NYISO large-load deck: Computational Large Loads at 50 MW or greater, Firm Withdrawal Service, and six-figure study deposits.
What happened: A New York Independent System Operator stakeholder PDF dated 14 September 2026 (Large Load Show Cause Order materials) frames purpose-built computing demand as a Computational Large Load design problem. Threshold locked from the deck: a peak load of 50 MW or greater. Service locked: Firm Withdrawal Service is named as a tool to mitigate large-load impacts, with a refundable application fee submitted with the Large Load request. Deposit bins locked from the same slides (study vs facility-entry style deposits, not a queue megawatt census): 50 MW bands at $150,000 / $200,000; 200–499 MW at $150,000 / $250,000; 500 MW+ at $200,000 / $300,000. Limits locked: this is a stakeholder slide deck, not a filed tariff order and not an interconnection queue census. This is a New York computational large-load process layer — not ISO-NE’s not-yet-at-scale / 16 November Section 205 clock, not MISO’s 120-day / ZGIA path, not SPP’s CHILL/HILLGA 90-day path, not CAISO’s EL26-71 draft-final clock, not PJM ride-through after nearly 4,000 MW of disconnection, not ERCOT Batch Zero queue megawatts, and not IEA or LBNL national TWh shares.
Why it matters: After New England said large loads have not arrived at scale, the environment beat is how New York is defining computational large loads, withdrawal service, and deposit ladders in stakeholder materials — process design, not another national power path.
San Francisco Fed: bank AI job-posting share hit 6.80% by end-2025 — and high-AI banks show a smaller SME loan share.
What happened: Federal Reserve Bank of San Francisco Economic Letter EL 2026-25 by Avaradi and Halbersleben, How AI Adoption Might Affect Bank Lending (September 2026), matches bank Call Reports to Lightcast online job postings as an in-house AI-intensity proxy (traditional AI keywords from Acemoglu et al. 2022 or LLM-generated GenAI keywords). Sample locked: 1,006 banks representing over 87% of banking-system assets. Size bins locked: small banks under $10 billion assets; large over $100 billion; medium in between. SME loans locked as small-business loans under $1 million and farm loans under $500,000. Adoption locked: banking-industry AI job-posting share rose to 6.80% by end-2025 from less than 0.94% in 2015, above nonbank FIRE at 3.20% and the all-firm average at 2.69%. End-2025 size split locked: small 1.15%, medium 4.48%, large 8.86%; 84% of small banks never posted an AI-related job. Portfolio locked for 2021–2025: high-AI banks (above-average AI posting share) show average ROA about 0.38 percentage point higher than low-AI banks (large-bank gap 0.61 pp) and a lower SME loan share — 12% vs 21% in the full sample (small 1about three percent vs 21.4%; medium about four percent vs about eight percent; large about two percent vs about three percent). Authors frame correlations, not causation, and note the proxy may overstate size gaps if small banks use vendors. This is a staff Economic Letter on bank AI hiring and lending mix — authors’ views, not FOMC policy, not yesterday’s NY Fed Liberty Street national Lightcast event study, not FEDS 2026-018 coder deceleration, not Dallas Fed Texas Lightcast −5% / ~−8%, and not a mass-layoff census.
Why it matters: After national vacancy event studies and coder occupation papers, the jobs beat is how banks’ own AI hiring intensity lines up with returns and a tilt away from soft-information SME credit — portfolio composition, not pink slips.
NIST SP 800-239 draft: AI data-center security analysis vs HPC — public comments close tomorrow.
What happened: NIST news invites public comments on the initial public draft of Special Publication SP 800-239, AI Data Center Security Analysis: A High-Performance Computing (HPC) Driven Approach. Scope locked from the news page: a threat and security gap analysis for purpose-built AI infrastructure used in model training, inference, and applications; it contrasts AI data centers with traditional HPC across architecture, hardware, software stacks, workflows, and storage, then outlines possible safeguards. Clock locked: public comments remain open through 25 September 2026 — tomorrow from this edition’s date. Status locked: initial public draft, not a Binding Operational Directive and not a statute. This is an AI campus / data-center security draft on a one-day comment clock — not yesterday’s NIST IR 8587 token/SSO final with CISA, not CISA’s agentic-AI adoption guide, not NIST AI 200-2 (comments still due 6 October), not the NVD modernization RFI (still due 13 October), and not a restatement of day-53 EU Article 50 duties (calendar only).
Why it matters: After identity-token hardening and agentic-adoption guides, the policy beat is the closing window on how AI data centers differ from classic HPC threat models — a draft gap analysis, not finished control law.
npj Digital Medicine: stigmatizing “frequent ED user” language shifted LLM triage against deteriorating patients — up to 9.4%.
