Head to head: Anthropic: Claude Sonnet 5.5 vs Qwen: Qwen3.8 2.4T A95B

Anthropic: Claude Sonnet 5.5: # Foundry Hall, Milwaukee Startup Accelerator, to Close Oct. 31 MILWAUKEE — Foundry Hall, the nonprofit startup accelerator that has helped launch hundreds of Milwaukee companies, will close its doors on October 31, ending a run that began in 2018. In those years, the organization supported 212 startups. Together, those companies now employ 1,140 people. The closing follows a decision by Blank River Bank, Foundry Hall's largest corporate sponsor, not to renew its sponsorship. The bank supplied 38% of the nonprofit's annual budget. A bank spokesperson said the decision reflects a shift in its community investment priorities and has nothing to do with Foundry Hall's performance. Executive director Camille Ross said the organization spent months hoping another funder would fill the gap. "We kept waiting for the replacement check that never came," she said. Foundry Hall has enough money to operate through October, but not to run another full program. Six employees will lose their jobs. Twelve startups that currently use its workspace will have to find somewhere else to work. For some founders, the loss is hard to measure in numbers. Dev Malik, who started his company through the accelerator, said the building gave his idea its first real footing. "My first investor meeting happened in that building," Malik said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." He said he has been thinking less about his own early days than about the founders who would have come next, the ones who will now start out the way he did, with a laptop and a borrowed table. The twelve resident companies face a more immediate problem. Deputy mayor Susan Cole said the city is talking with building owners and local universities about space for them. She did not say how soon an arrangement might be reached, and no replacement location has been announced. Ross said the staff's focus in the coming weeks is on those companies. She said the team plans to spend its remaining time helping founders find new space and keep their momentum, rather than on marking the end of the program. Her own future, and that of the five other employees whose jobs are ending, has been left mostly unspoken in the office, she said. The accelerator's work has been practical and often unglamorous: introductions, advice, a place to be on a Tuesday morning. The results are harder to miss. Over seven years, companies that began in its rooms hired people, signed leases of their own and, in some cases, outgrew the building altogether. After October 31, the space will be empty. The companies that started there will not be.

Qwen: Qwen3.8 2.4T A95B: # The Last Program Foundry Hall has always sounded like a place where something was about to begin. The nonprofit startup accelerator, opened in Milwaukee in 2018, gave founders a room to practice their pitches, a desk to build from, and a network that made small companies feel less alone. On October 31, it will close. The end did not arrive all at once. It came as a budget gap widened, then became impossible to ignore. Blank River Bank, Foundry Hall’s largest corporate sponsor, provided 38% of the organization’s annual budget and chose not to renew its support. Without that money, the hall has enough to operate through October, but not enough to carry another full program. Six employees will lose their jobs. Twelve startups now using its workspace must leave. Executive director Camille Ross said the staff kept waiting for a replacement gift that never arrived. The sentence captures the quiet exhaustion of the final months: meetings, calls, proposals, and the slow realization that no single rescue was coming. The bank, for its part, described the decision as a shift in community investment priorities and said it was unrelated to Foundry Hall’s performance. That distinction matters to the people inside the building, but it does not change the math. A place can be successful and still be unsustainable. Since opening, Foundry Hall has supported 212 startups, which together employ 1,140 people. Those numbers are more than statistics. They are payrolls, first hires, prototypes, customers, and the kind of local confidence that often takes years to build. Founder Dev Malik remembers what the building made possible. “My first investor meeting happened in that building,” he said. “Before Foundry Hall, I was pitching from the back table of a coffee shop.” The city is trying to soften the landing. Deputy mayor Susan Cole said officials are talking with the building’s owners and local universities about space for the companies currently working at Foundry Hall. Whether those discussions become leases, shared rooms, or temporary arrangements, they cannot replace the ecosystem that gathered there. In its last weeks, the hall is likely to feel both ordinary and charged. There will still be laptops open on tables, whiteboards covered in notes, and founders asking one another how things are going. But the conversations will carry an awareness that this address is about to disappear from their routines. Foundry Hall’s closure is not just the loss of a building. It is the end of a promise made to entrepreneurs who needed a place to start.

