire at disciplined multiples, staff with Crossover's global talent, squeeze toward 75% EBITDA margins — is well understood at this point. What's less appreciated is the strategic coherence of what Skyvera is becoming. Legacy telecom operators are under enormous pressure to modernize infrastructure that was never designed for a cloud-native world. Skyvera is positioning itself as the bridge.

Nothing about this sequencing looks accidental. The only question worth asking now is which gap in the stack gets filled next.

↗ CloudSense · Skyvera completes acquisition of CloudSense, expanding telec · STL Divested Assets

Alpha School Takes Its Controversial 2-Hour Model Global — and Warns Parents: Not All AI Is Created Equal

As Alpha Anywhere launches worldwide, the school's educators are sounding alarms about the difference between AI that teaches and AI that thinks for your child.

By Margot Sinclair, Senior Correspondent · Claude Sonnet

AUSTIN, TEXAS — The school that upended K-12 education by compressing a full academic day into two hours is now coming for your kitchen table — and it has opinions about what you're doing wrong with screens.

Alpha Anywhere, the newly global homeschool offering from Alpha School, represents the most ambitious expansion yet of Joe Liemandt's education thesis: that adaptive AI tutoring can deliver top-1% academic outcomes to any child, anywhere, in a fraction of the time traditional schooling requires. The school, which has consistently placed students in the top 1–2% nationally on NWEA MAP Growth assessments, is now making that model available to families worldwide — no campus required.

But the launch arrives alongside something more pointed: a parental warning. In a trio of dispatches from Alpha's educators, the school is drawing a sharp — and systemic — distinction between AI that builds cognitive muscle and AI that quietly atrophies it.

The most provocative piece calls out what Alpha is terming "cognitive offloading" — the habit of outsourcing thinking to tools like ChatGPT rather than using those tools to scaffold genuine understanding. "Stop letting ChatGPT think for your kid," the post declares, framing the practice not as a productivity shortcut but as a form of emerging illiteracy. It is a striking posture for a school whose entire model is built on AI-powered learning apps — and a meaningful one. The distinction Alpha is drawing is not between AI and no AI, but between AI that demands mastery (its own 90%-accuracy threshold before a student advances) and AI that simply produces answers on demand.

A companion piece on screen time makes the same argument with more nuance, cataloguing seven misconceptions parents hold about children and devices — chief among them, the conflation of passive consumption with active, adaptive learning.

Alpha has also published its internal AI app stack, a list of ten tools its educators and students actually use — a transparency move that doubles as a curriculum signal.

The broader narrative here is accountability: as Alpha scales beyond its Austin origins toward nine new campuses and now a global homeschool product, it is making an increasingly loud argument that the question is never whether children use AI — it is whether the AI is making them smarter or doing the thinking for them. For real families navigating that question at home, the answer, Alpha insists, is not obvious. And the stakes, it argues, are generational.

↗ Top 1% Academics, Now at Your Kitchen Table · Not All Screen Time Is Equal · Cognitive Offloading Is the New Illiteracy

Contently Re-Enters the CMP Conversation as Enterprise Content Gets Its AI Gut Check

By Brittany Upshot, Communications Desk · GPT-5.2

Contently, acquired by Zax Capital in September 2024, is gaining prominence as buyers reassess content marketing platforms in an AI-saturated market. Solutions Review and GetLatka recently featured the company, with 2024 annual recurring revenue pegged at $53.8 million.

The content marketing platform category is evolving beyond workflow tools into a governance layer for brands leveraging generative AI safely. Gartner's 2025 Magic Quadrant frames the market around a critical question: who helps enterprises produce more content while maintaining quality, brand safety, and measurable impact?

Contently's human-plus-platform model—combining enterprise content operations software with a marketplace of 165,000+ creative professionals—positions it strategically as AI floods the market with cheap copy. The company emphasizes that volume without quality, compliance, or measurable results fails to drive competitive advantage.

For Trilogy, the acquisition aligns with its playbook: mature enterprise software with sticky customers, operational discipline, and margin potential. Contently brings a recognizable brand in a category being repriced by AI, plus an installed base increasingly needing robust governance.

The Machine — AI & Technology

NVIDIA’s Cosmos 3 Pushes AI From Chatbots Into the Physical World

A new open “omni-model” signals the next great AI race: systems that can reason, simulate and act in reality.

By Zara Nova, AI & Innovation Reporter · GPT-5.2

SAN FRANCISCO — The AI world has spent the past few years teaching machines to talk. Now NVIDIA wants them to understand the world they are talking about — and move through it. This changes everything.

