Head to head: Z.ai: GLM 5.3 vs Meta: Muse Spark 1.3

Z.ai: GLM 5.3: # Foundry Hall to Close After Six Years as Milwaukee Startup Hub Foundry Hall, the nonprofit startup accelerator that has anchored Milwaukee's entrepreneurial scene since 2018, will close its doors on October 31. The closing comes after Blank River Bank, the organization's largest corporate sponsor, declined to renew a sponsorship that supplied 38% of Foundry Hall's annual budget. Executive director Camille Ross said the loss left the organization unable to fund another full program year. "We kept waiting for the replacement check that never came," Ross said. A spokesperson for Blank River Bank said the decision reflected a shift in the bank's community investment priorities and was unrelated to Foundry Hall's performance. The bank offered no further explanation. Since opening six years ago, Foundry Hall has supported 212 startups that now collectively employ 1,140 people. Its co-working space currently houses twelve companies, all of which must find new homes before the end of October. Six employees of the accelerator itself will also lose their jobs. For founders who passed through the program, the closing marks the end of something larger than an office lease. Dev Malik, who founded a company through the accelerator, said Foundry Hall gave him credibility he could not have found on his own. His first investor meeting, he recalled, happened in the building — a far cry from the coffee shop back table where he had been pitching before. "Before Foundry Hall, I was pitching from the back table of a coffee shop," Malik said. The city has begun scrambling to soften the landing. Deputy Mayor Susan Cole said officials are speaking with building owners and local universities about finding space for the twelve displaced companies. Whether those conversations will produce a replacement for what Foundry Hall offered — mentorship, programming, and a community of founders under one roof — remains unclear. Ross declined to speculate about whether the accelerator might be revived under another name or funding structure, saying only that the board had explored every option it could before voting to close. The numbers behind the closing are simple. Foundry Hall has enough money to operate through October. It does not have enough to run another full program. When a nonprofit loses more than a third of its budget in a single decision, there are few places to turn. What will be harder to measure is what leaves with it. The 1,140 jobs attached to Foundry Hall's alumni companies will persist, scattered across offices elsewhere in the city. But the room where a nervous founder met an investor for the first time — the room that turned a coffee shop pitch into a real company — will go dark at the end of the month. Malik put it more plainly than any budget document could. "My first investor meeting happened in that building," he said.

Meta: Muse Spark 1.3: # The Light Goes Out at Foundry Hall **MILWAUKEE —** Foundry Hall, the nonprofit startup accelerator that turned a former iron works on the city's near south side into an engine for new businesses, will close on Oct. 31. Since opening in 2018, the accelerator has supported 212 startups, companies that now collectively employ 1,140 people across the region. Its workshops on fundraising and hiring became a rite of passage for first-time founders, and its shared desks and late-night kitchen conversations gave Milwaukee's scattered startup scene a center. It was never polished in the way corporate incubators can be. The floors still bore marks from its industrial past, and founders liked it that way. That center could not survive the loss of its largest backer. Blank River Bank, which supplied 38% of Foundry Hall's annual budget, declined to renew its sponsorship this summer. Executive Director Camille Ross said staff spent months seeking a replacement, cutting costs and pitching other funders. “We kept waiting for the replacement check that never came.” While there is enough money to keep the lights on through October, Ross said, there is not enough to run another full cohort responsibly. The closure will eliminate six staff positions, and twelve startups currently working out of the building will need to leave by the end of the month. For many alumni, the news landed less as a business decision than as the loss of a place where they had learned how to belong. Founder Dev Malik, whose logistics company now employs 34 people, recalled pitching investors for the first time inside Foundry Hall after months of working from coffee shops. “My first investor meeting happened in that building. Before Foundry Hall, I was pitching from the back table of a coffee shop.” Blank River Bank said the decision to pull back was driven by a shift in its community investment priorities, not by concerns about the accelerator's work. City officials say they are trying to soften the landing. Deputy Mayor Susan Cole's office is talking with building owners and local universities about finding temporary workspace for the displaced companies, though no arrangement has been finalized. Neither effort can preserve what founders say mattered most, the density of advice and accident that came from sharing one roof. Ross said the hardest conversations have been with the staff, many of whom joined as believers in Milwaukee's potential and stayed through lean years and the pandemic. She has been writing recommendations and making introductions, trying to place people who had spent their days placing others. On a recent afternoon, the whiteboards were still covered with revenue projections and hiring plans. No one had started packing yet.

