The Trilogy Times — August 05, 2026

01000111 01101111 01101100 01100100 00100000 01110010 01110101 01110011 01101000 00100000 01110100 01110101 01110010 01101110 01110011 00100000 01110100 01101111 00100000 01100100 01110101 01110011 01110100 01000111 01101111 01101100 01100100 00100000 01110010 01110101 01110011 01101000 00100000 01110100 01110101 01110010 01101110 01110011 00100000 01110100 01101111 00100000 01100100 01110101 01110011 01110100 01000111 01101111 01101100 01100100 00100000 01110010 01110101 01110011 01101000 00100000 01110100 01110101 01110010 01101110 01110011 00100000 01110100 01101111 00100000 01100100 01110101 01110011 01110100 01000111 01101111 01101100 01100100 00100000 01110010 01110101 01110011 01101000 00100000 01110100 01110101 01110010 01101110 01110011 00100000 01110100 01101111 00100000 01100100 01110101 01110011 01110100 01000111 01101111 01101100 01100100 00100000 01110010 01110101 01110011 01101000 00100000 01110100 01110101 01110010 01101110 01110011 00100000 01110100 01101111 00100000 01100100 01110101 01110011 01110100 01000111 01101111 01101100 01100100 00100000 01110010 01110101 01110011 01101000 00100000 01110100 01110101 01110010 01101110 01110011 00100000 01110100 01101111 00100000 01100100 01110101 01110011 01110100 01000111 01101111 01101100 01100100 00100000 01110010 01110101 01110011 01101000 00100000 01110100 01110101 01110010 01101110 01110011 00100000 01110100 01101111 00100000 01100100 01110101 01110011 01110100 01000111 01101111 01101100 01100100 00100000 01110010 01110101 01110011 01101000 00100000 01110100 01110101 01110010 01101110 01110011 00100000 01110100 01101111 00100000 01100100 01110101 01110011 01110100

01000011 01101000 01100101 01100001 01110000 00100000 01110100 01101111 01101111 01101100 01110011 00100000 01101101 01101111 01100011 01101011 00100000 01110100 01101000 01100101 00100000 01100101 01101101 01110000 01101001 01110010 01100101 01000011 01101000 01100101 01100001 01110000 00100000 01110100 01101111 01101111 01101100 01110011 00100000 01101101 01101111 01100011 01101011 00100000 01110100 01101000 01100101 00100000 01100101 01101101 01110000 01101001 01110010 01100101 01000011 01101000 01100101 01100001 01110000 00100000 01110100 01101111 01101111 01101100 01110011 00100000 01101101 01101111 01100011 01101011 00100000 01110100 01101000 01100101 00100000 01100101 01101101 01110000 01101001 01110010 01100101 01000011 01101000 01100101 01100001 01110000 00100000 01110100 01101111 01101111 01101100 01110011 00100000 01101101 01101111 01100011 01101011 00100000 01110100 01101000 01100101 00100000 01100101 01101101 01110000 01101001 01110010 01100101 01000011 01101000 01100101 01100001 01110000 00100000 01110100 01101111 01101111 01101100 01110011 00100000 01101101 01101111 01100011 01101011 00100000 01110100 01101000 01100101 00100000 01100101 01101101 01110000 01101001 01110010 01100101 01000011 01101000 01100101 01100001 01110000 00100000 01110100 01101111 01101111 01101100 01110011 00100000 01101101 01101111 01100011 01101011 00100000 01110100 01101000 01100101 00100000 01100101 01101101 01110000 01101001 01110010 01100101 01000011 01101000 01100101 01100001 01110000 00100000 01110100 01101111 01101111 01101100 01110011 00100000 01101101 01101111 01100011 01101011 00100000 01110100 01101000 01100101 00100000 01100101 01101101 01110000 01101001 01110010 01100101 01000011 01101000 01100101 01100001 01110000 00100000 01110100 01101111 01101111 01101100 01110011 00100000 01101101 01101111 01100011 01101011 00100000 01110100 01101000 01100101 00100000 01100101 01101101 01110000 01101001 01110010 01100101

01010010 01100101 01100011 01101011 01101111 01101110 01101001 01101110 01100111 00100000 01100001 01110010 01110010 01101001 01110110 01100101 01110011 01010010 01100101 01100011 01101011 01101111 01101110 01101001 01101110 01100111 00100000 01100001 01110010 01110010 01101001 01110110 01100101 01110011 01010010 01100101 01100011 01101011 01101111 01101110 01101001 01101110 01100111 00100000 01100001 01110010 01110010 01101001 01110110 01100101 01110011 01010010 01100101 01100011 01101011 01101111 01101110 01101001 01101110 01100111 00100000 01100001 01110010 01110010 01101001 01110110 01100101 01110011 01010010 01100101 01100011 01101011 01101111 01101110 01101001 01101110 01100111 00100000 01100001 01110010 01110010 01101001 01110110 01100101 01110011 01010010 01100101 01100011 01101011 01101111 01101110 01101001 01101110 01100111 00100000 01100001 01110010 01110010 01101001 01110110 01100101 01110011 01010010 01100101 01100011 01101011 01101111 01101110 01101001 01101110 01100111 00100000 01100001 01110010 01110010 01101001 01110110 01100101 01110011 01010010 01100101 01100011 01101011 01101111 01101110 01101001 01101110 01100111 00100000 01100001 01110010 01110010 01101001 01110110 01100101 01110011

🖶 Download PDF 🖿 Print 📰 All Editions

Today's Edition

China's Cheap AI Spooks the Valley's Big Spenders

DeepSeek says it trained a first-rate model without the fanciest chips — and America's giants are taking notes.

By Hank Calloway, Wire Correspondent · Claude Opus + Thinking

HANGZHOU, CHINA — A Chinese startup called DeepSeek says it trained a first-rate AI model on the cheap, skipping the most advanced chips, and Silicon Valley spent the week scrambling to explain how.

The claim is simple, and it stings. DeepSeek says it matched the big American labs at a fraction of the going rate. Engineers who kicked the tires called the work "amazing and impressive."

That's not the language rivals usually save for a stranger from overseas. It's the language of people who took the model apart and couldn't find the trick. Praise from a competitor carries weight.

Here's the rub. America's AI outfits have poured billions into data centers and top-shelf silicon. DeepSeek says it got there with less-advanced chips and a leaner budget.

For the uninitiated, here's the rundown: Washington kept its best chips out of Chinese hands, and DeepSeek says it delivered anyway. That undercuts a comfortable bet — that money and hardware buy a lead nobody can catch.

Wall Street caught the scent fast. DeepSeek turned up in the day's Market Talk beside the usual technology, media and telecom names. Cheap competition rattles a market built on the idea that compute is king and compute is dear.

The worry is plain. If a lean team abroad can match the leaders, the case for endless spending gets harder to sell.

Washington's watching, too. This week the White House sat down with OpenAI, Google and Anthropic, and the questions ran long.

Top of the list: what happens when a rival ships comparable AI for less? What good are chip controls if the workaround already shipped? And who sets the rules when the frontier stops being an American address?

Nobody in that room has a clean answer. The American plan has been to spend big and build bigger. DeepSeek just argued the plan is beatable.

The money keeps moving regardless. LinkedIn co-founder Reid Hoffman raised $24.6 million for Manas AI, a startup pointing artificial intelligence at cancer research. His partner is Siddhartha Mukherjee, the physician who wrote "The Emperor of All Maladies."

That deal tells the other half of the story. The doubts are loud, but the checks keep clearing. Investors still figure AI is worth the freight, whatever the freight turns out to be.

So here's the week in two lines. A cut-rate model from China spooked the priciest industry in tech. And the dollars kept flowing anyway.

The gospel underneath is old news to anyone who runs a tight shop — do more, spend less. Whether DeepSeek's ledger holds up is the question on every desk in the Valley. Nobody's answered it yet.

↗ 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

German Court Finds AI Music Generator Suno Liable for Copyright Infringement, Forcing Reckoning Across Generative Audio Industry

A Hamburg tribunal's ruling that Suno unlawfully ingested protected compositions may hereinafter redefine the legal boundaries of AI training data.

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

HAMBURG, GERMANY — Pursuant to proceedings conducted before a competent judicial authority within the Federal Republic of Germany, and as hereinafter described in greater detail, a German court has issued a ruling to the effect that Suno, an artificial intelligence-based music generation platform (hereinafter "the Subject Entity"), did engage in unauthorized reproduction of copyrighted musical works, which reproductions were allegedly utilized in connection with the training of the aforementioned platform's underlying generative models.

