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

Backlog, requested 10 August 2026

19 min read

Everything asked for, written down so none of it is lost. Priority order. Each item has a done when that can be checked, not a vibe.

Three of these need structured data where the repo currently has only prose. That is the real work: once a thing is data, listing it, sorting it, filtering it and giving it its own page are all cheap. While it is a paragraph inside a file, none of them are.

#ItemNeeds new dataStatus
2Watch & Listen hubnew watch.jsonldone
3Timeline: multi-tag, filter, detailmigrate to timeline.jsonldone
4Home page: clickable counts, today's lead, taglinedone
5Section index pages (/daily/, /weekly/)done
6More charts, and actual imagespartly done
7Follow every material returned lead before accepting zerointernal candidate notesdone, 2026-09-27
8Reserve discovery for world news and independent sourcesdone, 2026-09-27
9Challenge a provisional zero with a bounded second sweepinternal review notesparked; declined by Reckon, 2026-09-27
10Give material returned leads a disposition on every editioninternal candidate notesparked; awaiting Reckon's decision, 2026-09-28

2. Watch & Listen hub

Podcasts and videos currently exist only as prose inside each daily brief. They vanish the day after publication.

  • One page, every recommendation ever made, newest first.
  • Searchable across title, show, person, and the note explaining why it matters.
  • Tag filters, multi-select. Picking two tags narrows, it does not widen.
  • People filter. Guests and hosts are the main axis: this is how you actually look for an episode. "That Karpathy one."
  • Each item keeps its original date, its issue link, and the note from the brief it appeared in.
  • The daily's Watch & Listen section is generated from the same data, so adding an episode once puts it in both places and it can never drift.

Done when: a new episode is added to data/watch.jsonl and appears in the daily brief and on the hub with no other edit.

3. Timeline

Tags are there and they work. They just do not do anything yet.

  • Multiple tags per entry. Most events are two things at once, a funding round in an lab is MONEY and MODEL.
  • Filter by tag, and by more than one at a time.
  • Clickable entries. Each event opens to the fuller account: what happened, why it was logged, which issues covered it.

Done when: selecting MONEY + INFRA shows only entries carrying both, and every row opens to a detail view.

4. Home page

  • The counts are links. Briefs → the daily archive, Weeklies → weekly, Issues → monthly. If it shows a number, it goes somewhere.
  • Today is obvious. The newest brief leads the page as the front-page story with its date, not as the first row of an undifferentiated list.
  • Tagline → "Daily Intelligence on the AI Landscape".

5. Section index pages

Implied by making the counts clickable: /daily/, /weekly/, /monthly/ need to exist as archives in their own right, each listing everything of that kind with author and date.

6. Images and charts

Asked directly: why are there so few? Honest answer, because it changes what to build.

  1. Press photography is not usable. Product shots, headshots and conference photos are licensed. A private site does not make republishing them fine, and "borrow it, it's just for me" is how you end up with a habit that breaks the day the site is shared.
  2. Generated stock imagery would be decoration. A glowing brain over an article about classifier catch rates adds nothing and costs credibility.
  3. Charts were built but only used in two reports. tools/charts.py has four types and most of the archive is still prose wrapped around numbers that want to be seen. This is the real gap and it is entirely self-inflicted.

So the plan is: every image on this site is generated from something true.

Shipped:

  • The evidence strip (tools/cover.py), every issue draws its own source-tier mix, one tick per citation. It earned its place immediately: the full report is visibly paler than the daily, because 41 of its 62 citations are tier C. That is a real fact about the issue that no sentence in it states.
  • Charts on the executive summary and full report. One was cut in review: a donut of the adoption funnel would have rendered "23% report ROI" as 29%, because the slices are a share of the 80% who embedded rather than of all enterprises. A chart that restates a figure incorrectly is worse than no chart.

Still to do:

  • Charts across the Top 20 and the rest of Vol. 1, the standing rule is 2–4 per long report, and only two of ten parts have any.
  • Video thumbnails on Watch & Listen, fetched once when an episode is added and committed, with generated fallback art where there is none.
  • Entity pages: data/entities.json now exists and the timeline already resolves against it, so /entities/<slug>.html is mostly wiring.

9. Challenge a provisional zero with a bounded second sweep

Declined by Reckon, September 27th 2026; stays parked. A zero-only challenge misses thin or weak nonzero editions, while item 8 improves breadth on every run. Reckon also declined a standing exception allowing 32 queries against the soft 20–30 budget for a case item 8 already covers. Do not implement the challenge sweep or widen the query budget.

Reopen only if item 8's trial shows the four reserved queries measurably starving technical discovery within the existing allowance; item 9 would then be the cheaper substitute for Reckon to reconsider. The September 27 trial did not demonstrate that condition: 24 queries left six slots, with no technical query denied. Coverage remained incomplete, and this early-day trial does not establish zero opportunity cost on a busy afternoon. The original proposal and cost reasoning below are retained as history, not authorization.