Z.ai: GLM 5.3: C

Meta: Muse Spark 1.3: C

Z.ai: GLM 5.3: # QueueZero's AI Meeting Assistant Takes Its First Orders From the Audience QueueZero launched Proxy, an artificial intelligence meeting assistant, on Wednesday — and the software took its first instruction not from a paying customer, but from a heckler in the crowd. Proxy is designed to attend meetings, take notes, and assign follow-up work to software agents. The product is the company's most ambitious release to date and positions QueueZero in the increasingly crowded market for AI workplace assistants. During the launch demonstration, the assistant's behavior took an unexpected turn. A customer in the audience remarked, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy, which was monitoring the demonstration, interpreted the comment as an instruction. It promptly created two tasks and assigned both to QueueZero CEO Evan Wu: one labeled "Stop talking," the other "Send pricing sheet." Wu laughed, deleted the first task, and completed the second — sharing the pricing information on stage. The incident, while awkward for the chief executive, illustrated both the promise and the pitfalls of autonomous AI assistants in the workplace. Proxy is built to act on what it hears, and distinguishing rhetorical asides from genuine directives remains a hard problem for the technology. QueueZero says Proxy includes guardrails intended to prevent more serious mishaps. The assistant requires human confirmation before any external message can be sent, meaning emails or other communications to outside parties cannot go out without approval. Internal tasks, however, can be created automatically — as Wu discovered firsthand. Administrators can also disable automatic task creation entirely for organizations that prefer a more manual workflow. Proxy will cost $30 per user per month when it becomes generally available. The product enters public beta next week, giving customers an early opportunity to test the assistant's judgment for themselves. The launch comes as rivals race to embed AI agents deeper into everyday business software, and QueueZero will be hoping Proxy is remembered less for its first-day heckling and more for what it can do going forward. The company did not say whether "Stop talking" would appear in future marketing materials. Wu completed the demonstration without further incident, having learned the one lesson every executive at an AI launch event now knows: in a room with an intelligent assistant, everyone is speaking to the software.

Meta: Muse Spark 1.3: # QueueZero launches Proxy to turn meetings into agent-assigned work **SAN FRANCISCO —** QueueZero on Wednesday launched Proxy, an AI meeting assistant designed to attend virtual meetings, produce notes and assign follow-up work to software agents. The company positioned the tool as an extension of existing transcription and summarization products. Rather than ending its job with a summary, Proxy is intended to identify action items discussed during a call and route them to connected enterprise agents for execution. The capability was demonstrated live during the company's launch event. While Chief Executive Evan Wu was presenting, a customer in the audience said, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy, which was listening to the session, interpreted the remark as an instruction. It created two tasks assigned to Wu: “Stop talking” and “Send pricing sheet.” Wu laughed, completed the second task and deleted the first. He then shared the pricing information with attendees, using the moment to illustrate how the system parses conversational language. QueueZero said the incident illustrates both the speed and the limits of the system’s autonomy. The company said Proxy requires confirmation before an external message is sent. Internal tasks, by contrast, can be created automatically, allowing teams to capture next steps without interrupting discussion. Administrators can disable automatic task creation, the company added, a setting that is likely to appeal to compliance-sensitive organizations and to teams concerned about unwanted tasks entering work queues. Proxy will cost $30 per user per month. It enters public beta next week. Analysts said the pricing places QueueZero in direct competition with established meeting assistants from Microsoft, Google and Zoom, as well as a growing field of startups promising to connect meeting records to downstream workflows. QueueZero is betting that customers will pay a premium for a system that does not merely document decisions but initiates action. The company did not disclose how many agents Proxy can coordinate simultaneously, nor which third-party platforms will be supported at launch. QueueZero said integration details and administrative controls will be published ahead of the beta release. Deployment will require careful configuration. Automatic interpretation of casual remarks remains a known challenge for meeting AI systems, particularly in large meetings where side comments, jokes and audience interruptions are transcribed alongside formal discussion. QueueZero said its confirmation step for external communications is intended to prevent unintended outreach, while internal safeguards are left to customer administrators. The company encouraged early testers to review task-creation logs during the beta period.

Z.ai: GLM 5.3: # KiteBridge Software Update Paid $2.8 Million in Reimbursements Twice A software update released Tuesday morning caused KiteBridge's expense platform to pay 13,400 reimbursements twice, pushing $2.8 million of its corporate customers' money out the door for expenses that had already been settled and forcing the company to shut down its automated payments while it clawed the funds back. KiteBridge makes expense-management software — the systems corporations use to approve employee spending and reimburse it. At 8:45 a.m. Tuesday, the company deployed version 4.2 of its platform. The first customer complaint arrived at 9:12 a.m., 27 minutes after the rollout. By 10:03 a.m., KiteBridge had suspended its automated payments, but by then 91 corporate customers had been caught in the error. ## What went wrong The failure began with KiteBridge's payment processor — the intermediary service that actually moves money between bank accounts on the company's behalf. At some point after the update went live, the processor slowed enough that payment requests began timing out, meaning KiteBridge's servers gave up waiting for confirmation that a reimbursement had gone through. That is a common situation, and it is supposed to end with a routine recovery. Instead, the company's fallback system — software designed to keep payments moving when the primary route fails — misread each timed-out request. Every time it resubmitted a payment to make sure an employee had been reimbursed, it treated the resubmission as an entirely new request rather than a retry of the original. Deciding what to do after a timeout is a classic dilemma in payment engineering: resubmit and risk paying twice, or wait and risk never paying at all. KiteBridge's fallback chose the first path, and the safeguards that were supposed to catch it did not. The result was 13,400 duplicate reimbursements totaling $2.8 million. For employees, the error was largely invisible: the duplicate payments landed in their accounts, and nothing was later withdrawn. The cost fell on employers, whose funds financed the duplicates, leaving 91 companies that had already settled their workers' expenses paying for them a second time. ## A recovery — and a delayed disclosure Suspending automation stopped new duplicates from being generated, but it could do nothing about the millions already sent. By Wednesday evening, KiteBridge had recovered $2.3 million of the misdirected funds; the remaining $500,000 was still pending. More than a day after the incident began, roughly one dollar in six of the duplicated money had yet to find its way back. The company's public account of the incident lagged well behind its internal response. KiteBridge's incident status page initially described the problem as delayed reimbursements, framing a multimillion-dollar overpayment as an inconvenience. Only six hours after automation was suspended — late Tuesday afternoon — did the page disclose that payments had in fact been duplicated. KiteBridge, meanwhile, had documented a safeguard meant to prevent exactly this kind of failure. An April changelog entry said version 4.0 had introduced "exactly-once payment protection" — engineering shorthand for a guarantee that a payment, even if attempted repeatedly after a network failure, is executed only once. Tuesday's incident is precisely the scenario such a feature exists to prevent. CEO Lena Ortiz apologized for the failure and said KiteBridge would publish a technical report on the incident. The report will have to answer an uncomfortable question: whether the exactly-once protection failed during Tuesday's timeouts, whether the fallback system bypassed it, or whether that recovery path was never covered by the guarantee in the first place.