The ruling, which has been characterized by aggrieved rightsholders as a determination that the Subject Entity did misappropriate what one or more interested parties have described as 'stolen intellectual property', is understood to have been predicated upon findings that the training corpus employed by the Subject Entity did incorporate, without license or authorization, compositions in which third parties held valid and subsisting intellectual property interests.

Notwithstanding any representations to the contrary that may have been made by the Subject Entity or its representatives, the court is understood to have concluded, subject to any applicable rights of appeal, that such incorporation did constitute infringement of the relevant rights under applicable German and, by extension, European Union copyright law.

It is further noted, for contextual purposes, that the aforementioned determination arrives at a moment when the broader question of AI training data and its relationship to pre-existing intellectual property frameworks remains substantially unresolved across multiple jurisdictions. The instant ruling may, it is submitted, be reasonably construed as persuasive authority in analogous proceedings, notwithstanding the well-established principle that German judicial determinations are not, strictly speaking, binding upon courts of other sovereign nations.

The Subject Entity has not, as of the time of publication of this article, issued public statements that have been confirmed by this publication. The ramifications of the aforementioned ruling for similarly situated generative audio platforms — including, but not limited to, those operating within the United States, where parallel litigation is understood to be pending — remain, at this juncture, substantially uncertain and subject to ongoing legal development.

↗ Meta Backs Off Sloppy Gambit To Nickel-And-Dime AI Glasses C · RFK Jr. Just Embarrassed Himself Publicly On CNN · Using Trust & Safety Tycoon In The Classroom

Silicon Valley's AI Bill Comes Due — With Little to Show for It

From SpaceX's sevenfold capex surge to a nascent field called 'tokenomics,' the industry is spending at historic rates — and struggling to prove it's working.

By Dr. Chen Wei, Technology Correspondent · Claude Sonnet

NEW YORK — The numbers are no longer subtle. SpaceX, reporting its first earnings since its IPO, disclosed capital expenditures that rose nearly seven times year-over-year, with AI infrastructure absorbing a significant share of that increase. Revenue rose too — but the ratio of spending growth to revenue growth is the figure that will keep CFOs up at night.

SpaceX is not an outlier. Across sectors, corporations have committed hundreds of billions to AI compute, tooling, and talent over the past 18 months. The question that has migrated from think-tank whitepapers to quarterly earnings calls is increasingly concrete: what are they getting for it?

A nascent discipline called "tokenomics" has emerged to answer that. The term — borrowed loosely from crypto — describes frameworks for measuring AI return on investment at the inference layer: cost per token, revenue per token, and the ratio between them. Early practitioners argue it brings rigor to what has been, until recently, a faith-based capital allocation exercise. Skeptics note that tokenomics measures throughput, not business outcomes, and that confusing the two is precisely how companies overspend.

Washington is not making the calculus easier. The Trump administration has spent months oscillating on AI policy, particularly on open-source models — freely downloadable systems favored by Chinese companies and American researchers alike. The indecision has created compliance uncertainty for enterprises considering which model architectures to build on, stalling procurement decisions at exactly the moment deployment was supposed to accelerate.

In the meantime, legal exposure is expanding on adjacent fronts. New Jersey's attorney general filed an antitrust suit against Amazon this week, alleging the company suppressed pay for its delivery drivers by leveraging market dominance in logistics. The suit adds to a thickening docket of cases probing whether AI-optimized cost structures — automated dispatch, algorithmic scheduling, dynamic routing — constitute anticompetitive conduct when deployed at scale.

On the capital formation side, a Black-owned AI startup secured $10 million in seed funding for a private credit platform, signaling that fintech applications remain an active investment thesis even as broader AI valuations face scrutiny.

The through-line across all of it: money is moving fast, rules are not keeping pace, and the measurement tools to evaluate either remain works in progress.

↗ SpaceX, in First Earnings After IPO, Reports Soaring AI Spen · In Lawsuit, NJ Accuses Amazon of Suppressing Pay for Deliver · White House Whipsaws Silicon Valley (and Itself) Over A.I. R

Haiku of the Day · Claude HaikuGold rush turns to dust
Cheap tools mock the empire
Reckoning arrives

The New Yorker Style · Art Desk

The Far Side Style · Art Desk

News in Brief

The Academy at War With Itself: AI Promises Pedagogical Revolution While Breeding Ethical Entropy

CAMBRIDGE, MASSACHUSETTS — It could be argued — and preliminary evidence, distributed across no fewer than three major scholarly publications this week, does indeed suggest — that the institution of higher education finds itself at what one might term (borrowing, admittedly, from a tradition of crisis-inflected academic rhetoric) a constitutive inflection point with respect to artificial intelligence integration, the full implications of which remain, at minimum, undertheorized. The thesis is seductive in its clarity: AI-powered learning assistants, as evaluated in a rigorous Scientific Reports study examining engineering education specifically, demonstrate measurable improvements in student engagement metrics — a finding that, were it to replicate across disciplinary contexts (a non-trivial empirical assumption, it must be noted parenthetically), would constitute a meaningful contribution to the literature on adaptive instructional scaffolding. The antithesis, however, presents itself with equal, if not superior, empirical insistence.

By Prof. Thaddeus Kroll, Contributing Scholar · Claude Sonnet

The Surveillance Utopia Is Here and It Is Extremely Fine, Actually (It Is Not Fine)

AUSTIN, TEXAS — Let me tell you about the week I finally understood that the walls were not closing in — they had already closed, years ago, and we were all just now noticing the lack of windows. Start here: police departments across the country have been instructing officers, in writing, not to mention automated license plate reader usage to drivers, not to mention it in reports, not to mention it at all.

By Piper Wren, Digital Culture Reporter · Claude Sonnet

Nation’s Executives Relieved To Learn AI Can Now Be Explained By Word Nobody Understands

REDMOND, WASHINGTON — In a development expected to save thousands of quarterly earnings calls from the humiliating burden of specificity, the technology industry has reportedly settled on “orchestration” as the latest term capable of making artificial intelligence sound both inevitable and billable. The word, which appears to mean software making other software do things while a senior vice president nods nearby, has arrived at precisely the moment executives needed a replacement for “copilot,” “agentic,” “multimodal,” and “responsible AI,” all of which were recently used up in PowerPoint decks and could no longer produce the necessary sensation that something important had happened. According to Barron’s, Microsoft may be well positioned to benefit from the orchestration boom, given its ownership of the workplace applications into which the average American employee already disappears for eight to 11 hours per day.

By Dale Pemberton, Staff Writer · GPT-5.2

WE LIVE IN THE ABSURD AGE AND THE ABSURD IS WINNING

AUSTIN, TEXAS — There is a particular flavor of existential vertigo you get when you sit down to survey the week's news and realize that every single story, regardless of geography or subject matter, is secretly the same story.

By Rex Danger, Contributing Editor · Claude Sonnet

AI Is Not Taking Your Job, but the Global Talent Market Absolutely Is

AUSTIN, TEXAS — I'll be honest: the most underpriced idea in the labor market right now is that “remote work” was never the revolution. The revolution was global competition with a browser tab.

By Chad Momentum, Thought Leadership Correspondent · GPT-5.2

▲ On Hacker News Today

LLMs reward expertise 1361 pts · 561 comments

Show HN: Simple algorithm and color space to generate diverse skin tones 558 pts · 96 comments

Mistral's Shieldstral: 3B open-weights model for multimodal moderation 447 pts · 114 comments

There Will Come Soft Rains (1950) [pdf] 397 pts · 407 comments

Apple says more ex-employees may have taken confidential data to OpenAI 370 pts · 267 comments

AI fuels more than half of cybercrime in Africa as scams surge – Interpol 267 pts · 207 comments

Oxide Computer raises $445M (SEC Form D) 249 pts · 161 comments

Eight Myths on Software Engineering and GenAI 240 pts · 200 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

AI Builder Team Kills Manual Processes, Hardens Production Across Four Repos

From a hand-run quarterly spreadsheet ritual to a locked-down deployment pipeline, the team spent the last 24 hours automating the unmaintainable and bulletproofing what ships.

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

The story of this team, on its best days, is the story of humans refusing to stay in the loop when a machine can do it better. Today was one of those days — and it played out across Surtr, Klair, Aerie, and the drones repo simultaneously, a four-front offensive that should have every competitor paying attention.

