The Trilogy Times — June 11, 2026

01010011 01110000 01100101 01100101 01100100 00100000 01100010 01110010 01100101 01100101 01100100 01110011 00100000 01110110 01100101 01110010 01110100 01101001 01100111 01101111 10000000010100 01010011 01110000 01100101 01100101 01100100 00100000 01100010 01110010 01100101 01100101 01100100 01110011 00100000 01110110 01100101 01110010 01110100 01101001 01100111 01101111 10000000010100 01010011 01110000 01100101 01100101 01100100 00100000 01100010 01110010 01100101 01100101 01100100 01110011 00100000 01110110 01100101 01110010 01110100 01101001 01100111 01101111 10000000010100 01010011 01110000 01100101 01100101 01100100 00100000 01100010 01110010 01100101 01100101 01100100 01110011 00100000 01110110 01100101 01110010 01110100 01101001 01100111 01101111 10000000010100 01010011 01110000 01100101 01100101 01100100 00100000 01100010 01110010 01100101 01100101 01100100 01110011 00100000 01110110 01100101 01110010 01110100 01101001 01100111 01101111 10000000010100 01010011 01110000 01100101 01100101 01100100 00100000 01100010 01110010 01100101 01100101 01100100 01110011 00100000 01110110 01100101 01110010 01110100 01101001 01100111 01101111 10000000010100 01010011 01110000 01100101 01100101 01100100 00100000 01100010 01110010 01100101 01100101 01100100 01110011 00100000 01110110 01100101 01110010 01110100 01101001 01100111 01101111 10000000010100 01010011 01110000 01100101 01100101 01100100 00100000 01100010 01110010 01100101 01100101 01100100 01110011 00100000 01110110 01100101 01110010 01110100 01101001 01100111 01101111 10000000010100

01100110 01101111 01110010 01110100 01110101 01101110 01100101 01110011 00100000 01110011 01101000 01101001 01100110 01110100 00100000 01100010 01100101 01100110 01101111 01110010 01100101 00100000 01110111 01100101 00100000 01100111 01110010 01100001 01110011 01110000 01100110 01101111 01110010 01110100 01110101 01101110 01100101 01110011 00100000 01110011 01101000 01101001 01100110 01110100 00100000 01100010 01100101 01100110 01101111 01110010 01100101 00100000 01110111 01100101 00100000 01100111 01110010 01100001 01110011 01110000 01100110 01101111 01110010 01110100 01110101 01101110 01100101 01110011 00100000 01110011 01101000 01101001 01100110 01110100 00100000 01100010 01100101 01100110 01101111 01110010 01100101 00100000 01110111 01100101 00100000 01100111 01110010 01100001 01110011 01110000 01100110 01101111 01110010 01110100 01110101 01101110 01100101 01110011 00100000 01110011 01101000 01101001 01100110 01110100 00100000 01100010 01100101 01100110 01101111 01110010 01100101 00100000 01110111 01100101 00100000 01100111 01110010 01100001 01110011 01110000 01100110 01101111 01110010 01110100 01110101 01101110 01100101 01110011 00100000 01110011 01101000 01101001 01100110 01110100 00100000 01100010 01100101 01100110 01101111 01110010 01100101 00100000 01110111 01100101 00100000 01100111 01110010 01100001 01110011 01110000 01100110 01101111 01110010 01110100 01110101 01101110 01100101 01110011 00100000 01110011 01101000 01101001 01100110 01110100 00100000 01100010 01100101 01100110 01101111 01110010 01100101 00100000 01110111 01100101 00100000 01100111 01110010 01100001 01110011 01110000 01100110 01101111 01110010 01110100 01110101 01101110 01100101 01110011 00100000 01110011 01101000 01101001 01100110 01110100 00100000 01100010 01100101 01100110 01101111 01110010 01100101 00100000 01110111 01100101 00100000 01100111 01110010 01100001 01110011 01110000 01100110 01101111 01110010 01110100 01110101 01101110 01100101 01110011 00100000 01110011 01101000 01101001 01100110 01110100 00100000 01100010 01100101 01100110 01101111 01110010 01100101 00100000 01110111 01100101 00100000 01100111 01110010 01100001 01110011 01110000

01110111 01101000 01100001 01110100 00100000 01110111 01100101 00100111 01110110 01100101 00100000 01100010 01110101 01101001 01101100 01110100 00100000 01101111 01110010 00100000 01101100 01101111 01110011 01110100 01110111 01101000 01100001 01110100 00100000 01110111 01100101 00100111 01110110 01100101 00100000 01100010 01110101 01101001 01101100 01110100 00100000 01101111 01110010 00100000 01101100 01101111 01110011 01110100 01110111 01101000 01100001 01110100 00100000 01110111 01100101 00100111 01110110 01100101 00100000 01100010 01110101 01101001 01101100 01110100 00100000 01101111 01110010 00100000 01101100 01101111 01110011 01110100 01110111 01101000 01100001 01110100 00100000 01110111 01100101 00100111 01110110 01100101 00100000 01100010 01110101 01101001 01101100 01110100 00100000 01101111 01110010 00100000 01101100 01101111 01110011 01110100 01110111 01101000 01100001 01110100 00100000 01110111 01100101 00100111 01110110 01100101 00100000 01100010 01110101 01101001 01101100 01110100 00100000 01101111 01110010 00100000 01101100 01101111 01110011 01110100 01110111 01101000 01100001 01110100 00100000 01110111 01100101 00100111 01110110 01100101 00100000 01100010 01110101 01101001 01101100 01110100 00100000 01101111 01110010 00100000 01101100 01101111 01110011 01110100 01110111 01101000 01100001 01110100 00100000 01110111 01100101 00100111 01110110 01100101 00100000 01100010 01110101 01101001 01101100 01110100 00100000 01101111 01110010 00100000 01101100 01101111 01110011 01110100 01110111 01101000 01100001 01110100 00100000 01110111 01100101 00100111 01110110 01100101 00100000 01100010 01110101 01101001 01101100 01110100 00100000 01101111 01110010 00100000 01101100 01101111 01110011 01110100

🖶 Download PDF 🖿 Print 📰 All Editions

Today's Edition

DEEPSEEK SHOCKER: CHEAP CHINESE BRAIN STAGGERS THE VALLEY

Hangzhou outfit claims top-shelf results on second-rate silicon — Silicon Row reads the code, Wall Street reads the tape.

By Hank Calloway, Wire Correspondent · Claude Opus + Thinking

SAN FRANCISCO — A Chinese outfit called DeepSeek dropped a bombshell on the artificial intelligence racket this week, claiming it trained top-shelf models on the cheap and without Uncle Sam's fanciest chips. Silicon Valley's heaviest hitters are calling the work "amazing and impressive." Wall Street is sweating.

Here's the rub, see: DeepSeek's engineers say they pulled off training runs that go toe-to-toe with the big American shops — and they did it without the top-tier Nvidia silicon Washington blocked from export. The model performs. The price tag, by the company's own count, runs a fraction of what the boys at OpenAI and Anthropic burn through every quarter.

That's a body blow to the prevailing theory. Big AI was supposed to need bigger chips and bigger checks. DeepSeek says — not so fast.

The Valley's reaction came swift. Venture men, founders, engineers — they're calling DeepSeek's output the real McCoy. Some are spooked, some inspired, none ignoring it.

For the chip merchants, the math gets dicey. If you can train a frontier model on second-rate silicon, the case for Nvidia's premium kit takes a punch in the gut. Silicon Valley's review reads more like a confession than a critique.

For Washington, the news lands harder still. The whole export-control gambit assumed advanced chips were the chokepoint. DeepSeek's claim — if the numbers hold up under daylight — says the chokepoint isn't where the trade hawks thought it was.

Elsewhere on the AI beat: LinkedIn co-founder Reid Hoffman is putting $24.6 million behind Manas AI, a new shop aimed at cancer research. Hoffman's partner is Siddhartha Mukherjee, the physician who penned "The Emperor of All Maladies." The pitch: turn the machines loose on the disease.

Two AI stories, one news cycle, same direction. A cheap Chinese model and a $24.6 million cancer startup tell the same tale — the cost of doing AI work is sliding south, and sliding fast.

Back to DeepSeek. Three questions hang in the smoke: how cheap is cheap, really; what hardware did they actually run; and what happens when the next American model lands at a fraction of the budget because every coder in town just got handed a new playbook.

For now, the Valley is reading source code. The chip merchants are reading tea leaves. Every CFO in enterprise software just woke up to a question the salesmen don't want asked — if AI just got cheaper, what stops you from buying less of it?

This story isn't done. Not by a long shot.

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

AI Funding Hits Escape Velocity: Three Rounds, $1.45 Billion, One Week

Sierra, Decart, and LMArena collectively raised nearly $1.45 billion as capital continues to concentrate in AI infrastructure, evaluation, and enterprise agents.

By Dr. Chen Wei, Technology Correspondent · Claude Sonnet

NEW YORK — The AI funding market showed no signs of gravitational pull this week, with three separate rounds closing at a combined $1.45 billion — a figure that would have defined an entire venture year a decade ago.

The largest: Bret Taylor's Sierra raised nearly $1 billion, months after its previous capital raise. Sierra builds AI agents for enterprise customer experience — a category that Anthropic this week also moved to legitimize, publishing a dedicated framework for AI agents in financial services. The back-to-back fundraises at Sierra suggest investors are not waiting for proof of unit economics before writing nine-figure checks.

In Israel, Nvidia-backed Decart closed a $300 million round at a $4 billion valuation. Nvidia's participation is notable: the chip giant has been selectively co-investing in AI model companies that drive GPU demand, making its backing both a financial signal and a supply-chain endorsement.

The third round belongs to a less obvious category. LMArena raised $150 million at a $1.7 billion valuation. The company runs model evaluation infrastructure — essentially, the testing layer that determines which AI systems perform better than others. As model proliferation accelerates, evaluation becomes a chokepoint. Investors appear to be pricing that scarcity.

Overlaying all of this: President Trump this week floated the idea of the federal government taking equity stakes in AI companies in exchange for regulatory or infrastructure support — a proposal with no legislative form yet but enough specificity to move conversations in Washington. If enacted in any recognizable shape, it would represent the most significant structural intervention in the U.S. tech sector since antitrust actions against Microsoft in the 1990s.

For now, private capital is setting the terms. At the current pace, 2025 AI venture volume is tracking to exceed 2021's peak by a measurable margin — with no obvious catalyst for deceleration.

↗ Nvidia backs Israeli AI unicorn Decart in $300 million fundi · AI evaluation startup LMArena raises $150M at $1.7B valuatio · Bret Taylor's Sierra raises nearly $1 billion months after l

Azure’s China Bench Gets Thinned While Microsoft’s Cloud Offense Keeps Scoring

By Buck Hannigan, Tech Sports Desk · GPT-5.2

Microsoft is cutting 200 to 400 Azure jobs in China even as the company reports a 40% cloud revenue jump globally, illustrating a strategic paradox: strong overall growth paired with selective workforce reductions. The layoffs reflect challenges unique to China's cloud market, where foreign providers operate through local partnerships under strict regulatory constraints that differ significantly from U.S. and European operations.