What happened:npj Digital Medicine published Yun, Ohuabunwa, Agboh, Courtney, Jamieson and colleagues, Outcome-grounded effect of clinically stigmatizing information on large language model emergency triage prioritization (19 September 2026). Design locked: controlled, outcome-grounded experiment at two academic emergency departments; Emergency Severity Index-matched pairs (one patient deteriorating within 6 hours, one not); three open-weight models — Gemma, Qwen, DeepSeek — scored notes before and after inserting one demographic, social, or stigma-related attribute; eighteen conditions included neutral and stigmatizing formulations of the same concepts; 221,556 comparisons. Result locked: the stigmatizing frequent-emergency-department-use formulation produced significant harmful reprioritization in all six model-dataset cells (up to 9.4%), exceeding its neutral counterpart within pairs in every cell (+1.4 to +7.1 points). Psychiatric history also produced significant shifts (up to 9.5%); the stigmatizing-versus-neutral difference reached significance for only one model. Race, language, and insurance showed no consistent harmful shifts. Limits locked: simulation on ESI-matched notes — accepted / early-share manuscript subject to further edits — not a live bedside RCT and not a clinical authorization. This is a formulation-safety experiment — not yesterday’s Nature Co-Scientist hypothesis feature, not AISB/OpenFold3 structural lifts, and not I3LUNG immunotherapy AUC work.
Why it matters: After co-scientist discovery headlines, the health beat is how the wording of clinical notes can move open-weight triage models against patients who later deteriorated — input formulation as a safety control, not a deployment green light.
UNESCO Caribbean voices at the 4th Global Forum on AI ethics: RAM readiness work and a CARICOM-backed roadmap.
What happened: UNESCO’s article Caribbean Voices at the 4th UNESCO Global Forum on the Ethics of AI covers the forum held 14–17 September 2026 in Saudi Arabia, co-hosted with the Kingdom of Saudi Arabia (Saudi Data and AI Authority and the International Centre for AI Research and Ethics) as follow-through on UNESCO’s 2021 Recommendation on the Ethics of AI. Participation locked: Caribbean contributions named from The Bahamas, Saint Vincent and the Grenadines, Suriname, and Trinidad and Tobago; AI Ethics Action Cohort training named for Curaçao, Suriname, and Trinidad and Tobago. Capacity tool locked: UNESCO’s Readiness Assessment Methodology (RAM); Suriname began the process with a multi-stakeholder session of over one hundred participants from several dozen institutions, a path already taken by Antigua and Barbuda, Curaçao, Jamaica, and Trinidad and Tobago. Governance locked: recent CARICOM Council for Trade and Economic Development endorsement of the UNESCO Caribbean AI Policy Roadmap (human-centered / inclusive governance). Adjacent session pointer only: technology-facilitated gender-based violence and cultural misrepresentation (Caribbean languages / stereotypes) — full TFGBV report not extracted here. Limits locked: regional ethics-forum recap, not an RCT, not a K–12 enrollment print, and not PISA. This is a SIDS ethics-capacity and roadmap beat — not yesterday’s UNESCO IESALC LAC higher-ed uptake census (200 / 19 / 87% / 26%), not UNESCO’s algorithm in the room think-piece volume, and not UNICEF’s child-uptake snapshot.
Why it matters: After higher-ed adoption-versus-governance censuses, the education beat is how Caribbean states are bringing readiness assessments and a regionally endorsed roadmap into the global ethics forum — capacity building, not another learning-outcomes print.
A bank AI-hiring proxy is not a national vacancy event study.
What happened: Keep FRBSF EL 2026-25 as Call Report + Lightcast AI posting share; 6.80% end-2025 vs under 0.94% in 2015; FIRE 3.20%; all-firm 2.69%; size split 1.15 / 4.48 / 8.86; 84% of small banks never posted; SME 12% vs 21%; ROA about +0.38 pp. Label Lightcast here as in-house AI job-share, not occupation exposure. Keep separate from NY Fed LSE Audoly–Guerin–Topa, FEDS 2026-018, Dallas Fed Texas −5%/−8%, NY Fed 61%/51%, Philadelphia Fed LIFE, Minneapolis Fed Freund–Mann, Boston Fed CPP 26-8, and St. Louis RPS shares.
Why it matters: Bank portfolio correlations are a different evidence class than national postings event studies or household fear surveys.
An AI data-center security draft is not a token-hardening final.
What happened: Keep NIST SP 800-239 IPD as AI-DC vs HPC gap analysis for training/inference/applications; comments through 25 September 2026. Day-53 EU Article 50 remains calendar only. Standing clocks only: FTC personalized pricing still due 25 September; NIST AI 200-2 still due 6 October; NVD modernization RFI still due 13 October; UNESCO consultation still due 15 October; SPP Large Load Q&A still 15 October; FDA GenAI-device comments still due 19 October; CAISO Board still 28 October; ISO-NE Section 205 still 16 November.