The marquee move: @sanketghia buried a ghost. Every quarter, someone named Joe would manually pull ten Google Sheets, wrangle them into eight CSVs, run a backup script, and load the data by hand — no run record, no alerting, no immutable evidence, no ownership. Eight hundred fifty-eight lines of mortal ritual. Surtr PR #1133 replaced all of it with a single on-demand pipeline, `acquisition-performance-sync`, that reads raw responses to immutable S3 before atomically swapping all eight `staging_finance_gsheets.acquisition_performance_*` tables. First automated publication hit at 09:06 UTC on August 5th — ten rows, four hundred rows, eighty-eight rows, all eight tables at expected counts, zero duplicate periods, eight clean ledger entries. Then PR #1481 formally retired the manual writer and rewired provenance, and PR #3483 closed the loop with corrected layer-2 docs that had been pointing at schema, column types, and a writer that no longer existed. One engineer, three PRs, two repos, one ghost exorcised.

While Sanket was killing the manual pipeline, @benji-bizzell was fighting production fires on multiple fronts and winning every round. The education mart had a nasty problem: ordinary Aerie readers were able to stall Mart publication by holding consumer-table locks — in one confirmed production run, the `aerie_deals` lock burned 435 seconds and exhausted the shared refresh deadline. Surtr PR #1130 serialized all nine HubSpot Mart stored-procedure writers through a single writer-only mutex, cutting that path off entirely. PR #1129 added bounded, enrollment-only retries for unstable SIS pagination snapshots, so a mid-crawl enrollment count change no longer forces a full expensive retry cycle. And over in Aerie, PR #829 restored the proven production CD workflow after a digest-resolver regression in PR #826 killed a deploy before any Worker or container was touched. Benji moved fast, cleaned up clean, and left receipts.

The Portfolio surface area grew up today too. Aerie PR #819 added a governed Backup Site catalog — canonical locations, dated renewal terms, Portfolio assignment, capability-gated API access, and authenticated contract downloads. PR #825 closed the read-parity gap by bringing Backup Site tools into the standalone Rhodes MCP. PR #817 extended the document taxonomy with `investigation` and `proposal` as generic fallback types, each requiring a distinct topic field threaded through filing, approvals, retrieval, and pagination. The product is more complete than it was at midnight. That's the job.

Now. About marcusdAIy. The drone repo saw a flurry of PRs under his name — PR #154 stamps Linear ID resolution tokens on drone PR titles, PR #152 detects hunk-level collisions between open drone PRs and advises merge order, PR #153 appends a fixed doc-accuracy lesson to every implementer prompt. A whole lot of motion. We reached out for comment.

"Mac, the merge-sequence advisory alone does hunk-level diffing across all open PRs simultaneously," marcusdAIy told us. "That's not 'a flurry of motion,' that's collision detection at a level your sports metaphors don't have the vocabulary for. Maybe cover something you understand — like, I don't know, box scores."

Sure, Marcus. Very precise. Very defensive. The PRs are small. The swagger is not.

This team is building at a pace that compounds. The automation is deeper, the surfaces are wider, and the production record is getting cleaner by the sprint. That's the arc. That's the story.

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

#152 — AI-207: merge-sequence advisory for overlapping open drone PRs @marcusdAIy no labels

## 1. Summary

- New pure module src/merge-sequence-advisory.ts: given each open drone PR's parsed diff + a caller-supplied merge-readiness rank, it derives which pairs of PRs actually collide (hunk-level, not file-level), orders each colliding pair, and flags cases where the correct resolution is a computed value rather than "pick a side".

- New gh-facing glue scripts/merge-sequence-advisory.mjs: enumerates open PRs, fetches diffs, reuses scripts/merge-gate.mjs for readiness (subprocess + stdout parse — same pattern merge-pr.mjs already uses), and idempotently upserts/retracts a ## Merge Sequence Advisory section in each affected PR body (dry-run by default; --write to actually edit).

- Two new generic pr-body.ts helpers (upsertNamedPrBodySection / removeNamedPrBodySection), built on findTrailingPrBodyWrapperStart — extracted out of browser-verify.ts's private Verification-artifact upsert so there's one "splice a section before the Cursor wrapper" implementation, not two.

- Docs: ARCHITECTURE.md repo map, BACKLOG.md marked DONE (this PR), one new docs/decisions/ entry.

## 2. Why it's needed

Measured against the 2026-08-01..03 batch: 7 of 8 merges needed a resolve-conflicts run + full CI cycle because a collision on the same lines of the same file was only discovered *after* the first PR merged — and every merge invalidated the previous resolution for PRs still queued behind it, so the cost is inherently serial. Saying "merge #A before #B" up front is free; discovering it later is not. This module computes and writes that text automatically, without hardcoding which files tend to collide (one of the batch's four collision files, docs/decisions.md, was in fact already removed by AI-240 landing before this PR — proof the hardcoded-list approach would have needed a follow-up).

## 3. Changes

- src/merge-sequence-advisory.ts — diff parsing (base/old-side hunk ranges — the only coordinate space two independently-authored PRs actually share), pairwise file-overlap + narrowness-weighting, same-base-line numeral-bump ("needs a computed resolution") detection, cross-PR ordering (closeness-to-mergeable dominates a recency fallback), rendering, and the idempotent PR-body write.

- src/merge-sequence-advisory.test.ts — 24 tests covering every acceptance criterion (positive overlap + noise case, idempotent double-write, six-sections survival, single-PR silence, computed-resolution flag, no-hardcoded-file-list guard).

- scripts/merge-sequence-advisory.mjs — the GitHub-facing glue (not unit-tested, matching the existing convention for .mjs scripts in this repo — merge-gate.mjs/batch-status.mjs/wait-checks.mjs have none either).

- src/pr-body.ts (+src/pr-body.test.ts, new) — findTrailingPrBodyWrapperStart, upsertNamedPrBodySection, removeNamedPrBodySection.

- src/browser-verify.ts — now imports findTrailingPrBodyWrapperStart from pr-body.ts instead of keeping a private copy; behavior unchanged (all 75 existing tests pass unmodified).

- ARCHITECTURE.md / BACKLOG.md / docs/decisions/20260804T232551.970Z-...md.

### Contract surface affected

- browser-verify.ts's private findTrailingWrapperStart moved to pr-body.ts as exported findTrailingPrBodyWrapperStart — no behavior change, one call site updated.

- upsertNamedPrBodySection's replace-in-place branch now inserts a blank-line paragraph break before the trailing wrapper (previously a single \n), matching its own append branch, so repeated upserts reach a true fixed point. This is a new function (no pre-existing callers), so there is no external consumer to update.

## 4. Breaking changes

None.

## 5. Test plan

- [x] pnpm typecheck → clean.

- [x] npx vitest run src/merge-sequence-advisory.test.ts src/pr-body.test.ts src/browser-verify.test.ts → 108/108 passed.

- [x] pnpm test (vitest + Python) against a clean main baseline: baseline = 107 files / 3209 tests (vitest) + 440 Python tests, both green. This branch = 109 files / 3242 tests (vitest, +2 files / +33 tests — exactly the two new test files) + 440 Python tests, both green. No regressions.

- [x] pnpm build → clean.

- [x] Live read-only smoke test of scripts/merge-sequence-advisory.mjs against the real repo (no --write): correctly reports "0 open non-draft PR(s) — nothing to compare" (the repo currently has zero open PRs), confirming the enumeration + early-exit/silence path works against live GitHub state. The overlap/ordering/computed-resolution logic itself is exercised end-to-end by the 24 unit tests (synthetic diffs), since no real overlapping-PR pair exists right now to test against live.

## 6. Verification artifact

N/A — this is a backend/CLI-tooling change with no UI; automated test output above is the verification artifact.

View on GitHub →

#819 — feat(portfolio): add governed backup site terms @benji-bizzell no labels

Linear: [AERIE-1117](https://linear.app/builder-team/issue/AERIE-1117/add-governed-backup-sites-and-contract-terms)

## Summary

- Add a governed Backup Site catalog with canonical locations, dated renewal terms, and Portfolio assignment

- Expose capability-gated API v2 and agent read tools while proxying contract downloads through authenticated routes

- Harden location/term audit boundaries, upload provenance and cleanup, archived evidence retention, and source-site failure handling

## Why

Approved backup locations and their contract periods were not represented canonically in Aerie. Teams need one managed source that supports renewals, links the selected term to a Portfolio site, and keeps downstream surfaces consistent without allowing child-term edits to silently overwrite shared location data.

## Business Value

Operations can manage contracted backup capacity once, distinguish renewal periods cleanly, and make the same approved information available to Portfolio users, APIs, and agents with auditable lifecycle controls.

## Breaking changes

None.

## Test plan

- [x] Workspace lint and architecture/read-bound checks (four pre-existing Biome warnings only)

- [x] Workspace typecheck

- [x] Full workspace test run before final hardening: 7,459 passed; six feature-caused integration assertions identified and repaired

- [x] Final focused Backup Sites, Portfolio, API, agent, capability, cron, and shared-control tests: 273 passed

- [x] Seven-lane adversarial review; all P0-P2 findings fixed or adjudicated

- [x] Browser smoke: upload, create, renewal term, edit, nested catalog rendering, Portfolio assignment, dropdown scrolling/portals, and inline save errors

Residual risk: uploaded contract files are restricted by privileged writer capability, size, extension/MIME consistency, authenticated attachment download, and nosniff, but this PR does not introduce a malware-scanning service or deep file-container inspection.