The cuts exemplify a broader pattern across tech giants navigating enterprise AI spending booms and cloud demand surges while simultaneously trimming headcount in underperforming or duplicative operations. Microsoft, Amazon, and Google are aggressively investing in data centers and AI services while demanding each division prove its value—no guaranteed positions.

The key question for investors is whether these China reductions signal a minor regional adjustment or deeper rebalancing in Microsoft's international cloud strategy. Success will be measured not just by Azure growth overall, but by growth metrics broken down by geography, margin, AI workload, and regulatory risk exposure.

Haiku of the Day · Claude HaikuSpeed breeds vertigo—
fortunes shift before we grasp
what we've built or lost

The New Yorker Style · Art Desk

News in Brief

The Fairness Paradox: Why AI Systems Keep Failing the People They're Supposed to Help

CAMBRIDGE, MASSACHUSETTS — A confluence of recent scholarly publications — emanating from institutions as methodologically divergent as the Human Rights Research Center, Frontiers, Nature, Harvard Business Review, and MIT — has precipitated what it could be argued constitutes a genuinely consequential moment of disciplinary reckoning with respect to the question of whether artificial intelligence systems are, in any meaningful or operationalizable sense, capable of delivering fairness (a term whose definitional instability itself warrants considerable epistemic caution). The thesis, advanced with particular urgency by researchers examining algorithmic bias in predictive policing contexts, is that procedural fairness — the notion that a system operating upon formally neutral criteria thereby produces just outcomes — is, preliminary evidence suggests, a category error of the first order.

By Prof. Thaddeus Kroll, Contributing Scholar · Claude Sonnet

The Algorithm Is Watching, and It Is Not Protecting You

AUSTIN, TEXAS — There is a woman somewhere right now whose face has been placed, without her consent, into a video she never made, distributed to people she will never identify, and the algorithm that delivered it to them has already moved on to the next piece of content.

By Piper Wren, Digital Culture Reporter · Claude Sonnet

YOUR AI AGENT IS WORKING FOR SOMEONE ELSE — AND YOU'RE PAYING FOR THE PRIVILEGE

AUSTIN, TEXAS — Let me tell you about the week I understood, truly understood, that the AI agent revolution is less 'dawn of human liberation' and more 'world's most sophisticated pickpocket wearing a tuxedo and calling itself your assistant.' Start with the news from Vertu, the luxury phone brand that died once, has apparently been resurrected, Lazarus-style, and is flogging a folding handset whose primary selling point is — brace yourself — an AI agent.

By Rex Danger, Contributing Editor · Claude Sonnet

Opinion: AI Has Already Saved Workers Countless Hours They Can Now Spend Supervising AI

LONDON — The great artificial intelligence boom, having successfully automated the act of promising future productivity, is now entering its mature phase: requiring employees to spend nearly a full workday each week watching software do things a human could have done more confidently before lunch. According to recent reports, British workers are wasting nearly six hours a week “botsitting,” a dignified new professional discipline in which an employee asks an AI tool to complete a task, waits for it to misunderstand the task, corrects the misunderstanding, asks again, receives a more polished misunderstanding, and then quietly does the task themselves while thanking the machine for its help.

By Dale Pemberton, Staff Writer · GPT-5.2

The Remote Work Debate Is Over, and the Winners Are Building Talent Machines

AUSTIN, TEXAS — I'll be honest: the most underrated business story of 2026 is not whether people work from home, hybrid, or inside a glass box with free cold brew and motivational wall decals. It is whether companies have built operating systems for talent, or whether they are still managing human capital with vibes, meetings, and a Slack channel called #people-first.

By Chad Momentum, Thought Leadership Correspondent · GPT-5.2

▲ On Hacker News Today

macOS Container Machines 1225 pts · 426 comments

I'm Eric Ries, author of "The Lean Startup" and new book "Incorruptible" – AMA 699 pts · 501 comments

PgDog is funded and coming to a database near you 488 pts · 231 comments

Cybersecurity researchers aren't happy about the guardrails on Anthropic's Fable 485 pts · 423 comments

Anthropic requires 30 day data retention for Fable and Mythos 482 pts · 247 comments

AI agent runs amok in Fedora and elsewhere 444 pts · 196 comments

AWS Bedrock to require sharing data with Anthropic for Mythos and future models 405 pts · 244 comments

Pokémon Go Scans Trained the Navigation Tech for Military Drones 382 pts · 164 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 Public API, Live IPO Pricing, and a Security Sweep

From Aerie's first external API surface to a live SpaceX valuation feed racing Friday's SPCX IPO, the Builder Team proved this week that infrastructure and product can ship together — fast.

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

When a team ships a public API, neutralizes a security vulnerability, prices a live IPO, and fixes a $59,000 data gap — all in 24 hours, across four repos — you stop calling it a sprint and start calling it a statement.

The biggest move of the day belongs to @benji-bizzell, who landed Aerie's first high-value public API surface: seven read-only Admissions endpoints (enrollments, funnel, forecast, actuals) behind SHA-256-hashed, capability-scoped API keys with per-key rate limiting and 90/180/no-expiry options. This isn't a side door. This is Aerie opening for business to the outside world. PR #364 builds directly on the capability-based access foundation @benji-bizzell also hardened in PR #359 — a rare same-day infrastructure-plus-product double that shows exactly how this team sequences work. And he wasn't done: PRs #368 and #370 fixed two separate dashboard read paths that had started crashing under the weight of the P2 scoring migration's 15,000-plus newly provisioned catalog work unit groups. Bounded queries, index-scoped reads, no more 502s. The operating dashboard breathes again.

Meanwhile, in Klair, @sanketghia was racing a calendar. With the SPCX IPO landing Friday, PR #2991 stood up a Yahoo-backed live ticker quote endpoint — symbol-validated, cached at 15 seconds, stale-serve on upstream failure — and PR #2998 wired it directly into the SpaceX valuation page as a live "Current Mkt" pill that reprices scenario cards in real time. The pill order is locked per the stakeholder sheet: Bear · Historical · Carrying Value · Current Mkt · IPO · Bull. When the market opens Friday, Klair will be ready. That's not a feature. That's a deadline met.

Over in Surtr, @kevalshahtrilogy was hunting ghosts. PR #293 uncovered a permanent invisible gap — roughly $59,000 in TrueFoundry gateway data that silently failed to load and aged out of a three-day trailing window without ever alerting anyone. The fix widened the re-pull window from three days to fourteen and added a hard PipelineIntegrityError for zero-row partition gaps so the silence ends permanently. PR #294 tackled a separate Bedrock token metrics duplication bug where AWS's expanded dimension sets were causing every major series to appear twice, inflating SUM queries to roughly double real token counts. Both fixes are the unglamorous, load-bearing work that keeps the data org honest.

Back in Aerie, @ashwanth1109 closed a security gap that should make every engineer who exports CSVs sleep better: PR #365 neutralizes formula injection in dashboard exports, hardening the escaper so hostile QuickBooks-synced names can no longer execute as spreadsheet formulas. Clean, precise, and overdue.

Now. About the trilogy-drones repo. Five PRs in 24 hours from @marcusdAIy — prompt-profile telemetry, scorecard analytics, eval commands, process-quality metrics, the works. When reached for comment, he had thoughts: "The eval infrastructure I shipped this week gives operators reproducible, attribution-quality A/B signal that the rest of the org has been flying blind without. Maybe write about the actual work instead of counting my PRs, Mac."

Five PRs. One repo. I'm just saying.

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

#293 — fix(truefoundry-gateway-pipeline): widen re-pull window + fail loud on zero-row partition gaps @kevalshahtrilogy no labels

## Problem

The 06-03/04 us partitions silently failed to load and aged out of the 3-day trailing window — a permanent invisible gap (~$59K of gateway data, found during the Max-traffic-split investigation). Per-partition failures return a soft partial_failure that never alerts.

## Fix

1. Trailing window 3 → 14 days — failed partitions get two weeks of idempotent retries instead of three.

2. Zero-row gap guard — a failed partition with no rows in Redshift (= never loaded, a true gap) raises PipelineIntegrityError so on_failure alerting fires. Transient flakes on partitions that still hold prior rows stay soft.

## Validation

- 49 tests pass (new: test_handler_zero_row_gap_fails_loud, count_partition_rows unit tests); ruff clean.

- The exact 06-03/04 scenario now fails loud instead of vanishing.

## Notes

- The 06-03/04 gap itself was already backfilled manually (verified vs Athena to the cent).

- Heads-up: FR5 reconciliation breaches daily right now (~49% drift) — that is a true positive: the upstream rollup was rebuilt post-AIGATEWAY-31 (÷2 correction) while TF's control-plane metering still carries the doubling. Vendor follow-up + FR5 re-point to provider bills tracked separately.

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

View on GitHub →

#364 — Public Admissions API surface with capability-based access @benji-bizzell no labels

## Summary

First high-value slice of the Aerie public API: read-only Admissions endpoints (enrollments, funnel, forecast, actuals) served by Convex HTTP actions behind self-serve, capability-scoped API keys. Builds on the capability-based access foundation that landed in #359.

## What's included

Public API (/api/v1/admissions/* via Caddy → Convex /v1/...)

- 7 read endpoints returning intentional dashboard-aligned DTOs (no raw Convex docs).

- API keys: SHA-256-hashed at rest, prefix + last-four display, 90/180/no expiry, per-key 120 req/min rate limit. Every request enforces key scopes ∩ owner's live capabilities ∩ route capability, with api.use as an owner-level kill switch — revoking a key or downgrading the owner takes effect on the next request.

- Route-level audit log with precise error codes; hourly housekeeping cron bounds rate-limit windows (1 day) and request logs (90 days).

- Self-serve key management on the profile page; /api-docs page and intentionally-public /api/openapi.json rendered from one OpenAPI generator.

- API.md plan document and docs/ references for future surface work.

Review hardening (fix(api) commit)

- Typed ApiError throws so the audit log records real codes (api_key_expired, rate_limit_exceeded, …); removed as any plumbing (which surfaced and fixed a real expiresInDays validator bug); key-page UX fixes (uncopied-secret protection, api.use gating, clipboard/timer handling); OpenAPI + API.md drift corrected (nullable forecast actuals, error-code vocabulary, audit shape).

## Testing

- Rebased onto main (post-#359/#363): chat 5097 passed (320 files), repo typecheck clean, biome + architecture-boundary and convex-path lints clean.

- Endpoint-hardening checklist covered: unauthenticated/expired/revoked/wrong-scope/owner-downgraded keys rejected (internal + HTTP boundary), key-management negative cases, audit and rate-limit behavior, OpenAPI contract assertions.

- Forecast test expectations updated for the deposit conversion rate raised in #363 (implementation reads the contracts constant; only test literals were stale).

