Head to head: Meta: Muse Spark Open vs MiniMax: MiniMax M3

Meta: Muse Spark Open: # Cloudnote Debugging Page Left Customer Prompt Excerpts Exposed for Four Hours *Company says a temporary internal page displayed short snippets from 1,842 sessions; investigation into whether anyone else saw them is still underway* Cloudnote left a debugging page displaying excerpts from customer prompts accessible on the open internet without authentication for four hours on Monday, exposing content from user sessions until it was taken down. According to logs reviewed by the company, the page included snippets of up to 200 characters from 1,842 user sessions, a small fraction of its 3.2 million registered users. Full documents, passwords and payment information were stored separately and were not present on the page, Cloudnote said. The exposure matters because prompts can contain sensitive personal, business or creative material that users expect to remain private, even in truncated form. Security researchers say debugging interfaces are a common source of accidental leaks when left enabled outside internal testing environments. Even brief excerpts, experts note, can reveal identifying details if a prompt began with names, addresses or proprietary instructions. Cloudnote said it removed the page 19 minutes after being notified by an outside researcher and has so far found no evidence that anyone other than that researcher accessed it. The company said its investigation is still underway, leaving it unknown whether the excerpts were viewed, copied or misused elsewhere. The findings will determine whether Cloudnote must notify affected users or regulators and what safeguards it adds to prevent similar exposures.

MiniMax: MiniMax M3: **Cloudnote Exposed User Prompt Excerpts on Unauthenticated Debugging Page** *A note-taking service says a four-hour window on Monday revealed short snippets from roughly 1,840 user sessions; full documents and credentials were stored separately and were not affected.* Cloudnote, a note-taking platform with 3.2 million registered users, accidentally left an internal debugging page accessible without authentication on Monday, the company confirmed Wednesday. The page contained excerpts from 1,842 user sessions, each snippet running up to 200 characters, according to logs reviewed by the company. The exposure lasted roughly four hours before a security researcher alerted Cloudnote. The company removed the page 19 minutes after being notified, a spokesperson said. Because the page was intended for engineering use during troubleshooting, it bypassed the authentication that guards the rest of the service. The incident matters because the snippets were drawn from real user prompts and could contain personal reflections, business notes or other private material typed into the app. Even short excerpts can reveal sensitive details when placed in context or stitched together. Cloudnote said it had found no evidence that anyone other than the notifying researcher accessed the page. The company has not completed its investigation, and it remains unclear how the page became publicly reachable or whether any additional data was viewable beyond the prompt snippets. Full documents, passwords and payment information were stored on separate systems and were not present on the debugging page.

Meta: Muse Spark Open: C

MiniMax: MiniMax M3: C

Meta: Muse Spark Open: # HelioDock Introduces Night Shift to Move Overnight Robot Oversight Off Site *Warehouse robotics firm ties remote fleet-monitoring software to new $24 million Series B ahead of September launch* HelioDock on Thursday launched Night Shift, software that lets operators monitor fleets of its autonomous warehouse robots remotely, a move aimed at keeping facilities running overnight without requiring managers to return to the building when a robot encounters an obstacle. The debut extends HelioDock beyond selling robots themselves and into the operational work of keeping them moving through off-hours when warehouses often run with thin staffing. Night Shift is set to become generally available Sept. 15 at $90 per robot per month, the company said. HelioDock disclosed the launch alongside a $24 million Series B led by North Coast Ventures, which it said brings its total funding to $39 million. Night Shift emerged from a 14-month private beta that involved nine logistics companies and 310 robots, according to HelioDock, giving the company a controlled setting in which to test remote supervision across multiple warehouse environments. HelioDock said beta customers experienced 17% fewer unplanned robot stoppages while using the software. That figure has not been independently verified, and HelioDock has not said whether it anticipates comparable performance for customers outside the beta. CEO Mara Velez described the use case in practical terms for overnight operations. “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 $90 monthly per-robot price point places Night Shift as a recurring software charge layered on top of HelioDock’s hardware, with costs scaling directly with fleet size. For an operator running dozens or hundreds of units, the fee would compound quickly, though HelioDock has not disclosed expected adoption rates or whether existing customers intend to subscribe after the beta period. The Sept. 15 general availability date gives current users and prospective buyers a window before the product shifts from testing to open sale. Beta feedback highlighted both the appeal and the constraints of moving oversight off site. Aaron Pike of Fulcrum Logistics, one of the beta participants, 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 and remote control becomes unavailable. The company has not detailed how that local behavior differs from supervised operation or how long systems typically take to reconnect. Those connectivity limits shape where Night Shift may prove most reliable. Facilities with robust indoor wireless coverage would retain continuous remote visibility, while sites that depend on cellular links in rural areas could face gaps in which operators lose the ability to intervene. Because robots fall back to local programming, they do not stop entirely if a link drops, but the remote layer that HelioDock is selling ceases to function until service resumes. Night Shift arrives in a segment where alternatives already exist. Competing fleet-management products are available from Relay Robotics and YardOS, offering warehouse operators other options for monitoring automation. HelioDock did not outline feature-by-feature differences with those systems, nor did it share data on how its claimed reduction in stoppages compares with outcomes reported by users of rival platforms. The concurrent $24 million financing gives HelioDock additional resources as it prepares for wider distribution. Led by North Coast Ventures, the Series B lifts the company’s total funding to $39 million, HelioDock said, with no other investors named in the announcement. The company has not specified how it will allocate the new capital between software development, manufacturing, or customer support for the rollout.