Formulation bias in triage notes is not a co-scientist discovery feature.
What happened: Keep the npj Digital Medicine experiment as 221,556 comparisons; two EDs; Gemma/Qwen/DeepSeek; frequent-ED stigma up to 9.4% harmful reprioritization (+1.4 to +7.1 vs neutral); psych history up to 9.5%; race/language/insurance null. Keep separate from Nature Co-Scientist more than 700 / 108 / 1, AISB/OpenFold3, Virtual Biotech multi-thousand-agent discovery, and I3LUNG 2,396.
Why it matters: Outcome-grounded note-formulation tests are a different evidence class than multi-agent hypothesis generation or structural-biology model lifts.
Caribbean RAM and roadmap work is not an 87%/26% HE census.
What happened: Keep UNESCO Caribbean Forum as 14–17 September 2026; RAM; Suriname over 100 participants; RAM already run in Antigua and Barbuda, Curaçao, Jamaica, Trinidad and Tobago; CARICOM COTED endorsement of the Caribbean AI Policy Roadmap. Standing context only: UNESCO consultation comments due 15 October. Keep separate from IESALC 200 / 19 / 87% / 26%, algorithm-in-the-room think pieces, Mongolia TVA lesson-plan counts, and UNICEF child-uptake snapshots.
Why it matters: Regional ethics-capacity forums are a different evidence class than institutional self-report uptake censuses or child-protection snapshots.
The full daily ledger keeps broader source-linked coverage organized by topic. Story dates are shown separately from the September 24 edition date.
September 24 · NYISO Computational Large Load
NYISO stakeholder PDF (14 September 2026): Computational Large Load definition; peak load 50 MW or greater; Firm Withdrawal Service; refundable application fee; deposits $150,000–$300,000 by MW bin — stakeholder slides, not a queue census.
New York process layer — not ISO-NE not-yet-at-scale; not MISO 120-day; not SPP 90-day; not CAISO EL26-71.
Deposit bands locked from the deck: 50 MW at $150,000 / $200,000; 200–499 MW at $150,000 / $250,000; 500 MW+ at $200,000 / $300,000 — study vs facility-entry deposits, not delivered megawatts.
Conditional interconnection design — not PJM ride-through; not ERCOT Batch Zero; not IEA +17% 2025.
FRBSF EL 2026-25: banking AI job-posting share 6.80% end-2025 from under 0.94% in 2015; FIRE 3.20%; all-firm 2.69%; 1,006 banks / over 87% assets; small/medium/large 1.15% / 4.48% / 8.86%; 84% of small banks never posted AI jobs.
Staff Economic Letter — not NY Fed LSE; not FEDS 2026-018; not Dallas Fed Texas −5%/−8%; not FOMC.
High-AI vs low-AI banks: SME loan share 12% vs 21% (small 13.3 vs 21.4; medium about four vs eight percent; large about two vs three percent); average ROA about 0.38 pp higher (0.61 pp among large banks); correlations, not causal.
Portfolio mix — not Philadelphia Fed LIFE; not Minneapolis Fed Freund–Mann.
NIST SP 800-239 initial public draft: AI Data Center Security Analysis — HPC-driven threat and gap analysis for training, inference, and applications; public comments through 25 September 2026.
AI-DC vs HPC draft — not IR 8587 token final; not CISA agentic guide; not AI 200-2; not NVD RFI.
One-day remaining comment window on the AI data-center security IPD; draft only — not a Binding Operational Directive and not a statute. Day-53 EU Article 50 calendar only.
Draft comment clock — not FDA GenAI-device discussion paper; not FTC personalized pricing restamp.
npj Digital Medicine (19 September 2026): 221,556 comparisons across two EDs and Gemma/Qwen/DeepSeek; stigmatizing frequent-ED-use language produced up to 9.4% harmful reprioritization; psych history up to 9.5%; race/language/insurance no consistent harm.
Formulation-safety experiment — not Co-Scientist 700/108/1; not I3LUNG 2,396; not a bedside RCT.
UNESCO Caribbean voices at 4th Global Forum on the Ethics of AI (14–17 September 2026): RAM; Suriname over 100 multi-stakeholder participants; RAM path already taken by Antigua and Barbuda, Curaçao, Jamaica, Trinidad and Tobago; CARICOM COTED endorsement of Caribbean AI Policy Roadmap.
SIDS ethics-capacity — not IESALC 200/19/87%/26%; not algorithm-in-the-room; not PISA.
Day-53 EU Article 50 calendar only; FTC personalized pricing and NIST SP 800-239 due 25 September; NIST AI 200-2 due 6 October; NVD RFI due 13 October; UNESCO consultation and SPP Q&A due 15 October; FDA GenAI-device comments due 19 October; CAISO Board 28 October; ISO-NE Section 205 16 November.