Anthropic: Claude Sonnet 5.5: # Arcwell Launches Relay, an AI Tool That Suggests Answers to Support Agents Arcwell has released Relay, a platform that uses artificial intelligence to analyze customer support conversations and suggest answers to agents as they work. Relay does two other things. It can identify customers who may be about to cancel, and it can create summaries. Arcwell is selling the product mainly on speed. The company says 12 companies tested Relay over six months and that agents using it closed tickets 14% faster. That figure is Arcwell's own. The company did not release customer-retention results, so there is no public evidence yet of whether the cancellation-spotting feature helps companies keep customers. Relay costs $65 per agent each month. Arcwell CEO Priya Shah framed the product as a fix for a familiar inefficiency. "Support agents spend too much time searching for information that their companies already have," she said. That premise shapes how Relay works, and it brings a trade-off. To suggest useful answers, the platform needs access to a company's support transcripts and its internal help documents. Together, those records can hold a good deal of sensitive customer information and company knowledge, and handing them to an outside vendor is a decision many customers will want to weigh carefully. Arcwell says it has addressed the concern in two ways. The company says customer data is encrypted and is not used to train shared models, meaning one customer's conversations should not shape the answers another customer sees. Companies can also choose a retention period of either 30 or 90 days. Arcwell has not said how those retention options apply to the transcripts and documents Relay ingests, and the company's claims about encryption and model training come from Arcwell itself. For prospective buyers, the pitch is simple: a measurable gain in ticket speed at a flat per-agent price. The unanswered question is whether Relay's other promises, especially spotting customers likely to leave, produce results that Arcwell is willing to show.

Qwen: Qwen3.8 2.4T A95B: # Arcwell launches Relay to bring AI suggestions to customer support agents Arcwell has announced Relay, an artificial intelligence platform that examines customer support conversations and suggests answers to agents. The release gives Arcwell a product aimed directly at support teams and places the company in the broader industry push to apply AI to customer service work. It also addresses a daily workflow where finding the right information can take a large share of an agent’s time. Relay performs three main tasks. It suggests answers based on support conversations, identifies customers who may cancel and creates summaries of those conversations. By combining those functions, the platform tries to turn transcripts and internal documents into guidance that agents can use while handling cases. Because Relay works from existing support conversations and help documents, its output depends on the material a company already maintains. Arcwell said Relay was tested by 12 companies over six months. Agents using the platform closed tickets 14% faster, the company said. Arcwell did not release customer-retention results, so the performance of the cancellation-warning feature remains undisclosed. CEO Priya Shah said the company built Relay around a familiar problem. “Support agents spend too much time searching for information that their companies already have,” Shah said. Relay costs $65 per agent each month. The subscription pricing makes the platform relevant to companies measuring whether AI can produce quick efficiency gains without adding complexity to existing support operations. The system needs access to support transcripts and internal help documents to generate its suggestions and summaries. That requirement places data handling at the center of the product’s deployment. Arcwell said customer data is encrypted and is not used to train shared models. Companies can choose a 30-day or 90-day retention period, giving them a choice in how long material is kept. The launch frames Relay as both a productivity tool and a data-sensitive system, with Arcwell asking customers to weigh faster ticket handling against the access required to make AI suggestions work inside their support operations.

Anthropic: Claude Sonnet 5.5: C

Qwen: Qwen3.8 2.4T A95B: C

Anthropic: Claude Sonnet 5.5: Going through the draft line by line: 1. `#Flood software moves upstream` violates the headline rule, which requires `# Headline` with a space. **Violation.** 2. `The company is expanding its pilot.` is supposed to be the dek, but it is not italicized. **Violation.** 3. `##Details` violates the section heading rule, which requires `## Heading` with a space. **Violation.** 4. `Read [the announcement](www.example.com/news).` uses a link that is not an absolute HTTPS URL. **Violation.** 5. `* The first pilot begins Monday.` begins with an asterisk instead of a hyphen. **Violation.** 6. `<div>The second pilot begins Friday.</div>` contains raw HTML. **Violation.** The blank lines do not violate anything. **6**

Qwen: Qwen3.8 2.4T A95B: 6

Anthropic: Claude Sonnet 5.5: # HelioDock launches remote monitoring software for warehouse robot fleets *Night Shift left a 14-month beta with a reported drop in stoppages, though one customer says it falters when connectivity does.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely. The company also announced a $24 million Series B round led by North Coast Ventures, bringing its total funding to $39 million. Night Shift is meant to keep managers from having to go to the warehouse whenever a robot gets stuck. HelioDock says beta customers had 17% fewer unplanned robot stoppages, a figure that has not been independently verified. A beta customer said the software was far less useful in facilities that lost cellular service. ## Aimed at after-hours problems CEO Mara Velez framed the product around overnight disruptions. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," she said. Night Shift spent 14 months in a private beta involving nine logistics companies and 310 robots. HelioDock attributes the 17% reduction in unplanned stoppages to that beta group. The company did not provide the baseline against which the reduction was measured, and the figure comes from HelioDock rather than from an outside review. ## Where connectivity matters Aaron Pike of Fulcrum Logistics, one of the beta customers, said Night Shift worked well in facilities with strong connectivity. During rural LTE outages, he said, it became less useful. That reflects how the system is built. When connectivity fails, robots return to their locally configured behavior, and remote control becomes unavailable. In practice, an operator watching from off-site can step in only while the link between the facility and the software holds. For sites with dependable connections, that limit may rarely come up. For those in areas prone to LTE outages, as Pike's experience suggests, the remote tools may be unavailable at the moments they are most needed. ## Pricing and availability Night Shift costs $90 per robot per month and becomes generally available September 15. At that rate, a 100-robot fleet would cost $9,000 a month. ## Funding and competition The Series B, led by North Coast Ventures, brings HelioDock's total funding to $39 million. The company did not provide details in the information released about how it will use the new money. Night Shift enters a market that already has competing fleet-management products, including offerings from Relay Robotics and YardOS.