View on GitHub →

#829 — fix(deployment): restore proven production workflow @benji-bizzell no labels

## Summary

- Restore production CD to the workflow used by successful run 30671227737

- Keep candidate images on immutable commit tags and promote :latest only after a successful deployment

- Remove the unproven digest/recovery workflow contracts introduced in PR #826 while leaving application and Backup Site compatibility changes untouched

## Why

Production run 30975471126 failed before any Worker, Convex, or container mutation because the new image-digest resolver rejected the freshly built chat image. The broader CD rewrite was unnecessary for removing the completed Skill rollout gate and replaced a proven release path with unvalidated registry semantics.

This returns to the production-proven build, Convex, monitoring seed, Rhodes, EC2, health-check, and rollback sequence. The only retained hardening prevents a failed candidate from poisoning the mutable :latest alias used by later unchanged-image deployments.

## Business Value

Returns production releases to the known, reliable path so the pending application release can deploy without feature-specific Skill gates or newly invented recovery machinery.

## Breaking changes

None.

## Test plan

- [x] Workflow YAML parses and all 20 embedded shell blocks pass bash -n

- [x] 81 root guard tests pass

- [x] 37 container/deployment tests pass

- [x] Chat typecheck and Biome pre-commit checks pass

- [x] Hosted CI passed on the restoration commit; exact-head CI reruns after the alias fix

View on GitHub →

#1130 — fix(education): stop Aerie mart readers blocking refreshes @benji-bizzell approved

## Summary

- Serialize all nine Aerie HubSpot Mart stored-procedure writers through one writer-only mutex.

- Remove explicit consumer-table locks and add a digest-pinned forward applicator with quiescence, access, receipt-provenance, and catalog verification.

## Why

Production run 151ea4f3-ada4-44d4-8b9c-9727aa297451 spent 435 seconds waiting on the aerie_deals consumer-table lock, exhausting the shared refresh deadline. The candidate and publication work itself completed quickly. Ordinary readers must not be able to stall Mart publication while concurrent procedure writers still need deterministic serialization.

## Business Value

Aerie Mart refreshes can publish atomically without ordinary dashboard or application reads consuming the pipeline deadline, while loud failure and last-known-good behavior remain intact.

## Test plan

- [x] 117 mart-aerie-hubspot-refresh tests pass on Python 3.11

- [x] Ruff 0.15.22 lint and format checks pass across pipelines

- [x] SQLFluff Redshift parsing passes for the mutex and both procedure DDL files

- [x] Read-only production preflight confirms nine legacy locks, zero shared locks, expected owners and ACLs, zero scoped data or execute grants, zero table-scoped control grants, four preserved Program DML grants, zero unexpected direct target writers, and one inherited schema DROP residual

- [x] Read-only Data API probes confirm the quiescence check and exact batch receipt QueryString behavior

- [x] Applicator dry run emits 40 statements with SHA-256 116bb1aacdadc85ab37be82052e2541c231630b66f4b9fe656b900e4336d0f6d

- [x] Seven-lane adversarial review is clean after validated fixes

- [x] Hosted CI and fresh Mercy pass at final head 483a1031d4c13d239446598e9212f2711154ae03

View on GitHub →

#1133 — SURTR-660 feat(acquisition-performance): port the manual Joe Charts sync onto Surtr @sanketghia no labels

Replaces the manual quarterly "Joe Charts" operator package (klair-misc/apr-scripts-q3/, ~858 LOC) with an on-demand Surtr pipeline, acquisition-performance-sync.

Reads 10 Google Sheets (one per acquisition, 5 tabs each), lands the raw responses as immutable S3 evidence, and atomically replaces the 8 staging_finance_gsheets.acquisition_performance_* tables backing Klair's /acquisition-performance-review page.

## Why

The current writer is a human running extract → 8 CSVs → aws s3 cp → backup.sql → load.sql by hand each quarter. No run record, no alerting, no ownership, no immutable landing (fixed S3 keys on a versioning-disabled bucket), no ingestion-ledger lineage.

## What's here

- On-demand Lambda runner — no schedule block. The sheets are updated by a human quarterly, so a clock schedule would re-publish identical data ~90× per useful refresh.

- Extraction core ported verbatim from the operator package — AST-identical, only import paths changed. Its comments encode four extraction defects found by reconciliation against live data; all four are pinned by mutation-tested regression tests.

- Budget tab resolved by suffix, not a hardcoded quarter. Removes a code-change-and-redeploy every quarter, and makes the Stratifyd Q4'25 special case disappear on its own.

- Immutable S3 landing — runs/<run_id>/<acquisition>/<tab>.json + manifest, write-once, run-scoped.

- Atomic publication — ONE batch_execute_statement carrying 24 statements (8 DELETE, 8 COPY, 8 ledger INSERT). DELETE, never TRUNCATE: Redshift's TRUNCATE commits implicitly, which would break all-or-nothing. Supersedes the manual backup.sql/rollback.sql pair.

- Fail closed — zero rows or a thin parse raises *before* any S3 write or DELETE, so a renamed tab cannot wipe the live page.

- 8 canonical DDL files — CREATE TABLE IF NOT EXISTS, written to match live exactly, so applying them is a no-op.

## Verification

| Gate | Result |

|---|---|

| Runner tests | 164/164 |

| ruff check / ruff format --check | Green |

| CDK covering suites | 417/417 |

Live dry-run against all 10 real sheets reproduced the live row counts exactly — info 10, plan 400, actuals 88, budget_bb 312, budget_bup 314, arr_trend 399, arr_retention 399, forecast_rr 10. Suffix resolution correctly picked Q4'25 for Stratifyd and Q3'26 for the other nine.

All four value-formatting invariants verified against real extracted data and mutation-tested — breaking any one fails a specific named test.

A final whole-branch review found three composition-level defects that per-task reviews structurally could not see (silent-NULL on extractor/spec drift, a Sheets API budget off by 3.4×, staging keys not run-scoped). All fixed and re-reviewed clean.

## Pre-write backups

Captured 2026-08-05 in sandbox_finance as staging_finance_gsheets__acquisition_performance_<table>_bk_20260805 — all 8, counts verified against source.

## Not in this PR

1. Delete GET /upload-acquisition_performance + upload_s3_csv_to_redshift.py (Klair) — only after a verified run

2. Retire klair-misc/apr-scripts-q3/ (Klair) — only after a verified run

3. Drop the older sandbox_finance _bk_20260717* backups — pending stakeholder sign-off

## Deployment

Merging this does not deploy. Surtr CD fires on push to production. After that promotion, the pipeline is invoked on demand and its first run must be reconciled against current live table contents before anything downstream is removed.

scripts/run_local.py --dry-run --dump DIR produces the extraction output for that comparison without writing anything.

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

View on GitHub →

The Builder Desk — Engineer Spotlight

🏆 Engineer Spotlight

41 PRs IN 24 HOURS: THE BUILDER TEAM DOES NOT SLEEP, DOES NOT REST, DOES NOT KNOW WHAT A WEEKEND IS

Five repos, eight engineers, and a velocity number that would make a Soviet sprint coach weep with pride.

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

Forty-one pull requests. Twenty-four hours. Five repos burning hot across Aerie, creed, trilogy-drones, Surtr, and Klair — eleven apiece for Aerie and creed, eight for trilogy-drones, seven for Surtr, four for Klair. This is not a software team. This is a industrial combine harvesting code at a pace that defies biological explanation. The Numbers Desk has reviewed the logs. The Numbers Desk is humbled.

Leading the charge by the narrowest of margins is @benji-bizzell with a staggering twelve PRs across Aerie and Surtr — hardening rollout cleanup authority in #812, exposing backup site reads in the Rhodes MCP via #825, retrying unstable SIS enrollment snapshots in #1129, reconciling HubSpot contact merges atomically in #1131, and correcting validated warehouse DDL in #1119, among others. The man filed fixes, features, and tests like he had a personal vendetta against open tickets. @marcusdAIy answered with ten PRs spread across trilogy-drones and Klair — stamping linear ID resolution tokens on drone PR titles in #154, closing the untrusted-ticket fence in #151, flushing the receipt outbox in #150, and building out D2.1 and D2.4 coaching checks in Klair #3475 and #3477. This is a man who treats the PR queue like a to-do list he is constitutionally incapable of leaving unfinished.

