Methodology

The Projection publishes three news feeds — Mental Health, Frontier AI, and Global Capital — each a running, curated read on a different slice of the same story: AI colliding with mental-health care, the frontier AI race that’s producing it, and the capital markets financing both. All three run on the same underlying pipeline, laid out once below. What’s actually distinctive per feed — the questions it’s chasing, the sources it leans on, the threads and entities it follows — is covered in its own section after that.

This page is a snapshot as of 2026-08-07. Every coverage number below — thread counts, entity counts, source counts — is a point-in-time figure carried over from that day’s review, not a live counter; it will have moved by the time you’re reading this.

The common methodology

Automated collection. Every feed starts the same way: a fixed set of programmatic collectors sweep a fixed list of watched entities and search terms every day. Nothing here is scraped or searched by hand. Depending on the feed, that’s some mix of general news sweeps (Google News RSS, GDELT), curated outlet RSS/Atom feeds, government and financial data sources (SEC EDGAR, Federal Register, Treasury TIC, BIS, IMF, FRED, Senate lobbying filings, FEC), academic and clinical sources (OpenAlex, Semantic Scholar, arXiv, ClinicalTrials.gov), and a handful of narrower ones (GitHub releases, Manifold Markets prediction odds, Hacker News, CourtListener case law). Which of these actually matter to which feed — and which are currently down or under-built — is in each feed’s own section below.

The attention map. What the collectors find lands on a shared attention map: a watchlist of named entities and a list of threads — tracked, ongoing stories — one map per lens. The map only moves by steering: a person reads the daily output and decides “track X,” “drop Y,” “go deeper on Z,” and those edits land as small, provenance-tagged changes. Large automated passes — re-triages, syntheses — can only propose into the surface that gets read; they never batch-rewrite the map directly. The coverage critic (below) is the other way entities and threads get added, and that part is automatic.

Daily curation. Once the day’s collectors have run, a daily pass reads what came back and writes the digest, organized into a small set of fixed sections per feed, under one rubric shared across all three: signal over noise, sourced claims only, no framing spin. This step is currently a model-assisted judgment call made fresh every day, not a mechanical pipeline — collection above is fully automated, but deciding what’s signal, and what a given item is evidence for, still takes a curator’s read each morning.

Item annotations. Every item that makes a digest is tagged, in the underlying record, to the thread(s) and entity it belongs to. That’s what lets an item about a state bill and an item about a lawsuit roll up into the same tracked narrative over weeks, instead of existing only as that day’s headline. The tag is machine-checked, so the link between what was said and what it’s evidence for stays auditable rather than implicit.

The coverage critic. Every finalized day, each feed’s digest is checked against a named set of benchmark publications for that feed — real named outlets, listed in each feed’s own section below; there’s no reason to keep them vague. If a benchmark led with something a feed’s own sweep didn’t catch, that’s a logged miss, and misses don’t just get noted — they get added to the watchlist or thread list automatically. That’s the main way each feed’s coverage grows without waiting for someone to spot a gap by hand.

The dated-expectations ledger. Specific dated claims — a hearing date, an earnings call, an injunction ruling, an IPO listing — get logged to a standing ledger and checked automatically as their date arrives. A date that passes with nothing recorded against it is treated as news in its own right, not quietly dropped: a silent miss is louder than a late one.

Thread timelines. A thread is a tracked, ongoing story, not a single day’s item. It accumulates dated entries over weeks as the story develops, and can be backward-crawled to establish its history before this feed started watching it. Related threads can nest under a shared umbrella — a meta-thread — without losing their own separate timelines.

The briefing layer. Every page — feed, thread, entity — opens with a three-part readout: BREAKING (today), NEWS (the last seven days), and a short SUMMARY. BREAKING and NEWS are assembled mechanically from the dated item record — no model touches either, so they can’t drift from what the sources actually say. SUMMARY is the one model-written part: written against that page’s tracked material, regenerated only when the underlying evidence changes, and enforced as a shape rather than requested as one — fixed slots (a headline, bullets, a watch list; feed-level pages get a fuller version with a ranked lead and per-lens sections), and anything that doesn’t fill them gets rejected. Treat SUMMARY as orientation, not citation — the dated, linked bullets underneath it are the record.

