Martice

There is potential disruption on the horizon for the LLM market! TypeSafe's Jev is a structured-decision model: unstructured state in, a typed answer out, no free-form strings. They put those calls at 70-500ms end-to-end. I read the "can't hallucinate" line with skepticism but I'm holding out hope that it proves itself in production! That being said, the schema (the shape of fields the call is allowed to return) is the output, so it cannot invent a third bucket. That is not the same as being right about the world. Right now, we are bound by latency and cost. Shipping and LLM product still waits on strings you parse and babysit. This makes it hard to integrate LLMs directly into operational software without degrading user latency experience. It's a hurry up and wait problem. On a hot-path route, a score, or a check on another model's output, I'd want an output the code can take, not a paragraph to re-parse. typesafe.ai/blog/introducing…

This looks like a game changer!

A fluent answer can look backed by the docs when it only shares words with a retrieved passage, while the right fact sits unused further down in the same context. That is not empty retrieval. The sources were there. The match was cheap. EAEV (Evidence-Aligned Entity Verification; Jia et al., arXiv 2609.08267) is useful for one move: poke that match. Drop the strongest passage the check leaned on, or scrub easy formatting tricks, and see whether the support still holds. Shared-word luck falls over. Support that was actually tied to the fact tends to stay. On their runs, removing that poke hurt the check more than removing any other single piece of the method. Shared words with a retrieved passage are not the same as the right fact supporting the claim.

This is a great opportunity Danti has secured! I'm looking forward to what this brings to the technology base. What are we doing for the Analysts? They ask for a finished report they can stand behind. Too often the system still hands back another pile to sort and calls that help. What we found works better is a focused ask, then our specialized AI analysts working it in parallel the way a real desk would. They argue the sources, mark what is thin, and land one spatially and temporally grounded picture you can question. From there the report or briefing comes out of the same workspace. That is the AI analyst workforce path. Space Force TacSRT is where we are expanding it.

I've watched desks burn senior people on fluent model answers. The prose sounds finished, so everyone starts line-editing like it is a finished brief. That is a terrible use of the people you actually trust. The better habit is smaller. Break the answer into checkable claims. Separate what already sits inside what you trust from what is contested or new. Then put senior attention only where the uncertainty is worth the cost of a real look. That is the budget. You are not paying expert time for every sentence. GRACE (Yi et al., arXiv 2609.04442) is one recent example of that shape: claim-level grounding against what you already trust, with an explicit frontier label for what still needs a person. I am not pitching their stack. The desk already knew the problem. The paper just named the rationing clearly. I've seen a senior spend the afternoon on sentences that never needed them, while the contested claim sat unmarked in the middle.

One sensor is never the picture. That's the whole claim of multi-phenomenology, and it's why fusion exists. Each mode answers a different question, on its own clock, at its own grain. Radar is an instant. A daily ship count is continuity. A notice is a decision to the minute. A ledger is the cost, weeks later. Held apart they're partial. Held together they are what happened over time and space. I walked that through one closed port (Algeciras, Storm Kristin): linkedin.com/pulse/one-senso…

Gemini started slewing across video this week instead of stuffing a 1 FPS stream into the window, and that's the collection pattern I've been arguing for since we started Danti. Index coarse. Then point the sensor at the window that actually answers. Don't ingest every frame just because the context window will take them. We've watched that fail with documents for years: dump everything in, hope retrieval sorts it, and call the bigger dump a better analyst. This landed on 1 September, on Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite, in the Gemini API and the Gemini Enterprise Agent Platform. The old path was static 1 FPS, or whatever FPS you set. The new path is an agentic loop with native video tools: search, scan, inspect a segment, across pixels, audio, and transcript. Google reports up to 88% fewer tokens, up to 66% lower cost, and up to 7% higher accuracy versus static processing. No extra feature fee. You turn it on with processing: "agentic". Those are "up to" numbers from their post, so I treat them as a ceiling, not a budget line. What I believe is the pattern, not the receipt. The miss most teams will make is treating this as cheaper 1 FPS. It isn't. It's a controller that spends tokens only on the moments that answer the question. Long video finally has a slew. That does not mean it has a stuffed window, and a stuffed window was never a watch. If you're already ingesting clips at 1 FPS, that's the comparison I'd actually trust: agentic versus static, token count, and whether the answer still cites the right second. Google already has that comparison going the right way: fewer tokens, higher accuracy. That's the drop I actually want. The model is spending less because it stopped stuffing frames that never get cited, not because it got sloppy about the second.