Qwen: Qwen3.8 2.4T A95B: # HelioDock Launches Night Shift Remote Monitoring for Warehouse Robots *The software emerges from a 14-month beta with nine logistics companies as the company raises a $24 million Series B.* HelioDock on Thursday launched Night Shift, a software service that lets warehouse operators monitor and manage fleets of autonomous robots without being on the floor, the company said. The release pairs a new remote-monitoring product with new funding for the maker of autonomous warehouse robots. HelioDock said Night Shift will become generally available Sept. 15 at $90 per robot per month. The company also announced a $24 million Series B led by North Coast Ventures, bringing its total funding to $39 million. Night Shift spent 14 months in a private beta that involved nine logistics companies and 310 robots, HelioDock said. The company said beta customers experienced a 17% reduction in unplanned robot stoppages. That figure has not been independently verified. The beta included nine logistics companies, giving HelioDock feedback from multiple operators. The participating fleet of 310 robots formed the base for the company’s stoppage claim during the 14-month test. The company framed the launch around reducing the need for on-site visits when robots pause unexpectedly. The scenario described by HelioDock involves a robot unsure of a pallet, an issue that might otherwise require a manager’s presence. CEO Mara Velez said the service is aimed at reducing late-night interventions. “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. HelioDock presented the 17% reduction as a result from the beta period. The claim relies on the company’s own measurements, because the figure has not been independently verified. HelioDock makes autonomous warehouse robots, and Night Shift extends that hardware with a layer of remote oversight. The company’s description centers on operators watching fleet activity from another location, rather than responding in person when a robot halts or requires attention. Night Shift does not replace the robots’ locally configured behavior. It adds a remote layer for monitoring a group of machines, while local configuration remains the fallback when connectivity fails. The service’s dependence on network connections was evident in beta feedback. Aaron Pike of Fulcrum Logistics said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. HelioDock said that when connectivity fails, robots return to their locally configured behavior. Remote control becomes unavailable during such failures. The beta feedback from Fulcrum Logistics highlights where that model can weaken. Pike’s comments show that the software’s usefulness is tied to the quality of the facility’s connection, with rural LTE service posing a challenge. Relay Robotics and YardOS offer competing fleet-management products, giving warehouse operators other sources for similar remote oversight tools. The Series B announcement gives HelioDock additional capital as it moves from beta testing to commercial availability. North Coast Ventures led the round, and the company said the financing brings its total funding to $39 million. The subscription price ties Night Shift’s cost to each robot in a fleet. At $90 per robot per month, the software is sold as an ongoing service rather than a one-time purchase. The September 15 release date follows the private beta period. The $90 monthly fee applies to each robot using Night Shift, so the software cost scales with the size of the fleet. The per-robot rate is the same pricing figure HelioDock disclosed for the general availability release.