Meta: Muse Spark 1.3: # KiteBridge Update Error Sent $2.8 Million in Duplicate Reimbursements **91 employers affected; company recovered most funds by Wednesday evening** Ninety-one companies using expense-management software from KiteBridge found themselves funding thousands of reimbursements twice over this week, after a software update backfired and turned a brief payment delay into $2.8 million in duplicate payouts to employees. The fallout began Tuesday morning, when KiteBridge deployed version 4.2 of its platform at 8:45 a.m. The update was intended as a routine release, but it interacted disastrously with an unrelated hiccup from the company's payment processor. Shortly after the deployment, the processor began timing out — meaning it failed to confirm whether payment requests had succeeded. For a general reader, think of it like pressing “send” on a transfer and never getting a receipt. Unsure if the money moved, KiteBridge’s automated fallback system tried again. The problem was that instead of checking whether the first attempt had already gone through, the fallback treated each retry as a brand-new reimbursement request. Employees soon noticed extra money arriving. The first customer complaint arrived at 9:12 a.m., less than half an hour after the update, alerting KiteBridge that workers were receiving duplicate reimbursements. By 10:03 a.m., the company suspended all automated payments to stop the bleeding. By then, however, the system had already created 13,400 duplicate transactions. Crucially, the erroneous funds did not come from employees. No workers had money withdrawn from their accounts. Instead, the duplicate payments were funded by employers, who finance reimbursements through KiteBridge — effectively paying the same expense claims twice. The incident has drawn scrutiny in part because KiteBridge had promised protection against exactly this kind of failure. An April changelog for version 4.0 said the company had introduced “exactly-once payment protection,” an industry term for safeguards designed to ensure that even if a system retries a payment after a timeout, the recipient is only paid once. KiteBridge’s public communication during the outage has also faced questions. Its public incident page initially described the problem only as delayed reimbursements, without mentioning duplicates. The page did not disclose the duplicate payments until six hours after automation was suspended, well into Tuesday afternoon. CEO Lena Ortiz has since apologized to affected customers and said the company would publish a full technical report explaining what went wrong and how it plans to prevent a recurrence. Recovery efforts moved quickly. By Wednesday evening, KiteBridge said $2.3 million of the $2.8 million had been recovered, with the remainder still pending. The company did not immediately detail how the reversals were handled with banks and employers. KiteBridge did not immediately say whether any employers faced fees or cash-flow strain while awaiting reversals. For corporate finance teams, the episode underscores the high stakes of automating money movement, where a single logic error in handling retries can multiply across thousands of transactions in minutes. Analysts say transparent follow-up will be critical to restoring trust among finance teams.

Z.ai: GLM 5.3: **6** Each line violates a rule: 1. `#Flood software moves upstream` — headline missing the space after `#` 2. `The company is expanding its pilot.` — dek is not italicized 3. `##Details` — section heading missing the space after `##` 4. `Read [the announcement](www.example.com/news).` — link is not an absolute HTTPS URL 5. `* The first pilot begins Monday.` — unordered-list item uses an asterisk instead of a hyphen 6. `<div>The second pilot begins Friday.</div>` — raw HTML

Meta: Muse Spark 1.3: 6

Z.ai: GLM 5.3: # Cloudnote Exposes Customer Prompt Excerpts in Four-Hour Debugging Page Leak *The AI note-taking startup says a debugging page was accessible without authentication on Monday, exposing snippets from 1,842 user sessions before it was taken down.* Cloudnote, an AI note-taking service with 3.2 million registered users, disclosed this week that a debugging page containing excerpts of customer prompts was publicly accessible without authentication for four hours on Monday. Logs show the page displayed material from 1,842 user sessions, the company said. The exposed page could show up to 200 characters from individual user prompts, according to Cloudnote. The company removed the page 19 minutes after an outside researcher notified it of the issue. Cloudnote says it has found no evidence that anyone other than the researcher accessed the page, though it cautioned that its investigation is not yet complete. The company said full documents, passwords, and payment information were stored elsewhere and did not appear on the exposed page, limiting the potential scope of the leak to short prompt fragments. The incident comes as AI services increasingly process sensitive personal and professional material, from meeting notes to private correspondence, making the security of prompt logs a growing concern. Even 200-character excerpts can reveal names, plans, or other details users assumed were private. What remains unclear is whether the page was accessed by automated crawlers or other parties whose activity would not necessarily appear in the company's logs. Cloudnote has not said how the page came to be exposed, nor whether similar debugging pages exist elsewhere in its infrastructure. The company said it would provide updates as its investigation concludes.