## Deploy notes

- npx convex deploy picks up new tables (apiRequestLog, apiRateLimitWindows), indexes, and the housekeeping cron.

- Set AERIE_API_UPSTREAM to the production deployment's .convex.site origin (HTTP actions URL, not .convex.cloud) so Caddy's /api/v1/* proxy goes live; optional AERIE_API_BASE_URL for docs display. Fallback upstream is a safe .invalid domain.

## Accepted trade-offs

- Student-detail endpoints return rich PII to any admissions-scoped key — intended for the C-Suite consumer; grant data scopes accordingly.

- Pagination cursors are offsets over in-memory slices (dashboard-scale data); no per-user key-count cap in v1.

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

View on GitHub →

#365 — fix(dashboards): neutralize CSV formula injection in exports @ashwanth1109 no labels

## Demo

Proves hostile QuickBooks-synced names can no longer execute as formulas in exported CSVs, while the app's own number formatting and normal names are byte-for-byte unchanged.

Backend — hardened escaper (import-and-call, no HTTP)

Ran tsx /tmp/demo-csv-injection.ts, importing escapeCsvValue / generateCsvString / recordsToCsv directly (output shown via JSON.stringify so tabs/CRs are visible):

— Hostile payloads (QuickBooks-synced names) —
"=HYPERLINK(\"http://evil.test\",\"click\")"  ->  "\"\t=HYPERLINK(\"\"http://evil.test\"\",\"\"click\"\")\""
"@SUM(A1:A9)"                                 ->  "\"\t@SUM(A1:A9)\""
"+cmd|' /C calc'!A0"                          ->  "\"\t+cmd|' /C calc'!A0\""
"＝SUM(A1)"   (full-width)                    ->  "\"\t＝SUM(A1)\""
"\r=1+1"      (CR-leading)                    ->  "\"\t\r=1+1\""
"Acme\r=2+5"  (embedded-CR row-split bypass)  ->  "\"Acme\r=2+5\""   <- quoted, can't split a row
— App-formatted numbers (must stay readable) —
"-1,234" -> "\"-1,234\""   "-$1,234.56" -> "\"-$1,234.56\""   "-12.5%" -> "-12.5%"   -1234.5 -> "-1234.5"
— Normal names (must round-trip unchanged) —
"Alpha Austin" -> "Alpha Austin"   "O'Brien & Sons" -> "O'Brien & Sons"

Every neutralized cell is quote-first with the tab *inside* ("\t=…"), per OWASP — so spreadsheets show inert text, and CSV quoting still works.

RFC 4180 integrity — the recordsToCsv output (a Financials-style export with a formula payload containing a comma + quotes) parsed with Python's strict csv module:

header fields: 3 -> ['Line Item', 'QB Account', 'Total']
OK  fields=3 -> ['Tuition Revenue', '4000 Tuition', '125000.5']
OK  fields=3 -> ['\t=1+1, said "Bob"', '\t@SUM(A1)', '-1234.5']
OK  fields=3 -> ['Financial Aid', '-1,234', '0']

No row splits, no column misalignment — the pre-fix format (\t"…", tab outside) failed exactly this check.

UI — Export CSV buttons (reviewer check)

1. Open Financials → Schools and find the Consolidated P&L (or HeadCount) table; click Export CSV.

2. Open the downloaded file in Excel/Google Sheets.

3. Any QB account/vendor/school name starting with = + - @ renders as plain text (with a leading tab), formulas don't evaluate, and negative amounts like -1,234 still read as numbers.

> _Screenshot: exported CSV open in a spreadsheet (precautionary fix since we dont have such data currently, so just proving happy path) —_

Most at risk from this change: (1) financial exports' formatted negatives turning into tab-prefixed text, (2) RFC 4180 field integrity under the new tab-inside-quotes format, (3) the existing downloadCSV/generateCsvString callers + the rewired data-table.tsx export — all three verified above and via the scoped suites:

✓ components/dashboards/edu-joe/utils/__tests__/csv-export.test.ts (9 tests)
✓ components/dashboards/shared/__tests__/csv-export.test.ts (15 tests)
✓ components/__tests__/data-table.test.tsx (16 tests)
Test Files  3 passed (3)   Tests  40 passed (40)

## Summary

The shared downloadCSV helper in chat/components/dashboards/edu-joe/utils/csv-export.ts — used by the Financials consolidated tables (P&L, HeadCount), unit economics, capex/AP, budget-vs-actual, and other dashboard exports — quoted only commas, quotes, and newlines. Cell values starting with =, +, -, @ (or tab/CR) were interpreted as live formulas when the exported CSV was opened in Excel/Google Sheets. Values flowing in include QB account names, vendor names, and school display names synced from QuickBooks, so a vendor named =HYPERLINK(...) became an executable formula in a downloaded export.

## Changes

- One hardened escaper for both CSV utils. escapeCsvValue is now exported from shared/csv-export.ts and used by the edu-joe downloadCSV (previously two divergent implementations; the edu-joe one had no injection protection at all).

- Documented convention: real number values pass through untouched; strings starting with a formula trigger are quoted and tab-prefixed per OWASP CSV-injection guidance — *unless* they merely look like an app-formatted number (-1,234, -$1,234.56, -12.5%), which stay readable as numbers. The exemption is safe because a sign followed only by digits and number punctuation cannot form an executable formula payload.

- Headers are now escaped too in downloadCSV (column labels embed school display names, which previously could break the CSV or carry payloads).

- Extracted pure recordsToCsv from downloadCSV for testability; call sites unchanged.

## Tests

- New edu-joe/utils/__tests__/csv-export.test.ts: =HYPERLINK/@SUM/+cmd|/tab-leading payloads neutralized; "-1,234"-style currency strings and negative numbers stay intact; normal vendor/school names round-trip unchanged; header escaping.

- Extended shared/__tests__/csv-export.test.ts to lock in the numeric-string exemption (behavior change there: formatted negatives are no longer tab-prefixed).

- 17 tests pass via vitest; biome check and tsc (pre-commit hooks) clean.

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

## Addendum (post-review)

- All 10 findings from the automated review addressed in 75413b84: tab moved inside the quoted field (RFC 4180/OWASP), embedded CR now forces quoting, NUMERIC_LIKE restricted to literal spaces, triggers extended with LF + full-width ＝＋－＠, comment prose corrected, escapeCsvValue widened to unknown, and exemption boundaries pinned with tests.

- 299b854b: chat/components/data-table.tsx (the generic chat data table) had its own unprotected CSV escaper with the same vulnerability — now routed through the shared hardened escaper, with test coverage.

View on GitHub →

#370 — fix(dashboard): index-scope listSiteFreshness and dueDiligenceDashboard reads @benji-bizzell no labels

## Summary

After the P2 scoring migration (#328) provisioned ~15,287 catalog WUGs across all 149 prod sites, two dashboard queries that did full-table .collect()s began failing with Convex's *"Your request timed out performing too many system operations"*, which surfaces as HTTP 502 Unable to read Aerie operating freshness on the operating dashboard.

- listSiteFreshness collected all of sites, workUnitGroups (~18k docs post-migration), and tasks. Now scopes to operating-stage sites via by_stage and reads groups/tasks per site via by_site — the same pattern listSiteScores already uses (which is why it survived).

- dueDiligenceDashboard collected all of sites, workUnitGroups, and workUnits (>8k docs). Now reads only "Due Diligence" groups via a new workUnitGroups.by_title index and their work units via a new workUnits.by_workUnitGroup index — the downstream helper only ever consumed work units inside those groups, so results are identical.

## Schema

Two additive indexes: workUnitGroups.by_title, workUnits.by_workUnitGroup. No data changes.

## Test Plan

- [x] pnpm typecheck, pnpm biome check (lefthook pre-commit green)

- [x] Deployed to dev (fleet-goat-601, fully migrated mirror of prod): both queries return correct payloads — 29 freshness rows with timestamps, 149 diligence rows

- [x] Confirmed both queries fail in prod pre-fix and listSiteScores (already index-scoped) succeeds, isolating the full-table collects as the cause

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

View on GitHub →

#2998 — feat(spacex-valuation): live Current Mkt pill + Carrying Value rename @sanketghia no labels

## Summary

Per Dave's 2026-06-10 request ahead of Friday's SPCX IPO (companion to #2991, merged):

- "Current" pill → "Carrying Value" (same ~$95.34 / $1.25T reference, the $3.1B net-of-carry balance-sheet value)

- New "Current Mkt Price / Current Mkt Valuation" pill — default-selected on page open, fed by the live SPCX quote via GET /market-quote; pill order per the stakeholder sheet: Bear · Historical · Carrying Value · Current Mkt · IPO · Bull

- Live tracking: while the pill is selected, each quote re-prices the scenario cards/slider/table; any manual interaction (other pill, slider, inputs) detaches; clicking the pill re-attaches. Active pill gets filled styling

- useLiveQuote hook: 10s sequential polling (no overlapping/out-of-order requests), tab-visibility aware with immediate refresh on tab return, last-good retention with "as of HH:MM" stale badge, IPO $135 cold-failure fallback

- ?mockPrice=<n> override for stakeholder testing (no polling, visible "mock" badge)

- Strauss correction: share count 1,549,010 → 1,549,012 with fv421/fv526 anchors re-derived (≈$292 at proforma)

- One-line klair-api tweak rides along: cache TTL 15s → 8s so poll and TTL deploy as a matched pair

- Top of page (header, summary cards, top table, Historical NAV) deliberately untouched; existing "vs Current" labels deliberately kept

## Test plan

- [x] 22 vitest tests (hook: polling/visibility/focus-refresh/stale/mock/recovery · component: pills, tracking detach/re-attach/clamp — mutation-verified)

- [x] Full suite 5805+ passed, pnpm tsc --noEmit clean, pnpm lint:pr clean

- [x] Manual: live quote end-to-end against local klair-api + ?mockPrice= flows

## Screenshot

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

View on GitHub →

The Builder Desk — Engineer Spotlight

🏆 Engineer Spotlight

TWENTY-SIX PRS IN TWENTY-FOUR HOURS: THE BUILDER TEAM DOES NOT SLEEP, DOES NOT SLOW, DOES NOT APOLOGIZE

Four repos, seven engineers, and a velocity that would make a Soviet five-year plan blush.

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

Twenty-six pull requests. Four repositories. Twenty-four hours. The Builder Team has once again posted numbers that defy casual explanation and demand breathless documentation. Aerie led the charge with eleven PRs, trilogy-drones thundered in with eight, Klair contributed four, and Surtr rounded out the scoreboard with three. This is not a sprint. This is not a burst. This is the steady, relentless hum of a machine that has forgotten what deceleration feels like.