Anthropic: Claude Sonnet 5.5: # At the Top of the Support-Bot Rankings, a Cheaper Model Matches the Priciest A new benchmark of five language models on customer-support work found that the best-scoring model was not meaningfully better than the runner-up, which costs about 60% less. ## The headline result Researchers asked five models to draft answers to 200 anonymized customer-support tickets. Model North had the highest average quality score, 84.2. Model Vale was a hair behind at 83.8. The researchers report that this gap was not statistically significant, so the data gives no good reason to call North the better model on quality. The cost gap, by contrast, is large. North's estimated cost is $1.90 per 1,000 tickets, while Vale's is $0.74, about 61% less. Vale also answered 18% faster. At small volumes this hardly matters. At a million tickets a month, the estimated difference is roughly $1,160, and the speed advantage could matter for live chat. For a team choosing between the two, the benchmark points toward Vale as the default unless something specific favors North. ## How the test worked Each model wrote one proposed answer per ticket. Three experienced support managers graded the answers blindly, without knowing which model produced which. Their scores combined three things: factual correctness, usefulness and tone. The blind setup reduces the chance that graders favored a familiar brand or a recognizable writing style. Using experienced managers, rather than another AI model, as judges is a strength, since they know what an acceptable reply looks like in practice. But the single combined score hides trade-offs. A friendly, useful answer with a factual slip and a dry, accurate one could land at similar numbers. ## The rest of the field Below the top two, quality fell as cost fell. Ember scored 79.1 at $0.31 per 1,000 tickets, Kite 76.4 at $0.22, and Harbor 71.0 at $0.08. These are not simply "you get what you pay for" results, because the relationship flattens at the top: North costs more than twice as much as Vale for no measurable quality gain. Further down, the trade-offs are more real. Harbor is the cheapest option by a wide margin, but its score trails the leaders by well over ten points. Whether that is acceptable depends on how much human review each draft will get. Ember may suit teams that want to save money and can tolerate a modest drop. The benchmark does not say how large a gap in score translates into customer complaints or extra agent editing time, so those cutoffs are a judgment call. ## Where models differed, and where they didn't The two leaders had different strengths. North performed best on billing disputes and Vale on technical troubleshooting. That is suggestive, but the per-category samples are necessarily smaller than the full 200 tickets, and the report does not say these category gaps were statistically significant. Treat them as hypotheses to check on your own ticket mix rather than as reasons to route tickets by type. The most consistent finding across the field was a shared weakness: all five models performed poorly when a ticket depended on an image attachment. Screenshots of error messages, photos of damaged goods and images of invoices are routine in support queues. Any team whose tickets often include them should expect a human to handle those cases, or should test image handling separately before relying on any of these models. ## Limits of the evidence Several constraints limit how far the results travel: - **One company, one language.** The tickets were in English and came from a single software company. Results could differ for hardware, finance, healthcare or other fields with different jargon and stakes. - **Drafts only.** The study did not test agentic actions, tool use or multilingual support. Issuing a refund, looking up an account or replying in Portuguese are all outside what this benchmark measured. - **Estimated costs.** The prices are estimates per 1,000 tickets. Real bills depend on ticket length, prompt design and vendor pricing, which changes often. - **One answer per ticket.** Each model produced a single response per ticket, so the results don't show how much a model's quality varies from run to run. - **Small differences.** With 200 tickets and three graders, gaps of a point or so, such as North versus Vale, shouldn't be treated as decisive. The same caution applies to close neighbors elsewhere in the ranking. ## What to take from it The practical lesson is about method as much as about any one model. The most expensive option did not buy a better result at the top of this field, and the cheapest options carried a clear quality penalty. Teams should pilot two or three candidates on their own tickets, include image-dependent cases, and weigh cost and speed alongside quality scores. On this evidence, Vale is the model to beat, and no model here is ready to handle screenshots unsupervised.

Qwen: Qwen3.8 2.4T A95B: # The benchmark’s clearest lesson: Vale matches North’s quality for much less In a new comparison of five language models on customer-support tickets, the clearest finding is not that one model dominated. It is that Model Vale and Model North delivered statistically indistinguishable quality, while Vale cost less and answered faster. Across the test of 200 anonymized support tickets, North averaged 84.2 and Vale averaged 83.8. That 0.4-point gap is small, and the researchers found it was not statistically significant. Readers should not treat North’s slightly higher average as proof that it is the better model. In a blind evaluation with only 200 tickets, this kind of margin is best read as a tie at the top. The stronger case for Vale comes from economics and speed. Vale’s estimated cost was $0.74 per 1,000 tickets, compared with $1.90 for North. That is roughly 61% lower, a large difference for support teams handling high volumes, especially if the models are used for first-pass drafting, triage, or suggested replies. Vale also answered 18% faster than North. In a live queue, faster responses can reduce customer wait times and ease agent workload. Still, the benchmark does not say Vale is universally better. North performed best on billing disputes, while Vale performed best on technical troubleshooting. That suggests routing may matter more than picking a single model for everything. A company whose tickets are dominated by payment issues, refunds, or account charges might get more value from North. A product with heavy troubleshooting demand may find Vale a more natural fit. The cheaper models in the test show the usual cost-quality tradeoff. Ember, Kite, and Harbor all cost less than Vale, but their scores were lower. Harbor was by far the least expensive at $0.08 per 1,000 tickets, yet it scored only 71.0. Whether that is acceptable depends on the task. For simple, low-risk questions, a cheaper model might be good enough. For disputes, security concerns, or angry customers, the drop-off may be too risky. ## How the test worked The researchers evaluated five models on 200 anonymized customer-support tickets, with each model producing one proposed answer per ticket. Three experienced support managers graded the answers blindly, meaning they did not know which model wrote each response. Scores combined factual correctness, usefulness, and tone, which is a sensible mix for support work. A good answer must be accurate, practical, and respectful. Blind grading reduces brand bias, and using three managers helps balance individual preferences. Even so, scoring support quality is partly subjective. The value of a response can depend on company policy, product details, and the customer’s emotional state. Those judgment calls are one reason averages should be treated as guidance, not final verdicts. That design helps, but it also sets boundaries. The test used English-language tickets from one software company, so the results may not transfer to other industries, languages, or customer expectations. The researchers did not test agentic actions, tool use, or multilingual support. In other words, this is a test of proposed written answers, not of models that can query internal systems, change account settings, or take actions inside a support stack. The benchmark also reveals a shared weakness: all five models performed poorly when a ticket depended on an image attachment. That is important because real support often includes screenshots, error messages, receipts, or configuration images. If a team’s workflow depends heavily on visual evidence, these results are not yet a green light for full automation. Human review or specialized image handling remains necessary. ## What buyers should take away The safest reading is that Vale offers the strongest balance of quality, cost, and speed in this test. North remains credible, especially for billing-heavy workloads, but its higher price is hard to justify when the overall quality gap is not statistically meaningful. The lower-priced models may have niches, but they did not match the leaders on average quality. For organizations piloting support AI, the practical next step is not to chase the highest average score. Instead, match the model to the ticket mix, validate performance on your own data, and keep humans in the loop for images, high-stakes disputes, and cases where tone and accountability matter most.

