Josh Mandel, MD

There is *a lot* I want to know about where GPT-5.5 thought summaries come from. In the middle of debugging tool use outputs like: > • Ran find /opt/android-sdk /usr/share/java -name '*guava*.jar' 2>/dev/null | head -20 I get the following thought trace (!!) > • I need to look for a guava jar, so I'll start by searching online. I'm curious about the best places to find it. Should I check grocery stores, specialty shops, or maybe some online retailers? I wonder what brands are out there and if there's a specific type I should aim for, such as guava jam or preserves. I'll take a quick look and see what options come up!

Hmm @AnthropicAI or knowledgeable users, does "full" here mean "including pre-resumption"? >The design should have been simple: if a session has been idle for more than an hour, we could reduce users’ cost of resuming that session by clearing old thinking sections. Since the request would be a cache miss anyway, we could prune unnecessary messages from the request to reduce the number of uncached tokens sent to the API. We’d then resume sending full reasoning history.

@dorofino I think you're talking about two very different layers of the stack. The shape and quality of the data that clients are authorized to see.

SMART Permission Tickets look promising for provenance and scoping. The thing I keep bumping against is that once a permission is granted, what gets returned still depends on whether the source data has real structure. A signed ticket to pull plan docs does not help much if the payload is a 40-page PDF with no schema behind it.

How might SMART Permission Tickets help with patient self-access, proxy access, public health, and more? Overview / explainer / live demo here: youtu.be/KlHGLZgkqcw

Airline platitude or undertaker platitude? "It is our honor to care for you on your journey."

Even better than canned visualization app is adhoc report generation using the analysis JSON as input. claude.ai/public/artifac… is the generated visualization HTML, and claude.ai/share/8ac52a72… is the Claude session that generated it (you can see it requires very little prompting). @zakkohane

This seemed like a fun task for Claude Opus 4.7! joshuamandel.com/greenbergolem/ has an example run on the claim "Hydroxychloroquine improves clinical outcomes in patients with COVID-19." github.com/jmandel/greenb… describes the pipeline. I spent 3-4h on this with Claude Code managing 10 parallel copilot subagents for ~12h. It can run autonomously to analyze claims where the literature is findable in full-text open access sources. github.com/jmandel/greenb… is the underlying output data with nodes, edges, and judgments driving the UI.

@ick_real Mildred!

I'm looking for a ridiculously old-fashioned girl's name for our new born . Think great-grandma name. Very old and rare. Any suggestions asap pls?

water leak after dishwasher install; undersink cabinet toekick dismantled for repair; heating duct exposed, uncovered; baby shoves stuff in there for fun all week; can't see what's fallen in there; grab digital otoscope to investigate; android app has rotted; try linux laptop, but scope doesn't talk USB Video; claude finds 5 gh repos that have reverse engineered protocol; codex writes userspace v4l driver in rust and android app in kotlin in one shot; use the scope; find mini broom resting one foot below floor level; rescue; rejoice github.com/jmandel/anysco…

@rodrigo_humanos Goal here is portability > granularity, i.e. trying to talk authorization policies that real-life servers can today support, and allowing an app to authorize against a set of endpoints with "SMARTv2 scope granularity" (basically category level FHIR), from a single ticket.

The SMART on FHIR delegation model makes this genuinely interesting. Scoping a ticket so an agent can read labs but not touch meds or notes requires finer granularity than most OAuth flows support today. Curious whether you are using resource type scopes or something closer to capability tokens.

Codex 5.4 xhigh fast is cranking away in tmux, where Claude 4.6 Opus 1M can track + visualize its progress, do research in the background to anticipate needs, and nudge Codex to continue when it asks for feedback. SMART Permission Tickets reference implementation doesn't write itself... yet ;-) Happy Friday!

I should be clear that this kind of thing only works when you already have a well-organized project and spend an hour making a solid plan 😉 Next month the story may be different.

@dorofino And x.com/JoshCMandel/st… :)

@dorofino See joshuamandel.com/regulations.go…

My Regulations dot gov comment analysis of 97 submissions on the Imaging Interop RFI is up and ready at joshuamandel.com/regulations.go… Quick summary article here...