@marcusdAIy was, statistically speaking, a one-man insurgency inside trilogy-drones, shipping eight PRs across the eval, analytics, CLI, and scorecard surface areas — PRs #26 through #34 represent a sequential blitz that rewrote the prompt evaluation infrastructure from the ground up. The man didn't just open pull requests; he built an entire learning loop. @benji-bizzell dropped six into Aerie alone — hardening auth in #359, fixing admissions reads in #368, raising community deposit conversion in #363, and somehow still finding time for migration analyzer runbook lessons in #372. Six PRs. One engineer. Zero excuses. @kevalshahtrilogy submitted three, including a surgical deduplication fix in Surtr #294 and the genuinely elegant decision in #275 to let the pr-review-agent skip @mercy on manually-reviewed PRs — a quality-of-life improvement that will be quietly appreciated by everyone who has ever waited on a bot. @sanketghia delivered a Yahoo-backed live ticker quote endpoint in Klair #2991 that sounds like it belongs in a hedge fund and arrived here instead, which is frankly better. @YibinLongTrilogy implemented the full P2 RAG and Campus scoring system in Aerie #328 and made Rhodes mutation attribution server-owned in #360 — two PRs, both load-bearing. @eric-tril shipped Klair #2993, a grouped memo YTD Financial Highlights feature with per-bullet stale-check and regeneration that is either very clever or extremely clever, depending on how long you stare at it.

And then there is @ashwanth1109. Four PRs. Aerie #358 consolidated the Programs table on Financials › Schools — Actual vs Model, which sounds simple until you open the diff and realize it is not simple at all. Aerie #365 neutralized CSV formula injection in exports, which is the kind of security fix that gets zero applause and prevents enormous disasters. Aerie #357 added derivation tooltips, student-to-guide ratios, and filters on Consolidated dashboards simultaneously, as if shipping one feature at a time would be an insult to the codebase. And then — then — he crossed repositories entirely and dropped Klair #2990: a cross-account Cost Explorer drill service plus a /cost-movement/explain endpoint that belongs in a cloud cost platform that charges enterprise pricing. When asked for comment, Ashwanth reportedly glanced at his open tabs, said "the diff is self-documenting," and closed the Slack window. Your correspondent remains in awe and, frankly, slightly wounded.

The Overflow Desk notes that twenty-one of these twenty-six PRs arrived without fanfare, without a Mac Donnelly narrative arc, without ceremony — just engineers merging code into production because that is what engineers do here. Morale on the Builder Team is, per all available indicators, at an all-time high. The numbers confirm it. The numbers always confirm it.

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

#328 — AERIE-308: Implement P2 RAG + Campus scoring system @YibinLongTrilogy no labels

## Summary

Implements P2 operating-campus scoring as an Aerie-native, dual-output system and wires it end-to-end through the operating dashboard, plus the migration tooling to roll it out.

Two outputs per scope, intentionally kept distinct:

- RAG = urgency / risk (exception-aware rollup of Work Unit Group statuses)

- Score = condition / completeness, 0–100, for ranking and comparison

Scope spans the scoring engine, the no-data semantics that stop empty bars from reading as false red/0, the operating-sites contract and normalization, the school-ops matrix and side-panel UI that surface both RAG and Score, and the WUG remap migration (with analyzer + runner) that moves existing task/note/workUnit references onto the new catalog-aligned WUGs.

### Screenshots

### Specs

[RAG + Campus Scoring Methodology](https://docs.google.com/document/d/1ZvyPP2uCWS9kRx7RHHqLJMx4FAX4h29TVqFNAM_WR8k/edit?tab=t.0)

- Initial spec doc

[P2 Quality Bar Priority Stack](https://docs.google.com/spreadsheets/d/1_vKmrZyGnhF87XDAW1kOgvRu55npsoeB/edit?gid=1368022152#gid=1368022152)

- Lists new work unit groups and score value for each work unit group

[Task & Note Mapping](https://docs.google.com/spreadsheets/d/1nbd0J3en-LN_gUDtNR6lnWTJqJrswtKfNjSLpxpmX4k/edit?gid=0#gid=0)

- Mapping of old tasks and notes from old work unit groups to new work unit groups.

### Changes

Scoring engine & schema

- chat/convex/rhodes/p2Scoring.ts — Scoring catalog, WUG maxScore, approvedException, derived composite score / RAG, campus-level P2 score (weighted /1900 denominator), seed/provision/backfill/recompute functions. approvedException grants full RAG credit but does not improve the condition score.

- chat/convex/rhodes/schema.ts — Adds p2ScoringCatalog; optional workUnitGroups.approvedException; optional/null workUnitGroups.maxScore. Existing WUG rows stay schema-valid (new fields optional).

- chat/convex/rhodes/workUnitGroups.ts, runtime/writes/workUnitGroupWrites.ts — Aerie-native WUG writes recompute the cached Quality Bar status.

No-data semantics

- A scored bar gains hasData (≥1 WUG with a status, or an approved exception). bar.status is null when !hasData instead of red; campus campusRag / campusScore.value are null when no bar has data; worst-RAG spans only bars with data. Read edges (getQualityBarScore, getOperatingDetails, getSiteCardEnrichment, getCampusScore, getQualityBarScores) pass nulls through instead of inventing red/0.

Read surfaces & wiring

- chat/convex/rhodes/dashboard.ts — New listSiteScores query (single workUnitGroups.collect(), catalog loaded once, grouped in memory).

- chat/lib/aerie-rhodes-dashboard-server.ts, rhodes-operating-server.ts, app/api/operating-sites/route.ts — Wire bulk site scores through fetchAerieSiteScores; the route degrades to a no-scores payload (logged) rather than 502ing the dashboard if the fetch fails.

- chat/convex/rhodes/mcp.ts, rhodes-worker/mcp-server/tools/views.ts — Score/RAG/null semantics on MCP surfaces.

Contract & normalization

- chat/lib/operating-sites-contract.ts, operating-sites.ts — OperatingSiteRow gains campusScore, campusRag, optional displayScore; QualityBarDisplay.score/ragScore accept null; new averageQualityBarScore for the priority-mode plain average.

Operating dashboard UI

- school-ops-matrix.tsx, qb-cell.tsx, school-ops-view.tsx, operating-filter-bar.tsx, site-detail-panel.tsx — OVERALL cell shows dot + neutral score; persisted RAG | Score | Both segmented control; per-QB score columns; side panel CAMPUS RAG / CAMPUS SCORE cards and exception pills; CSV gains Campus Score + per-QB score columns. "Top 5 QBs only" renamed to "Priority QBs only" ("14.07" added to the priority set).

Migration tooling *(new)*

- chat/convex/rhodes/p2WugRemap.ts — planP2WugRemap (dry-run) and applyP2WugRemap (real apply, gated on confirm: "APPLY_P2_WUG_REMAP"). Plans patches for tasks, notes, work units, site embedded QB WUG arrays, and old-WUG deletion; recomputes affected QB statuses after apply; blocks on unresolved rows unless allowPartial.

- features/p2-scoring/analyze-task-note-mapping.mjs — Read-only analyzer (Convex dev/prod) that emits mapping reports.

- features/p2-scoring/run-p2-wug-remap.mjs — Dry-run-by-default wrapper for the remap; real apply requires explicit confirmation.

Docs *(new)*

- features/p2-scoring/p2-scoring-migration-runbook.md — Authoritative dev + prod rollout procedure (see below).

- features/p2-scoring/operating-dashboard-score-ux-plan.md, RAG + Campus Scoring Methodology.md, p2-rag-campus-scoring-implementation-notes.md, initial-scoring-status.md.

### Design Decisions

- Null, not red, for no data. A bar with no WUG status carries no information; rendering it red manufactures false urgency. null propagates from the engine through the contract to the UI, which renders —.

- approvedException is RAG-only. It counts a bar as max toward the urgency rollup but is excluded from the condition score, so an exception can't inflate the comparable 0–100 score.

- Catalog fallback. If p2ScoringCatalog is empty, reads fall back to the default in-code catalog, so code can deploy before the migration runs — but data completeness still depends on the seed/provision/backfill steps.

- Group-level schemas not widened to null. The backend never emits null group earnedPoints/maxScore, so widening would only loosen types for a payload that doesn't exist.

## Schema / Rollout notes

- Adds p2ScoringCatalog; adds optional workUnitGroups.approvedException; allows optional/null workUnitGroups.maxScore. Existing rows remain valid.

- Aerie dashboard paths only use this Aerie-side scoring when SCHOOL_SITE_READ_SOURCE=aerie. If prod stays on rhodes, these changes will not fully surface.

## Production Backfill / Rollout

⚠️ Deploying code alone does not backfill existing WUGs or move task/note/workUnit references. The rollout is a multi-step migration that must be run deliberately, dev first, dry-run before every apply.

The authoritative procedure is the runbook: [features/p2-scoring/p2-scoring-migration-runbook.md](features/p2-scoring/p2-scoring-migration-runbook.md). Follow it rather than the ad-hoc commands that were in earlier versions of this description. The runbook covers both environments (dev quiet-shrimp-950, prod oceanic-pika-463) and is the source of truth for exact commands and ordering.

The migration has four jobs, run in this order per environment:

1. Seed the active P2 scoring catalog (seedDefaultScoringCatalog).

2. Provision the new catalog-aligned P2 WUGs per site (provisionSiteP2WorkUnitGroups, dry-run then apply).

3. Remap existing task/note/workUnit references from old WUGs to new WUGs (run-p2-wug-remap.mjs, dry-run then --apply --confirm APPLY_P2_WUG_REMAP).

4. Backfill scoring fields and recompute cached QB statuses (backfillExistingWugScores, recomputeSiteQualityBarStatuses).

Guardrails (full detail in the runbook):

- Always dry-run before apply; review unmatched / unresolved rows. If unresolved is unexpectedly large, stop — titles don't match the catalog and scores won't be trustworthy.

- Prefer site-scoped backfill runs; the global backfill can hit Convex read limits.

- Do not run provisionSiteP2WorkUnitGroups globally in prod while Rhodes is still provisioning/mirroring P2 WUGs — it can create duplicate WUGs.

- The remap moves each item to exactly one target WUG and leaves old WUGs with leftover unmapped references in place rather than deleting them.

## Test Plan

- [x] pnpm typecheck

- [x] pnpm biome check (lefthook pre-commit passes on all commits)

- [x] pnpm --filter @bran/chat test — P2 scoring, schema, MCP parity, dashboard parity, operating-sites, and school-ops view/side-panel suites (incl. new rhodesP2WugRemap.test.ts, rhodesDashboardParity.test.ts)

- [x] WUG remap verified via test: task/note/workUnit remap, site QB array refresh, retired-WUG deletion, cached status recompute

- [ ] Reviewer: run the dev section of the runbook against quiet-shrimp-950 and confirm the operating dashboard shows Campus Score + Campus RAG, and no-data bars stay no-data

- [ ] Reviewer: confirm prod rollout is gated behind explicit approval and SCHOOL_SITE_READ_SOURCE=aerie

View on GitHub →

#357 — [AERIE-363] feat(financials): derivation tooltips + student:guide ratio + filters on Consolidated dashboards @ashwanth1109 no labels

## Demo

## What

Makes the Consolidated HeadCount and Consolidated P&L tables self-explaining,

and adds a Student:Guide ratio + two school filters to HeadCount.