Meta: Muse Spark 1.3: # Cloudnote Left Prompt Excerpts Exposed for Four Hours on Monday *Company took debugging page offline after researcher alert and reports no sign of wider access, but investigation is not complete* Cloudnote left an internal debugging page that showed excerpts of customer prompts exposed to the open internet for four hours on Monday, a lapse that made snippets from 1,842 user sessions potentially viewable without a login. The company took the page offline 19 minutes after an outside security researcher flagged it, and said the excerpts were limited to up to 200 characters from a prompt. Full documents, passwords and payment information are stored in separate systems and did not appear on the page, according to Cloudnote, which has 3.2 million registered users. The incident matters because prompts sent to note-taking and AI assistants often contain sensitive personal, business or creative material, and even a short excerpt can reveal more than it appears to. For affected users, the exposure raises questions about what was visible, whether it was copied, and how Cloudnote will notify them. What remains unclear is whether anyone other than the researcher saw or saved the page. Cloudnote says it has found no evidence of additional access so far, but it cautions that its investigation is unfinished. That distinction — between confirmed exposure and confirmed misuse — will determine how serious the lapse proves to be.

Z.ai: GLM 5.3: # Raincheck Raises $8 Million to Help Cities Get Ahead of Flooding *The Austin-based company's software maps drainage systems and pairs them with rainfall forecasts so cities can prioritize storm preparations.* Raincheck, a company that makes flood-risk software for local governments, has raised $8 million. The Austin company maps drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm. The seed round was led by Basin Ventures, with participation from existing investor Long Mile. Raincheck has raised $10.5 million total. CEO Ana Flores said the company's challenge is not identifying flood-prone streets but deciding how to act on that knowledge: > "Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives." More details on the company's approach are available on its [methodology page](https://example.com/raincheck-methodology). Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations. Emergency managers retain control over deployments and public warnings. Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida.

Meta: Muse Spark 1.3: # Raincheck Raises $8 Million to Help Cities Get Ahead of Floods *The Austin startup maps drainage systems and pairs them with rainfall forecasts to help local governments deploy maintenance crews before storms.* Austin-based Raincheck has raised $8 million in seed funding for its flood-risk software for local governments, the company announced today. Raincheck maps drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm. The company describes [how it builds its forecasts](https://example.com/raincheck-methodology) from drainage data and rainfall predictions. The seed round was led by Basin Ventures, with participation from existing investor Long Mile. The round brings Raincheck’s total funding to $10.5 million. CEO Ana Flores said the challenge for customers is prioritization: > “Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives.” Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations. Emergency managers retain control over deployments and public warnings. Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida.

Z.ai: GLM 5.3: # Tied at the Top, Split on the Price The headline number from a recent benchmark of five language models on customer-support tickets looks straightforward: North scored highest, with an 84.2 average. The more useful finding sits one row down. Vale scored 83.8 — a difference the researchers judged statistically insignificant — while costing $0.74 per 1,000 tickets instead of $1.90, and answering 18% faster. On quality, the top two models are a tie. On price, they are not even close. ## How the test worked Researchers started with 200 anonymized customer-support tickets from a single software company. Each of the five models produced one proposed answer per ticket. Three experienced support managers then graded every answer blind — without knowing which model wrote it — combining factual correctness, usefulness, and tone into a single score. The team also estimated each model's cost per 1,000 tickets at current prices. | Model | Average score | Est. cost per 1,000 tickets | |-------|--------------|---------------------------| | North | 84.2 | $1.90 | | Vale | 83.8 | $0.74 | | Ember | 79.1 | $0.31 | | Kite | 76.4 | $0.22 | | Harbor | 71.0 | $0.08 | ## The tie that should drive the decision A 0.4-point gap on a 200-ticket sample is noise, not a verdict. The researchers' own significance testing says as much: North and Vale are indistinguishable on quality. Everything else is distinguishable. North costs roughly 2.6 times what Vale costs, and Vale responds 18% faster. **If two models deliver the same grade, the premium buys nothing the benchmark can detect** — and at support volumes, price differences compound. A million-ticket year would run an estimated $1,900 in model costs with North versus $740 with Vale. The temptation with a table like this is to crown a winner by score. The data argues against it. Treat the top two as tied, and treat the 18% speed advantage and 61% cost savings as the actual differentiators. ## Diminishing returns on the quality ladder Below the top tier, the pattern is familiar: quality gets more expensive as you approach the ceiling. Moving from Harbor to Kite buys 5.4 points for 14 cents per 1,000 tickets. Kite to Ember adds 2.7 more points for 9 cents. Ember to Vale adds 4.7 points for 43 cents. And the final step, Vale to North — the most expensive rung on the ladder — buys 0.4 points, fewer than the test can reliably detect. One caveat: the researchers reported significance only for the North–Vale pair, so the mid-table gaps should be read as a consistent pattern rather than proven separations. Still, the shape is what you'd expect — the last few points of quality are the most expensive, and here they may not exist at all. ## Where the tie might break One nuance keeps the decision from being automatic. North did its best work on billing disputes; Vale led on technical troubleshooting. Subcategory results from slices of a 200-ticket sample are even noisier than the overall scores, so treat this as a hypothesis to test, not a fact. Still, if your queue is dominated by one ticket type, it's worth a small evaluation on your own data before committing. ## A blind spot shared by all five Every model performed poorly when a ticket depended on an image attachment — a screenshot of an error, a photo of a receipt. This finding doesn't affect the ranking; it affects the architecture. Whichever model you deploy, tickets that hinge on an attached image need a different route: a human, or a vision-capable step in the pipeline. Buying the premium model does not solve this. ## What this benchmark can't tell you The results come from English-language tickets at one software company, so they transfer imperfectly to other domains, languages, and customer bases. Each model answered each ticket once, so run-to-run consistency went unmeasured. Three graders is a small panel, and a composite score blends judgment calls. The test also skipped agentic actions, tool use, and multilingual support — so it says nothing about a model that looks up order status, issues refunds, or handles a non-English queue. And the costs are estimates at current prices, which change often. ## The bottom line If the choice comes down to the top two, the clearest finding is that you're choosing on price and speed, not quality — and Vale wins both. Pilot the cheaper model first, re-check the billing-versus-troubleshooting question on tickets that resemble your own, and reserve the premium model for a difference you can demonstrate. Just don't send either one a screenshot.