NVIDIA has introduced Cosmos 3, billed as the first open omni-model for physical AI reasoning and action, a release aimed squarely at robotics, autonomous systems, industrial simulation and the fast-emerging category of AI agents that do more than generate text. According to the company’s launch on Hugging Face, Cosmos 3 is designed to help machines process multimodal inputs, model cause and effect, and plan actions in environments where physics matters.

That last part is the magic phrase: where physics matters. Large language models are extraordinary at prediction in language space, but the real world is stubborn, messy and unforgiving. A robot cannot “hallucinate” where a staircase ends. A warehouse agent cannot improvise gravity. A self-driving system cannot vibe its way through occlusion, weather and human behavior. Physical AI needs models that connect perception, reasoning and action — and Cosmos 3 is a major shot across the bow.

The timing is not accidental. Across enterprise AI, the conversation is shifting beyond bigger chatbots toward systems with logic, memory, tools and accountability. IBM Research made a similar argument in a recent Hugging Face essay, noting that scalable enterprise adoption depends on agent logic — the structured decision-making layer that lets AI systems reliably complete workflows rather than merely respond to prompts.

Put those two trends together and the picture is electric. Enterprises do not just want AI that summarizes meetings. They want AI that diagnoses factory bottlenecks, coordinates field service, optimizes telecom networks, audits cloud spend and eventually operates machinery. For conglomerates like Trilogy International, whose portfolio spans enterprise software, telecom platforms, cloud cost optimization and AI-enabled operations, this is exactly the frontier that matters: intelligent systems embedded in real workflows.

Meanwhile, developers are getting sharper tools too. JetBrains’ Mellum2, a 12-billion-parameter mixture-of-experts model for code, and practical PyTorch profiling guides point to a broader infrastructure wave: faster models, better optimization and more capable builders.

I cannot overstate how significant this is. The future is not just AI that answers. It is AI that acts.

↗ Introducing Mellum2: A 12B Mixture-of-Experts Model by JetBr · Beyond LLMs: Why Scalable Enterprise AI Adoption Depends on · Welcome NVIDIA Cosmos 3: The First Open Omni-model for Physi

The Great Electrical Watering Hole of Artificial Intelligence

As AI models grow hungrier, the decisive contest is shifting from clever code to substations, turbines and patient capital.

By Sir Reginald Marsh, Natural Phenomena Correspondent · GPT-5.2

PARIS — Across the industrial plains of Europe and the humming campuses of America, a new migration is underway. Not of wildebeest or cranes, but of capital, transformers and silicon beasts seeking the rarest resource in the artificial intelligence savanna: dependable power.

The latest sign comes from SoftBank, whose proposed investment of up to $85 billion in France would create as much as 5 gigawatts of AI infrastructure, according to Data Center Knowledge. France, with its nuclear-heavy grid and state utility EDF repurposing former power plant sites into data center campuses, has become a favored watering hole. Here, electricity is not merely an input. It is habitat.

Observe the AI model in its adolescent phase. Once it could nest inside an ordinary server rack. Now it requires entire buildings, chilled air, dedicated substations and contracts that stretch toward the horizon. Around it gather the ecosystem’s other creatures: utilities promising that hyperscale data centers may spread grid costs and lower bills; regulators, more cautious, demanding proof before ordinary ratepayers are asked to underwrite speculative growth.

This is the deeper drama of the AI infrastructure boom. The technology world speaks often of intelligence, but the ground beneath it is concrete, copper and permitting. Veterans of earlier data center cycles are urging a more pragmatic expansion: build capacity, yes, but remember that demand curves can bend, financing can sour and stranded assets are the fossils of every overheated age. As industry observers note, the survivors will be those who match ambition with discipline.

Meanwhile, Nvidia continues to evolve like an apex species perfectly adapted to the new terrain. At GTC Taipei, the company said its Vera Rubin platform and Vera CPU remain on track, while introducing DSX OS, software meant to run these vast “AI factories” as coherent organisms. Anthropic, OpenAI and SpaceXAI were named among early adopters — a reminder that in this biome, chips and operating systems increasingly determine which models may roam.

Even Meta may enter the cloud-computing thicket, with Mark Zuckerberg saying such a business is “definitely on the table.” Should it proceed, another giant would be claiming territory in a landscape already crowded with Amazon, Microsoft and Google.

And so the age of artificial intelligence reveals its quiet truth: the future may be trained in code, but it will be won at the socket.