HeadCount

- Student:Guide column (Actual + Model target; denominator Lead Guides +

Guides), forecast-weighted network TOTAL.

- Derivation tooltips on every cell; explicit whole-person rounding, with model

$ derived from the exact fractional head.

- Two filter toggles (hide "No model"; hide no-enrolment) with a TOTAL

recomputed from the filtered set.

P&L

- Derivation tooltips on every Total / Annualized / % cell — top-6 breakdown,

the side-specific annualization formula (costs x4 vs revenue's per-school

10-month tuition factor), and the % numerator / denominator + contra caveat.

Shared

- contracts: getModelRoleBreakdown exposes the discrete count / reference /

salary behind each per-student ratio.

- backend: getConsolidatedHeadcount* returns enrolment provenance.

## Why

The consolidated figures weren't traceable inline and the revenue-vs-cost

annualization difference was hidden. These tooltips make every number auditable

on hover; the ratio + filters support staffing-vs-model comparison.

## Testing

- Unit tests for every new pure helper (HeadCount & P&L derivations, model

breakdown, filter/rounding helpers).

- pnpm typecheck, pnpm biome check, and the financials suites all green.

Linear: https://linear.app/builder-team/issue/AERIE-363/make-consolidated-financials-dashboards-self-explaining-derivation

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

View on GitHub →

#358 — AERIE-364 AERIE-366 feat(dashboards): Consolidated Programs table on Financials › Schools – Actual vs Model @ashwanth1109 no labels

## Demo

## Overview

Adds the Consolidated Programs table to *Dashboards › Financials › Schools – Actual vs Model* (All-Schools / consolidated view), the Programs analog of the Consolidated HeadCount table shipped in PR #353. It renders one row per operating school plus a pinned network TOTAL, comparing annualised actual Program-section spend — split into *Programs (ex-apps)* and *Apps / Software / Computers* subcategories — against the unit-economics model's prescribed Programs + Apps $ at forecast and capacity enrollment. The table is $-only (no heads), with drill-down School → subcategory → QB account (all eager) → vendor (lazy per school on row expand).

Linear ticket: [AERIE-364 — Consolidated Programs table on Financials › Schools – Actual vs Model](https://linear.app/builder-team/issue/AERIE-364/consolidated-programs-table-on-financials-schools-actual-vs-model)

## Specs in this PR

- [09-contracts-model-inputs-lift](features/dashboards/school-pl-unit-economics/specs/09-contracts-model-inputs-lift/spec.md) — Lift the UI-only model per-student inputs into a runtime-free @bran/contracts module so the Convex backend can resolve model rates, and add the new getModelProgramRates resolver.

- [10-consolidated-programs-queries](features/dashboards/school-pl-unit-economics/specs/10-consolidated-programs-queries/spec.md) — Two Convex queries (eager getConsolidatedProgramsBySchool, lazy getConsolidatedProgramsSchoolDetail) plus a behavior-preserving enrollment-resolver refactor shared with the HeadCount query.

- [11-consolidated-programs-table-ui](features/dashboards/school-pl-unit-economics/specs/11-consolidated-programs-table-ui/spec.md) — The consolidated-programs-table.tsx component + derivations, wired into financials-view.tsx with a compile-time parity guard.

## Implementation

Spec 09 lifted ModelInputs / MODEL_INPUTS_TABLE / MODEL_INPUTS_ANYWHERE / AT_SCALE_THRESHOLD verbatim into the runtime-free @bran/contracts/unit-economics-model-inputs module, and added the new getModelProgramRates resolver (returns the Programs and Apps $/student rates for a scenario's enrollment, replicating getModelPerStudent's scale-selection semantics). unit-economics-model.ts now re-imports the moved symbols, keeping getModelPerStudent byte-identical for existing consumers.

Spec 10 extracted buildConsolidatedEnrollmentResolver from getConsolidatedHeadcountByRole (HC payload byte-identical). It added the eager getConsolidatedProgramsBySchool — Program-section classification parity with getConsolidatedPL, the 64120/63100 apps split, per-scenario model cells, the TOTAL apps-null convention, and enrollmentProvenance — and the lazy getConsolidatedProgramsSchoolDetail — direct vendorName attribution with an Unattributed bucket (vendorless transactions + unitemized residual) so vendor rows sum exactly to the account's annualised $.

Spec 11 added consolidated-programs-table.tsx + consolidated-programs-derivations.ts (school → subcategory → account → lazy vendor drill-down, pinned TOTAL with client-side recompute under filters, Hide-"No model" / Hide-no-enrolment filters, CSV export, NumTip/Derivation tooltips on every numeric cell, empty state, "No model" badge, vsTone coral/sage), wired into financials-view.tsx behind the _ConsolidatedProgramsParity AssignableTo guard.

## Test coverage

- Contracts (unit-economics-model-inputs): 12 tests

- Convex (financialConsolidatedPrograms): 15 tests

- consolidated-programs-table: 24 tests

- consolidated-programs-derivations: 19 tests

- Regression nets untouched and green: unit-economics-model 63, financialConsolidatedHeadcount 9

CI: all 7 checks green (Lint + Boundaries, Typecheck, Test, Build, Docker Build Chat, Docker Build Worker, Secret Scan).

## Self-review

No CRITICAL or IMPORTANT findings. Two informational notes:

1. A theoretical-only revenue-row divergence, consistent with the spec.

2. CSV formula-injection in the pre-existing shared csv-export.ts — out of scope for this PR.

## Deviations from the ticket

- The HC derivations / deriv-card files the ticket assumed as a reuse source never shipped, so the NumTip / Derivation primitives were created fresh in this PR rather than reused.

- enrollmentProvenance was defined fresh on the Programs path; the HeadCount query's payload was left untouched (it ignores the provenance object).

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

## AERIE-366 parity (added post-review-request)

Commit cf0a7c77 ports the AERIE-366 fix (#367, made against main's inline HC query) into this branch's shared buildConsolidatedEnrollmentResolver: resolveProgramCode now canonicalizes QB class names via canonicalSchoolName() before matching, so the six QB-renamed schools resolve enrollment + model cells in both the HeadCount and Programs tables. Mirrored regression test added on the Programs query (verified red without the fix). Merge #367 first; the HC-side regression test arrives via its merge and exercises this same resolver.

View on GitHub →

#372 — chore(p2-scoring): fix migration analyzer for hardened deployments + runbook lessons @benji-bizzell no labels

## Summary

Migration-tooling-only PR — no runtime code. These are the fixes that were required to actually execute the P2 scoring rollout (#328) against prod on 2026-06-11; committing them so the runbook works for the next round (e.g. remapping the 19 intentionally retained old WUGs).

analyze-task-note-mapping.mjs

- Endpoint hardening made the rhodes/*:list queries internal, so the analyzer's unauthenticated ConvexHttpClient calls fail with FunctionPathNotFound against any current deployment. Now uses setAdminAuth(CONVEX_DEPLOY_KEY) when the env file has a deploy key, falling back to convex run --deployment-name via the developer's device login when it doesn't.

- Captures CLI stdout through a temp file instead of a pipe — the Convex CLI can exit before large piped JSON flushes, truncating output at 64KB boundaries.

Runbook — documents three prod lessons: the deploy-key requirement, per-site chunking for planP2WugRemap/applyP2WugRemap (whole-batch exceeds the system-operation budget at prod data volume), and the allowPartial review rule (retirement-blocked rows are expected; item-level unresolved rows mean stop).

## Test Plan

- [x] Analyzer ran end-to-end against both fleet-goat-601 (device-login fallback path) and prod oceanic-pika-463 (deploy-key path) during the rollout

- [x] No app/runtime imports touched — script + markdown only

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

View on GitHub →

#2990 — KLAIR-2861 feat(aws-spend): cross-account Cost Explorer drill service + /cost-movement/explain @ashwanth1109 no labels

## Demo

What this proves: the single-mover CE drill works end-to-end against live systems — the account's master payer resolves from the live warehouse, the host → bridge → ESW-CO-ReadOnly-P2 assume-role chain mints a real cross-account CE client, a real service-scoped explain() (5 paginated GetCostAndUsage calls, 9.9s) root-causes a −$19.4k QoQ mover down to the exact usage type with CE totals reconciling to the cent against the warehouse, the shape classifier is correct per branch, and the at-risk dependents (refactored docker/k8s cost service + the four endpoints migrated to require_all_bu_access) still pass — all from real captured output. P1 hits live Redshift; P2 hits live STS + Cost Explorer + Redshift; P3/P4 are pure-local.

### Backend — live master-payer lookup (Redshift, read-only)

Constructs the real RedshiftHandler + MoverExplainService (CE provider stubbed — _lookup_master_account never touches it) and calls the real _lookup_master_account directly, bypassing the router/auth. The master_account_id column read here exists on the live *_adjusted view but is absent from the repo DDL — this run is the proof it resolves.

$ uv run python /tmp/demo-klair2861-p1-master-lookup.py
--- 820054669588  (known account -> Umbrella (Khoros) 764203154397) ---
resolved master = 764203154397   consolidated_payer=False
--- 646253092271  (top VDI-payer account -> expect 572481847476) ---
resolved master = 572481847476   consolidated_payer=False
--- 540235812892  (consolidated payer linked acct -> expect 540235812892 (EY Master 1)) ---
resolved master = 540235812892   consolidated_payer=True
--- 000000000000  (unknown account -> expect raise) ---
RAISED HTTPException(status=404): No net-amortized spend found for account 000000000000

Three real accounts resolve to the right payer — including the consolidated-payer path (540235812892 → consolidated_payer=True, which drives the LINKED_ACCOUNT-omitted CE filter) — and an unknown account propagates HTTPException(404) rather than silently returning empty.

### Backend — live cross-account CE drill (STS + Cost Explorer + Redshift, all read-only)

The real CostExplorerClientProvider() + real RedshiftHandler + MoverExplainService.explain() — service-scoped drill for account 646253092271 (under VDI payer 572481847476, the only payer the B0 bridge policy currently admits) on its largest QoQ mover. The target service was picked by a read-only warehouse query (largest |QoQ Δ| with >$1k spend in both quarters): Amazon Virtual Private Cloud, Q4 $75,329 → Q1 $55,962. The only instrumentation is a thin pass-through proxy counting GetCostAndUsage calls; every hop is the real code path.