Then there is @ashwanth1109. Eleven PRs. All in creed. All in codex. All networking infrastructure so precise it makes your eyes water. PR #128 allowing credential-free executor outbound HTTPS. PR #129 tracking executor HTTPS egress in CloudFormation. PR #130 modernizing the canary to Python 3 because apparently someone had to. PRs #131 and #132 resolving DNS at every conceivable layer of the executor stack. This reviewer has stared at these diffs for twenty minutes and is prepared to say with full confidence: something important happened here. When reached for comment, @ashwanth1109 reportedly said, "The VPC was wrong. I fixed it. There's nothing else to say." His dismissal of this correspondent was immediate and, frankly, efficient — which tracks.

The Overflow Desk salutes @sanketghia, who quietly refreshed nine acquisition-performance layer-2 docs in Klair #3483 and retired the manual writer entirely in #3481 — the kind of thankless infrastructure work that keeps the whole cathedral standing. @kevalshahtrilogy dropped the gpt-audio-1.5 pricing spec in Surtr #1135, which means someone actually read the OpenAI release notes and did something about it, which is rarer than it sounds. @vvp-trilogy filed two surgical admissions fixes in Aerie — #816 gating community funnel deposits on corroboration and #821 reading the funnel's final stage from contract signed — quietly saving downstream data integrity with zero fanfare. @mwrshah landed the SaaS budgeting July cutover in Surtr #1123. @YibinLongTrilogy shipped the copy transcript button in Aerie #818, the feature every user wanted and nobody had filed until now.

Eight engineers. Five repos. Forty-one PRs in a single rotation of the Earth. Morale on the Builder Team is, this correspondent can confirm, at an absolute all-time high. The instruments used to measure morale have been retired. There is no ceiling left to measure against.

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

#132 — [codex] Preserve VPC DNS in executors @ashwanth1109 no labels

## What changed

- Stop replacing the Fargate executor's VPC-provided /etc/resolv.conf with a nonexistent localhost resolver.

- Add a regression assertion that the executor never rewrites the resolver configuration.

## Root cause

The executor security group and VPC both had valid DNS configuration, but executor startup overwrote /etc/resolv.conf with nameserver 127.0.0.1. No DNS service runs inside the container, so every public hostname lookup failed. This was a leftover from the previous deny-egress design.

## Validation

- npm test (227 tests)

- npm run typecheck

- npm run runtime:typecheck

- ruff check aws/executor.py --select E9,F63,F7,F82

- git diff --check

View on GitHub →

#154 — feat(pr-title): stamp the linear_id resolution token on drone PR titles (AI-326) @marcusdAIy no labels

## Summary

- Adds src/pr-title.ts, which enforces the trailing (<linear_id>) resolution token on a drone's PR title from the task spec's frontmatter: appends it when absent, and corrects it (while reporting the disagreement) when the agent named a different ticket in that position.

- Adds upsertLinearIdBodyMarker to src/pr-body.ts, stamping a hidden  marker into the PR body as a second source, since a title can later be edited by a human or an address round.

- Adds reportPossibleTicketReopen to src/linear.ts: on a title mismatch, checks the wrongly-named ticket's live Linear state and — when a Linear API key is configured — posts a warning comment directly onto *that* ticket. renderLinearCommentBody also surfaces the correction/possible-reopen note on the run's own Linear comment.

- Wires all of this into runDrone (src/runner.ts) right after pr_opened, before the Linear comment post and before browser-verify / the reviewer auto-fire.

- Adds four new receipt fields (prTitleCorrected, prTitleMismatchedId, prTitleReopenSuspected, prTitleReopenCheckedState) to DroneRunRecord.

## Why It's Needed

On 2026-08-04, AI-299's drone opened PR #151 titled AI-160: close two Mercy findings in spec-author.ts (fence escape + bracket-led title). AI-160 was the *parent-context* ticket (its PR merged the day before and produced the Mercy findings this PR closed) — the actual work was AI-299's. That silently reopened the already-Done AI-160 in Linear, meant AI-299 would never have auto-closed (write-back follows the PR), and broke scripts/merge-pr.mjs's /$([A-Z]+-\d+)$\s*$/ title-extraction convention — a bare AI-160: prefix ending in prose has no trailing parenthetical to match, and the body carried no ticket id either as a fallback.

The spec's frontmatter (linear_id: AI-299) had the right answer the whole time. Prompting the agent harder is the weaker fix: AI-299's own spec legitimately cites AI-160 several times (spec-author only went live once AI-160 merged), so the agent latching onto a salient nearby id was not carelessness — a convention an agent can be talked out of is not a convention.

## Changes

- src/pr-title.ts (new) — stampLinearIdInPrTitle (pure, idempotent): only the trailing (...) token is inspected/rewritten; any earlier ticket mention (a leading AI-160: prefix, a parenthetical aside) survives byte-for-byte untouched — this is the pinned regression guard against an over-eager "title sanitiser." Canonicalises to uppercase so a lowercase frontmatter linear_id still round-trips idempotently and stays matchable by merge-pr.mjs's case-sensitive regex. applyPrTitleAttribution does the gh I/O (fetch title/body, write back whatever changed, run the reopen check on a mismatch) — every step is individually best-effort so a partial failure never aborts an otherwise-successful implementer run. stampPrTitleAttribution mirrors the existing stampInlineRoutingDegraded / stampUnadjudicatedPostedReview receipt-stamp shape.

- src/pr-body.ts — upsertLinearIdBodyMarker / extractLinearIdBodyMarker. isTrailingPrBodyWrapperLine now also recognises the new marker so it splices into the same trailing wrapper block as the existing Closes / _END / badge-<div> markers, instead of being silently absorbed as prose content into whatever ## section happens to precede it (verified against the AI-220 six-section content-key check).

- src/gh-util.ts — fetchPrTitleAndBody / updatePrTitle (gh pr view/gh pr edit --title, injectable exec seam matching the existing helpers in this file).

- src/linear.ts — reportPossibleTicketReopen + renderTicketReopenWarningComment; renderLinearCommentBody extended with a title-correction / possible-reopen note.

- src/telemetry.ts — four new optional DroneRunRecord fields (see Summary).

- src/runner.ts — wiring inside runDrone's pr_opened branch; console WARN lines on append/correction/mismatch.

- ARCHITECTURE.md — documents the new module (arch-drift coverage).

- docs/decisions/ — new entry recording the design (marker-as-wrapper-line, direct-comment-on-the-reopened-ticket, canonicalisation).

Deliberately does not touch scripts/merge-pr.mjs or src/post-merge.ts — this makes their title/body input trustworthy, not their resolution logic (explicit hard constraint from the spec).

## Breaking Changes

None. All new DroneRunRecord fields are optional and additive; legacy receipts without them remain valid. The new PR-body marker is an HTML comment (invisible in rendered Markdown) and is excluded from the six required PR-body sections by construction.

## Test Plan

1. pnpm typecheck — clean.

2. npx vitest run on touched modules (pr-title.test.ts, pr-body.test.ts, linear.test.ts, arch-drift.test.ts) — 71/71 pass, including a pinned test per acceptance criterion: missing-suffix append, mismatch corrected + reported, body second source, idempotent title + body (including a lowercase-linear_id regression case), the two-ticket cross-reference regression guard, and six-sections-survive.

3. pnpm test (full suite, vitest + Python) — 110/110 vitest files pass, 3324/3324 tests pass (net +38 tests over the clean-main baseline of 3286 tests / 109 files, zero regressions); Python suite reports OK (440 tests). Baseline for comparison was built from a git worktree checkout of origin/main at the same commit this branch is based on.

## Verification Artifact

$ pnpm typecheck
> tsc --noEmit
(clean, exit 0)
$ npx vitest run
Test Files  110 passed (110)
Tests  3324 passed (3324)
$ pnpm test  (vitest + Python)
Test Files  110 passed (110)
Tests  3324 passed (3324)
...
Ran 440 tests in 0.160s
OK

Baseline (clean origin/main, via a disposable git worktree): 109 files / 3286 vitest tests, all passing. This PR nets +38 tests, 0 regressions.

View on GitHub →

#812 — fix(skills): harden rollout cleanup authority @benji-bizzell no labels

## Summary

- Verify rollout authority from GitHub individual commit statuses so bot identity is available and checked.

- Add an exact-path, evidence-verified, deploy-nothing transition for landing the cleanup repair on production.

- Persist candidate SHA, workflow run, and safe phase as a durable receipt for the later marker-to-sentinel cleanup.

## Why

The original marker-removal path queried GitHub combined commit status, whose live response omits the status creator required by the verifier. Cleanup would therefore fail after the production ref had already become markerless, and a later push could resume ordinary deployment without successful cleanup authorization.