LangChain dropped a pre-release on 1 September where a connector that needs a human refuses the call instead of quietly guessing and moving on. That's langchain==1.4.0a3. You install it with pip --pre, plus langchain[mcp] and fastmcp 4.0.0 or newer. It is not GA. I wouldn't put it on a production pager, and I wouldn't let anyone on my team treat it as shipped. I'm looking at it because of the elicitation behavior. Most agent demos I've sat through do the opposite. An MCP server asks a mid-call question, nobody's there to answer, and the run just continues. Hope nobody notices the skipped confirm. Default-off is the other shape: if you haven't opted into a pause, the connector that requires a person stops. That's a product decision I wish more harnesses would make. When you do want the pause, you set elicitation="interrupt" and a LangGraph run actually waits. Someone accepts, declines, or cancels, and the graph resumes. We've been gluing that together with custom callbacks for years: confirm this collection, confirm this classification, confirm this source. Interrupt puts it in the run instead of a side channel you forget to check. MCPAdapter folding MCP servers into core LangChain tools is fine, and list_tools can cache the catalog with cache_mode use, refresh, or bypass, per SEP-2549. That's useful if you're tired of rediscovering the tool list every turn. It is not the reason I'm reading a pre-release. The reason is the refuse. A human-required tool that keeps going isn't autonomous. It's unsupervised. That's the distinction I care about when we wire analysts at Danti, and it's why I'd trial this pre-release even though I wouldn't ship it yet. If it silently proceeds with elicitation off, or fails to pause with interrupt on, I wouldn't ship the connector.

Google Research published MAPL-EMIT on 1 September, and I don't want this read as another AI-found-methane post. A Swin transformer reads EMIT hyperspectral radiance to detect, outline, and locate methane point sources. EMIT flies on the ISS: hundreds of bands, 60 m spatial, 80 km field of view, 7.4 nm spectral sampling. It was built for minerals, not gas. The score that matters is 84% recall on expert-annotated plumes. Against NASA's matched-filter L2B, MAPL-EMIT finds about 50% more plausible plumes across about 1100 EMIT granules. It detects at 24 of the world's 25 top-emitting landfills. Pixel-wise matched filters smear overlapping industrial plumes into one blob. A spatial transformer plus the spectrum can pull the sources apart. That's not "AI found methane." That's phenomenology plus scene structure, which is a very different claim. Google is releasing a global plume database on Earth Engine, with physics-based spectral-fit confidence and lower/higher tags you can threshold. The comparison figure is visible vs NASA L2B vs MAPL-EMIT. There's a separate overlapping-plume source-localization figure. Those two figures are the argument. I don't need another methane gallery. I need to know whether the pipeline can split two sources that a matched filter glued together.

Claude Code shipped restricted mode on 27 August, and it's the first-party read-only harness I actually want for analyst work. v2.1.248 added --restricted and CLAUDE_CODE_RESTRICTED=1, which strips command and code tools, and WebFetch unless you named them in --tools. File tools stay inside the working directory, it refuses bypassPermissions, and it ignores user, project, and local settings. That isn't a softer coding agent. That's a different job: read the files in this directory, don't become a shell. Four days later, v2.1.257 on 1 September added Containment Escape. Auto mode no longer auto-approves cloud metadata-credential fetches, egress evasion, or cross-tenant reach unless the environment marks them expected. Analyst workbenches fail here as tradecraft, not as UX. An auto-mode that can hit instance metadata is a credential incident sitting on a default. Restricted mode is the flag air-gapped imagery and document review actually needed. In that room, Bash is the incident, not the feature. Containment Escape is what you still need if you aren't in restricted mode and auto mode is on. They are not substitutes for each other. I don't think of a read-only analyst as a coding agent with the dangerous tools left on another flag. It's a harness, and the harness has to say so.