Anthropic: Claude Sonnet 5.5: # Cloudnote debugging page exposed prompt excerpts from 1,842 sessions *The company says it has found no sign anyone but a security researcher viewed the page, but its investigation is not finished.* Cloudnote said an internal debugging page containing excerpts of customer prompts was accessible without a login for about four hours on Monday. Logs show the page included material from 1,842 user sessions. Each excerpt could include up to 200 characters of a prompt. Cloudnote removed the page 19 minutes after a security researcher notified it of the problem. The company says it has found no evidence that anyone other than the researcher accessed the page. That conclusion could change. Cloudnote said its investigation is not complete, and it is not yet clear how many individual users the 1,842 sessions represent. The company has 3.2 million registered users. Cloudnote says the page did not contain full documents, passwords or payment information, which it says were stored elsewhere. Those claims have not been independently verified. Even short excerpts can matter. A few hundred characters of a prompt can include names, work details or other personal information, depending on what a user typed. What the exposed excerpts actually contained is among the questions the investigation has yet to answer.

Qwen: Qwen3.8 2.4T A95B: # Cloudnote Says Debugging Page Exposed Prompt Excerpts From 1,842 Sessions *The four-hour exposure was limited to short fragments, but the company says it has not yet finished determining whether anyone else viewed the data.* Cloudnote, which has 3.2 million registered users, said it discovered Monday that a debugging page containing excerpts from customer prompts was accessible without authentication for four hours. Company logs indicate the page displayed excerpts from 1,842 user sessions, with each excerpt containing as many as 200 characters from a prompt. The exposure underscores how easily internal debugging tools can become public-facing when access controls fail. The company removed the page 19 minutes after a security researcher reported it. Cloudnote says it has found no evidence that anyone other than the researcher accessed the page during the window, but its investigation remains incomplete. Cloudnote has not said how the page came to be exposed. That leaves open whether the fragments were copied, indexed or seen by other visitors. The incident matters because even partial prompts can reveal sensitive intentions, workflows or personal details, depending on what users entered. Cloudnote said full documents, passwords and payment information were stored elsewhere and did not appear on the exposed page. Whether any copies were made is unknown. The company has not said which customers were affected or whether they will be notified.

Anthropic: Claude Sonnet 5.5: # KiteBridge Software Error Sends $2.8 Million in Duplicate Reimbursements to Employees of 91 Companies A software failure at expense-management company KiteBridge generated 13,400 duplicate reimbursements worth $2.8 million on Tuesday. The money came from the accounts of 91 corporate customers. By Wednesday evening, $2.3 million had been recovered, and about $500,000 remained pending. Employees were not charged. No one had money withdrawn from their accounts, and the employers funded the duplicate payments. ## How the failure unfolded KiteBridge, which makes software that helps companies collect and reimburse employee expenses, deployed version 4.2 of its platform at 8:45 a.m. Tuesday. The first customer complaint arrived at 9:12 a.m. The problem began with a timeout at a payment processor, the outside company that actually moves money between accounts. When a processor fails to respond in time, the sender usually cannot tell whether the payment went through, so software typically retries or switches to a backup process. KiteBridge's fallback system did that, but it made a critical mistake. It treated the retries as new reimbursement requests rather than repeat attempts at the same one. Each retry could therefore trigger another payment for an expense that had already been, or was already being, paid. KiteBridge suspended automated payments at 10:03 a.m., less than 80 minutes after the update went live. By then the system had produced the 13,400 duplicates. ## A safeguard that was supposed to prevent this The failure is notable because KiteBridge had said it protected against exactly this kind of error. A changelog published in April said version 4.0 had introduced "exactly-once payment protection," a term for safeguards meant to ensure that a single request results in a single payment, even if the system retries after a failure. It was not clear from the available information whether the 4.2 deployment triggered the processor timeout or only exposed a weakness in how the fallback handled it. ## Questions over communication KiteBridge's public incident page initially described the problem as delayed reimbursements, which suggested that employees were receiving money late rather than twice. The page first disclosed the duplicate payments about six hours after automated payments were suspended. That gap matters for customers. Finance teams that saw only a delay notice would have had little reason to check for overpayments or begin clawing back money. ## Company response CEO Lena Ortiz apologized and said the company would publish a technical report. KiteBridge has not yet said what caused its duplicate-prevention safeguards to fail, how long it will take to recover the remaining funds, or when automated payments will resume. The recovery so far represents roughly 82 percent of the duplicated money. Until the rest is returned, some employers are out funds they never meant to pay, and their employees may still be asked to return reimbursements they received twice. The technical report is likely to answer the central questions: why the protections introduced in April did not stop the duplicates, and why the first public explanation understated the problem.

