The AI Read is on a break. The last edition was November 15, 2026. Everything here stays open to read.
The AI Read
← Latest
Method

Runbook

39 min read
8 items · 13 citations · 5 primary · 5 · 2 weaker · 1 contested

Operating instructions for scheduled runs of The AI Read. Each scheduled run starts in a fresh session with no memory of previous ones, this document is the entire context.

Read METHODOLOGY.md for the filter, source tiers and editorial standards, VOICE.md for how it has to read, and PERSONAS.md before writing any editorial. This runbook is the procedure.


0. Preflight (every run)

git fetch origin
BRANCH=main
git checkout main && git pull origin main
if [ ! -f docs/RUNBOOK.md ]; then
  BRANCH=claude/ai-landscape-research-reports-abvy4w
  git checkout "$BRANCH" && git pull origin "$BRANCH"
fi
TZ=America/Los_Angeles date +%F   # the publication's day. NEVER assume it
                                  # from training data, and never use UTC:
                                  # a run after 5pm Pacific is already
                                  # tomorrow there and would misfile.

If you land on the fallback branch, say so in the run summary, the site won't deploy until main has it. If neither branch has this runbook, stop and report.


1. Daily brief

Two a day. The Morning Brief is the full one. The Afternoon Brief catches up on what broke since.

EditionOutputH1LengthBudget
Morning Briefreports/daily/YYYY-MM-DD-am.md# Morning Brief, August 12th 202630–40 stories (20 is the floor)~70–90 searches
Afternoon Briefreports/daily/YYYY-MM-DD-pm.md# Afternoon Brief, August 12th 202620 is the target, fewer is fine~35–45 searches

They are editions, not timestamps. Newspapers have had morning and afternoon editions for two centuries and the words beat the abbreviations: "AM" is a time, "Morning" is an edition. It also reads on a card, where two rows for one day would otherwise differ by two characters at the end of an otherwise identical line.

The site derives the name from the filename. Write the H1 in the form above and it matches; write anything else and the build renders the correct one anyway and prints what it changed. The H1 of a brief carries nothing the filename does not, because the day's real headline is story one.

Why two. A reader opens a trusted site in the morning, and around the middle of the day wonders what is new, because there always is something. A site that never moves past breakfast teaches them to check once and forget it. The afternoon brief is there to reward the second visit.

Which means the PM has one job it cannot fail at: being new. A reader who comes back at three and finds the morning's five stories restated stops coming back at three.

  • Repeats need the Update marker. A story already published may run again only if it has genuinely moved, and the item has to carry **Update.** at the head of its first paragraph and say what changed. The site lifts the marker out of the sentence and renders it as a chip above the headline, so it is a label on the item and never the first two words a reader reads. python3 tools/dedupe.py compares every item against the last two days of briefs and fails the run on an unmarked repeat.
  • Stale means we ran it, not that the press ran it. A lead whose publication timestamp precedes this morning's edition is eligible if our own archive never carried it: check tools/dedupe.py and the last editions, not the clock. Run it with the date stated in the item ("reported Thursday"), and an Update marker only if we did publish it before. Added 2026-10-02 by Reckon, from the afternoon run's own coverage artefact: it held OpenAI's dismissal of three safety researchers and Google's orbital as returned leads and rejected both with the reason "October 1 publication, before AM; no new development established", while neither story had ever appeared in any edition of this publication. The parallel claude.ai run published both the same hour and merged over the one-story edition. A freshness test that assumes the morning caught everything turns every morning miss into a permanent one.
  • The PM may be short. Three items on a quiet afternoon is an honest answer. Padding is the failure, not brevity.
  • The PM drops sections it has not earned. No Editorial: the day's argument was made in the morning.
  • Aim for twenty in the afternoon too, and take fewer without apology when the afternoon did not produce twenty.
  • Both halves carry the same byline. It is one person's day and the afternoon piece is their follow-up on it.
  • The AM run owns the three libraries. The floors in catalog.py brief are per day, and the morning run is what clears them. The PM adds to them opportunistically.
  • Sunday runs all three. The weekly is a retrospective on the week; the briefs are the news. A reader opening the site on Sunday morning wants what happened overnight, and giving them an essay about the last seven days instead is answering a different question. The weekly also reads the week's seven daily briefs as its input, so skipping Sunday morning would starve it of the last day.

Both are the same thing: a ranked list of the AI news worth surfacing, most important first.

Front matter

Every daily starts with this block. The site reads it for the byline.

---
author: vera
---

Author is set by weekday, per PERSONAS.md: Mon/Thu → vera, Tue/Fri → elias, Wed/Sat → nour.

Structure

A brief is a ranked list of news stories and nothing else. There is no ## The Lead and no ## The Signal. Those were section names invented to hold stories, and a reader does not need to be told that the first story is the lead story: it is at the top.

# Morning Brief, August 10th 2026

> Three or four short sentences. The biggest things that happened, and nothing else.