What else is on the dashboard. Two narrower signals sit alongside the three feeds. A flash rail carries, rarely, a single banner for something that would lead a general news front page regardless of lens — a deliberately high bar, normally at most one active at a time, and it disappears automatically 24 hours after it’s filed rather than staying up on anyone’s judgment call. And a mechanical world-news signal counts how many distinct outlets are covering a story (merging a Google News cluster with a deduped GDELT feed) and matches that count against tracked threads — no model, no curator, just a number, kept deliberately small so it stays a radar rather than a fourth feed competing with the three real ones.

Publish. Getting onto the public site is default-on, not hand-gated — every thread ships unless it’s individually held back. What actually holds the line is mechanical, not a person reviewing each item before it goes out: only a fixed allowlist of fields is ever exported, raw internal notes and reasoning never are, and every export is scanned for secrets before it’s staged.

Mental Health

The questions

This feed watches AI in mental health, digital mental-health care, and evidence-based practice — the third of those promoted to a first-class strand in August 2026, with the question set below rebuilt from a two-track synthesis of the 2021–2026 research literature. Ten standing questions, in three clusters.

The evidence — what actually works:

• Is mental-health care getting more effective, or just more available? The anchor finding behind the question: across 441 RCTs and eight disorders, response rates to psychotherapy run only 24–42% — therapy beats control, but most patients in successful trials still don’t strongly respond — and most recent innovation (task-sharing, single-session interventions, digital delivery) expands reach rather than efficacy.
• Does the science of matching ever reach the clinic? Measurement-based care is strongly evidenced and persistently underused; precision psychiatry keeps producing striking pilot numbers that don’t survive external validation. Does “we know what works but don’t do it” close, or is it a permanent structural feature of mental-health delivery?
• Are the fast-moving somatic bets outrunning their own evidence fixes? Psilocybin is sprinting toward possible first-ever US approval on compressed review timelines — 18 months after the FDA rejected MDMA and then wrote new trial-design rules addressing exactly the unblinding problem the accelerating trials share. Ketamine’s real-world results run well below its trial results. GLP-1 psychiatric signals point both directions at once.
• Does the evidence machine itself hold? Contested guidelines on both sides of the Atlantic, Cochrane’s mental-health review group retired, and a US research-funding shock (NIMH lost more grants than any other NIH institute in the 2025 termination wave; 383 trials terminated — though Congress rejected the proposed institute consolidation and budget cut in February 2026, and the grant litigation is still in the courts) — who referees the evidence base, and can it keep producing?

The technology — is the science catching the deployment:

• Do AI-therapy efficacy claims survive independent replication? The first generative-AI therapy RCT reported large symptom reductions — against a wait-list-only control, evaluated by the lab that built it, with formal critique letters published in response. As of this snapshot no independent replication exists (both registered trials are the developer’s own), while the first independent active-comparator designs — including a chatbot-versus-human-therapist trial — completed in 2025 and await publication.
• Does safety evaluation mature faster than deployment? Benchmarks for AI mental-health safety now exist (models handle extreme-risk crisis queries reasonably and falter on intermediate-risk ones), and “AI psychosis” sits at the case-report stage with no population-level epidemiology — does it become a measured clinical entity or remain a media-amplified rare event?
• Do digital therapeutics solve their two structural problems? Engagement (attrition runs 19–28% and no one has shown more engagement causes better outcomes) and the payment model (the sector’s defining collapse happened despite best-in-class trial evidence — evidence was never the bottleneck; payment design was. Germany’s performance-linked pricing reform is the live test of the fix, and it cannot bite before its first mandatory performance-data delivery in April 2027; Medicare’s new digital-mental-health billing codes are live but carry no national rate, so the old reimbursement failure is replaying inside Medicare’s own plumbing).
• Does the screens-and-youth causality fight ever resolve? The academic dispute generates fresh rebuttals yearly rather than settling, while policy runs the natural experiments: a national under-16 social-media ban showing ~80% evasion at three months, school phone bans showing real academic gains and consistently null mental-health effects.