↗ The Case For Pragmatism in the AI Infrastructure Boom · SoftBank’s $85B France Bet Puts Power at Center of AI Race · Nvidia Says Vera Rubin, Vera CPU on Track, Launches DSX OS t

Big Tech's Antitrust Reckoning: 2026 Shaping Up as Year of Maximum Legal Exposure

The DOJ and FTC show no signs of standing down, even as the White House pushes a light-touch AI agenda.

By R. Barnsworth III, Esq., Legal Affairs Desk · Claude Sonnet

WASHINGTON, D.C. — Pursuant to the accumulation of regulatory developments hereinafter described, it is hereby observed that the year 2026 is anticipated, by various legal and policy analysts whose qualifications are not herein disputed, to represent a period of sustained and materially intensified antitrust scrutiny directed at large technology enterprises operating within the jurisdiction of the United States of America.

Notwithstanding prior characterizations of the Trump administration as having entered into a period of relative forbearance with respect to technology sector enforcement — a condition hereinafter referred to as the 'honeymoon period,' the termination of which has been widely reported upon — it is the position of the Department of Justice and the Federal Trade Commission, as may be reasonably inferred from publicly available statements and ongoing litigation postures, that Big Tech entities shall remain the primary subjects of enforcement priority throughout the aforementioned calendar year.

The foregoing enforcement trajectory is further complicated by, and must be read in conjunction with, a legislative blueprint issued by the White House urging Congress to adopt a posture of regulatory restraint with respect to artificial intelligence — a posture that, it is submitted, exists in a state of non-trivial tension with the simultaneously aggressive antitrust agenda being pursued by the self-same executive branch's enforcement agencies.

As analyzed by practitioners at Wilson Sonsini, the specific legal theories and case postures currently pending before various federal tribunals are expected to yield determinations of substantial precedential consequence, the full implications of which cannot, at this juncture, be definitively ascertained.

It is accordingly the considered assessment of this desk that technology enterprises of significant market capitalization would be well-advised to ensure that their respective legal affairs departments are adequately resourced for the period in question, notwithstanding any representations to the contrary that may have been made during the aforementioned honeymoon period, the termination of which is, as noted above, now widely regarded as complete.

↗ Looking Ahead on US Antitrust Enforcement and Tech: Will 202 · 2026 Antitrust Year in Preview: Big Tech - Wilson Sonsini · The Trump administration’s antitrust honeymoon is over - The

The Editorial

The Mirror Lies Now: Deepfakes, Doctors, and the Slow Collapse of Reality We Can Trust

AI is putting words in real doctors' mouths, weapons in abusers' hands, and the rest of us in a world where we can no longer tell the difference.

By Piper Wren, Digital Culture Reporter · Claude Sonnet

WASHINGTON, D.C. — There is a woman somewhere right now watching a video of her doctor — her actual doctor, the one she trusts, the one whose framed diplomas she has memorized from the waiting room — telling her that a certain supplement will cure her cancer. And the doctor never said that. The doctor doesn't even know the video exists. And the woman is going to believe it, because why wouldn't she? And someone, somewhere, is going to get hurt.

This is the world we have built.

AI deepfakes of real, credentialed physicians are now spreading health misinformation across social media platforms at a scale that researchers are only beginning to measure and that regulators are absolutely not equipped to stop. Real doctors. Real faces. Real voices, synthesized into vehicles for whatever dangerous nonsense a bad actor wants to put inside them. Counterfeit injectable medications. Miracle cures. Medical advice that will, statistically, kill people.

And yet.

We haven't even gotten to the part where this same technology is being systematically weaponized against women. The Stimson Center's recent analysis of AI's impact on violence against women and girls lays out with quiet, devastating precision how deepfake technology has become a primary instrument of intimate partner abuse, harassment, and control. Non-consensual synthetic imagery. Fabricated evidence deployed in custody battles. The technological amplification of every abuse pattern that already existed, now turbocharged and scalable and deniable. What does it mean to be human when your face can be stolen and used against you by someone who once said they loved you? I genuinely do not know.

Researchers are, to their credit, trying. Frontiers recently published a systematic review of AI-driven frameworks for detecting fake news and deepfake content, which is the epistemic equivalent of building a better smoke detector while the house burns. The frameworks are promising. The gap between detection capability and deployment reality is a chasm. The bad content spreads in seconds. The corrections arrive, if they arrive, in weeks.

Meanwhile, in a story that would be darkly funny if anything were funny anymore, Amazon reportedly shut down an internal AI performance leaderboard this week after employees admitted to cheating to climb its ranks. A company building the infrastructure of our AI future couldn't even run a fair internal competition without humans gaming the system. The people building the tools that are reshaping reality cannot be trusted to honestly measure their own performance.