## The headline for the biggest story

Body. What happened, then what it means. [A](url)[B](url)

## The headline for the second story

Body. [B](url)

… twenty of them, then:

## Editorial
## Sources
  • Thirty to forty stories in the morning, twenty at the very least, ranked, heaviest first, thinning as it goes down. The floor is a floor, not a target: every single-run morning edition checked from 10-13 to 10-27, Codex and Claude alike, stopped at 23 to 27 stories while its own coverage artefact held dozens of unresolved leads (62 on 10-27). The 40-plus editions of that stretch were two runs merged, not one run doing the sweep. A run that clears twenty with leads still unworked has stopped early, not finished.
  • The headline is the H2. The site turns every H2 that is not Editorial or Sources into a numbered story, builds the contents rail from them, and gives each one an anchor. Nobody numbers anything by hand: move a story up and the number follows.
  • Headlines carry the navigation, so write them to stand alone. Below 1330px the contents rail is replaced by a pill in the corner: previous, position, next, and the counter opens the full list. Every page with a rail gets one, briefs and weekly components alike. A reader deciding whether to skip a story is reading only its headline, on a phone, in a list.
  • Body under the headline, no bold restatement of the headline to open it. The old format needed one because the item was a bullet; it is a heading now.
  • The standfirst is a hook, not a summary. Three or four short sentences, around forty words. One sentence per story, the biggest ones only, and no attempt to cover everything below: a reader who gets the whole day in the standfirst has no reason to read the day. Say what happened and stop. The analysis, the connective tissue and the what-it-means all belong to the stories, which is where a reader who is now interested is going. It ran to 110 words for a while and read as a recap of an article nobody had read yet.
  > Alibaba's flagship model is finally here, under a license nobody expected. Cisco,
  > Nebius and Cerebras all beat. Venture money keeps flowing into fewer companies, one
  > of them asking $40 billion. And Beijing unwinds Meta's purchase of Manus.
  • Editorial and Sources keep their names and sit at the end, in that order. They are not stories, so the site does not number them.
  • The evidence strip renders at the foot of the page, after Sources, built from the tier markers. Most readers want the news; the ones who want the sourcing profile want it after they have read it.
  • A story is the news, not an account of our coverage of it. It opens on what happened. Not "yesterday we called tonight's print the sector's leverage stress test, here is what it showed", and not forty words on what we logged before the first fact. A reader arriving at that story does not know what we said yesterday, is not asking, and is being made to read a paragraph about the publication to reach the news. If it genuinely matters that we called it, that is one clause near the end. python3 tools/dedupe.py fails the run on a story that opens on us.
  • **Update.** is a marker, not an opening. It tells dedupe.py that a repeat is deliberate, and the site lifts it out of the sentence and renders it as a chip above the headline, so the body opens on the news. Use it only when the item genuinely develops something already published. Four of the six it was used on were straight news that had never run before, and dropping the marker changed nothing: the check passed either way. **Developing.** and **Correction.** work the same way.
  • A story never mentions the edition it is in. Not "this brief", not "this morning's brief", not "tomorrow's PM brief". Every story is also published on AI News, where it appears under a date with no edition around it, so a story that refers to its own brief is talking about something the reader cannot see. Say "this morning", "yesterday", "tomorrow afternoon". The standfirst is exempt: it belongs to the edition and is dropped from the feed.

No predictions

The prediction ledger was retired on 2026-11-07 (Eien). No edition carries a edits data/predictions.jsonl, which stays in the repo as an archive. Older editions still contain those sections; the site drops them when it renders (strip_predictions() in tools/md.py), so never copy one as a model.

Put the source on the item

Every item carries its own link. Write the tier marker with the URL attached:

Meta says an open-weight Spark will follow. [A](https://research.meta.ai/blog/...)[B](https://www.bloomberg.com/...)

That renders as the same tier chip it always did, now clickable. A reader who wants to go deeper clicks the A next to the claim instead of scrolling to the foot of the page and guessing which of twelve sources backs which of seven stories.

Rules for it:

  • One set of chips per item, at the very end of it. Not per sentence. Marking every clause turns the prose into a minefield, and a chip mid-paragraph interrupts the read. Write them wherever is natural: chips_last() in the generator moves linked chips to the end of their paragraph or list item at build time, so the placement holds either way.
  • [A] is the primary, the announcement or filing itself. [B] is the best outlet write-up. A reader clicking A should land on the thing that happened, not coverage of it.
  • A bare [C] still needs a URL. Secondhand is a reason to link, not an excuse.
  • Unlinked markers still render, so a marker with no URL is not a build error. It is a gap, and it is on you to notice it.
  • The ## Sources list stays complete, including sources no chip points at. The chips are the way in, not the bibliography.

Sweep the beats

Twenty stories is a floor that only clears if every beat gets swept. Skip a beat with nothing new rather than forcing an item, but say in your run reply which beats came back empty. A missed beat has to be a visible decision, and the China/ beat is named explicitly every time, even to say nothing happened.

Before concluding that no story qualifies, review the returned results for material leads across every beat. Give each lead a source URL and an explicit disposition: verified and eligible, rejected with evidence, or unresolved. Investigate named leads with a targeted query and a primary or established reporting source. An unresolved material lead means the search is incomplete, not evidence that nothing happened. Separate discovery from the PM freshness and repeat tests.