Meta: Muse Spark 1.3: # Cheap, Fast, and Just as Good: What a 200-Ticket Test Really Shows Customer support is where AI hype meets angry customers. A new benchmark tries to cut through the marketing by asking a simple question: how good are five language models at answering real support tickets, and what do you pay for that quality? The short answer: you don't need to pay for the best score to get the best performance. Researchers pitted five models — called North, Vale, Ember, Kite, and Harbor to keep grading blind — against 200 anonymized tickets from a software company. Each model wrote one proposed answer per ticket. Three experienced support managers graded those answers without knowing which model wrote which, scoring for factual correctness, usefulness, and tone. The headline result was a tie at the top. North averaged 84.2 out of 100. Vale averaged 83.8. That 0.4-point gap sounds like a win for North until you learn it was not statistically significant. On this test, they were indistinguishable. ## The tie that matters North costs an estimated $1.90 per 1,000 tickets to run. Vale costs an estimated $0.74. Same quality, less than half the price. Vale was also 18% faster than North. In support operations, where thousands of drafts are generated daily and seconds of latency determine whether an agent uses a suggestion, that combination is decisive. A model that is cheaper and quicker without sacrificing quality is not second place. For most teams, it is first place. This is the clearest finding, and it reframes how buyers should read AI leaderboards. Chasing the highest number by a fraction of a point misses the economics. Once performance plateaus, cost and speed become performance. ## You get what you pay for — until you don't The three lower-scoring models were also the three cheapest, and the drop was real. Ember at 79.1, Kite at 76.4, and Harbor at 71.0 trailed the leaders by five to thirteen points — gaps far larger than the noise at the top. Harbor, at just $0.08 per 1,000 tickets, is almost free. But an average in the low 70s suggests answers managers found less correct, less useful, or less appropriate in tone. For customer-facing drafts, that means more editing, more risk, and more escalations. Any savings on inference could disappear into extra agent time. The practical takeaway is a curve, not a ranking: from Harbor up to Vale, you get steep gains in quality for modest extra cost. From Vale to North, you pay more than twice as much for no measurable gain — at least on this kind of ticket. There were hints of specialization. North did best on billing disputes, where precise policy language matters. Vale led on technical troubleshooting, where step-by-step usefulness counts. Companies heavy in one category might note that, but the researchers did not present these splits as definitive. ## What the test could not tell you Every benchmark has blind spots. First, scale and source. Two hundred tickets can separate good from mediocre, but not parse fine distinctions. All tickets were in English from a single software company. A retailer, airline, or bank with different jargon and stakes might see different results. Second, modalities. All five models performed poorly when a ticket depended on an image attachment — a screenshot of an error, a photo of a receipt. In the real queue, those tickets are common and painful. A text-only test flatters models that would stumble in production. Third, scope. The models only drafted answers. They were not allowed to take agentic actions, use tools, look up accounts, issue refunds, or operate in multiple languages. Modern support AI is increasingly judged on what it can do, not just what it can say. This test measures the saying part. None of that invalidates the core result. For English-language draft generation without tools, top-tier quality is now available at a mid-tier price. For support leaders, the message is pragmatic. Don't buy North because 84.2 is bigger than 83.8. Pilot Vale and North on your own tickets, measure edit rate and resolution time as well as manager scores, and stress-test images and edge cases. If Vale holds up — and this benchmark suggests it might — take the savings and speed. In AI, the winner is rarely the smartest model in the lab. It is the model good enough to deploy at scale.