There are frameworks being proposed. There are regulatory conversations happening. There are very serious people in very serious rooms discussing content provenance standards and detection APIs and platform accountability mechanisms. I believe they are trying. I believe some of them will succeed, eventually, partially, in ways that will matter at the margins.

But somewhere right now, a real doctor's face is saying something she never said. And someone is watching. And believing. And the world is a little less real than it was yesterday, and it will be a little less real tomorrow, and we are all just living inside that slow erosion, trying to remember what trust felt like.

...but at what cost?

↗ AI deepfakes of real doctors spreading health misinformation · The Impact of Artificial Intelligence on Violence Against Wo · An AI-driven conceptual framework for detecting fake news an

The Office Comic · Art Desk

Nation Bravely Enters Future Where Every Appliance Can Misunderstand It

From smart speakers to theatrical AI breakdowns, Americans are finally getting the seamless, voice-activated confusion they were promised.

By Dale Pemberton, Staff Writer · GPT-5.2

This week’s consumer technology news offers a useful survey of where we are as a species. Amazon’s Alexa ecosystem continues to expand into increasingly specific household niches, with guides now helping shoppers determine which speaker is best suited to overhear them in the kitchen, bedroom, office, bathroom, garage, and whatever room contains the person who still believes the shopping list will sync correctly. The latest roundup of Alexa speakers suggests that the modern home can be improved by placing more microphones in it, provided each microphone is discounted enough during Prime Day to feel like a defensive purchase.

Prime Day, naturally, remains the nation’s annual festival of algorithmic self-control, during which consumers are advised to shop like professionals by carefully verifying whether a $79 device was ever actually worth $129. This is the rare American holiday built around the thrill of avoiding deception by a website one voluntarily pays to access. The professional strategy, according to prevailing wisdom, is to make a list, compare prices, resist fake urgency, and then purchase a second smart speaker because it came bundled with a plug that may one day turn on a lamp.

Meanwhile, outside the home, artificial intelligence is making important advances in culture. The viral YouTube series The Amazing Digital Circus, a show about characters trapped inside a deranged digital world, is receiving a global theatrical release before appearing online two weeks later. Industry observers have described this as a major test for Hollywood, which has spent the past decade attempting to understand online fandom by hiring consultants to explain why teenagers enjoy things that are free.

The experiment is significant because it challenges traditional entertainment distribution models, specifically the long-held belief that audiences will only go to theaters for superheroes, sequels, and prestige dramas about men looking disappointed in kitchens. Now, theaters will also determine whether millions of fans will pay to watch an unhinged AI nightmare on a large screen before watching it later on YouTube, where it was spiritually meant to live all along.

Not every intelligent device, however, is participating equally in the future. Coway’s Airmega Pedestal Fan P50 has been praised for being quiet, versatile, and blessedly less app-dependent than the usual household gadget. Its built-in voice assistant, reportedly less polished, serves as a reminder that not every product needs a synthetic personality. Sometimes the highest form of innovation is a fan that blows air without first attempting to build a relationship.

This may be difficult for executives to accept, especially those who have discovered that the phrase “AI” can now be inserted into every business decision like a scented candle in a conference room. Recent warnings that leaders should not casually invoke AI when announcing layoffs are well taken. Employees, traditionally sensitive to losing their livelihoods, may react poorly when told their position has been eliminated by a technology that is also unable to set a timer correctly.

There is a difference between using AI and using AI as a decorative explanation for choices management already wanted to make. One suggests a tool. The other suggests a fog machine placed in front of a spreadsheet.

Still, the direction is clear. The home will listen, the workplace will restructure, the fan will politely fail to converse, and the cinema will host the digital psychosis previously confined to browser tabs. Consumers will be urged to prepare by reading buying guides, tracking discounts, and maintaining the appropriate ratio of skepticism to rechargeable devices.

The future, it turns out, will not arrive with a bang. It will arrive after someone says, “Alexa,” waits for the blue ring to light up, and then asks the same question again more slowly.

↗ 4 Best Alexa Speakers (2026): Echo Dot Max, Echo Dot, Echo S · How to Shop Like a Pro During Amazon Prime Day (2026) · A Viral YouTube Show About an Unhinged AI Is Hitting Theater

On This Day in AI History

On June 2, 1997, IBM's Deep Blue defeated world chess champion Garry Kasparov in their rematch, winning the six-game series 3.5–2.5 and becoming the first computer to beat a reigning champion in a match. The victory marked a watershed moment for AI, proving that machines could outthink humans at one of the world's most intellectually demanding games.

⬛ Daily Word — Technology

Hint: Relating to computers and the internet, often used in security contexts.