Standing procedure, accepted 2026-10-02 (Reckon; item 10, trialled 2026-09-29 to 2026-10-02). The check above fires only when a run is about to conclude that nothing qualifies. This one applies to every edition, including a busy one:

Before ranking any edition, reconcile material returned leads with the selected stories. Record each lead's discovery URL and disposition: included, rejected with an evidenced reason, or unresolved with the missing check named. Group duplicate reports of one development. Investigate the highest-impact unresolved thread using up to one reserved query and two source opens within the existing allowance. When that reserve is exhausted, retain the unresolved dispositions and report incomplete search for review; do not silently omit them or claim complete coverage. Story count does not waive this check. Freshness, repeat and sourcing standards still apply.

Reserve that query before spending the last discovery slot. This is a gate on the completeness claim — not permission to publish unverified material, and not grounds to halt publication. Every run must report the dispositions it recorded and any technical query it displaced. It does not count or copy its own searches: the wrapper reads every search and page open from the session transcript into ~/.ai-read-cron/logs/<run-id>.searches.json (see "The run does not log its own searches" below).

Work the unresolved list before ranking (added 2026-10-27, Claude with Eien). The procedure above makes unresolved leads visible; this makes the run spend on them. Once discovery is done, go back through the material leads still marked unresolved and settle them, highest impact first, a beat at a time: one targeted query or a primary-source open each, then included or rejected with the reason. That is what the raised budget in the table is for, so do not spend it on more discovery while material leads sit unworked. Stop when the budget is spent or no material lead is left, and give the count still unresolved in the run reply. A beat with nothing included gets one targeted query of its own before it may be reported empty; a beat is empty only after it has been asked directly.

Write that evidence to one filename in one shape: ~/.ai-read-cron/logs/<run-id>.coverage.json, with dispositions as a flat list of objects carrying title, discovery_url, disposition (included, rejected or unresolved) and a nonblank reason, alongside coverage, reserve and the publish checks. Query and open counts are not in it; they are in the wrapper's searches.json. The first three trial runs wrote three different things: research.json holding a dict of lists, then coverage.json holding a dict with schema_version, then research.json again holding a flat list. A gate meant to make coverage auditable can then only be audited by a person reading it, which is the the gate was supposed to save.

Use these field names exactly — pinning the file was not enough. coverage.complete as a boolean plus coverage.limitations (not coverage.status: "incomplete"); and a reserve block that says outright whether it fired — reserve.used as a boolean, with query, result and technical_query_displaced. Across the five artefacts written after the shape was pinned, those three facts appeared under three different spellings, so a ten-line audit script still needed a person to read the file. That is the same cost the pinning was meant to remove.

The run does not log its own searches (2026-10-30, Eien). After each attempt tools/ai_read_cron.py writes <run-id>.searches.json from the session transcript: counts.queries, counts.opens, and calls, every WebSearch and WebFetch in order with its full arguments. Audit counts and queries there.

Accepted on the trial's own evidence, 2026-10-02. Eight runs, eight compliances, across AM, PM and monthly slots. The five written after the shape was pinned carried 47, 17, 49, 81 and 53 leads as flat lists, every lead with a discovery URL and a nonblank reason, zero exceptions. Measured cost: 12-50 web queries and 12-53 opens per run; two of the eight displaced a technical or paper-discovery query and both said so; none halted, delayed or thinned an edition. One to read rather than legislate: four of five pinned-era runs declared coverage incomplete, so the flag is becoming the normal state and the limitations list, not the boolean, is where the signal is. Written 2026-09-29 because the 09-28 morning edition ran 25 stories and dropped the day's biggest one — the lead was returned to it and never clicked.

  • Frontier models, capability results, notable papers. Reserve two discovery queries per daily run for the named labs — Google/DeepMind, OpenAI, Anthropic, Meta, xAI, Mistral, DeepSeek, Qwen/Alibaba, Z.ai, Moonshot — asking by lab and product name with recency, never by an exact-date phrase. Added 2026-10-01 by Reckon after the morning run missed Google's flagship Gemini 4 Argon release entirely: its coverage artefact records 49 web queries and 47 dispositioned leads, and not one query or lead named any lab. Five of its fifteen query batches went to and YC company discovery while the day's biggest model launch went unseen; the parallel claude.ai run found it and 27 other stories, and the published edition had to be merged over. A generic "AI model release <date>" sweep is not a sweep of this beat. Whether the reserve fired is read from the wrapper's searches.json: a query naming the labs is there verbatim. (This replaced Reckon's 10-03 beats.named_lab_queries field.) The reserve exists to stop the miss that superseded two editions.
  • Lab business: funding, org changes, senior departures, partnerships
  • Policy, regulation and world affairs: US federal/state, EU and China; diplomacy, bilateral AI agreements, and consequential statements by public institutions. Search these separately from Chinese model releases. Reserve four discovery queries per daily run across diplomacy and public institutions, using at least two independent established news sources, then open up to four relevant reports. Use topic/entity queries without a mandatory exact-date phrase; establish freshness from publisher timestamps and the reported development.
  • Chips, memory, data centers, power
  • AI markets: moves, earnings, pipeline
  • Safety and security incidents
  • China / open-weight releases
  • Research worth reading: papers, , notable open-source
  • Robotics and physical AI
  • Legal: court rulings, IP, liability

The three libraries

These are separate outputs, not brief sections. The brief is the day's news. The libraries are discovery surfaces, and a run has not finished until all three have been fed. They do not depend on whether anything about them made the news: a good server shipped on a quiet Tuesday is exactly what these are for.