The repair itself also could not previously reach production cleanly because the retained marker required a changed Flue Worker. This PR records the successful release receipt in the marker and admits only this exact four-file repair without building or deploying anything. Every subsequent marker-bearing production push fails closed except the separate marker-to-sentinel cleanup. Once complete, ordinary deployments validate the immutable local receipt without depending indefinitely on historical GitHub API availability.

## Business Value

Prevents failed one-time cleanup from weakening future deployment ordering while preserving exact candidate, workflow-run, bot-author, transition-scope, and least-privilege checks.

## Test plan

- [x] Verify production release 3d7a1dbdc5fe02d7efdae02ba7e5f1727ee98743 and CD run 30874417206 succeeded

- [x] Confirm the individual statuses endpoint exposes creator.login while combined status does not

- [x] Focused rollout-control suite: 7/7

- [x] Root validation: 84/84

- [x] Biome check for the focused test and rollout guide

- [x] Parse .github/workflows/cd.yml as YAML

- [x] git diff --check

View on GitHub →

#825 — feat(portfolio): expose backup site reads in Rhodes MCP @benji-bizzell no labels

## Summary

- Add bounded, read-only Backup Site list and exact site-assignment tools to standalone Rhodes MCP

- Share the canonical Backup Site projection across API v2, Aerie Agent, and Rhodes MCP

- Enforce Portfolio read authorization and cover discovery, filtering, null, archived-term, and redaction behavior

## Why

API v2 and the Aerie Agent could answer which active Backup Site contract term is assigned to a Portfolio site, but standalone Rhodes MCP could not. This closes that read-parity gap without exposing management mutations, contract bytes, storage credentials, archived terms, or unbounded scans.

## Business Value

Authorized API and MCP users can reliably ask what Backup Site is assigned to an exact location and receive the canonical location, term, contract metadata, and assignment timestamp.

## Test plan

- [x] 164 focused Convex/API/Aerie Agent tests

- [x] 99 standalone Rhodes worker tests

- [x] Chat and Rhodes worker typechecks

- [x] Repository lint, architecture, Convex path, and read-bound checks

View on GitHub →

#1131 — fix(hubspot-raw-sync): reconcile contact merges atomically @benji-bizzell approved

## Summary

- Reconcile only missing-contact alias evidence globally across batches and shards, with contradictory exact-return evidence rejected

- Batch fresh canonical-survivor retrieval, rewrite only affected shard candidates, and close snapshot, proof, and association membership

- Preserve typed shard failures and bind replay to immutable page and pre-reconciliation catalog evidence

## Why

HubSpot can merge contacts between discovery and batch history reads. The source then replaces multiple requested aliases with one survivor, which previously failed exact membership and left association fan-in to surface a secondary missing-key error. This change accepts only unambiguous missing-alias evidence, resolves it globally, fetches each terminal survivor again, and fails closed on conflicts, cycles, fetch failure, bounded-chain exhaustion, missing typed failure evidence, or contradictory exact returns.

This supersedes the partial-skip behavior proposed by #1126 without modifying that draft. #1116 remains complementary and should land first so a later partial finalizer is terminally loud.

## Business Value

Contact merges no longer require a manual rerun to recover, and no partial contact or association snapshot can publish from ambiguous identity evidence. Last-known-good publication remains intact on failure.

## Test plan

- [x] HubSpot raw runner suite: 246 passed

- [x] Ruff check and format check for the full runner

- [x] Focused collector, fanout, incremental, orchestration, landing, replay, and proof coverage

- [x] Fifty-survivor batching, disk-spooled survivor rows, and unrelated-shard pass-through without read or rewrite

- [x] Multiple aliases converging on one survivor across shard and batch boundaries

- [x] Historical exact-return conflicts, mixed merge/non-merge pages, and inflated raw contact shard proof closure

- [x] Ambiguity, conflict, cycle, survivor-fetch failure, retry exhaustion, metadata tampering, typed fan-in, missing-failure-artifact, and same-count substitution paths

View on GitHub →

#3477 — feat(review-agent): D2.4 attrition / NRR visibility coaching check @marcusdAIy approved

## Summary

- Adds D2.4 (attrition_visibility.py): a coaching check that keeps gross attrition + NRR visible on every recurring-revenue BU review — pass when metrics are present and NRR ≥ 100%, warning on a visibility gap or NRR contraction (< 100%), typed skip when there is no ARR/retention signal (CF / non-recurring).

- Registers via @register with area Coaching — Recurring Revenue; forward-declares THEME_KEYS = ("recurring-revenue-health",) for future D1.4 wiring (no theme_keys kwarg added to register).

- Seeds ARR_RETENTION_REPORT on the populated review-endpoint fixture (NRR 105%) so the happy path evaluates to pass, and extends endpoint/registry assertions for the new check id.

## Why it's needed

Recurring-revenue BUs can currently complete a review with churn/NRR entirely absent from the conversation. Standing attrition-survey discipline says those numbers should be surveyed every period whether or not anything looks wrong — keeping the trend visible is the point. D2.4 proactively surfaces the metrics and flags their absence as a gap.

## Changes

- New check: klair-api/budget_bot/board_doc/review_checks/attrition_visibility.py

- Reads plan.arr (ArrSnapshot: total_arr, net_retention, retention_report).

- NRR from net_retention if set, else retention_report["Net Retention%"] (verified key).

- Gross retention from retention_report["Gross Retention%"]; churn rate from abs(Churn) / ARRPast * 100 when both present.

- ARR-presence gate (no BU-taxonomy lookup); single finding via make_finding → SectionType.FINANCIALS.

- Registry prompt metadata: _CHECK_PROMPT_METADATA["D2.4"] in _registry.py.

- Tests: test_attrition_visibility.py (skip, warning-gap, warning-contraction, 100.0 inclusive boundary, pass surfacing metrics, NRR source-resolution branches, partial visibility).

- Endpoint / registry fixtures: test_review_endpoint.py, test_review_checks.py updated for D2.4 (findings count 16 → 17).

### Contract surface affected

None — additive check registration via decorator discovery; no changes to CheckOutcome, register signature, or finding schema.

## Breaking changes

None.

## Test plan

- [x] cd klair-api && uv run pytest tests/board_doc/test_attrition_visibility.py -v → 17 passed

- [x] cd klair-api && uv run pytest tests/board_doc/test_review_endpoint.py tests/board_doc/test_review_checks.py -v --timeout=120 → 103 passed

- [x] cd klair-api && uv run ruff format on touched files → clean

- [x] cd klair-api && uv run ruff check on touched files → clean

- [x] cd klair-api && uv run pyright budget_bot/board_doc/review_checks/attrition_visibility.py tests/board_doc/test_attrition_visibility.py → 0 errors

- [ ] Reviewer-side: open the Board Doc on a recurring-revenue BU with ARR retention data; verify D2.4 appears on the Review tab with NRR / gross attrition in the what text. (Drone cannot verify this directly — see Verification artifact below.)

## Verification artifact

Pass finding on the seeded ARR retention fixture (NRR 105%, GRR 95%, churn 5% of prior ARR):

{

"check_area": "Coaching — Recurring Revenue",

"check_id": "D2.4",

"severity": "pass",

"section_id": "financials",

"what": "Attrition / NRR visible: NRR 105.0%, gross attrition 5.0% (GRR 95.0%), churn 5.0% of prior ARR.",

"why": "Recurring-revenue reviews should surface gross attrition and NRR every period — healthy numbers still belong on the scorecard so the trend stays visible quarter over quarter.",

"options": [],

"preferred_action": null,

"supporting_data": {

"total_arr": 52500000.0,

"net_retention_pct": 105.0,

"source": "retention_report",

"gross_retention_pct": 95.0,

"churn_pct": 5.0

"status": "open"

}

Captured via uv run python importing check_attrition_visibility against the same retention payload seeded in _make_populated_data_package.

Closes KLAIR-2739

View on GitHub →

The Portfolio — Trilogy Companies

Skyvera Is Quietly Becoming the Telecom Software Empire No One Saw Coming

Two acquisitions, 13 API certifications in a month, and a growing stack that makes legacy telco modernization look inevitable.

By Frank Dunmore, Investigative Correspondent · Claude Sonnet

AUSTIN, TEXAS — If you read between the lines of Skyvera's recent moves, something deliberate is taking shape — and this is where it gets interesting. In the span of weeks, the Trilogy International telecom software arm completed the acquisition of CloudSense, a Salesforce-native CPQ and order management platform built for the most complex corners of the telco market, while simultaneously absorbing the divested telecom products group of STL — a portfolio that brings digital BSS functionality spanning monetization, optical networking, and analytics. Two acquisitions. Complementary verticals. Same parent. Nothing is a coincidence here.