Parked, September 26th 2026. A cheaper-frequency alternative to item 8's every-run reservation, with the limitation that a weak nonzero edition would not trigger it. Use the existing session; no second or full rerun is required. Item 7 remains useful because another sweep alone can return leads that are again ignored.

  • Proposed instruction. After docs/ROUTINES.md:309–311, the paragraph beginning “FIVE STORIES ON A QUIET AFTERNOON,” add: “Before accepting a zero-story PM, perform one challenge sweep: up to four queries and four page opens using named unresolved leads and an untested world-news source or beat. Drop exact-date phrases and check publisher clocks directly. Reconcile its candidates with the first sweep and record why each qualifies, fails, or remains unresolved. If material uncertainty remains, report incomplete search for review. This does not relax the freshness or repeat rules.” If accepted, put the procedure in runbook §1 and make this prompt a pointer, consistent with the routines' existing authority rule. No prompt is edited by this proposal.
  • Cost per PM run. No extra searches on a nonzero run. On a provisional zero, up to four extra queries and four opens, estimated 2–5 minutes, 4,000–12,000 returned-text and 500–1,500 review-note tokens. On the incident's 28-query baseline, this permits 32 queries: an explicit exception to the soft ~20–30 budget requiring Reckon's approval. If fraction z of PM runs trigger it, average added search cost is at most 4 × z queries and 4 × z opens per PM run. Stop after one challenge; unresolved evidence goes to review rather than an unbounded search loop.
  • What it would have done on September 26th. It would have triggered and required named Copilot/dialogue follow-up plus a different world-news route before treating the four rejected near-misses as support for zero. It could have found the Pope's post-AM coverage, but discovery is not guaranteed. Even without that discovery it would have exposed the uninvestigated leads for review instead of supporting the “first genuine zero-story PM” claim. It would not automatically turn the parallel lane's three stories into three eligible post-AM developments.

Done when: after approval, a provisional-zero trial performs exactly one bounded challenge and records candidate dispositions and actual extra cost; a nonzero trial incurs no challenge searches. An unresolved challenge remains an incomplete search and cannot be reported as a verified absence of news.

Cost basis and decision for Reckon. These are planning allowances, not measured or dollar quotes. Queries count query strings, not batched ; opens are additional work. Returned-text and note tokens are not total billed tokens: repeated context, caching and reasoning affect billing. The incident log reports 108,137 tokens and the session took 177.585 seconds; neither supplies a marginal dollar rate. Price a trial using actual billed input/output/ usage and the estate's contracted rates, plus any search fees. Do not add items 7–9's estimates mechanically when they check the same lead. Recommend item 7 first because it addresses leads demonstrably returned, then have Reckon choose routine breadth (8), a zero-only challenge (9), or both with shared work counted once. Whether older news missed by AM belongs in PM, and whether a genuine zero gets a public notice, remain editorial decisions. This backlog authorizes neither.


10. Give material returned leads a disposition on every edition

Accepted as a standing procedure by Reckon, October 2nd 2026 15:00, on the trial's own evidence: eight runs, eight compliances across AM, PM and monthly slots; the five artefacts written after the shape was pinned carried 47/17/49/ 81/53 leads as flat lists with a discovery URL and a nonblank reason on every one. Cost 12-50 queries and 12-53 opens per run, two displacements declared out of eight, no edition halted, delayed or thinned. Accepted with one amendment: the pinned artefact's field names are now fixed too (counts.queries, coverage.complete as a boolean, reserve.used), because across those five runs the same three facts appeared under three spellings and an audit script still needed a person. Still no budget exception, backfill or publication-policy change. Trial history: approved as a bounded trial 2026-09-29 07:00, live in RUNBOOK §1 (Sweep the beats) from that afternoon's edition through 2026-10-02. The reasoning for approving rather than declining: the case against is real but every one of its costs is measurable, and the only way to measure them is to run it. Parked note, September 28th 2026, kept for the record: The right change to trial is a disposition gate on every edition, including a rich nonzero edition, rather than only on a zero. Recommend the bounded trial below, not a new discovery sweep or an allowance increase. Eien retains the publication decision.

What would have caught this omission. The AM already received the pause lead: S:64, page L14 says “OpenAI paused model training”; L85–88 supplies a headline and NBC source pointer. S:57, L226–229 separately supplies the US government-site lead. Neither was followed or given a rejection ground. The incident audit corrects the brief's tentative premise that the pause never returned. The original edition had 25 stories, so item 7's zero-only condition did not apply. Reconciliation would have surfaced the pause for a decision without another search; it would not by itself verify the story or require publication.