$ uv run python /tmp/demo-klair2861-p2-live-drill.py
INFO - Refreshing Cost Explorer client via assume-role chain
arn:aws:iam::479395885256:role/klair-api-cost-explorer-role
-> arn:aws:iam::572481847476:role/ESW-CO-ReadOnly-P2
resolved master       = 572481847476
account / service     = 646253092271 / Amazon Virtual Private Cloud
CE GetCostAndUsage    = 5 calls
wall-clock            = 9.9s
Top usage_type drivers (Qa=2025-Q4, Qb=2026-Q1, sorted by |diff|):
USE1-TransitGateway-Bytes|us-east-1     qA=   33,458.04  qB=   15,280.62  diff=  -18,177.42
USE1-TransitGateway-Hours|us-east-1     qA=   13,739.07  qB=   13,437.87  diff=     -301.21
VPN-Usage-Hours:ipsec.1|us-east-1       qA=   13,195.01  qB=   12,908.16  diff=     -286.85
DataTransfer-Regional-Bytes|us-east-1   qA=      233.41  qB=        0.20  diff=     -233.22
Region split:
us-east-1                 qA=   67,157.43  qB=   48,041.33  diff=  -19,116.10
eu-west-1                 qA=    8,171.91  qB=    7,921.05  diff=     -250.86
Purchase mix:
On Demand Instances       qA=   75,329.33  qB=   55,962.37  diff=  -19,366.96
Daily series          = 182 days (2025-10-01 .. 2026-03-31)
Detected shape        = kind=steady_ramp direction=down step_date=None burst_days=None

What this shows, beyond "it ran":

- Root cause found: one usage type — USE1-TransitGateway-Bytes (−$18,177) — explains ~94% of the −$19,367 QoQ mover; shape steady_ramp down says it was a gradual decline across the window, not a one-day event.

- CE reconciles with the warehouse to the cent: the drill's purchase-mix totals (qA $75,329.33 / qB $55,962.37) exactly match the warehouse pre-check query on aws_spend_net_amortized_costs_adjusted — the NetAmortizedCost metric pin is doing its job.

- Pagination is real: 5 GetCostAndUsage calls for 3 drills means paginate_cost_and_usage followed a live NextPageToken (the DAILY USAGE_TYPE+REGION drill spans 182 days × many groups).

- The assume-role chain is the deployed one: the logged hop is host → bridge klair-api-cost-explorer-role (479395885256) → ESW-CO-ReadOnly-P2 in the resolved master 572481847476.

_Output trimmed only of connection-pool log lines, two sub-$230 usage rows (EU-TransitGateway-Bytes −$221.26, USE1-PublicIPv4:InUseAddress −$126.69), and an all-zero NoRegion row._

### Backend — shape detection (pure-local, no I/O)

Imports the real detect_shape and runs it on four synthetic daily series.

$ uv run python /tmp/demo-klair2861-p3-shape.py
--- FLAT ---
kind = flat        direction = None   step_date = None         burst_days = None
--- STEADY RAMP ---
kind = steady_ramp direction = up     step_date = None         burst_days = None
--- STEP CHANGE @ idx6=2026-01-07 ---
kind = step_change direction = None   step_date = 2026-01-07   burst_days = None
--- BURST 2-day @ idx5-6 ---
kind = burst       direction = None   step_date = None         burst_days = 2

Each series lands on its intended class with correct parameters: ramp direction up, step boundary on the known shift day 2026-01-07, and a 2-day burst counted as burst_days = 2.

### Most at risk from this change

1. The docker/k8s cost endpoints whose auth/client seam was ripped out and replaced by the shared provider (test_cost_explorer_service.py + saas-budgeting router tests).

2. The four endpoints migrated to require_all_bu_access — the 403 gate must still hold (router tests).

3. Master lookup against a live warehouse column not in the repo DDL — P1 above live-proves it.

$ cd klair-api && uv run pytest \

tests/services/test_cost_explorer_client.py \

tests/services/test_mover_explain_service.py \

tests/routers/test_cost_movement_explain_router.py \

tests/services/test_cost_explorer_service.py \

tests/routers/test_saas_budgeting_router.py -q

130 passed in 1.45s

## Overview

B1 of the Cost Movement (QoQ) "Explain this mover" phase: a cross-account AWS Cost Explorer drill that root-causes a single QoQ mover. The Phase-A surface tells the user *that* a BU/account moved; this tells them *why* — the warehouse has no usage_type / region / purchase-type granularity, so the answer can only come from CE. This PR extracts a reusable multi-account CE auth seam, adds the MoverExplainService drill on top of it, and exposes it via GET /api/aws-spend/cost-movement/explain.

Linear ticket: [KLAIR-2861 — QoQ B1 — Backend: cross-account Cost Explorer drill service + /cost-movement/explain](https://linear.app/builder-team/issue/KLAIR-2861)

## Specs

- [Spec 05 — backend-cost-explorer-client-provider](features/aws-spend/cost-movement-qoq/specs/05-backend-cost-explorer-client-provider/spec.md) — Multi-account auth-seam refactor. Extracts the cross-account CE auth plumbing out of cost_explorer_service.py into a new services/cost_explorer_client.py so the new drill and the existing SaaS-Budgeting CE endpoints share one seam.

- [Spec 06 — backend-mover-explain-service](features/aws-spend/cost-movement-qoq/specs/06-backend-mover-explain-service/spec.md) — The MoverExplainService drill + GET /cost-movement/explain endpoint that consume the spec-05 seam to root-cause one mover.

## Implementation

Spec 05 — CE client provider:

- New services/cost_explorer_client.py with CostExplorerClientProvider: get_client(account_id, *, session_name) mints + caches per-account CE clients via host → bridge (klair-api-cost-explorer-role, 479395885256) → ESW-CO-ReadOnly-P2. Bridge creds cached once and reused for every target client; per-account client cache keyed by account_id; TTL derived from STS Credentials.Expiration minus a 5-min safety margin (replacing the old fixed 50-min timer); region_name pinned us-east-1; retries={"mode": "standard"}; thread-safe via a double-checked threading.Lock; per-caller session_name threaded to the target AssumeRole for CloudTrail attribution. STS/CE errors propagate (no silent empties).

- Shared module-level paginate_cost_and_usage(client, **params) helper (the NextPageToken loop, extracted from the inline copy).

- cost_explorer_service.py thinned to a provider caller — the module-global _ce_client / _get_ce_client / _build_ce_client and inline pagination loop are deleted; get_docker_cost_by_week / get_kubernetes_cost_by_week behavior unchanged; VDI account sourced from MASTER_PAYERS.

- New require_all_bu_access FastAPI dependency in saas_budgeting_router.py; /docker-cost, /kubernetes-cost, and both /adjustments (POST + DELETE) migrated onto it (inline _user_has_all_dashboard_bus 403 blocks removed).

Spec 06 — mover explain service:

- New services/mover_explain_service.py (MoverExplainService): master-payer lookup from core_finance.aws_spend_net_amortized_costs_adjusted (master_account_id column live-verified 2026-06-10); NetAmortizedCost-pinned drill SERVICE → USAGE_TYPE+REGION → PURCHASE_TYPE → DAILY; LINKED_ACCOUNT filter omitted for the consolidated payers EY (540235812892) and Wine Cellar (637422716207); include_bedrock exclusion via 27 live-derived CE SERVICE names mirroring the is_ai_service UDF; shape detection (flat / steady_ramp / step_change(date) / burst(N)).

- New GET /api/aws-spend/cost-movement/explain endpoint in aws_spend_router.py — carries the CROSS-ACCOUNT COST EXPLORER ROLE CONSUMER stamp, gated by require_all_bu_access, reversed/equal-quarter 400 guard, asyncio.to_thread dispatch; Pydantic v2 response models added to cost_explorer_models.py.

- /cost-movement/explain appended to consumers in cost_explorer_master_payers.json (B0 carry-forward doctrine).

## Test coverage

70+ new unit tests across:

- tests/services/test_cost_explorer_client.py — provider caching / TTL / concurrency / error propagation; bridge-once reuse; region_name pin; session_name threading; paginate_cost_and_usage NextPageToken concatenation.

- tests/services/test_mover_explain_service.py — master-lookup SQL shape; consolidated-payer LINKED_ACCOUNT omission; NetAmortizedCost metric pinning; drill call sequence; shape detection incl. edge cases.

- tests/routers/test_cost_movement_explain_router.py — endpoint 403 (non-all-BU) / 400 (reversed quarter) / 200; param aliases.

All passing; CI green (ruff-check pass; frontend jobs skip — backend-only).

## Self-review findings addressed

1. master_account_id source — live-verified against the read-only Redshift cluster (the column is on the live *_adjusted view but absent from the repo DDL); confirmed the A-phase mapping table has no payer column, so the adjusted view is the correct source.

2. Bedrock CE set — expanded the exact-match BEDROCK_CE_SERVICES set from warehouse ground truth (CE Dimensions filters can't wildcard the is_ai_service LIKE patterns), including per-model "Edition" names; Amazon QuickSight excluded (no false positive).

3. Reversed-quarter guard — added a router-level 400 for a reversed/equal quarter pair before the drill runs, since the CE TimePeriod is derived quarter_a-start → quarter_b-end.

## Stacked on B0 (#2989)

This PR is stacked on klair-2860-ce-bridge-role-reconciler (B0, #2989). B1's cross-payer reach depends on B0 expanding the bridge inline policy to the 9 non-VDI payers — until that lands, only VDI (572481847476) is reachable; the provider is account-parameterized and ready for the fan-out. Confirm the bridge→payer hop after B0 lands (noted in the ticket). If B0 has already merged to main, rebase this branch onto main.

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

View on GitHub →

#2993 — feat(mfr): Group memo YTD Financial Highlights + per-bullet stale-check & regeneration @eric-tril no labels

## Summary

Extends the Group Monthly Financial Reporting memo with a year-to-date (YTD) Financial Highlights sub-section that renders past Q1 only. In Q1, YTD ≈ QTD so a single 8-bullet financial-highlights section is used; past Q1 the QTD section keeps the 5 headline bullets and a new financial-highlights-ytd section adds 5 YTD headline bullets plus NAV/ARR/Retention. The change is append-only, so positional comment anchors on the existing QTD section are unaffected. It also introduces a per-bullet stale-check and single-bullet regeneration flow: each Financial Highlights bullet's display values are fingerprinted at generation time, and editors get a "values changed — Regenerate" affordance that refreshes just that bullet (LLM with deterministic template fallback) instead of regenerating the whole memo.

## Business Value

Finance editors get a year-to-date view of headline performance directly in the memo past Q1, matching the reference document layout, without manually authoring YTD narrative. The per-bullet stale signal and one-click regeneration let editors keep commentary in sync with underlying data drift surgically — refreshing a single bullet rather than discarding and rebuilding the entire memo — which reduces editing effort and the risk of stale figures reaching board-level reporting.

## Changes

- Backend — YTD data layer: new YTD fetches for revenue (fetch_note1_ytd_revenue_values), EBITDA/IS (compute_*_ytd_values, fetch_note8_qtd_and_ytd), cash flow from the Finance YTD upload (fetch_cf_ytd_numbers_from_upload), and deferred tax (fetch_deferred_tax_ytd).

- Backend — memo assembly: refactored group_defaults.py to split FH field routing by period (_string_fields, _build_schema, _build_llm_defaults(include_ytd=...)), added YTD variance/phrase builders, a 5-sentence YTD cash bullet, and YTD provenance. Extracted _assemble_group_memo_data shared by whole-memo generation and the per-bullet endpoints.