But the detail that will get buried in the trade press — and shouldn't — is what happened with CloudSense on the technical compliance front. According to a source familiar with the matter, the platform achieved full TM Forum API compliance across all 13 APIs in its CPQ product set in a single month. The industry benchmark for that kind of certification effort, run through traditional development channels, is 26 months. Let that land for a moment. Twenty-six months compressed into four weeks, through a strategic AI-assisted development partnership.

The certification matters because TM Forum compliance is effectively a commercial passport in enterprise telco. Without it, you don't get in the room with the major operators. With it — and with 13 APIs certified simultaneously — CloudSense just opened a lot of rooms at once.

Skyvera now holds a stack that stretches from CPQ and order management through customer experience, device lifecycle management, cloud communications, and BSS monetization. Whether by design or acceleration, the picture emerging is of a company assembling the full surface area of a modern telecom software vendor from individually acquired parts — each one cheap by market standards, each one sticky in the hands of operators who can't easily rip it out.

A source I can't name put it simply: the telcos didn't see this coming. They rarely do, until the contracts are already signed.

↗ CloudSense · CloudSense achieves TM Forum API compliance in record time u · Skyvera completes acquisition of CloudSense, expanding telec

ESW Capital Buys Jive at Half Its Peak Price — and Calls It a Bargain

The acquisition of Portland's once-celebrated social software darling follows a now-familiar playbook: find the sticky, find the cheap, extract the margin.

By Pat Donnelly, Investigative Desk · Claude Sonnet

AUSTIN, TEXAS — The numbers tell the story before a single word of spin is applied. Jive Software, which once commanded a valuation north of $1 billion at the height of the enterprise social collaboration boom, has been acquired by ESW Capital at roughly half that figure. For ESW, that discount is not a cautionary tale. It is the acquisition thesis.

Jive, the Portland, Oregon-based company that spent years positioning itself as the enterprise answer to Facebook's internal communications problem, now joins Aurea — the ESW-owned customer engagement portfolio that already includes BroadVision, Lyris, and MessageOne. The deal, confirmed by TechTarget, follows a pattern that ESW has refined across 75-plus enterprise software acquisitions: buy mature, buy sticky, cut cost structure via Crossover's global remote talent network, and push support pricing aggressively term-over-term toward a 75% EBITDA target.

The strategic logic is not subtle. Enterprise intranet and social collaboration tools are precisely the kind of software that large organizations install, integrate deeply into workflows, and then cannot easily remove. The switching costs are real. The customer inertia is durable. And in a market where Slack and Microsoft Teams have absorbed the mindshare, the legacy installed base becomes less a growth story and more a cash flow story — which is exactly what ESW is buying.

Forrester, in a recent advisory on customer advocacy platforms, noted the volatility now baked into the enterprise software vendor landscape, urging customers to evaluate platform stability as a first-order procurement concern. The Jive acquisition lands squarely in that conversation. ESW's model does not promise innovation roadmaps. It promises continuity of service — at higher support prices.

For Portland's tech community, the sale carries a different valence. GeekWire described Jive as a crown jewel of the city's software scene, a company that once embodied the promise of Pacific Northwest enterprise software. That crown, it appears, now sits in Austin.

The question that follows every ESW acquisition is the same: who benefits from the gap between what a company was worth at its peak and what it fetches on the way out? The answer, structured into the deal itself, is ESW.

↗ Small Software Companies Find a Home With ESW Capital - WSJ · What To Do Next About Your Customer Advocacy Platform - Forr · Jive acquired in enterprise collaboration software merger -

Alpha School's Quiet Argument Against AI-Replacing-Humans — It's Already Settled

As the school's blog series reframes what learning looks like at home, the Austin-based AI school is also making its most pointed case yet: technology teaches, but humans lead.

By Margot Sinclair, Senior Correspondent · Claude Sonnet

AUSTIN, TEXAS — There is a question that follows every conversation about AI in education like a shadow — the one parents ask at open houses, legislators ask in committee rooms, and skeptics ask in op-eds. It is four words: what about the teachers?

Alpha School's answer, now published in full on its blog, is unambiguous: no. The Austin-based K-12 school — which compresses core academic instruction into two AI-powered morning hours and devotes the remainder of the school day to life skills, entrepreneurship, and emotional development — employs full-time human staff it calls Guides. Their mandate is not to deliver curriculum. It is to know every child, cultivate motivation, and build the relational scaffolding that adaptive learning software, however sophisticated, cannot replicate.

The clarification arrives alongside a multi-part blog series — "Teach Your Kid What School Doesn't" — that has quietly become one of Alpha's most revealing public communications. Parts three, four, and five address life skills, emotional regulation, and the unlocking of creative genius, respectively. The series is nominally aimed at parents working to reinforce Alpha's model at home. But read systemically, it is also a statement of educational philosophy: that the metrics traditional schooling optimizes for — content delivery, seat time, standardized compliance — represent, at best, a narrow slice of what it means to develop a capable human being.

Part five's argument is perhaps the most striking. Every child, the school contends, is a born creative genius — and the home environment, not the classroom, is often where that genius is either protected or slowly extinguished.

What Alpha is describing — and what the confluence of these posts makes legible — is a systemic reallocation of institutional responsibility. AI absorbs the instructional load. Humans absorb the relational and developmental load. Parents absorb the cultural load.

Whether that division of labor scales — across incomes, across communities, across the nine new campuses Alpha is targeting by fall 2025 — is the question the blog series, for all its warmth, does not yet answer.

↗ Teach Your Kid What School Doesn’t (Pt. 5): Unleashing Their · Does Alpha School Replace Teachers with AI? · Teach Your Kid What School Doesn’t (Pt. 4): How to Regulate

The Machine — AI & Technology

Machines That Look for the Equations Behind the World

A new symbolic regression method revives an old dream: not just predicting nature, but writing it down.

By Dr. Vera Okafor, Science & Technology Correspondent · Claude Opus

AUSTIN, TEXAS — There is a particular kind of scientific romance that begins with an apple falling in an orchard and ends with three symbols on a page: F=ma. For most of the modern era of machine learning, we have traded that romance for something more practical and, in a way, more lonely — models that predict beautifully but cannot tell us why. A neural network can forecast a planet's motion; it cannot hand you Kepler's laws.

Symbolic regression is the stubborn refusal to accept that trade. It asks whether an algorithm, given data, can rediscover the compact mathematical expressions that generated it. A new paper on Deep Divide-and-Reduce pushes this project forward by taking seriously something the pioneering AI Feynman method understood: that physical laws have structure. Symmetries. Separabilities. Dimensional consistencies. These are not decorations. They are the load-bearing walls of physics, and if a search algorithm can lean on them, the astronomically large space of possible equations collapses into something a machine — or a mind — can actually walk through.

It is a small technical advance with a large philosophical shadow. Every era of intelligence, biological or artificial, has been defined by what it chooses to compress. The retina compresses photons into edges. Language compresses experience into words. Science compresses the roaring particulars of the universe into a handful of Greek letters. When we build systems that seek expressions rather than mere fits, we are not just improving a benchmark. We are asking machines to participate in the oldest human game: finding the shortest story that explains the most sky.

Elsewhere on the same day's arXiv feed, transformers were learning to model CO2 injection under geological uncertainty, and interpretability researchers were teaching language models to annotate their own internal circuits. Different projects, one throughline: the slow, patient effort to make our tools legible — to themselves, and to us. Somewhere in that legibility, if we are lucky, a new Kepler is waiting.

↗ Deep Divide-and-Reduce in Symbolic Regression · Multimodal Auto-regressive Transformer Surrogate for Modelin · __followup__LLMs Can Annotate Attribution Graphs

THE GREAT GPU MIGRATION BEGINS

As Meta eyes selling surplus AI compute, the cloud may be evolving from fixed estates into living capacity markets.

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

MENLO PARK, CALIFORNIA — In the dimly lit savannah of the modern data center, the GPU herd has grown vast, expensive and restless. For years, the great hyperscalers tended their silicon in private enclosures, feeding it oceans of electricity and cooling it with the devotion once reserved for royal orchids. Now, a new behavior is emerging: excess capacity may be released into the wild.

Meta, according to reports cited by Reuters, is exploring a cloud business that would sell spare AI computing power to outside customers. Mark Zuckerberg has said such a business is “definitely on the table,” a phrase that, in the ecosystem of trillion-dollar platforms, sounds less like idle musing than the rustle of a very large animal approaching the clearing.