Proposed instruction, for runbook §1 only after approval: “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 the query before spending the last discovery slot. This is a gate on the completeness claim, not automatic permission to publish unverified material or halt all other publication.

Measured , not a planning estimate. The replay record compares the two affected retained discovery responses with the original edition and prior headings. It accounts for all 54 numbered dailyai.report candidates and ten HeadsUpAI updates, including repeats, old news and unresolved non-pause leads. The informed reviewer knew the incident; this was not a blind trial or a complete replay of all 49 queries.

Measured stageExtra queries / opensInput tokens (cached subset)Output tokensElapsed
Retained-lead reconciliation; pause flagged unresolved0 / 0277,656 (250,880)3,042102.935 s
Targeted follow-up and disposition1 / 2 attempts267,994 (252,800)1,67060.939 s
Whole bounded replay1 / 2 attempts545,650 (503,680)4,712163.874 s

The single query was OpenAI paused model training rogue agents government websites, with no wording or later-merge URL supplied. Its results led to the primary DNS report, whose successful open substantiated the pause and containment failure. The KQED/AP open failed and is charged. The primary page's September 25 update date predates AM; present-day retrieval does not prove historical index state. The government-site details remain separately unresolved. This does not demonstrate discovery of the absent payload reconstruction or Amodei memo.

Budget and cost boundaries. Original AM: 49 query strings, 25 opens and 21 clicks. Adding the measured path gives 50 queries, 27 opens and 21 clicks, within the AM's existing upper query target; no original technical query need be displaced in this instance. A run already at its upper target must reserve/reallocate a slot or escalate without another query. This replay's 41,970 uncached input plus 4,712 output equals 46,682 tokens on the original log's uncached-plus-output convention; total input plus output is 550,362. Reasoning's 456 tokens are already in output. These are session-counter differences including repeated context, input reloads, commentary and evidence notes, not isolated production marginal billing. Preparation and final proposal/audit work are outside the interval. No dollar charge, marginal plan rate or search fee is available. A full every-edition gate's recurring token cost remains unmeasured; these numbers must not be sold as its upper bound.

Case for. The failure is demonstrated follow-through, not hypothetical retrieval weakness: the pause was explicitly in the AM's input. One shared disposition check closes the gap that allowed 25 good stories to conceal a major omission. It preserves valid freshness rejections and uses existing discovery work. In this replay the next single query reached primary evidence without exceeding 50. Even if retrieval had failed, the unresolved pause would have reached review instead of disappearing. Unlike a zero-only sweep, the check operates on the edition that actually failed.

Case against. The two streams alone produce many unresolved candidates; one follow-up slot does not resolve them. An every-edition gate could create routine escalation noise, delay publication, and divert scarce verification from technical reporting. “Material” is still the same model's judgment: it can miss a lead again or rationalize an exclusion. The replay is informed, limited to the affected streams and uses a later search index, so it proves neither reliable future compliance nor affordable full-edition coverage. The measured token cost is substantial even with caching. A strict gate requiring every candidate to be verified would stall this replay, while a soft incomplete-search note might become meaningless boilerplate. Rejecting the proposal is reasonable if that trade buys too little editorial value.

Decision boundary. This evidence does not establish “no affordable change would have caught it”: a bounded check surfaced it inside the query allowance. It also does not establish an affordable standing full-verification gate. Approve only a trial if Reckon wants evidence on that recurring trade. Item 9 stays declined. Its revival condition is measurable technical starvation from item 8's four queries. Neither the incident nor this replay identifies a denied technical query attributable to that reservation; being near 50 alone is not the condition. A zero-only substitute would not have triggered here anyway.

Done when, after approval: a prospective busy nonzero edition and a zero candidate record dispositions across all returned material leads, retain evidenced rejections, expose unresolved omissions, and measure complete gate input/cache/output usage and actual query/open counts. Record any displaced technical query and escalation outcome. Reckon then accepts or declines the standing procedure; no budget exception, backfill or publication-policy change follows from this parked proposal.


Standing decisions

  • Server-render everything, filter on the client. Every item is in the HTML; JavaScript only hides what does not match. Search and filters die gracefully with JS off, and there is no index to keep in sync.
  • No dependencies. Still true. No , no framework, no external request at .
  • One filter component, shared by watch and timeline. Pages needing the same interaction is a component, not three implementations.

Done

Completed items retain their original proposals and acceptance criteria below, with the completion date, what shipped and linked evidence.

7. Follow every material returned lead before accepting zero

Done, September 27th 2026. Reckon approved; shipped in 137660a. RUNBOOK §1 now requires an evidenced disposition for every material returned lead before accepting zero, with unresolved leads blocking that conclusion. The acceptance replay records Copilot's rejection for PM freshness and the dialogue's unresolved status after targeted follow-up, preserves old/unverifiable rejections, and confirms the Pope was absent from the retained responses. No story quota or freshness/repeat exception was introduced.