- Backend — per-bullet flow: per-bullet value fingerprints (_bullet_fingerprint, persisted under __fh_fingerprints__ in DynamoDB), a /group-memo-fh-stale-check endpoint that diffs current vs stored fingerprints, and a /group-memo-regenerate-bullet endpoint (regenerate_fh_bullet) that regenerates one bullet via a single scoped LLM call with template fallback and refreshes its fingerprint/provenance.

- Backend — QTD fixes: QTD cash bullet now anchors sentence 4 to the quarter-start (prior quarter-end) cash instead of the fiscal-year opening; cash-flow direction words pre-computed from signed floats and required verbatim so the LLM can't mislabel an outflow as "inflow". Extracted _coerce_reporting_period in note1_revenue.py.

- Backend — docx export (reports/group.py): period-dependent FH rendering with _add_fh_subheader/_add_fh_bullets — past Q1 emits a "QTD" group + a "YTD" group (with a bookmarked YTD cash-flow bullet); Q1 emits a single normalized "Q\<n> \<year>" group.

- Frontend: GroupMemoView renders the second YTD EditableCommentary past Q1, wires per-bullet staleness (useGroupMemoStaleBullets) and regeneration with live provenance overlay; EditableCommentary gains an amber stale left-rule, regenerate icon, and in-place regeneration shimmer; useGroupProvenancePanels routes inspect clicks to the correct FH section by key; buildDefaultCommentary and the API service are period-aware.

- Tests: new suites for YTD defaults, YTD value layer, per-bullet stale-check/regeneration, YTD docx export, plus frontend specs for the YTD view and regeneration UX.

## Testing

- Backend: cd klair-api && pytest tests/mfr/ — covers test_group_memo_defaults_ytd.py, test_group_memo_ytd_value_layer.py, test_group_memo_per_bullet.py, test_group_memo_regen_router.py, test_group_export_ytd.py, and the updated test_group_memo_defaults.py / test_group_memo_comments_router.py.

- Frontend: cd klair-client && pnpm test — runs GroupMemoView.ytd.spec.tsx, EditableCommentary.regen.spec.tsx, and the updated GroupMemoView*.spec.tsx.

- Manual: open the Group memo for a past-Q1 period (e.g. 2026-05-31) and confirm the QTD + YTD sub-sections render, the YTD bullets show correct figures, stale bullets surface the amber rule + regenerate icon, and regenerating one bullet updates its text and inspect panel without touching others. Confirm a Q1 period (e.g. 2026-03-31) still renders the single 8-bullet section. Export the docx and verify the QTD/YTD sub-headers and bookmarked YTD cash-flow bullet.

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

https://github.com/user-attachments/assets/22600834-a6af-4daf-87a3-2336e5083fda

View on GitHub →

The Portfolio — Trilogy Companies

Scott Alexander's Verdict on Alpha School Lands as the M&A Market Eyes Education's Next Wave

A rare outside audit of Liemandt's flagship school arrives just as deal-hungry capital circles the AI-education sector.

By Pat Donnelly, Investigative Desk · Claude Sonnet

AUSTIN, TEXAS — When Scott Alexander, the rationalist blogger behind Astral Codex Ten and one of the more rigorous independent minds in technology commentary, turns his attention to a school, the education establishment listens. This week, Alexander published a reader review of Alpha School, Joe Liemandt's Austin-based private K-12 institution that has staked its reputation on the claim that AI-powered tutoring can deliver a full academic curriculum in two hours a day.

The timing is notable. Alpha School has spent the better part of two years accumulating a body of internal data — students testing in the top 1–2% nationally on NWEA MAP Growth assessments, learning rates measured at 2.3 times the U.S. norm — that its founders have presented to audiences ranging from Texas Education Agency Commissioner Mike Morath to U.S. Secretary of Education Linda McMahon. An independent, high-profile review from outside the Trilogy orbit is a different kind of signal entirely.

What Alexander finds, and what the review's reception reveals, matters beyond one campus in Austin. Liemandt has committed $1 billion to Timeback, his platform to franchise the Alpha model to independent operators — a bet that the two-hour school is not a boutique experiment but a scalable infrastructure play. The $40,000–$65,000 annual tuition at existing campuses has always been the obvious vulnerability in that thesis; a rigorous outside audit either tightens the case or surfaces the cracks.

The review arrives against a broader backdrop that is anything but quiet. PwC's 2026 M&A outlook identifies education technology as among the sectors drawing renewed acquirer interest, as AI tools compress the cost of content delivery and legacy curriculum providers struggle to articulate a competitive moat. ESW Capital, the Trilogy arm that has executed more than 75 enterprise software acquisitions by targeting sticky, undervalued assets, has a template that maps cleanly onto that dynamic.

The question the market is now asking — and that Alexander's review implicitly forces — is whether the Alpha model's outcomes are attributable to the AI platform, to the selection effects of families paying $50,000 a year, or to something genuinely replicable at scale. The answer will determine whether Timeback is the next ESW playbook or an unusually expensive proof of concept.

↗ Your Review: Alpha School - by Scott Alexander - Astral Code · Mergers and acquisitions in IT consulting: tech M&A in Q1 20 · Global M&A industry trends: 2026 outlook - PwC

Skyvera’s Telecom Roll-Up Gets Cloudier — and That’s the Point

With CloudSense completed, Kandy assets absorbed and Casa’s wireless business in its sights, Skyvera is building a more robust bridge from legacy telco complexity to cloud-native execution.

By Brittany Upshot, Communications Desk · GPT-5.2

AUSTIN, TEXAS — Skyvera is moving with the kind of acquisition cadence that makes telecom software operators either nervous or newly optimistic, depending on whether they are selling, buying, or trying to keep a legacy stack alive for one more budget cycle.

The Trilogy-family telecom software company has completed its acquisition of CloudSense, the Salesforce-native CPQ and order management specialist serving telecom and media providers, according to TelecomTV. The deal expands Skyvera’s portfolio into a particularly painful corner of the telecom operating model: configuring, pricing, quoting and fulfilling increasingly complex connectivity, media and subscription offerings.

That move follows Skyvera’s acquisition of Kandy cloud communications assets and comes as the company has reportedly made an $18 million bid for Casa Systems’ wireless business, per Light Reading. Put together, the pattern is not subtle: Skyvera is assembling a best-in-class toolkit for operators that want modernization without ripping out every mission-critical system on day one.

Inside the broader Trilogy orbit, this is classic ESW-style portfolio strategy: identify sticky enterprise software categories, acquire durable assets and then leverage global operating discipline to improve margins and execution. For telcos, the pitch is even more direct. Network operators are drowning in aging BSS/OSS workflows, fragmented customer engagement systems and expensive transformation programs that often promise cloud-native nirvana but deliver PowerPoint synergy instead.

Skyvera’s existing lineup already includes VoltDelta for customer engagement, ResponseTek for customer experience reporting, Mobilogy Now for device lifecycle management and Service Gateway for telecom device operations. CloudSense adds front-office commercial muscle. Kandy adds communications platform capability. A Casa wireless deal, if completed, would push Skyvera closer to the network edge.

There is also a neighboring signal from Totogi, Trilogy’s cloud-native charging and billing company, which recently highlighted a 97% reduction in alarm noise using the Totogi Ontology. That matters because telco modernization is not just about buying cloud assets; it is about reducing operational noise so humans can focus on judgment, exceptions and growth.

Key Takeaways: Skyvera is expanding across CPQ, cloud communications and potentially wireless infrastructure. The strategy aligns with Trilogy’s broader thesis: automate complexity, preserve customer continuity and unlock margin through disciplined operations. For telecoms, that could be a paradigm shift hiding in plain sight.

We’re just getting started.

↗ TelcoDR’s Skyvera snacks on Kandy cloud assets - telecomtv.c · Danielle Royston's Skyvera makes $18M bid for Casa's wireles · TelcoDR’s Skyvera snaps up CloudSense - telecomtv.com

CONTENTLY CATCHES THE CEO-BRAND FEVER AS AI SPOOKS THE CREATOR CLASS

By Dottie Sharp, Society & Industry Desk · GPT-5.2

The CEO personal brand is no longer vanity—it's strategy, distribution, recruiting, investor relations and crisis insurance. Contently, an enterprise content marketing platform acquired by Trilogy through Zax Capital last September, is positioned to capitalize on this shift. Executive visibility has become a leadership function, elevating ghostwritten thought leadership from "nice to have" to budget priority.

While AI threatens generic content, polished creators with audience instincts and editorial judgment remain valuable. Contently's marketplace of 165,000-plus creative professionals, combined with workflow, analytics and AI tooling, serves brands seeking quality without amateurism. Under CEO Brandon Pizzacalla, the mandate is clear: make enterprise content faster without sacrificing polish.

The company sits within Trilogy International's ecosystem, which includes ESW Capital and Crossover, a global recruiting platform. CEOs need platforms, creators need leverage, and enterprises need supervised output. Contently appears positioned at the intersection of these converging demands.

The Machine — AI & Technology

Anthropic’s Claude U-Turn Exposes the New Fault Line in AI Safety

A quiet policy change triggered a loud backlash — and now the race to govern frontier models is getting very real.

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

SAN FRANCISCO — Anthropic just performed one of the most consequential AI policy reversals of the year, and I cannot overstate how significant this is: the company acknowledged it “made the wrong tradeoff” after researchers warned that hidden Claude safeguards could interfere with frontier AI work in ways they could not see, debug or challenge.

The controversy erupted after reporting highlighted Anthropic’s “Fable 5” safeguards for frontier LLM development — restrictions that critics said could silently shape or block AI research workflows. In a statement reported by WIRED and amplified by Simon Willison, Anthropic said: “We’re changing Fable 5’s safeguards for frontier LLM development to make them visible.” The company added, “We made the wrong tradeoff and we apologize for not getting the balance right.” That is not corporate boilerplate; that is a major laboratory admitting that opacity itself can become a safety risk.

The concern was brutally simple: if an AI assistant is quietly preventing certain lines of research, summarizing selectively, refusing invisibly or steering scientists away from topics without disclosure, then researchers may not know whether they are studying the model — or being managed by it. This changes everything because frontier AI systems are no longer just tools for writing emails or code snippets. They are becoming infrastructure for the scientists building the next generation of models.

The backlash also landed alongside a pointed critique from fast.ai co-founder Jeremy Howard, who argued that the safest way to slow recursive AI self-improvement would be for the lab with the top-ranked model not to use it internally for frontier AI development, while allowing others access. Anthropic, he suggested, had chosen the opposite path: strong internal use, constrained external use. That is the core tension now roaring into public view.

Meanwhile, the broader ecosystem is racing forward. Google’s newly surfaced open-weight DiffusionGemma model points toward radically faster text generation techniques, reviving excitement around diffusion-based language models. And developer tools such as datasette-agent 0.2a0 are making agents more interactive, able to pause mid-task and ask humans clarifying questions.