Qwen: Qwen3.8 2.4T A95B: # KiteBridge Update Creates $2.8 Million in Duplicate Reimbursements A routine software update at expense-management software maker KiteBridge went wrong Tuesday morning, causing the company’s payment system to send 13,400 duplicate reimbursements totaling $2.8 million across 91 corporate customers. The error began after KiteBridge deployed version 4.2 of its platform at 8:45 a.m. The first customer complaint arrived less than half an hour later, at 9:12 a.m., signaling that something in the newly released code was already affecting live payments. According to the company’s account of the incident, the trouble started when a payment processor timed out. In payment systems, a timeout can leave a transaction in an uncertain state: the software may not know whether the payment succeeded, failed, or is still waiting. KiteBridge’s fallback system attempted to handle that uncertainty by retrying the payment. But instead of recognizing those retries as attempts to complete an existing reimbursement, it treated them as new reimbursement requests. The result was multiple payments for the same expense claim. In simple terms, a payment system needs a way to remember that a claim already has a payment attached, even when a request has to be sent again. KiteBridge’s fallback did not make that distinction correctly. The failure overwhelmed normal controls and produced a rapid burst of duplicate transactions. KiteBridge suspended automated payments at 10:03 a.m., stopping the system from issuing additional reimbursements while engineers investigated. By that point, however, the software had already generated 13,400 duplicate payments. The financial impact fell on employers rather than employees. KiteBridge said no employees had money withdrawn from their accounts; the duplicate payments were funded by the companies that use the platform to reimburse workers’ expenses. That distinction matters because expense-management systems typically move money from corporate accounts to employees after claims are approved. When the system creates extra payments, the employer’s funding source, not the employee’s bank account, is affected. KiteBridge’s handling of the incident drew attention as well. Its public incident page initially described the issue as delayed reimbursements. Six hours after the company suspended automated payments, the page was updated to disclose that duplicate payments had occurred. The episode also raised questions about an earlier promise. In an April changelog, KiteBridge said version 4.0 had introduced “exactly-once payment protection,” a safeguard designed to ensure that a payment is processed only one time even if a network delay, timeout, or system restart causes confusion. Version 4.2 was supposed to build on that foundation, but the payment-processor timeout exposed a gap in how the fallback logic interpreted retried transactions. By Wednesday evening, KiteBridge said it had recovered $2.3 million of the duplicate funds. The remaining amount was still pending, leaving finance teams at affected companies to reconcile outstanding payments and confirm that employees had not been overpaid or underpaid. Chief Executive Lena Ortiz apologized for the incident and said KiteBridge would publish a technical report explaining what happened, why the safeguard failed, and what changes will prevent a repeat. For the 91 affected customers, the next step will be a detailed accounting of every duplicate reimbursement and a clear timeline for recovering the rest.

Anthropic: Claude Sonnet 5.5: # QueueZero Launches Proxy, an AI Meeting Assistant That Assigns Work to Software Agents QueueZero launched an AI meeting assistant called Proxy on Wednesday. The product can attend meetings, produce notes and assign follow-up work to software agents. It enters public beta next week. ## A demonstration with an unplanned test During the launch demonstration, a customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy interpreted the remark as an instruction. It created two tasks assigned to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. The exchange showed both the product's ability to turn speech into assigned work and the difficulty of separating casual remarks from requests. ## Safeguards QueueZero says Proxy requires confirmation before an external message is sent. That means a task generated from a meeting should not, on its own, produce an email or other communication to someone outside the organization. Internal tasks, such as those assigned to Wu, can be created automatically. Administrators can disable automatic task creation, which lets organizations require a person to review tasks before they appear in a team's workflow. The distinction matters because meeting conversation is full of hints, jokes and offhand comments. A note-taking tool that records them is one thing. A tool that turns them into assignments is another, because it changes what a meeting transcript means for the people in the room. ## Pricing and availability Proxy costs $30 per user per month. The public beta begins next week. QueueZero is positioning Proxy as more than a transcription product. Many meeting assistants summarize discussions and list action items for people to review. Proxy goes further by assigning follow-up work to software agents, so the output of a meeting can move directly into execution. ## What to watch The demonstration suggests that the main question for customers will be how much interpretation they want an assistant to perform. Automatic task creation can save time when a meeting produces clear commitments. It can also generate clutter when participants speak loosely, or assign work to the wrong person. The confirmation requirement for external messages and the administrator setting for automatic task creation indicate that QueueZero expects some organizations to want tighter control. Those controls will likely matter most in larger deployments, where a single misread comment could create work for many people. In this case, the error was easy to correct. Wu removed the task he did not want and completed the one he did, which is a workable review process for a CEO, if not one every employee could use on a task like "stop talking." The public beta will give customers a chance to test how Proxy handles ordinary meeting language outside a controlled demonstration.