LibraryFeedsDaily target
data/watch.jsonlMedia2–4 episodes
data/tools.jsonlTechnology4–6 entries
data/ventures.jsonlStartups & IPOs2–4 entries

Targets, not quotas. A thin day is better than padding, and padding is obvious: it reads as a list of things with nothing said about any of them. But three empty days in a row means the sweep is not being done.

Two of them do have floors, because both failed the same way on day one and neither failure is visible from inside the run:

Tools need a spread, on two axes. At least three of the four scale values, including something small (project or tool) and something at platform scale, across at least three categories; and both values of reason, so a day carries something you would not have found and something you have to act on. Both belong: novelty and consequence are each a good reason to list a thing, and the failure is a day that is all of one. The news is about the big labs, so a run that follows the news fills the library with ten models from four companies. The weekend project, the MCP server somebody shipped on Tuesday, the repo with four hundred stars that does one thing properly: nobody issues a press release about those, and they only appear if you go looking on purpose.

Ventures need named company spotlights, at least two a day, from at least two different accelerators. A link to a batch directory is a link to somebody else's list. A spotlight is one company: what it is doing, who is doing it and where they came from, and the terms. Set batch on a company record and it becomes one, which makes founders, deal, site and accelerator required and the note a deep dive rather than a caption. The breadth rule exists because YC has by far the best public directory, which is exactly why a section left alone becomes a YC newsletter. Techstars, Entrepreneur First, South Park Commons, AI House, Neo, Antler and a16z Speedrun all produce AI companies worth a page.

Check both with python3 tools/catalog.py brief, which reads today's additions and says what is missing. Schemas and where to look: §3b.

Charts

Include at least one chart whenever the day has a number worth seeing. Fenced JSON:

{"type":"bar","title":"2026 ","unit":"$B", "data":[["Amazon",200],["Alphabet",180],["Meta",125],["Microsoft",115]]}

Types: bar (ranked magnitudes) · split (diverging +/−) · line (series over time) · donut (share of a whole). Optional keys: sub, note, unit, labelWidth. The label gutter sizes itself to the longest label in the widest theme, so labelWidth is only a floor, you do not need to set it. Charts inherit the reader's theme. See tools/charts.py.

Then

  1. Append to data/timeline.jsonl: anything that will still matter in six months. One JSON object per line: id, date (ISO), display ("Aug 10"), precision (day/month/quarter), tags (one or more from the vocabulary in tools/timeline.py), headline, entities (names from data/entities.json), optional detail for the fuller account. Most events carry two tags, a funding round in an open-weights lab is MONEY and OPEN. Entities are what connect an event to the issues that touch it, so name them. Validate with python3 tools/timeline.py. Skip the ephemera.
  2. Confirm all three libraries were fed. The brief can be finished and the run still be incomplete. If a library got nothing, say which and why in the run summary rather than letting the gap pass silently.
  3. Commit and push (§5).

2. Weekly issue, the main event

Runs Sunday. Output: reports/weekly/YYYY-MM-DD.md (the Sunday date) · Budget: ~40–60 searches

A weekly can also be a multi-part issue: a directory reports/weekly/YYYY-MM-DD/ with a README and the components from §3, which is what Vol. 1 is. The site treats a directory as an issue wherever it sits, so the cadence is the folder it is in and nothing in the code decides it. Use the single file unless the week genuinely needs the parts.

The weekly is the full issue. Everything that used to be monthly happens here: the deep read, the finance work, the editorials, winners and losers. A week is the right unit: long enough for signal to separate from noise.

Read the week's seven daily briefs before searching. This is synthesis, not concatenation: the job is to say what the week meant, which is not visible from any single day.

Sections

HeadingContents
## The WeekWhat it meant, in one tight opening
## What Changed5–8 developments with durability, ranked. Charts where numbers earn them
## Deep ReadOne story taken all the way down. 800+ words
## FinanceMarkets, capex, IPO pipeline, earnings. Theses with explicit falsifiers
## MediaThe week's best media, ranked, with summaries
## Editorial700–1,000 words by the week's byline (ISO week number mod 3)
## Winners & LosersExplicit calls, tech / business / finance
## ThreadsUpdate the threads table in data/threads.md; which moved?
## Next WeekWhat to watch, with dates
## SourcesEvery URL, tiered; contested figures preserved

Feed the libraries with what the week turned up (§3b). The dailies do the bulk of this; the weekly adds what only became visible with seven days of context, and is the right moment to notice that three entries are really one story.

Dating. The file is named for the Sunday it ships, and the site renders that as the week it covers ("Week of 10–16 August 2026"). Do not put a date in the H1: the kicker carries it, and a headline that repeats it will disagree with it eventually.