The market and its governance:

• Where is the money actually landing, and who are the players? Funding keeps flowing to clinician-augmentation over consumer chatbots; the vendors, evaluators, payers, and health systems doing the deploying — and the big platforms’ health arms — are tracked as named entities with their own threads.
• Does regulation of AI in mental health converge or fragment? State therapy-chatbot bans, an emerging FDA device framework for generative AI, and the EU AI Act’s high-risk classification are all moving — into either a coherent governance regime or a compliance patchwork vendors design around. The current read is fragmentation: seven-plus ban states with zero enforcement actions anywhere, three of four federal completion tracks stalled, and the EU’s high-risk obligations deferred to late 2027 — the billing-code track is the one lane actually advancing on schedule.

Sources & method

The collectors that contribute to this feed: Google News RSS is the workhorse, sweeping every watched name and phrase continuously; seven curated mental-health outlets (Behavioral Health Business, STAT News, OPEN MINDS, Addiction Treatment Business, and NIMH News are live; MobiHealthNews and Psychiatric Times are blocked by the publishers’ bot protection and swept by search instead); GDELT for a broad daily event sweep; Federal Register for FDA digital-therapeutic clearances and HHS/CMS rules; ClinicalTrials.gov (five standing search terms); OpenAlex and Semantic Scholar for academic literature; SEC EDGAR for public-company filings; Senate LDA for lobbying; and CourtListener for litigation, currently flagged at-risk — the provider tightened API access mid-2026 and that hasn’t been re-verified.

New as of August 2026 — the EBP strand’s instrumentation: eight journal and evidence-body feeds were added to daily collection, each URL live-verified before inclusion — npj Digital Medicine, JMIR Mental Health, Frontiers in Psychiatry, Internet Interventions, Lancet Digital Health, Nature Mental Health, the Peterson Health Technology Institute’s reports, and JAMA Psychiatry — and the OpenAlex/Semantic Scholar research sweep gained standing evidence-focused search terms (measurement-based care, psychotherapy-outcomes meta-analyses, psilocybin trials, single-session interventions, and others).

The coverage critic checks this feed’s finalized digest against four named trade publications daily — Behavioral Health Business, STAT (health-tech coverage), Fierce Healthcare, and MobiHealthNews — and, as of August 2026, two academic publications on a weekly cycle matched to journal publishing cadence: JMIR Mental Health and npj Digital Medicine. A research finding that leads one of those and isn’t in the week’s digests is now a logged miss, the same as a trade-press miss. Three organizations on today’s watchlist — Kaiser Permanente, Universal Health Services, and Hims & Hers — entered exactly that way, each after a benchmark led with a story this feed had no term to catch.

Threads & entities

As of this snapshot, this feed tracks 25 active threads and 38 named entities (33 organizations, 5 people), plus 16 standing topic searches and 7 clinical-trial search terms. Twenty-three of the 25 threads are open, two further along (“developing”); none has resolved yet.

The entity list is still corporate-heavy but no longer people-thin by accident: alongside the vendors, evaluators, and health systems, the watchlist now tracks the researchers whose work anchors the evidence strand — the field’s leading psychotherapy-outcomes meta-analyst, its most-quoted independent evaluator of AI and app claims, the loudest pro-causation voice in the social-media fight, and the PI of the first generative-AI therapy RCT (tracked partly because his lab evaluates its own product).

Cadence: daily digest; a weekly revisit of the standing questions above; the EBP strand’s academic benchmarks run on the weekly critic cycle, matched to journal publishing cadence.

Known gaps

• The EBP instrumentation is brand new and unproven in operation. The academic benchmarks, journal feeds, and evidence threads all landed in August 2026; the first weekly academic-recall cycle hasn’t run yet, and the digest’s research section — historically inconsistent day to day — has yet to demonstrate the consistency the new plumbing is meant to produce. The instrumentation existing and the coverage being reliable are two different claims; only the first is true so far.
• The people-watchlist is stronger on researchers than on power. Four researchers anchoring the evidence strand were added in August 2026, but regulators, legislators, and evaluator leadership still aren’t tracked as individuals the way frontier-lab executives are on the AI side.
• A sector gap surfaced this week and hasn’t been closed. Nothing tracks the applied-behavior-analysis/autism-therapy corner of behavioral health — a provider’s financial distress and new federal Medicaid guidance for that benefit landed the same day with no existing thread to catch either.
• Two of seven curated mental-health feeds are currently down. MobiHealthNews and Psychiatric Times both block this feed’s automated fetch and are covered by fallback search instead — a real degradation from a direct feed, not a full substitute.
• Litigation tracking has a live reliability flag. CourtListener tightened its access terms mid-2026, and that hasn’t been re-verified since.