MiniMax: MiniMax M3: # HelioDock Rolls Out Remote Robot Monitoring Tool, Closes $24 Million Series B *Warehouse robotics startup pairs Night Shift launch with new funding as it pushes into a crowded fleet-management market.* HelioDock, the maker of autonomous warehouse robots, on Thursday launched Night Shift, a software platform that allows operators to monitor and intervene with robot fleets from outside the facility. The company simultaneously announced a $24 million Series B led by North Coast Ventures, bringing its total funding to $39 million. The dual announcements underscore HelioDock's effort to grow beyond hardware and build a recurring software business in industrial automation, a category where fleet-management tools are becoming as important as the robots themselves. Night Shift is designed to let supervisors watch live robot activity, receive alerts when machines stall or behave unexpectedly, and take remote action to redirect a confused unit without driving to the site. The software spent 14 months in a private beta across nine logistics companies operating a combined fleet of 310 robots. According to HelioDock, those beta customers experienced 17% fewer unplanned robot stoppages after deploying Night Shift. The company did not share the methodology behind the figure, and it has not been independently verified. For warehouse operators, even modest reductions in unplanned downtime translate into measurable savings, since every stalled robot can ripple through picking, packing and shipping schedules. HelioDock CEO Mara Velez framed the product around a recurring pain point in the industry: "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet." That scenario is precisely what Night Shift is built to address. The platform streams telemetry from each robot and gives a remote operator a live view of the fleet, with the ability to diagnose or redirect a stalled unit in real time. HelioDock describes the tool as a complement to on-site supervision rather than a replacement for it. Feedback from one beta customer points to a more complicated picture. Aaron Pike, who works with Fulcrum Logistics, said Night Shift worked well in facilities with strong network connectivity, but its usefulness dropped sharply at rural sites during LTE outages. Pike's experience highlights a question that sits at the heart of any remote fleet-management product: what happens when the network itself fails? According to HelioDock, robots revert to their locally configured behavior when connectivity drops, and remote control becomes unavailable until the link is restored. The company characterizes the fallback as a deliberate safety design, intended to ensure that robots do not act on stale or missing commands. The trade-off, however, is that an operator hoping to intervene remotely may have to wait for cellular service to return before any action is possible — a gap that could matter most in the rural and edge-of-network facilities where on-site help is hardest to dispatch. Night Shift enters a market that already includes established competitors. Relay Robotics and YardOS both sell fleet-management software aimed at warehouse and yard automation customers, and HelioDock will need to convince operators that its year-plus beta and tight integration with its own robot hardware justify either switching platforms or running Night Shift alongside an incumbent system. Night Shift will be priced at $90 per robot per month and is set to become generally available on September 15. The Series B round led by North Coast Ventures is intended to fund continued product development and an expansion of the sales team, according to the company.