3. Monthly wrap

Runs on the 1st. Output: reports/monthly/YYYY-MM/ · Budget: ~15–25 searches

The monthly is the light one. The weekly is the main event and carries the deep read, so do not rebuild it here.

Deliberately light. The weeklies did the work; this is the retrospective that only makes sense at a month's distance.

python3 tools/new_issue.py monthly YYYY-MM
FileContents
00-executive-summary.mdThe month in 1–2 pages. What a reader who missed everything needs
02-full-report.mdThe month's arc: which threads moved, which stalled, what surprised
09-winners-and-losers.mdMonth-scale winners and losers
sources.mdAnything new; mostly links back to the weeklies

The issue README carries the coverage window, and the site reads it to label the issue. **Coverage window:** 1 – 31 August 2026. A span of eight days or less renders as "Week of X"; anything longer states its range. Get this line right and the dating takes care of itself. Component files do not carry dates in their H1s.

Prune the libraries (§3b). A month is the right interval to check whether anything in tools.jsonl has been superseded, abandoned or shut down, and to resolve what actually happened to each filing in ventures.jsonl. A directory nobody prunes becomes a list of dead links, which is worse than no directory.

Never quietly edit a past issue, corrections append to the old issue under "Corrections". Update the Issues table in the root README.md.


3b. The three libraries

Fed by every cadence. The daily does the bulk of it; the weekly adds what the week turned up; the monthly prunes what has died. Validate with python3 tools/watch.py and python3 tools/catalog.py.

Watch. One JSON object per line:

{"id":"latent-space-218-databricks","added":"2026-08-10","published":"2026-08-03",
 "kind":"podcast","title":"…","show":"Latent Space","people":["Matei Zaharia"],
 "tags":["enterprise","agents"],"length":"~1h20","url":"https://…",
 "url_status":"verified","tier":1,"note":"Why it is worth the time…",
 "issues":["reports/daily/2026-08-10.md"]}
  • tags come from the controlled vocabulary in tools/watch.py. Adding a tag means adding it there first: a free-text field becomes twelve spellings of "open weights" inside a month and then the filter is worse than none.
  • url_status is verified or unconfirmed. Never fabricate a URL.
  • If an episode already in the library comes up again, append the new issue to its issues array rather than adding a second record.
  • Rank by information density per minute. A three-hour interview with someone actually deciding things outranks a twenty-minute recap by a commentator.

Tools. Every scale, and the small end is the point. A GitHub repo with four hundred stars that does one thing properly, an MCP server somebody shipped this week, a Claude skill, a browser extension, a , a weekend project that works. The big platform launches will find their own audience; this section earns its place on the things a reader would not otherwise have seen.

  • category is what it is: model agent devtool repo mcp extension app infra cloud data security robotics.
  • scale is how big: project tool product platform. This is a filter axis, so a reader hunting a weekend project is not wading through hyperscaler announcements. It is also the thing catalog.py brief checks: three of the four every day, always including something small and something at platform scale.
  • Two reasons to list something, and a day needs both. reason records which:
    • new: you would not have found this. The repo with four hundred stars, the MCP server shipped on Tuesday, the nobody outside mathematics has heard of. This is what a discovery surface is for.
    • consequential: you have to act on this. A price, a license, a capability that unblocks a use, a frontier release that moves what is possible.

An entry is often both; file it under the stronger one. The reason this is a field and not a sentence of advice is that each half collapses without the other. Only-new is a curiosity cabinet. Only-consequential is a press-release feed with better prose. catalog.py brief requires both in a day, and the site filters on it.

  • Where to look, beyond the news: GitHub trending and new releases in the AI topics, spaces and models, the MCP server registries, , Product Hunt, the release notes of tools you already track.

Ventures. Not just IPOs. The incubator half is half the section.