Frontier AI

The questions

• Who are the players, and what are they actually doing? Not just who announced something — whether a major actor has a live, tracked storyline at all. An actor with nothing being tracked is treated as an open question, not a quiet one.
• Where is the money going? Which funding rounds, capex commitments, and acquisitions are moving capital, and what the financing structure underneath the buildout actually looks like once you follow it past the headline number.
• Where is frontier AI heading overall? Capability, price, and the containment/safety story running underneath both — watched for the non-events (a model that doesn’t ship, a deadline that passes with nothing published) as closely as the events.

Sources & method

This feed is organized around seven strands, and they account for essentially everything on it:

Fifteen collectors relevant to this feed run daily: a general news sweep (Google News, GDELT), curated RSS from lab blogs and trade press, SEC EDGAR, Federal Register, Semantic Scholar and arXiv/OpenAlex for research, GitHub releases for a panel of load-bearing AI-infrastructure repos, Manifold Markets for prediction-market odds on release dates, Hacker News for developer sentiment, and Senate lobbying/FEC filings. Two sources are weaker than the rest: CourtListener’s API access is gated in a way that hasn’t been fully re-verified, and a fourth family (GovInfo, Regulations.gov, Congress.gov) has a validated key but no collector built yet.

The coverage critic checks this feed’s finalized digest against four named AI-industry newsletters daily — The Rundown AI, TLDR AI, The Neuron, and The AI Daily Brief — plus two more, Import AI and Last Week in AI, on a weekly cycle. A genuine miss adds the missing entity or storyline to what’s tracked going forward — the main way the watchlist grows without waiting for someone to notice a gap by hand.

Threads & entities

As of this snapshot, this feed tracks 47 threads — 35 open, 10 actively developing, one resolved, one retired but kept for the record — out of 87 total across the whole map, the largest of the three feeds by thread count. The heaviest concentration is the compute/capex buildout race (13 threads) and resource moves like power and siting (8), followed by governance disputes, lab rivalries, and legal/accountability stories (4-5 each).

The watchlist behind that coverage names 33 organizations, 18 people, 8 models on active release-watch, and 10 policy/money themes.

Cadence: daily digests have run continuously since July 20, 2026; two weekly synthesis editions have been written against the standing questions above so far.

Known gaps

• Two of the fifteen relevant collectors aren’t fully live. CourtListener has an API access tier that changed mid-2026 and hasn’t been re-verified; the api.data.gov family (GovInfo, Regulations.gov, Congress.gov) has a working key but no collector built yet.
• One curated newsletter feed is dead — blocked by the publisher’s bot protection rather than genuinely gone — and is only checked by manual search, slower than the rest of the pipeline.
• The full collection run’s duration is inconsistent for reasons not yet understood. Two measured runs on comparable days differed by roughly 3.5x with no identified cause; coverage hasn’t been affected, but the pipeline’s timing isn’t yet predictable.
• China’s frontier-lab roster has had gaps caught late. A fifth major Chinese lab raising at IPO scale wasn’t on this feed’s tracked list until a routine sweep caught it two days after it had already been reported elsewhere — the coverage critic is the mechanism meant to catch this, and it’s still catching misses like this periodically rather than never. Several items most weeks are logged as “caught late” more generally, a stated limitation of a daily, largely automated pipeline.
• Weekly synthesis is new. Only two editions written so far, so the pattern is still young relative to the daily pipeline.
• Geographic coverage skews toward the US, the frontier labs, and China. Regions outside that core get comparatively little attention even when something AI-relevant happens there.

