# Seven-frontier HAI evidence watch · 2026-11-19 KST

## Scope and authoritative inputs

Research window2025-01-01 through execution morning2026-11-19KST; recent14-day overlapNovember5–19. Original HAI-research-trends.md reread in full(53lines), alongside connectedGitHub baselineNov10,latestNov18 HTML,53-record register,scoped watch_state,_config.yml. Source authority is connectedGitHub/Library,not old scratch copies. Baseline blob62fef19e85fd336539628a48b749b982f610606d retained. Today target initially404; one stable target only,no additional automation.

Actual defaultbranchmaster,unprotected,rulesets[]. RootAGENTS.md,_posts/AGENTS.md,CONTRIBUTING.md404. Two recent postsNov17/18 inspected. Current live tag anchors#agentic+ai,#foundation+models verified. _config.yml timezoneAsia/Seoul,permalink/:year/:month/:day/:title.html. PriorNov18 publicHTTP200 and expected title independently verified. No previous post changed.

## Fresh searches

Recent engine with recency14:

- LLM sycophancy overreliance research November18 2026
- agent workflow control human oversight research November2026
- LLM uncertainty expression calibration November2026 paper
- contestable interactive AI research November2026 paper
- AI provenance contribution ownership research November2026 paper
- synthetic human interaction evaluation November2026
- multi agent coordination replication November2026 paper

Full-period/representative and version follow-up:

- 2025 2026 synthetic interaction evaluation human simulator validity papers site:aclanthology.org
- 2025 2026 uncertainty expression human AI research calibration papers
- 2611.12412 v2 no result changed;2611.17610 v2
- Primary exact-page/PDF check for PULSE-CARE.

Selective retrieval,not exhaustive recall or full Crossmark/retraction audit. Excluded planned empirical programs and consultations from completed findings. Microsoft agent controls announcement(2026-11-18,https://commandline.microsoft.com/introducing-agent-controls-an-enterprise-control-plane-for-agent-behavior/) is an implementation/product announcement,not a new paper or independent human-effect experiment; not added to evidence register.

## Newly reviewed evidence and unchanged status

N28: Srimonti Dutta,Dinky Mishra,Aditi Singh,Akshata Kishore Moharir,PULSE-CARE: A Measurement-Validity Audit of Simulated-Patient Clinical Agent Evaluation. AACL-IJCNLP2026LongPapers,5904–5923,DOI10.18653/2026.aacl-long.266. Official proceedings monthNovember2026;PDF conference dateNov6–10. First preprint/public-online date not established,so no exact first-public day inferred. Older formal paper newly reviewed today,not a Nov19 publication event or newly confirmed acceptance of an already tracked preprint.

Primary sources inspected:

- https://aclanthology.org/2026.aacl-long.266/ (official metadata/abstract,directHTTP200)
- https://aclanthology.org/2026.aacl-long.266.pdf (methods,repeated-sampling controls,limitations,ethics)
- https://github.com/mishradinky/pulse-care/blob/main/README.md (blob86bf6e0ae407d376b2874bdbd8742c61cdf68d00)

PDF sections4.3,6.3,6.6,6.9,9–10 checked for denominator/claim boundaries. 24synthetic cases,576main trajectories,1542total;392repeated-analysis trajectories comprise98reused+294additional,not392new independent seed runs. Repeated API calls use no API-level seed;“multi-seed” is nomenclature. Score sensitivity0.25 is normalized range,not clinical error rate. Nominal17% rank flips fall to0 under0.05practical margin. Demographic pair disparity0.257 versus same-input repetition0.255 does not establish a demographic effect. No real patient records/interactions. Scorers/rubric/composite weights lack clinical validation; extension model checkpoint changes disclosed. No independent reproduction performed.

README declares cases/,simulator_cards/,outputs/,scripts/run_full_analysis.py and stored trajectories. It claims CPU/offline regeneration without API keys,but this run did not execute it or time it. License section remains TODO;rootLICENSE404 through connectedGitHub. Public accessibility is not verified reuse permission. Local plans therefore use separately reusable/synthetic nonclinical substitutes until terms/schema/hash verified; do not copy restricted data or assume source/model compatibility withMPS.

ExistingN25 official arXiv comments reread: v2 source release,related work/threats and competing interests expanded;no result changed. Keep source/DOI availability change separate from numerical correction. ExistingN26 official history stillv1. No new tracked-paper official retraction/erratum/outside-team replication verified in bounded search. N28 repeated sampling is author robustness analysis,not independent replication. N25 two-implementation within-group replication distinction retained.

R6/R7 conditions supplemented;R1–R7 central judgments maintained. Human overreliance/agency/subjective ownership and actual clinical safety are not inferred from model/synthetic evaluation.

## Access,corpus and experiments

55direct initial requests(53evidence+home+priorarticle):53HTTP200,2HTTP403(P04,N19). AdditionalN28official pageHTTP200. OfficialPDF read through web service;GitHubREADME/LICENSE calls separate. HTTP success is not proof of fullbody/appendix/version verification. Historical access levels retained per row. No material claim that all53papers were fully re-read.

54records preserve2024-first-public backgroundP07/P08,untraced first-public dates,P16formal8/preprint9model note,N22author spelling difference and other prior limitations. IDs belong to the Nov10seven-axis review,not Nov8ten-topic program. All numeric facts are author-reported. No human/model study,independent result replication,clinical trial,RL training or data-artifact execution performed. New local tests are falsifiable proposals only.

## Editorial and publication QA

Current publisher instructions and all6mandatory references read. Four fullCSS assets embedded in required order. English title,Korean body,YAMLbyte1,layoutnull,date2026-11-19,categories[artificial intelligence],exactly2existing tags. Required font roles/fallbacks retained. Hero-immediate static inlineSVG has4meaningful evidence-path nodes,uniqueIDs,title/desc/caption/text explanation;local focusable minimum1020px scrolling preserves14px+ labels. Diagram is conceptual,not causal evidence. Four highlight levels match textual importance labels;importance is not evidence confidence.

Static checks passed:77perspectives in exact11-step order across7axes,54unique evidenceIDs,HTMLnesting,SVGXML,internal/ARIA anchors,uniqueIDs,CSScompleteness/order,highlight mapping,no script/Liquid dependency. Playwright launch attempted:Chromium executable absent;no1440px/390px visual inspection claimed.

Only today'sHTML,eight date-scoped companion files,and scoped seven-frontiers/watch_state intended for commit. OriginalNov10 and all other posts preserved. Content/state commits advanced together using expected-head lease. Search success,repository commit and publicPages deployment recorded separately.