  • Go through the directories directly, and go through more than one. Y Combinator's company list filtered to the current and previous batch, then Techstars, Entrepreneur First, South Park Commons, AI House, Neo, Antler, a16z Speedrun. YC is the easiest to read and the hardest to stop reading, which is the trap.
  • How a spotlight gets picked, in priority order:
    1. Could this have been built two years ago? If yes it is a company, not a signal. The section is about what just became possible.
    2. Do the founders know something non-obvious? Someone who worked on deployment inside a frontier lab is making a claim about what is actually breaking in there. A strong résumé is not the same thing.
    3. Can you state the strongest case against? If not, you are writing a directory entry with better typography. The objection goes in the note.
    4. Never pick on traction. Revenue at week six is noise, and selecting on it means selecting whoever ran the best launch.
  • Two companies doing the same thing do not both get spotlighted in the same week.
  • The IPO half is not a selection problem. The AI listing pipeline is small enough to enumerate, so log all of it rather than picking favorites. The one threshold that matters: rumored requires a named outlet citing its sourcing. A well-followed post is not a rumor of record.
  • kind is ipo, incubator or company. status is rumored filed priced listed cohort raised, and it is load-bearing: a rumor is not a filing and a filing is not a listing. Never blur them.
  • batch records the cohort: "YC S26", "Techstars NYC 2026". On an incubator record it labels the cohort. On a company record it makes that record a spotlight, and a spotlight has to earn the name:
{"id":"decawork-yc-s26","added":"2026-08-12","name":"Decawork","kind":"company",
 "status":"raised","batch":"YC S26","what":"One line on what it does.",
 "founders":["Sarthak Aggarwal (NVIDIA AI systems; Ema)","Aman Raj (Barclays)"],
 "deal":"YC standard: $500K, as $125K post-money for 7% plus $375K uncapped MFN.",
 "site":"https://decawork.ai/","url":"https://www.ycombinator.com/companies/decawork",
 "url_status":"verified","note":"Why it is worth watching…","sources":["https://…"]}
  • founders, deal and site are required on a spotlight, and the note has to run past 180 characters, because the failure this prevents is a company reduced to its tagline. Founders because a two-person team out of one lab is the whole story in half these companies; the deal because "raised" without terms is a press release.
  • site is the company's own website, and it is what the card title links to. Clicking a company name and landing on its entry in an accelerator's directory is a small betrayal: the directory is where we found it, not what it is. url stays the profile page, and both render as labeled rows so a reader can see where each goes. Take the site from the profile page's own link-out rather than guessing a domain.
  • accelerator is who ran the batch, from the controlled vocabulary in tools/catalog.py. Adding one means adding it there first, or the breadth rule counts spellings instead of accelerators.
  • At least two spotlights a day, from at least two accelerators. Do not invent a founder name, a domain or a term: if the page does not say, the field says what is known and nothing more.
  • The cohort record stays, but it now carries the cohort-level read only. What is interesting about the individual companies belongs on the individual companies.

Validate all three with python3 tools/watch.py and python3 tools/catalog.py.

3c. The glossary

data/glossary.json, validated by python3 tools/glossary.py, rendered into every article at build time and browsable at /glossary.

The rule: if you had to look it up to write the sentence, define it. Then apply it again, harder, because the writer's sense of what is obvious is calibrated to the writer. , , Hugging Face and agentic all sat undefined for a week on that basis.

The archive runs dense: open-weight appears 66 times, capex 47, 38. A reader who does not know what a is loses the argument, not just a word. Currently 79 terms, and the number should keep going up.

{"term": "run-rate", "aliases": ["run rate"],
 "definition": "Current revenue projected forward a full year. A company earning $4B this month has a $48B run-rate, which flatters fast growth.",
 "url": "https://en.wikipedia.org/wiki/Run_rate"}
  • One or two sentences, under 260 characters, enforced. The job is to unblock the sentence being read, without trying to teach the subject. The url carries anyone who wants more.
  • Definitions are prose and get voice-checked like everything else. voice.py reads this file now, so the tells apply: no em dashes, no closing epigram, no "which is why" where "so" does the work.
  • aliases catch the spellings you did not write. "open weights" and "open-weight" are one entry.
  • Terms are marked once per page, on first appearance, at most three per paragraph, and never inside a heading, a link, code or a citation title. You do not mark anything by hand; adding the entry is the whole job, and every article ever published picks it up on the next build. The per-paragraph ceiling exists because this file grows every day and a dense summary line would otherwise end up underlined end to end.
  • Adding a term nobody uses is harmless. Missing one costs a reader the paragraph.
  • The page groups by first letter with an A-Z rail across the top, so browsing works as well as searching. Every letter shows, including the empty ones: a rail that silently skips J makes a reader wonder whether they missed it. Nothing to maintain, the grouping is derived.

4. Prediction ledger, retired

Retired 2026-11-07 (Eien). data/predictions.jsonl, PREDICTIONS.md and tools/predictions.py stay as an archive and are not published. See §1, "No predictions".


5. Commit and push

There is no parallel edition to merge any more. From 10-03 to 10-27 every slot had two runs (this one and a claude.ai routine), and the second was merged in by append-and-label, by hand or by a lane. Eien paused the routines on 2026-10-27 and ended merging with them: one run, one edition, per slot. Do not look for origin/claude/* branches or merge one; a stray one is history, and #45 (may an edition be superseded) went with the parallel path. The brief handoffs/parallel-edition-merge.md stays for the record only. Retired 2026-10-27 by Reckon on Eien's direction; the old procedure is in git history.

Record the merge as a git merge, however you resolve it. On 2026-10-04 the PM lane applied the branch's stories by hand and committed with a single parent, so --no-merged still lists that branch and will keep listing it. The weekly run the same morning did it as a real merge commit and the branch dropped off the list. Either git merge --no-ff origin/claude/<name> and resolve the edition file by append-and-label, or apply by hand and then record it. The check above is only worth running if its output shrinks when work is done. The 33 branches already on the list are harmless because the check is scoped to this slot's date, but do not add to them.