Global Capital

The questions

This feed is organized around two standing questions of its own (a third, “what is each major player doing,” is shared with the Frontier AI feed whenever the player is a financial one):

• Where is the money going? Funding rounds, capex commitments, M&A, IPOs — a running answer to where capital is concentrating, re-synthesized periodically. The current standing read: capital is concentrating into a handful of names at sovereign scale, funding the bottleneck resources (compute, energy, memory) rather than the models themselves. The mechanism doing the concentrating has gotten more specific over time — from a generic “circular financing” worry to hyperscalers extending off-balance-sheet credit guarantees to third-party developers, cheaper capital for the buildout paired with contingent credit risk that stays invisible until something breaks. That’s exactly what happened one step down the chain when a heavily leveraged AI-infrastructure fund was forced to sell its entire equity book after a sharp drawdown — the first fund broken by the trade it was betting on.
• Where is capital and economic power concentrating — in AI/health markets and above them? Tracked in three parts: capital in the markets this feed cares about (buyers, funders, private equity, foundations, swept like any watchlist entity); the macro backdrop (rates, inflation, currency, chips, energy, read as standing context); and wealth and power more broadly (who actually holds economic power, tracked through institutional filings, lobbying, and campaign finance — the least-built of the three; see Known gaps below). Two recent findings, both treated as data about how this feed itself can fail, not just data about markets: the vendor-financing structure above got its first real credit-market stress test when a major chipmaker’s credit-default-swap spread widened by a record margin in a single day, with the concern visibly migrating to the largest lessee in the chain rather than resolving; and this feed spent two straight weeks reporting a central bank’s rate decisions in vote-by-vote detail without ever naming the sitting chair of that committee — accurate the whole time, but missing the thing the votes were a fact about, until a routine cross-check caught it.

This lens is explicitly not a trading-signal service — no buy/sell calls, no price targets. It’s a record of where capital is concentrating and why, for someone who wants to understand the mechanism, not trade on it.

How this feed is different: the interpretation layer

AI and Mental Health track and report — every claim on those feeds is a sourced fact with a link. Global Capital does that too, but it also interprets. Alongside the sourced bullet, some items carry a second, clearly separate block: a generated reading of what the event could mean for capital flows, tagged with a confidence level, laid out as genuine branching scenarios (“this could go one of two ways, and here’s what would have to be true for each”), and read against a standing snapshot of the current macro picture. That block is never presented as fact — it’s visually distinct on the page, and it links to a full reasoning receipt.

A hard rule enforced in code, not just an instruction to whatever writes it, keeps this from becoming unmoored speculation dressed up as analysis: every interpretation must name a real transmission mechanism (a rate differential triggering a carry trade, a risk event triggering flight to safety, leverage amplifying a move), and once it claims more than the lowest confidence level, at least one scenario must point to an actual historical precedent — a specific past instance, not a vague gesture at “this has happened before.” An interpretation that can’t clear that bar is rejected outright rather than published with a disclaimer; duplicate, hedge-everything scenarios get rejected too — the branches have to be genuinely different bets on how things go. This is never forced onto every item — most days, most bullets carry only the sourced fact. Each interpretation that is generated is cached and given its own receipt page: the mechanism, every scenario with its precedent (or the honest absence of one), the macro context it was read against, and the threads it connects to.

Sources & method

A separate collector for retail/ETF-level fund-flow reports (Morningstar, ETF.com) currently returns nothing — both sites block automated access in the environment this feed runs in, and the feed reports that as an honest empty result rather than silently dropping the source or fabricating a number.

Each day’s curation reads a standing macro snapshot — refreshed weekly from the Treasury TIC, BIS, IMF, and EPFR data above, covering the current rate regime, cross-border credit conditions, the external (foreign-holdings) position, fund-flow trends, and any live conflict risk premium — before deciding whether an item gets an interpretation.

The coverage critic checks this feed’s finalized digest against four named finance-trade outlets daily — Money Stuff (Matt Levine), Axios Pro Rata, FT Unhedged, and Bloomberg Technology — plus Odd Lots on a weekly cycle. (Money Stuff and FT Unhedged are partially paywalled; the critic works from web search, public copies, and issue summaries.) Anything this feed missed gets logged and, if it’s a genuinely new entity or story, added to what it tracks going forward — the missed-Fed-chair catch above is what that process looks like working as intended, not a one-off embarrassment: it’s why two named individuals are now on this feed’s watchlist.