NTT Docomo, NTT Data, and Synspective signed last week (28 August) to use satellite imagery when crews can't get to a damaged cell site. That's the story. Japan's largest mobile network, an optical shop, and a SAR constellation, agreeing to look from space when roads, power, or weather keep a truck in the yard. NTT Data supplies the optical, which is easier for a human to actually read. Synspective supplies StriX, which still images at night and in bad weather. Then they match those pictures against Docomo's own outage table and public-agency reports, and they want a restoration order out of it: priority, crew, access routes. Most people will write this up as a SAR-versus-optical bake-off. That's the wrong contest. Optical isn't better SAR, and SAR isn't sharper optical. Each one answers a question the other is blind to. The release says they'll be used complementarily, which is the right word if that's what you mean, and the wrong word if you mean average the two rasters. Docomo already knows which sites are dark. The satellites answer the questions that table can't: is the road still there, is the site underwater, can a truck get through at night. Leave the outage table out of the join and you have a mosaic. You do not have a restoration order. I keep seeing the same miss in GEOINT stacks. We fuse the pretty layers and forget the operational one. The image is never the product. What I actually want is whether you can connect the image to the thing you already know is broken.

Claude Code dropped Fable 5.1 on 1 September, and the number everyone is going to screenshot is $0.25 per million for cache reads. That's a 75% drop. List is still $10/$50 per million input/output. Anthropic says typical workloads land about 25% cheaper, and complex coding or highly agentic work can get to about 45%. I get why procurement will put that line on a slide. It isn't the constraint I actually care about. What classified shops need is true zero data retention with the detection logs sitting on the customer's side, not a vendor promise that they aren't keeping the prompts. Anthropic's path for that is Enterprise Frontier Safeguards (customer-held cloud storage for misuse detection), and those don't land until later this fall. Until then, eligible customers get zero data retention on Fable 5.1. That's a real permission. It is not the same permission as holding the logs yourself. If your authorizing official needs the second one, Fable 5.1 is blocked until EFS exists, and the cache price does not move that date. There's a better number on the launch table anyway. Fable 5.1 was evaluated with production safeguards on. On OSWorld 2.0, any task where those safeguards intervened scored zero. That's more useful to me than 52.6% on Terminal-Bench-Science 0.1. A scoreboard that zeros the work when the guardrail fires is telling you how the analyst actually runs, not how a science bench looks in a screenshot. Invitation-only Mythos 5.1 is the same weights with different safeguards. That's the product. I wish more launch tables led with that instead of the cache line.

Salesforce's agent cited a real 90-day return policy and a real purchase date, then still rejected a valid return. Both facts were accurate. Salesforce Engineering published the diagnostic on 31 August. The exception lived one relationship away: product class, then extended policy, then membership and price. Gold member. An $899 WM-500. 120 days instead of 90 if the order clears $500. None of that sat in the snippets the agent retrieved. The citations made the miss look sourced. A wrong answer with two true footnotes is harder to catch than a wrong answer with none. GEOINT agents fail the same way. Classification. Eligibility. Which source system owns the exception. Those paths never resemble the question, so search never brings them back. The answer still arrives with a clean citation list. Salesforce's fix is not more retrieval. They split the graph. TBox first: entity types and relationships, human-validated before any records load. ABox second: the Gold membership, the WM-500. Pointers from those entities onto the structured records so the agent does not rediscover table mappings at request time. They call this a diagnostic, not a benchmark. Treat it that way. Today: take one agent answer you already shipped. Cover the citations. Ask whether the exception path is in the graph, or only in the footnotes. True citations can make a wrong decision look finished.