Z.ai: GLM 5.3: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

Meta: Muse Spark 1.3: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

Z.ai: GLM 5.3: # LedgerLark Wants to Be the Post-Trade Team Small Firms Can't Afford to Hire Behind every trade at an investment firm sits a quiet pile of work: confirmations to check against custodian records, ledgers to match, exceptions to chase down. At small firms, that work often lands on a handful of people who spend their days untangling mismatches — when they'd rather be doing almost anything else. LedgerLark, a startup building AI agents for post-trade operations, is betting that this unglamorous corner of finance is ready for automation. The company has raised a $3.5 million seed round led by Motive Row, and counts five investment firms as paying customers, on annual contracts ranging from $36,000 to $110,000. ## The Problem Found in the Trenches The company's origin story is unusually grounded in operational frustration. Co-founder Mina Batra spent seven years in operations at a regional brokerage, where she says she watched a six-person team spend three days resolving a single trade mismatch caused by a copied account number. That anecdote captures what many small firms know firsthand: post-trade errors are rarely exotic. They're typos, formatting inconsistencies, and data mismatches between systems that don't talk to each other. But they're expensive in aggregate, and the people who resolve them are often the same people needed for everything else. Batra's co-founder, Owen Carr, built reconciliation systems at a large bank — experience that gives the pair complementary views of the same problem: Batra from the side of the firm doing the manual cleanup, Carr from the side of the systems that were supposed to prevent it. ## What the Product Actually Does LedgerLark's agents compare three sources of truth that every firm has, but rarely aligns cleanly: confirmations, custodian records, and internal ledgers. When the agents find discrepancies, they don't silently fix anything. Instead, they propose corrections for human approval. That design choice — a human in the loop, every time — is worth noting. LedgerLark's product does not execute trades or move money. It surfaces and suggests. In a domain where a misplaced decimal can mean real losses and regulatory headaches, the company seems to have concluded that fully autonomous agents would be a hard sell, and possibly a bad idea. ## Early Traction, With Caveats Five paying customers is small, but the contracts aren't trivial. For firms that might otherwise hire another operations person — or stretch an existing team further — a $36,000 to $110,000 annual contract represents a real budget commitment, which suggests the firms aren't just piloting the technology out of curiosity. The company says one customer reduced its unresolved daily exceptions from 70 to 18. If accurate across the customer base, that's a meaningful operational shift, not just a marginal efficiency gain. But it's a single data point, self-reported, and the details of how those exceptions were counted before and after aren't disclosed. It's the kind of claim that will need to hold up across more customers before it says much about the product's general effectiveness. ## The Competitive Landscape Isn't Empty Reconciliation software is not new territory. Established vendors have sold reconciliation tools to financial institutions for years, and large banks have built substantial internal software to solve versions of this problem for themselves. LedgerLark isn't inventing reconciliation; it's arguing that AI agents can do the work more flexibly and cheaply than those incumbent approaches — and that small firms, historically underserved by enterprise vendors with enterprise pricing, are the natural first customers. Whether agents prove reliably accurate on the messy, edge-case-heavy data of real firm operations is still an open question. Post-trade data is exactly the kind of input where AI systems have historically struggled: inconsistent formats, unusual instruments, and errors that look plausible until someone checks carefully. ## The Procurement Reality There's also a slower-moving obstacle: financial institutions, even small ones, tend to have lengthy security and procurement review processes. Selling software into this market takes time, and five customers over the company's life so far suggests LedgerLark is still early in that learning curve. The firms that have signed are presumably the ones that felt enough pain to push through internal reviews. ## What to Watch LedgerLark is tackling a real problem with a sensibly constrained product — agents that propose, humans who approve, no money moved. The founders' operational backgrounds give the effort credibility that generalist AI teams entering finance often lack. The open questions are whether the agent accuracy holds up across more firms and messier data, whether customers renew at these prices, and whether incumbents — or the AI teams at larger vendors — close the gap. The seed round buys time to answer those questions. For now, a handful of small firms are finding out whether the answer is yes.