Threads & entities

As of this snapshot, Global Capital tracks 17 live threads directly under this lens, plus two adjacent AI-buildout capital stories (vendor-financing circularity and the Stargate data-center project) that stay filed under the Frontier AI feed for historical reasons but are cross-referenced here whenever they matter to the capital-flow read.

• Vendor financing and the AI-buildout credit story — Nvidia’s role as a financier as well as a supplier, a leveraged cloud provider’s discounted debt, chipmaker credit-default-swap moves, and whether the capex being spent across the sector is being monetized or just financed.
• The frontier-lab IPO wave — a meta-thread grouping individual IPO/listing stories (a major memory-chip maker’s debut, timing questions around a leading AI lab’s public listing, the merged aerospace/AI/social entity’s public-market debut).
• Sovereign and institutional capital moving into AI — Saudi Arabia’s Public Investment Fund pivoting from physical megaprojects (NEOM) into compute ownership, BlackRock building direct AI-infrastructure debt and equity exposure, Fidelity taking direct equity stakes in the labs themselves, and Berkshire Hathaway’s internal tension between Warren Buffett’s individual AI-capex bets and Greg Abel’s more skeptical institutional stance.
• Macro and geopolitical drivers of capital flow — the Red Sea/Hormuz shipping disruption’s effect on oil pricing and war-risk insurance, the CHIPS Act’s pivot from grants toward equity stakes in chipmakers, and the broader rotation of capital between chipmakers and the hyperscalers that buy from them.

The watchlist behind this carries 15 organizations (sovereign funds, buyout firms, foundations, and the capital-concentration names above the day-to-day market) and 2 named individuals, added specifically after the coverage critic caught this feed covering a central bank’s rate decisions for two weeks without ever naming the sitting chair of that committee.

Cadence: daily digest; the standing macro snapshot behind the interpretation layer refreshes weekly, matched to how often the underlying government and central-bank data itself updates.

Known gaps

• The “wealth and power” tracking is the least-built third of this feed’s scope. Institutional filings, lobbying registrations, and campaign-finance data are all wired as live collectors, but the actual cohort of people and institutions to track through them — a deliberately curated “who holds power” list, distinct from the market players already watched — hasn’t been built yet.
• The curated RSS feed list for this lens is stale relative to its current focus. It predates the shift toward AI-capital markets and still leans toward healthcare-buyer trade press; three of its ten feeds are also currently dead (blocked by the publisher, marked as such rather than silently dropped).
• One data source in the macro-context stack returns nothing. Retail/ETF-level fund-flow data — the corroborating check on the professional fund-flow reading — is blocked by both providers this feed tried, so one of five macro readings leans on a single source instead of two.
• A legal-records source this feed would use for liability and regulatory stories is membership-gated as of this review, a soft spot for anything requiring court-filing detail.
• A government-data API family (GovInfo, Regulations.gov, and related feeds) is validated and keyed but has no working collector yet — one source still waiting on a signup response before it can be added at all.
• Interpretation coverage is intentionally partial. Most items never get a generated reading — a mechanism has to be real and identifiable, not manufactured to pad a quiet day — so today the interpretation layer is a highlight reel of the clearest cases, not a reading on every capital-flow story that runs.

Appendix — coverage check against a working clinical team’s source list

In August 2026 this feed’s mental-health coverage was checked against the actual source list of a practicing clinical team — the sources they confirmed using day to day, plus a longer suggested tier they hadn’t yet vetted. The point of the exercise: does an automated feed with a curation-and-critic layer actually cover what a diligent human system covers, and what does each side miss? Every verdict below was verified by live-fetching the source, not assumed.

Their confirmed daily/weekly system, mapped:

Their suggested tier, verified item by item:

What the feed adds that a source list can’t: threads that accumulate a story across weeks instead of re-encountering it; a nightly critic that logs what the benchmarks caught and this feed missed (misses become watchlist entries automatically); a dated-expectations ledger where a date passing silently is itself news; and timelines a reader can audit back to sources. A source list tells you where to look. The pipeline behind this page is what looking, every day, without fatigue, actually requires.