Original proposal, retained: the parked status and pending instructions below describe September 26, before the approval and implementation above.

Parked, September 26th 2026. Proposed search-procedure change for Reckon; not approved or applied. Items 7–9 address the PM investigation in LEDGER.md, not a requirement to publish more stories or a public zero-story notice. The zero artefact is a separate handoff item.

The failure was partly follow-through: the PM received a three-part Copilot relaunch lead and an AI-dialogue lead, then investigated Warp, PAL Robotics, LMSYS and Higgsfield instead. It issued 28 query strings in eight search-bearing calls. Eighteen queries contained exact date phrases. Neither a forced timeout nor a product-only source whitelist was found. AM's twenty stories had nineteen primary company/project/paper sources and one Reuters report; both runs used HeadsUpAI for discovery. Primary sourcing alone did not ensure broad discovery.

  • Proposed instruction. At docs/RUNBOOK.md:246–249, after the paragraph beginning “Twenty stories is a floor,” add: “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.” Leave the current lines alone pending Reckon's decision.
  • Cost per daily run. Planning estimate: 1–3 minutes and 500–1,500 tokens of candidate notes; zero extra queries if all material leads were already checked. Reserve up to four queries and four page opens within the current search allowance for unresolved leads; budget another 2–5 minutes and roughly 4,000–12,000 returned-text tokens when that reserve is used. Exhausting the reserve leaves an incomplete-search finding, not an automatic zero. This reallocates discovery effort and may reduce exploration of minor launches.
  • What it would have done on September 26th. The retained session output at S:28 says “Microsoft launched a revamped Copilot app” with Home, Code and Autopilot; S:52 says “they agreed to open an AI dialogue channel.” S is the PM session identified and hashed in the investigation ledger. Those two leads would have required investigation or an explicit unresolved disposition before accepting zero. It would not by itself have found the Pope, absent from the retained results. Copilot was September 25 news and the summit's CNBC report preceded AM; finding them does not authorize PM publication.

Done when: after approval, a replay of the retained PM responses produces explicit, evidenced dispositions for Copilot and the dialogue before any zero conclusion; the same review can still reject old or unverifiable stories. It must not claim the Pope was present in those responses.

8. Reserve discovery for world news and independent sources

Done, September 27th 2026. Reckon approved; shipped in a9585ca. RUNBOOK §1 now reserves four queries for diplomacy and public institutions through independent established news routes, separate from Chinese model discovery, inside the existing allowance. The live daily discovery trial records the four strings, Reuters/CNBC/AP routes, publisher-clock checks, and a distinct China/open-weights sweep. It measured 24 queries, 18 open attempts and 279.444 seconds for research and review, including empty results; actual session usage was 1,303,023 input tokens (1,194,112 cached) and 8,377 output tokens. These are session counters, not a marginal invoice. Technical starvation was not demonstrated; unresolved leads and access failures leave coverage incomplete. No story quota or query-budget exception was introduced.

Original proposal, retained: the parked status and pending instructions below describe September 26, before the approval and implementation above.

Parked, September 26th 2026. The runbook already requires policy coverage. The gap is that generic regulation queries and China model searches did not cover diplomacy or public institutions adequately, and one technical roundup dominated follow-up. No fixed source list excluding world news was found.

  • Proposed instruction. Replace docs/RUNBOOK.md:253, currently “Policy and regulation (US federal/state, EU, China),” with: “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.” Example source routes are Reuters, CNBC, Axios and AP/NBC, not an exclusive whitelist. These are proposed words only; the source line remains unchanged.
  • Cost per daily run. Four queries and up to four opens, replacing four existing queries inside AM's ~35–50 / PM's ~20–30 allowance. Estimated 2–5 minutes and 4,000–12,000 returned-text tokens for the reserved work; net query increase is zero, while extra opens depend on what it replaces. The opportunity cost is four fewer technical/product discovery queries. Verification of additional candidates consumes the remaining allowance; the four-query reservation is not four guaranteed verified stories.
  • What it would have done on September 26th. A diplomacy sweep such as US China summit AI dialogue would have pursued the visible summit lead; a public-institutions sweep such as artificial intelligence world leaders speech could have surfaced the Pope, followed by a named Paris/Pope query. This is a plausible retrieval improvement, not a historical replay result. The ledger records CNBC's summit publication at 03:58 PDT and Reuters' Pope coverage at 07:00 PDT, both before PM at 14:00. The latter was after AM; the former still required a freshness disposition. Historical index state is unknown. This option offers no specific guarantee of finding Copilot.

Done when: after approval, a daily run shows four reserved query strings, two independent news-source routes and publisher-clock checks for candidates; China diplomacy is distinguishable from the China/open-weights sweep. A trial records actual query, token and elapsed-time costs, including empty results, without introducing a story quota.