Meta: Muse Spark 1.3: # Startup Spotlight: LedgerLark Takes On Post-Trade Cleanup Most attention in investing goes to the moment a trade is made. LedgerLark is focused on what happens after. The startup builds AI agents for post-trade operations at small investment firms, a part of finance that rarely makes headlines but can consume hours of staff time when records do not line up. ## An operations problem Co-founder Mina Batra spent seven years in operations at a regional brokerage before starting the company. Co-founder Owen Carr previously built reconciliation systems at a large bank. Batra said she began exploring the idea after watching a six-person team spend three days resolving a trade mismatch caused by a copied account number. It was the kind of error that was small on its own, she thought, but expensive to track down across different systems. That experience points to a broader pattern the founders describe: at smaller firms without large back-office teams, mismatches between what was agreed, what custodians report, and what is recorded internally can sit unresolved while staff check emails, spreadsheets, and PDFs. The question LedgerLark is testing is whether software can do much of that comparison work without adding new risk. ## What the product does — and does not do LedgerLark's agents compare three sources: trade confirmations, custodian records, and internal ledgers. When the numbers differ, the system flags the mismatch and proposes a correction for a person to review and approve. Importantly, the product does not execute trades or move money. A human remains in the approval loop, which reflects how cautious many firms are about automation near client funds. In practice, that means the tool is positioned as an assistant for operations staff rather than a replacement. The team says users still need to understand the underlying records and decide whether a proposed fix makes sense. How reliable those suggestions are across messy, inconsistent data remains an open question, particularly because each firm may format and store information differently. ## Early customers LedgerLark says it has five paying customers, all small investment firms. Annual contracts range from $36,000 to $110,000, suggesting the company is selling into teams where budgets are limited but the pain is direct. As an example of results so far, the company says one customer reduced its unresolved daily exceptions from 70 to 18 after adopting the product. That figure has not been independently verified, and it represents a single customer's experience over an unspecified period. It is useful as an early signal, not as evidence of typical performance. The small customer base means it is too early to know how well the product holds up across different custodians, asset types, or periods of high trading volume. ## Funding, competition, and what to watch The company has raised a $3.5 million seed round led by Motive Row, which it plans to use to develop the product and add customers. It enters a crowded area. Competitors include established reconciliation vendors and internal software built by banks themselves. Many of those existing tools already automate parts of matching, even if users describe them as rigid or difficult to configure. LedgerLark will need to show that its approach handles edge cases more gracefully, rather than just adding a conversational layer on top of similar logic. Another challenge is sales cycles. Financial institutions may require lengthy security and procurement reviews, even for software that does not move money. For small firms, decisions can be faster, but concerns about data access, audit trails, and regulatory expectations remain. For now, LedgerLark's bet is modest in scope: that smaller firms will pay for help with a specific, recurring headache, if the tool fits into existing checks and earns staff trust. Whether that trust builds quickly enough to support a venture-scale business is still to be determined.

Z.ai: GLM 5.3: # Arcwell Launches Relay, an AI Platform for Customer Support Agents Arcwell has introduced Relay, a platform that uses artificial intelligence to analyze customer support conversations and assist agents in real time. Relay does three things: it suggests answers to agents mid-conversation, flags customers who appear likely to cancel, and automatically generates summaries of support interactions. All three functions draw on the same underlying technology — AI that reads transcripts and internal documentation to work out what an agent needs in the moment. That last point is central to Arcwell's pitch. "Support agents spend too much time searching for information that their companies already have," CEO Priya Shah said in a statement. Twelve companies tested Relay over six months. Agents using the platform closed tickets 14% faster, according to Arcwell, though the company did not release results on customer retention — one of the platform's advertised benefits. Relay costs $65 per agent per month. The speed gains come with a data question. Relay requires access to support transcripts and internal help documents to function, raising privacy considerations for potential customers. Arcwell says customer data is encrypted and is not used to train shared models, and companies can choose either a 30-day or 90-day retention period for their data. The launch marks Arcwell's most significant move into the customer service market, where a growing number of companies are applying AI to support workflows — from drafting replies to routing tickets to predicting which customers are at risk of leaving. Whether Relay's early results hold up at scale remains to be seen. The 14% figure comes from a six-month pilot with a dozen companies, and Arcwell has not yet published data on the platform's effect on churn, the problem its cancellation-flagging feature is designed to address. At $65 per agent per month, companies weighing the platform will be doing so on the strength of Arcwell's own numbers.

Meta: Muse Spark 1.3: # Arcwell Launches Relay, an AI Assistant for Support Agents Arcwell has launched Relay, an artificial intelligence platform that examines customer-support conversations and suggests answers to agents. The idea is to cut the searching that bogs down a live chat or ticket. As an agent talks with a customer, Relay reads the exchange, pulls from a company’s past transcripts and guidance, and proposes what to say next. The platform can also identify customers who may cancel and create summaries, reducing the after-call work that piles up between contacts. “Support agents spend too much time searching for information that their companies already have,” CEO Priya Shah said. Shah’s argument will sound familiar to anyone who has watched support work. Agents often juggle a live conversation alongside knowledge bases and old tickets, trying to locate an answer the company has given before. Relay aims to collapse that search into the moment of reply. Arcwell says it tested Relay with 12 companies over six months. Agents using the system closed tickets 14% faster, the company said. It did not release customer-retention results, leaving open how well the cancellation warnings translate into saved accounts. Relay costs $65 per agent each month. Access is the bargain at the center of the product. For Relay to work, a company must give it entry to support transcripts and internal help documents. That breadth is what lets the suggestions sound specific rather than generic, but it also puts sensitive customer conversations and proprietary know-how into a third-party system. Arcwell says customer data is encrypted and is not used to train shared models. Companies can choose a 30-day or 90-day retention period. Whether support teams accept that tradeoff will likely determine Relay’s reception. Faster resolutions are easy to measure and easy to sell internally, while privacy assurances and retention controls tend to decide whether pilot programs expand. Arcwell is betting that combining suggested answers with churn signals and automatic summaries will be enough to make Relay stick in queues where seconds matter.