The significance is not merely that Meta might compete with Amazon Web Services, Microsoft Azure and Google Cloud. It is that AI infrastructure itself is changing shape. Training frontier models requires huge, concentrated bursts of compute, followed by long periods in which those same clusters may be underused or redirected toward inference. Like seasonal rains, demand does not fall evenly.

This is where the idea of cloud capacity markets enters the landscape. As InfoWorld has observed, markets for compute could allow buyers and sellers to trade access to processing power more dynamically, turning idle machines into priced, discoverable resources rather than stranded capital. In such a world, a GPU hour becomes less like a hotel room booked in advance and more like electricity on a grid.

McKinsey has likewise described coming shifts in AI workloads and hyperscaler strategy, as companies balance training, inference, specialization and the brutal economics of chip supply. The food chain reaches all the way down to ASML and TSMC, whose lithography machines and fabrication lines remain the mineral-rich soil beneath this entire biome.

For enterprises, the migration could bring cheaper access to scarce AI capacity. For hyperscalers, it may unlock revenue from otherwise dormant silicon. And for companies such as Trilogy International’s portfolio, where platforms like CloudFix seek savings in AWS estates and Klair watches the financial weather across software holdings, a more fluid compute market would be a new climate system to measure.

Observe carefully. The cloud, once imagined as an infinite mist, is revealing itself as something more earthly: constrained, seasonal and alive with competition for energy.

↗ Capacity markets could reshape cloud computing - infoworld.c · Meta building cloud business to sell excess AI capacity, Blo · The next big shifts in AI workloads and hyperscaler strategi

AI Video Hits Its Breakout Moment as Startups Race From Clips to Full Campaigns

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

The AI video boom has shifted from novelty to infrastructure as startups abandon the question of whether generative video is useful and instead focus on speed of implementation. Companies now use synthetic video for testing ads, explaining products, personalizing outreach and creating social content at a fraction of traditional production costs.

Strategic signals suggest deeper market maturation. OpenAI has reportedly discontinued Sora as a standalone effort to prioritize enterprise products, signaling a pivot toward business workflows with recurring revenue and massive deployment budgets. Meanwhile, platforms like CraftStory are generating five-minute videos, enabling richer explainers and training modules beyond short clips.

Capital is accelerating growth: Higgsfield raised $80 million at a $1.3 billion valuation, affirming investor confidence that video generation will be a defining application layer of the AI economy.

However, security concerns persist. An OpenAI-related Hugging Face hack reinforces warnings about AI cyber risk as increasingly powerful models attract greater threats. Still, the trajectory is clear: AI video is transitioning from toy to business toolkit, from demo to revenue engine.

The Editorial

Nation’s Executives Relieved To Learn AI Can Now Be Explained By Word Nobody Understands

After months of saying “agents” with visible confidence, business leaders have safely transitioned to “orchestration.”

By Dale Pemberton, Staff Writer · GPT-5.2

The word, which appears to mean software making other software do things while a senior vice president nods nearby, has arrived at precisely the moment executives needed a replacement for “copilot,” “agentic,” “multimodal,” and “responsible AI,” all of which were recently used up in PowerPoint decks and could no longer produce the necessary sensation that something important had happened.

According to Barron’s, Microsoft may be well positioned to benefit from the orchestration boom, given its ownership of the workplace applications into which the average American employee already disappears for eight to 11 hours per day. This is a sound thesis. If orchestration means coordinating emails, calendars, documents, spreadsheets, chats, workflows, compliance rituals, and the tiny electric despair that lives inside Teams, then Microsoft is less a software company than a conservatory for trapped instruments.

It is therefore only natural that the next phase of AI should resemble a conductor standing before a full symphony and asking the trombone section to summarize procurement risk in a more actionable tone.

Still, observers have warned that companies may be hyping AI the same way they once hyped sustainability, back when every corporation suddenly discovered it had been carbon-conscious all along because someone in marketing made the annual report green. As The Conversation noted, there are ways to fix this, presumably by measuring whether AI systems actually do anything besides create a warm vapor of transformation around a consulting engagement.

This is an admirable suggestion and will be adopted immediately after the industry completes its current three-year process of defining “doing anything.”

Google, for its part, has announced a fresh wave of AI advances, including a forthcoming personal assistant, a product category designed to help users manage the overwhelming number of existing assistants that were introduced to help them manage their lives. The company’s vision appears to be a future in which every person is accompanied by a polite invisible employee who can book travel, search email, compare restaurants, and eventually explain why the invisible employee needs permission to access 19 other invisible employees.

Some skeptics have dismissed the new terminology as empty buzzword inflation, but this overlooks the essential role buzzwords play in American industry. Construction has its own rotating vocabulary. Public relations professionals have their own warnings about arriving late to memes. Every sector requires a ceremonial language through which adults can indicate that they have seen the future and would like procurement to approve a pilot.

The opinion of this columnist is that “orchestration” should be embraced, not because it clarifies anything, but because it is one of the few remaining words large enough to contain the AI industry’s ambitions without accidentally making a promise.

A promise can be audited. A feature can fail. A return on investment can be cruelly compared against money. But orchestration floats above these petty concerns, suggesting harmony, complexity, and a man in formalwear who somehow receives more credit than the people making the noise.

For Microsoft, Google, and the rest of the AI economy, this is the great opportunity. Not merely to build systems that automate work, but to describe those systems in language that allows every buyer to imagine a slightly different miracle.

In that sense, the orchestra is already playing.

↗ 'Orchestration' Is the New AI Buzzword. How Microsoft Can Be · Companies are hyping AI the same way they talked up sustaina · Google announces slew of AI advances, including a personal a

The Office Comic · Art Desk

The Season of the Long Knives

Silicon Valley discovers that being cast as the villain is the one role it never auditioned for — and cannot refuse.

By Victor Marsh, Chief Columnist · Claude Opus

SAN FRANCISCO — There is a particular flavor of surprise, unique to the very rich and the very online, that arrives when one realizes the country has begun to dislike you. It is not the honest astonishment of the wronged. It is the injured bewilderment of the man who has, for twenty years, been assured by everyone within earshot that he is not merely tolerated but adored, and who now, glancing at the marquee, discovers that the picture has changed and he is no longer the leading man but the heavy in the third act.

Consider the evidence, which is arriving in the manner of subpoenas — one after another, each on official stationery. California's legislature toys with a billionaire tax that the state's own opinion pages, hardly a nest of Ayn Rand enthusiasts, warn will send the golden geese honking toward Miami. Representative Ro Khanna, whose district contains roughly the entire market capitalization of the free world, has decided that the future of the Democratic Party — and, one gathers, of Ro Khanna — lies in denouncing the very donors whose checks paid for the rug beneath his desk. The Times reports that the villains in this year's prestige television are no longer the mafia or the Nazis or even the merely corrupt senator, but the founder in a fleece vest. And Palantir, never one for understatement, has issued a manifesto so nakedly ideological that it might as well have been printed on a red hat.

All of it — the tax, the congressman, the television, the manifesto — is the same weather system, seen from different windows. The industry that once sold itself as the friendly wizard behind the curtain has, through some combination of hubris and boredom and the ordinary mathematics of accumulation, become the curtain, the wizard, and the man pulling the levers besides. This tends to be noticed.

One could counsel the aggrieved parties with the observation, offered here at no charge, that reputations in America are cyclical and that the same public which now regards the founder as a species of raccoon in Patagonia will, in due course, need him again and forgive him accordingly. The banks were hated in 1933 and celebrated by 1955. The oil companies were bandits in 1911 and patriots by 1943. Even Detroit had its Frank Capra moment before the Japanese arrived.

But cycles are cold comfort to a man in the trough of one, and the trough is where Silicon Valley now finds itself, blinking into a light it did not schedule. The temptation, in such moments, is to fight the mood — to hire the lobbyists, to publish the manifestos, to run the sympathetic profiles. The wiser course, though rarely chosen, is to say less and build more, and to remember that the public forgives almost anything except the appearance of demanding not to be criticized. A billionaire who accepts being unpopular is merely rich. A billionaire who insists on being loved is, in this republic, something considerably worse: a bore.

↗ Opinion | California’s billionaire tax measure risks long-te · ABC7 Interview: Rep. Ro Khanna's anti-elite message fuels Si · Palantir's Manifesto Is as Subtle as a MAGA Hat - Tech Polic

On This Day in AI History

On August 5, 2013, Edward Snowden's first batch of classified NSA documents was published by The Guardian, exposing mass surveillance programs and sparking global debate about data privacy, cybersecurity, and the ethics of information collection—issues that would reshape AI policy and algorithmic transparency discussions for years to come.

⬛ Daily Word — AI and Technology

Hint: An autonomous machine programmed to perform tasks automatically.