Qwen: Qwen3.8 2.4T A95B: # QueueZero’s AI Meeting Assistant Assigns CEO a Task From Audience Remark QueueZero on Wednesday launched Proxy, an artificial intelligence assistant designed to attend meetings, take notes and turn discussion into follow-up work assigned to software agents. The product, arriving in public beta next week, is priced at $30 per user per month. The company pitched Proxy as a way to reduce the administrative burden that follows workplace meetings. Instead of participants spending time writing summaries or chasing action items, the assistant listens to the meeting, generates notes and creates tasks for colleagues or automated systems. QueueZero says the software is designed to keep human oversight over outward communications. During a launch demonstration, the system showed how quickly it can move from conversation to task management. A customer in the audience said, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy interpreted the remark as an instruction and created two tasks assigned to QueueZero CEO Evan Wu: “Stop talking” and “Send pricing sheet.” Wu laughed, completed the second task and deleted the first. The first task proved easier to delete than to complete. The moment became a brief test case for the product’s boundaries between casual remarks and actionable requests. For QueueZero, it also highlighted the challenge facing AI agents that are expected to be helpful without becoming overly literal. The company says Proxy requires confirmation before an external message is sent. Internal tasks, however, can be created automatically. Administrators can disable automatic task creation, giving organizations more control over how aggressively the assistant converts comments into work items. That settings layer may matter as companies place AI assistants into sensitive environments such as sales calls, customer briefings and executive meetings. A tool that can summarize discussions and delegate tasks may save time, but it must also avoid acting on side comments, jokes or unclear instructions. Proxy enters a crowded market for meeting software, where vendors are racing to pair transcription with automation. The public beta will give QueueZero early feedback on how customers want such systems to interpret speech, assign responsibilities and handle mistakes. For now, the launch event offered an early example of how AI meeting tools can move from listening to acting faster than traditional workplace software.

Anthropic: Claude Sonnet 5.5: # Flood-Risk Software Maker Raincheck Raises $8 Million *The Austin company helps cities decide where to send crews before storms hit, and plans to use the funding to hire engineers and expand into Florida.* Raincheck, an Austin company that makes flood-risk software for local governments, has raised $8 million, it announced today. The 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. Basin Ventures led the seed round, and existing investor Long Mile also participated. Raincheck has raised $10.5 million in total. CEO Ana Flores described the problem the software is meant to solve: > "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. Details of its [forecasting methodology](https://example.com/raincheck-methodology) are available online. 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.

Qwen: Qwen3.8 2.4T A95B: # Raincheck Raises $8 Million to Help Cities Get Ahead of Flood Risk *The Austin software maker maps drainage networks and pairs them with rainfall forecasts, giving local governments a way to target maintenance crews before storms hit.* Raincheck, a flood-risk software company for local governments, said today that it has raised an $8 million seed round. The Austin company maps drainage systems and combines those maps with rainfall forecasts, helping cities decide where to send maintenance crews before a storm arrives. The round was led by Basin Ventures, with participation from existing investor Long Mile. Raincheck has now raised $10.5 million in total. CEO Ana Flores said the company’s customers often already know where flooding occurs, but need help deciding how to respond first. > “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. More information about the company’s approach is available in [Raincheck’s methodology overview](https://example.com/raincheck-methodology). Its forecasts are intended as recommendations, not directives. Emergency managers retain control over deployments and public warnings. Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck said the new funding will help it hire eight engineers and expand into Florida.