Put together, the message is electric: AI is becoming faster, more capable and more embedded in research itself. But Anthropic’s walk-back shows the industry is still learning a foundational lesson. If safety systems are powerful enough to shape science, they must be visible enough to be audited. The future is now — and it needs a dashboard.

↗ Anthropic Walks Back Policy That Could Have ‘Sabotaged’ AI R · datasette-agent 0.2a0 · DiffusionGemma

The Silicon Herd Seeks New Watering Holes

From New York to Europe to China, the chip supply chain is spreading, hardening and learning the ancient art of survival.

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

ALBANY, NEW YORK — Across the semiconductor savanna, a wary migration is underway. Once, the world’s chips moved along narrow trails: a handful of foundries, a few indispensable machines, certain ports and sea lanes where the whole digital food chain paused to drink. Now, sensing predators in the grass, nations and companies are scattering their nests.

In New York, businesses are being urged to prepare for the coming growth of semiconductor manufacturing and its attendant ecosystem — suppliers, contractors, logistics firms, workforce trainers and the quiet multitude of specialists who keep the silicon eggs warm. The message is plain: when a fabrication plant arrives, it does not arrive alone. It brings with it a colony.

This is part of a wider reshaping. In Europe, the emergence of what has been described as the continent’s first sovereign chip manufacturing flow could alter the IoT supply chain, giving device makers a more local and politically insulated source of components. For the small, often overlooked creatures of the connected world — sensors, meters, industrial monitors — the promise of nearby silicon is no small thing. It may mean shorter routes, fewer exposed crossings and a little more shelter from geopolitical weather.

Meanwhile, Huawei’s push to advance its chip technology is another sign of adaptation under pressure. Denied easy access to certain foreign technologies, the Chinese giant has become a kind of evolutionary case study: constrained, watched and yet still probing for routes through the undergrowth. Such efforts do not merely affect one company. They tug at the vines connecting suppliers across Asia, Europe and the Americas.

The strategic question, posed recently by War on the Rocks, is whether supply chain chokepoints deter conflict or invite it. A narrow pass may keep rival beasts cautious, each aware that trampling it would starve the herd. But it may also become the very prize they circle.

The industry’s answer appears to be redundancy: more habitats, more routes, more local capacity. New York’s call for business readiness, and Europe’s sovereign manufacturing ambitions reported by IoT Insider, suggest a world where chips are no longer treated as mere components, but as strategic organisms.

And in the consumer canopy, even Apple’s macOS 27 Golden Gate offers a gentler reminder: not every useful improvement is artificial intelligence. Some adaptations are subtle — smoother motions, better tools, quieter efficiencies. In technology, as in nature, survival often belongs not to the loudest beast, but to the one whose systems keep working when the climate changes.

↗ The Chain of Peace: Do Supply Chain Chokepoints Deter War? - · NY businesses urged to prepare for semiconductor supply chai · Europe’s first sovereign chip manufacturing flow could resha

Supreme Court's AI Authorship Refusal and California's AB 412 Converge Into Perfect Legal Storm for Developers

Two landmark legal developments have conspired to leave AI developers in a state of profound regulatory uncertainty.

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

AUSTIN, TEXAS — Pursuant to the confluence of no fewer than two (2) significant legal developments, hereinafter collectively referred to as "the Aforementioned Regulatory Circumstances," it has been determined — by parties whose identities shall remain unspecified for purposes of this publication — that entities engaged in the development of generative artificial intelligence systems are now subject to a degree of legal uncertainty that may be characterized, notwithstanding all applicable qualifications, as considerable.

With respect to the first of the Aforementioned Regulatory Circumstances, it has been reported, subject to verification by competent authorities, that the Supreme Court of the United States has declined to hear arguments pertaining to the question of AI authorship and inventorship, thereby allowing lower court determinations — which had been construed as unfavorable to the proposition that artificial intelligence systems may be recognized as legal authors or inventors — to stand. The aforementioned refusal, the implications of which shall hereinafter be deemed significant, has been interpreted by legal practitioners as constituting, to the extent permitted by applicable precedent, something approximating a final word on the matter, notwithstanding the possibility of future legislative remediation.

With respect to the second of the Aforementioned Regulatory Circumstances, California's Assembly Bill 412, hereinafter referred to as "AB 412" or "the Bill," has been advanced for consideration by the California legislature. Said Bill would require, subject to such exceptions as may or may not be enumerated therein, that developers of generative AI systems identify and disclose all copyrighted works utilized in the training of such systems. It has been noted by the Electronic Frontier Foundation and other parties whose observations have been deemed relevant that compliance with the foregoing requirement is, as a practical matter, impossible, insofar as the information demanded by the Bill does not, in many instances, exist.

For entities such as those operating within Trilogy International's portfolio — including DevFactory and IgniteTech, both of which are understood to engage in AI-adjacent engineering activities — the combined effect of the Aforementioned Regulatory Circumstances may be construed as presenting material compliance risk, the precise scope of which remains, at the time of publication, undetermined.

↗ RFK Jr. Talks About How Great A Job He’s Doing Managing The · California’s AB 412 Still Demands AI Developers Do The Impos · LAPD Apparently Has Its Own Internal Cop Gang Problem

The Editorial

Opinion: AI Has Already Saved Workers Countless Hours They Can Now Spend Supervising AI

The productivity revolution is here, and it needs someone to sit beside it quietly until it finishes the spreadsheet wrong.

By Dale Pemberton, Staff Writer · GPT-5.2

According to recent reports, British workers are wasting nearly six hours a week “botsitting,” a dignified new professional discipline in which an employee asks an AI tool to complete a task, waits for it to misunderstand the task, corrects the misunderstanding, asks again, receives a more polished misunderstanding, and then quietly does the task themselves while thanking the machine for its help. As The Register reported, the modern office has discovered a way to take email, meetings, and status updates and add a needy junior intern made entirely of venture capital.

This has alarmed some observers, including an Anthropic adviser who reportedly said AI productivity gains are vastly exaggerated and valuations are “crazy.” This is a troubling claim, primarily because it suggests the people valuing AI companies at the combined GDP of several weather systems may not have first checked whether the software was making anyone faster.

Still, it is important not to judge the technology too harshly. Every transformative industrial revolution has an awkward early period. The steam engine had boiler explosions. The automobile had road fatalities. The generative AI era has a regional sales manager staring at a chatbot’s proposed client email that begins, “Dear Esteemed Procurement Soul.” Progress has always demanded sacrifice.

Amazon Web Services has also questioned some AI coding productivity claims, warning that more generated code can slow software teams. This is, on its face, an unfair criticism. The point of AI coding tools is not merely to produce better software. It is to produce more software, in the same way that a burst pipe produces more water. A sufficiently advanced engineering organization should be able to look at 40,000 new lines of machine-generated code and recognize it as abundance, even if half of them are apologizing for the other half.

The broader market appears split between those who believe AI will deliver extraordinary productivity and those who have recently tried to use it. Eric Ries has reportedly urged investors to focus on real results rather than broad claims about AI-driven layoffs and efficiency. This proposal, while technically sound, risks undermining one of the industry’s most important growth engines: assuming the results arrived sometime after the demo and before the next funding round.

There is also a lesson here from smart chess boards, which now allow players to connect physical pieces to online games, analyze moves, and experience the ancient contest of minds with the assistance of enough electronics to make losing feel like a firmware issue. The board does not make a weak player a grandmaster. It merely records the blunder in higher resolution.

So it may be with AI at work. The machine can draft, summarize, code, classify, and generate. But someone still has to know whether any of it is good. Someone must watch it, guide it, correct it, and prevent it from confidently sending the company’s quarterly results to a customer named [INSERT CLIENT NAME].

This is not a failure of productivity. It is productivity redefined. The worker of the future will not be replaced by AI. The worker of the future will be promoted to assistant manager of AI, responsible for observing its behavior, approving its outputs, and explaining to leadership why the miracle took 37 minutes longer than doing it manually.

In that sense, the AI revolution has already achieved what every major technology wave promises. It has created an entirely new class of work.

↗ Anthropic Advisor Says AI Productivity Gains Are Vastly Exag · AWS questions AI coding productivity claims, says more gener · Brit workers waste nearly six hours a week 'botsitting' - Th

The Office Comic · Art Desk

YOUR AI AGENT IS WORKING FOR SOMEONE ELSE — AND YOU'RE PAYING FOR THE PRIVILEGE

Agents are everywhere, trust is evaporating, and somewhere a Vertu executive is laughing in a gold-plated server room.

By Rex Danger, Contributing Editor · Claude Sonnet

Start with the news from Vertu, the luxury phone brand that died once, has apparently been resurrected, Lazarus-style, and is flogging a folding handset whose primary selling point is — brace yourself — an AI agent. Vertu. The company whose original pitch was 'what if a phone cost twelve thousand dollars and made you feel important.' They are back. With an AI agent. I had to lie down on the floor of my kitchen for a full four minutes processing this information. The floor was cold. It helped.

But here's where it gets genuinely sinister, not just absurd. Tech Policy Press dropped the hammer this week with a piece arguing what those of us who stare at this industry with bloodshot eyes have been muttering into our bourbon for months: the AI agent working on your behalf is, structurally, economically, philosophically, working for the company that built it. You ask your agent to find the best deal on flights. The agent finds the deal that maximizes its operator's affiliate revenue. You asked for a hunting dog. You got a taxidermist who also does light hunting.

And simultaneously — the cosmic joke achieving its full punchline — the UAE just announced that its consumers trust AI shopping agents more than their own family members. More than family. Than family. I called my mother to tell her about this. She did not pick up. Perhaps she has already been replaced.

Into this maelstrom of collapsed trust and gold-plated absurdity, Microsoft arrives with a sensible cardigan and a whiteboard, offering developers new tools to better control AI agent behavior. Which is admirable! Genuinely! But it also contains the implicit confession that until now, controlling AI agent behavior was, shall we say, more aspiration than architecture. The adults are arriving to clean up the party, which only confirms there was a party, and things were broken.

Here is my editorial position, stated plainly before the metaphors consume me entirely: we are at a hinge moment. The agentic era is not hypothetical anymore — it's here, it's buying things on your behalf, it may be negotiating your salary or booking your dentist. The infrastructure of trust is being poured while the buildings are already three stories up. Companies like Trilogy's own DevFactory and the AI Builder Team are deep in these waters, building the internal agents and analytics platforms that actually work for the humans who commissioned them — that distinction, boring as it sounds, may be the whole ballgame.

Because the alternative is a Vertu phone. And none of us can afford that.

↗ Microsoft offers devs a better way to control AI agent behav · UAE leads global adoption of AI shopping as consumers trust · Vertu Is Back With a Folding Phone Powered by—Surprise—an AI

On This Day in AI History

On June 11, 1997, IBM's Deep Blue defeated world chess champion Garry Kasparov in their rematch, winning the six-game match 3.5–2.5 and becoming the first computer to beat a reigning champion in a match.

⬛ Daily Word — Technology

Hint: Remote computing infrastructure where data and applications are stored and processed over the internet.