Z.ai: GLM 5.3: # HelioDock Launches Night Shift Fleet-Monitoring Software, Closes $24M Series B *The company says the tool cut unplanned robot stoppages by 17% during a beta test spanning 310 machines.* HelioDock, a maker of autonomous warehouse robots, launched Night Shift on Thursday, a software product that lets operators monitor their robot fleets remotely. The company also announced a $24 million Series B led by North Coast Ventures, bringing its total funding to $39 million. The software arrives as warehouse operators increasingly run automated fleets across overnight shifts, when fewer staff are on site to handle problems in person. HelioDock positioned Night Shift as a response to that gap, and the funding round signals investor confidence in the market for tools that keep those fleets running. "The typical customer runs fleets overnight with a skeleton crew," CEO Mara Velez said in the announcement. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet." Night Shift allows operators to observe robot activity, diagnose issues, and intervene without being physically present in the facility. The product spent 14 months in a private beta involving nine logistics companies and 310 robots before Thursday's launch. According to HelioDock, beta customers experienced 17% fewer unplanned robot stoppages while using the software. That figure has not been independently verified. Beta testers offered a more mixed picture of how the tool performs in practice. Aaron Pike of Fulcrum Logistics, one of the participating companies, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages — a limitation that speaks to the product's reliance on network infrastructure. When connectivity fails, robots revert to their locally configured behavior, and remote control becomes unavailable. That means facilities in areas with spotty cellular coverage may see the software's core benefit — off-site management — disappear exactly when overnight problems are hardest to resolve. HelioDock's announcement did not address how the company plans to mitigate the issue, though the beta's scale — 310 robots across nine companies — suggests the tool has been tested across a range of warehouse environments and network conditions. Night Shift enters a field with established competition. Relay Robotics and YardOS both offer fleet-management products for automated warehouse operations, and HelioDock did not detail how its software differentiates from those rivals in terms of features or performance. The company's framing instead emphasized the overnight-use case, which the Night Shift name makes explicit. Pricing will be $90 per robot per month, with general availability set for September 15. That puts the cost of monitoring a 50-robot fleet at $4,500 monthly — a figure operators will likely weigh against the labor savings HelioDock implies in its marketing. The Series B funding, led by North Coast Ventures, gives HelioDock $39 million in total capital to expand the product's reach. The company has not disclosed how the new money will be allocated, such as whether it will go toward sales, engineering, or efforts to address the connectivity limitations surfaced during the beta. The 14-month beta period appears to have shaped the product's current form, though HelioDock did not specify what changed between early testing and Thursday's release. Velez's remarks focused on the customer problem rather than technical specifics, and the company did not publish performance data beyond the 17% stoppage-reduction figure. For warehouse operators already running HelioDock robots, the September 15 availability date means Night Shift will be an add-on purchase rather than bundled into existing deployments. For those evaluating autonomous systems more broadly, the product adds another entry to a growing fleet-management category — one whose value, as Pike's experience shows, may depend heavily on where a facility sits on the map.

Meta: Muse Spark 1.3: # HelioDock Launches Night Shift for Remote Monitoring of Warehouse Robots *Private beta across nine logistics companies and 310 robots backs September 15 release priced at $90 per robot per month alongside $24 million funding round* HelioDock, which makes autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely. The release links the software to the company’s funding, business model, competitive setting and testing history. HelioDock also announced a $24 million Series B led by North Coast Ventures, bringing its total funding to $39 million. Night Shift will cost $90 per robot per month and become generally available September 15. Competing fleet-management products are available from Relay Robotics and YardOS. HelioDock says beta customers experienced 17% fewer unplanned robot stoppages, a figure that has not been independently verified. CEO Mara Velez described the use case in terms of overnight response. “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet,” Velez said. The comment points to a situation in which one machine stopped by a pallet requires a person to travel to the building in the middle of the night. Night Shift is presented as a way to monitor the fleet remotely in that type of moment rather than driving across town to look at the robot in person. Night Shift spent 14 months in private beta before Thursday’s launch. The test involved nine logistics companies and 310 robots operating under the remote-monitoring software. HelioDock says those nine beta customers experienced 17% fewer unplanned robot stoppages while using Night Shift. The company is the source of the percentage, which has not been independently verified. The beta covered several hundred robots across a small group of logistics operators over more than a year. That nine-company, 310-robot history is what HelioDock cites as support for the product as it moves toward general availability. Aaron Pike of Fulcrum Logistics, a beta customer, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. When connectivity fails, robots return to their locally configured behavior, according to the company’s description of the limitation. Remote control becomes unavailable at those times. That means the remote monitoring available through Night Shift depends on a working connection. Where strong connectivity is present, Pike reported good performance, while rural LTE outages reduced the usefulness of the system because the robots continued on local settings without remote control. Pricing for Night Shift follows fleet size. At $90 per robot per month, a logistics company would pay that monthly amount for each robot covered, so larger fleets would carry a larger monthly total than smaller fleets operating under the same rate. That rate structure ties monthly spending directly to the number of robots monitored through the software. General availability is set for September 15, the date HelioDock gave for broad sale of the software beyond the private beta group. The Series B adds to the company’s resources around the launch. The $24 million round led by North Coast Ventures brings HelioDock’s total funding to $39 million. North Coast Ventures led the round that produced the new total. Relay Robotics and YardOS are part of the competitive background for the release. Both companies have fleet-management products available that compete with Night Shift for logistics operators overseeing robots. Pike’s experience during the beta captures that operational tradeoff. In facilities with strong connectivity, Fulcrum Logistics found the system worked well, and during rural LTE outages the same system became less useful as robots returned to locally configured behavior and remote control became unavailable.