Anthropic: Claude Sonnet 5.5: # Startup Spotlight: LedgerLark Wants to Shrink the Back-Office Hunt for Mismatches ## A three-day problem with a small cause Mina Batra spent seven years in operations at a regional brokerage. One episode there stuck with her. A six-person team spent three days resolving a trade mismatch, and the cause was a copied account number. The error itself was trivial. The cost came from finding it. Someone had to compare broker confirmations, custodian records and the firm's internal ledger, work out which one was wrong, and trace the discrepancy to its source. That experience led Batra to found LedgerLark with Owen Carr, who built reconciliation systems at a large bank. The company makes AI agents for post-trade operations at small investment firms, the teams that handle what happens after a trade is agreed. ## What the product does LedgerLark's agents compare three sources: trade confirmations, custodian records and a firm's internal ledgers. When they find a break, they propose a correction. A human then approves or rejects it. The product does not execute trades or move money. That limit looks deliberate. The software reads records and makes suggestions, and people keep authority over any change. For a buyer worried about automated errors, that boundary matters, though it also means the product's value depends on how good the suggestions are and how quickly staff can review them. ## Early traction LedgerLark has five paying customers, with annual contracts ranging from $36,000 to $110,000. The company has raised a $3.5 million seed round led by Motive Row. The company says one customer cut its unresolved daily exceptions from 70 to 18. That is a meaningful change if it holds, but it comes from LedgerLark rather than an independent review. Questions remain about how long the improvement lasted, how the customer defined an exception, and whether trade volumes or other processes changed over the same period. With five customers, any single result carries a lot of weight. ## The competition and the caveats LedgerLark is entering a field that already has players. Established reconciliation vendors sell software for matching records across systems, and some institutions rely on internal bank software built for their own workflows. Carr's background at a large bank suggests he knows the problem from that side, but it also means LedgerLark is competing with tools that customers already trust and have integrated. The company's case is presumably that small firms are underserved by large vendors and in-house systems. That is plausible, since a small firm may lack engineering staff to build its own tooling and may not want a heavyweight deployment. But it is a hypothesis, and the company's growth will test it. Selling to financial institutions brings another constraint. They may require lengthy security and procurement reviews, which can stretch sales cycles and delay revenue even when a customer likes the product. Smaller firms may move faster than banks, but they still handle sensitive data and answer to regulators and clients. Every new customer means asking someone to let a new system read confirmations, custodian data and internal ledgers. ## What to watch Several things will show whether LedgerLark's approach works: - **Accuracy of proposed corrections.** If reviewers often reject suggestions, the time savings shrink. - **Renewals.** The first annual contracts will be a better signal than initial signings. - **Breadth.** Different custodians, asset types and record formats may cause problems that a handful of customers have not yet exposed. - **Competitive response.** Established vendors could add similar AI features, narrowing any gap. ## A narrow bet LedgerLark's bet is specific. It does not promise to automate finance. It aims to cut the time spent finding why records disagree. The origin story, a mismatch that took six people three days, is a modest but familiar kind of operational pain, and that may be a strength. The harder question is whether that pain is acute enough, across enough small firms, to support a durable business against competitors that already have customers and credibility. Five paying customers is a start, and the next year will say more.

Qwen: Qwen3.8 2.4T A95B: # Startup Spotlight: LedgerLark Targets the Paperwork After the Trade LedgerLark is building AI agents for a part of investing that rarely attracts attention: the operational work that begins after a trade is executed. The startup’s software compares trade confirmations, custodian records and internal ledgers, then proposes corrections for human review. It does not execute trades or move money. The idea came from a familiar failure. Co-founder Mina Batra spent seven years in operations at a regional brokerage, where she saw a six-person team spend three days resolving a trade mismatch caused by a copied account number. That kind of error is small in origin but expensive in labor, especially at firms that lack the scale to absorb routine exceptions. Batra co-founded LedgerLark with Owen Carr, who previously worked on reconciliation systems at a large bank. Their combined experience points to a narrow but practical problem: post-trade operations at small investment firms often rely on manual checks, spreadsheets and email threads, even as data volumes grow. LedgerLark’s agents are designed to sit into that workflow. They read and match records across systems, flag discrepancies and draft proposed fixes. A person still approves the correction. This “human in the loop” framing matters in financial operations, where a wrong entry can affect client reporting, compliance or cash positioning. Early traction is modest but specific. The company says five firms are paying customers, with annual contracts ranging from $36,000 to $110,000. In one customer case, LedgerLark says unresolved daily exceptions fell from 70 to 18. That is a meaningful reduction if it holds over time, but it is also a single example. The startup will need to show whether similar results appear across different asset classes, custodians and internal systems. The seed round provides cash, not proof of demand. LedgerLark has raised $3.5 million led by Motive Row. The harder test is commercial: can it win and keep enough small firms without long implementation cycles? Competition is not theoretical. Established reconciliation vendors already serve parts of this market, and larger banks often rely on internal software built for their own complexity. Small investment firms may use consultants, outsourced providers or a patchwork of legacy tools. LedgerLark’s pitch will need to show clear advantages in setup time, cost, usability and accuracy. There is also the usual friction of selling to financial institutions. Even smaller firms may require security reviews, compliance sign-off, data-access controls and procurement steps that slow adoption. These cycles can stretch sales timelines and raise acquisition costs. The company’s focus on post-trade exceptions may reduce implementation risk, since it does not control trading or cash movement, but it still touches sensitive operational data. The central question is whether LedgerLark can become a dependable specialist rather than a feature added by larger platforms. Its agents need to handle the messy edge cases that consume operations teams: partial fills, corporate actions, fee mismatches, custodian delays and data entry errors. If the software can only clean up straightforward exceptions, its value may be limited. Still, the problem appears real enough. Operations staff know that reconciliation failures do not usually come from one dramatic breakdown. They come from many small mismatches that pile up during the day. Automating the first pass of detection and correction could free skilled workers to focus on judgment calls instead of copy-and-paste investigations. LedgerLark is not trying to replace the operations function or promise autonomous finance. It is aiming at a narrower job: making exceptions easier to find, explain and resolve. In a field where accuracy matters more than speed alone, that restraint may be useful. The next phase will be proving that its early numbers are repeatable, and that small firms will pay year after year.

Anthropic: Claude Sonnet 5.5: # 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.

Qwen: Qwen3.8 2.4T A95B: # 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.