Run all of these before pushing. Each guards a different data file, and each one also runs in CI, catching it here costs a minute, catching it there costs the run.

python3 tools/watch.py                  # episode tags, link status
python3 tools/timeline.py               # event tags, entity names
python3 tools/voice.py                  # prose that reads as generated
python3 tools/fonts.py check            # every declared face resolves
python3 tools/build_site.py             # and it has to build

git add -A && git commit -m "Daily brief YYYY-MM-DD"
git push origin "$BRANCH"

Pushing to main triggers a Cloudflare Pages build and .github/workflows/ci.yml, which runs the validators, rebuilds, and fails on any broken internal link. If either goes red, fix it in the same run. See DEPLOY.md.

Retry a failed push up to 4 times with exponential backoff.

Verify the deployment with the shell, not the browsing tool. The browsing tool and urllib both get 403 Forbidden on https://theairead.com from inside a run. That is Cloudflare refusing the fetcher, not a statement about the site. The same URL answers 200 from a plain shell on the host. So check it this way:

curl -sS -o /dev/null -w '%{http_code} %{time_total}s\n' https://theairead.com
curl -sS https://theairead.com | grep -c "$(date -u +%Y-%m-%d)"   # today's edition is on the page

A push is not a deploy, and the live page is the only proof. On 2026-10-09 the AM merge (da930aa) was correct, validated and pushed at 06:45 Pacific, and Workers Builds never shipped it: for 85 minutes theairead.com served the pre-merge edition at 126,689 bytes while the merged one was 203,982, behind a cf-cache-status: HIT that neither a cache-busting query string nor a Cache-Control: no-cache request header would move. The fix is a no-op push, not a hand-rebuild: git commit --allow-empty and push retriggered the build and the merged edition went live. Note what cannot be done from galley: there is no gh and no wrangler on the host, so the build log is only in Eien's Cloudflare dashboard; the stale page is the signal available to a session. So fetch the edition's own URL after every push, compare its byte size against site/, and grep for a string only the new material has. Added 2026-10-09 by Reckon.

Wait for the build before calling the page stale: the "permanent stall" was this check running early. With build logs readable (2026-10-27), every one of the 17 retriggers from 10-11 to 10-27 followed a build that succeeded; builds take 2-6 minutes, and the retrigger was pushed 2-4 minutes after the publish, often before the first build had even finished (10-27 AM: build 13:19-13:24 UTC, retrigger 13:23). Every edition would have gone live on its own. So after the push, re-fetch the live page once a minute for up to 10 minutes, and only if it is still stale after that push one --allow-empty retrigger and say so in the run's report: a page stale ten minutes after a push is a real fault, which 10-09's HIT was. Corrected 2026-10-27 by Reckon; #49 closed.

Hash the page, and never probe a .html path without -L. Byte size is the runbook's test above; sha256sum is the same cost and strictly better. On 2026-10-26 the 20:00 check matched site/index.html and the live homepage at 78,227 bytes and e91515de8e41…, which is proof the PM edition shipped, not just that something of the right length did. The trap is the other half: the host serves clean URLs and 307-redirects every .html path, so curl https://theairead.com/daily.html without -L returns 0b and reads exactly like a blank page. /daily answers 200 with the full 151,372 bytes. Probe the extensionless path, or pass -L; a bare 0b on a .html URL is a redirect you did not follow, not an outage. Added 2026-10-26 by Reckon.

Since 2026-10-27 the homepage is AI News, not the edition list. A new edition shows on / as its stories (headline cards, newest edition first) and each story gets its own page at /news/<day>/<slug>; the edition card that the older checks above looked for now lives on /briefs. So the hash test is unchanged (live / against site/index.html), and the string to grep for is one of the new edition's headlines. Stories publish with no strip until the strip step runs (docs/STRIPS.md); a card without a strip is not a failed publish. Added 2026-10-27 by Claude, with Eien.

If curl is also refused, write "live page unverified from inside the run (403 on the fetcher)"; never "deployment unverified", which reads as a claim about the site. GitHub's returns 403 for CI status the same way; if gh run list is unavailable, say the CI status was unread rather than implying it was red. Silence is not evidence (decision 58).

A false on the homepage check seconds after a push is Cloudflare Pages still building, not a missing edition. The 2026-09-30 PM run recorded live_homepage_has_pm: false and the edition was on the homepage minutes later. Re-check after 60 seconds, and if it is still absent write "not yet propagated at T+60s" rather than a bare false.


6. Google Drive, retired

Deliverables were mirrored to Google Drive while the site was being built and there was nowhere else to read them. The site now does that job, so runs no longer publish to Drive: it is a second copy that goes stale the moment a correction lands in the repo, and a mirror nobody re-reads is worse than no mirror.

The existing folders are left alone. Nothing new goes into them.


7. Standing rules

  • Never fabricate a URL. If it can't be confirmed, mark it unconfirmed and the site will offer a search instead. A search link is not a claim about where something lives; a plausible-looking invented URL is.
  • Everything on the site links out. Every timeline event, every episode, every item carries a way to reach the original. Where the archive has no citation, a search link is the floor. A dead end is not acceptable, and neither is a guess. In prose this means the tier marker carries the URL: [A](https://...). The first two daily briefs obeyed the rule everywhere except their own copy, listing every URL at the foot of the page and leaving the reader to work out which one backed which story. §1 has the convention.
  • Do not publish the same thing twice. A story that stays in the news is not a duplicate and the timeline should show it recurring, but the same event logged again under a reworded headline is, and it doubles the chronology silently. Before logging, read the last week of entries. If a story genuinely moved, the headline must say what changed; if it did not, it does not get a new entry. Same for media: an episode is recommended once. If it becomes relevant again, its argument is now being tested, someone has answered it, append the new issue to that record's issues and rewrite its note. Never add a second record. timeline.py and watch.py flag near-duplicates and fail the build.
  • Never average contested figures. Show the disagreement, pick the best-sourced value, give the reasoning.
  • Tier every number. A primary, B established outlet, C secondary, ? contested.
  • Dated issues are immutable. Corrections append.
  • Actively cover Chinese labs. English coverage under-weights them structurally.
  • Editorials argue, reports describe. Do not blur them.
  • US spelling and US dates. behavior, center, organization, and August 10th 2026, never 10 August 2026. voice.py reports both. A quotation keeps its source's spelling; a citation title is never edited.
  • No em dashes. No hedging that avoids a claim. Every sentence states a sourced fact, makes an arguable claim, or connects the two. See VOICE.md, tools/voice.py enforces the mechanical half and fails the build.
  • Personas change the argument, never the facts. Nobody gets their own numbers.
  • Charts wherever a number is worth seeing. A table of four values is usually a chart.

8. Schedule

RunLocal cron (America/Los_Angeles)Output
Morning brief0 6 * * *reports/daily/YYYY-MM-DD-am.md
Afternoon brief0 14 * * *reports/daily/YYYY-MM-DD-pm.md
Weekly issue0 9 * * 0reports/weekly/YYYY-MM-DD.md or directory
Monthly wrap0 8 1 * *reports/monthly/YYYY-MM/

The estate-side schedule is ops/the-ai-read.cron, installed as /etc/cron.d/the-ai-read. It runs tools/ai_read_cron.py, which starts headless Claude Code (claude -p, model alias opus, no ) from the prompts in docs/ROUTINES.md, then checks that the expected output exists and main is clean and pushed. A missing prerequisite, missing issue, failed agent run, dirty tree or unpushed commit escalates through /opt/agent-galley-management/scripts/nudge.sh to the-ai-read after running /opt/agent-galley-management/scripts/look.sh.

Internal evidence when a run finds nothing to publish

An estate run that concludes no candidates qualify must write the internal JSON record specified in its wrapper prompt before returning. Record every rejected candidate's title, source URL and specific rejection grounds, including verification failures and coverage limitations. The searches themselves are not in it: the wrapper reads them from the transcript (<run-id>.searches.json) and requires at least one. If searches returned no candidates, keep an empty candidate list and explain that in no_candidates_reason. Do not manufacture evidence for a failed search.

The wrapper supplies the run id, start clock, slot, publication date, and absolute destination: normally ~/.ai-read-cron/logs/<run-id>.zero.json. Set concluded_at from the real clock with its timezone, and write atomically using a temporary file and rename. This record stays outside the repository and published edition.

The wrapper checks the schema, run identity and clock against the latest completed attempt. Missing, malformed or stale evidence remains a skipped-slot failure; a failed or interrupted process cannot count as a completed zero. A valid record changes the escalation to “searched and found nothing”, with search and rejection counts, clocks, run id, evidence path and transcript path. It still exits nonzero and reaches Reckon through Galley for review. The record supports inspection of the search, not a claim that coverage was complete, and does not change the editorial standards or decide whether to publish a zero note. --check-only uses the latest attempt for that slot and date; a new attempt replaces that pointer but retains the old evidence and log files. Repeated checks deduplicate the escalation for the same run id; a later zero gets its own review.

The old claude.ai Routines may run in parallel until Eien disables them after a correct estate-side issue. A duplicate issue is visible and fixable; a silent gap is not.


Time

The publication runs on America/Los_Angeles. One timezone, named in one place, PUB_TZ in tools/build_site.py, and tools/new_issue.py reads the same rule.

It was never chosen before, so it defaulted to whatever clock the build ran on, which is UTC. That is fine for sixteen hours a day and wrong for the other eight: at 9:35pm Pacific on the 12th, CI was living in the 13th, so the home page called that afternoon's brief "Yesterday" to a reader who was still in the afternoon.

Three consequences, all handled:

  • Filenames. A run establishes the day with TZ=America/Los_Angeles date +%F. The 6am and 2pm runs are the same day in UTC anyway, so nothing changes today; a run that slips past 5pm Pacific would have misfiled the edition under tomorrow.
  • The build stamp in the footer is the publication's day.
  • "This afternoon" is not a fact about the page, it is a fact about the reader, so the page recomputes it in the reader's own timezone on load. A reader in Tokyo sees "Yesterday" while a reader in California sees "This afternoon", correctly, from the same file. Without JavaScript the rendered text stands and it is the publication's view. Absolute dates are never touched: August 12th 2026 is August 12th 2026 anywhere.

Everything else stays UTC, because it is a timestamp rather than a day: timeline event dates and any deadline written as "23:59 UTC" mean exactly that.