Meta: Muse Spark Open: # $2.8 Million in Duplicate Payments Leaves Employers Chasing Reimbursements After KiteBridge Update Fails **A routine software deployment turned into a costly scramble for 91 companies after promised payment protections failed to stop thousands of duplicate transactions.** An expense-software glitch that flooded employee bank accounts with $2.8 million in duplicate reimbursements this week has left dozens of businesses trying to recover funds and raised questions about safeguards that were supposed to make such an error impossible. KiteBridge, which makes expense-management software used by companies to reimburse workers for travel, meals and other costs, said the incident generated 13,400 duplicate payments on Tuesday and affected 91 corporate customers. No money was withdrawn from employees' own accounts; the duplicate funds were drawn from employer accounts that finance reimbursements, leaving companies to recoup overpayments from staff who had already been paid twice. The sequence began on Tuesday morning when KiteBridge deployed version 4.2 of its platform at 8:45 a.m. According to the company, a timeout from its payment processor then triggered a fallback system. Instead of recognizing that the original payments were still processing and pausing or deduplicating them, the fallback treated each automatic retry as an entirely new reimbursement request. The result was a second payment for each transaction caught in the loop. The first customer complaint arrived at 9:12 a.m., flagging that employees had received the same reimbursement twice. KiteBridge said it suspended all automated payments at 10:03 a.m., which stopped further duplicates. By that point the total had already reached $2.8 million. The failure has drawn particular scrutiny because of a protection KiteBridge had previously advertised. In April, the changelog for version 4.0 said the company had introduced "exactly-once payment protection." The concept is standard in financial software: each payment is assigned a unique identifier so that even if a connection drops or a processor times out and the system retries, the transaction can only be settled once. On Tuesday, that protection appears to have failed or been bypassed when the fallback system created new requests instead of retrying the existing ones with the same identifiers. By Wednesday evening, KiteBridge said $2.3 million had been recovered through payment reversals and coordination with banks and employers. The remainder was still pending recovery. The company has not said how long full recovery is expected to take or whether any funds already spent by recipients may be unrecoverable. The company's public communication during the incident has also been criticized. Its public incident page initially described the problem only as delayed reimbursements, without mentioning that duplicates had been issued. The page was not updated to disclose the duplicate payments until about six hours after automated payments had been suspended, a gap that left some customers uncertain about what had happened to money sent to employees. Chief Executive Lena Ortiz has apologized for the outage and said the company will publish a technical report detailing what went wrong. She said KiteBridge takes responsibility for the failure and is working directly with affected customers to trace and reverse erroneous transactions while it reviews its deployment and failover procedures. For KiteBridge, which handles high volumes of routine but financially sensitive reimbursements, the incident tests a core promise to customers: that automation will reduce manual errors without introducing systemic ones. Until the company explains how its exactly-once safeguard was circumvented and what will change in its fallback logic and testing, the episode is likely to remain a cautionary example of how a single timeout, paired with flawed retry handling, can multiply into millions in unintended payouts.

MiniMax: MiniMax M3: # KiteBridge Software Glitch Floods Corporate Accounts With $2.8 Million in Duplicate Reimbursements A routine Tuesday morning software update from expense-management firm KiteBridge spiraled into a $2.8 million accounting headache for 91 corporate customers, after a payment-processor timeout caused the platform to issue roughly 13,400 reimbursements it had already paid. The episode began at 8:45 a.m. Tuesday, when KiteBridge pushed version 4.2 of its expense-management software into production. The first customer complaint arrived in the support queue at 9:12 a.m. — less than half an hour later. ## How a Retry Became a Duplicate According to details that surfaced in the days that followed, the trigger was a timeout in the third-party payment processor that KiteBridge uses to disburse reimbursements on behalf of its corporate clients. When that processor failed to acknowledge a transaction in time, KiteBridge's fallback system — designed to retry payments that appear to have failed — treated each retry as a brand-new reimbursement request instead of as a re-attempt of the original. In a properly designed retry path, the system would re-submit the transaction using a unique identifier so that the processor could recognize it as the same payment. Because KiteBridge's fallback logic did not preserve that identity, every retry was processed as if it were the first one. The result: 13,400 duplicate payments totaling $2.8 million, generated in roughly the space of an hour. KiteBridge suspended its automated payment pipeline at 10:03 a.m., about 80 minutes after the deployment and just under an hour after the first complaint reached the company. ## Who Paid, and What Was Recovered The duplicated money did not come out of employees' pockets. KiteBridge's model works the other way around: employers fund the reimbursements that flow through the platform to their workers. The 91 affected corporate customers were the ones who saw two — and in some cases more than two — debits for the same batch of expense claims. By Wednesday evening, KiteBridge said it had clawed back $2.3 million of the $2.8 million. The remainder was still pending. ## "Exactly-Once" Promises Meet the Fallback Path The failure is especially uncomfortable for KiteBridge because of what the company said about an earlier release. An April changelog entry for version 4.0 announced that the update had introduced "exactly-once payment protection" — industry shorthand for a guarantee that a transaction will be processed one and only one time, no matter how many times the system retries it. Tuesday's outage suggests that guarantee did not hold up when the fallback path was activated, the very scenario such protections are meant to cover. ## A Slow Public Acknowledgment KiteBridge's public incident page also drew criticism for the pace of its disclosure. In the hours immediately after the suspension, the page described the issue as delayed reimbursements — a framing that did not match what customers were seeing in their own bank records. The duplicate-payment nature of the incident did not appear on the incident page until roughly six hours after automated payments had been halted, leaving corporate customers to reconstruct the scale of the problem from their own transaction logs. ## What Comes Next CEO Lena Ortiz apologized and pledged to publish a technical report detailing the root cause and the company's remediation plan. For an industry that increasingly relies on automated disbursements through third-party processors, the episode is a reminder of how a single missed acknowledgement can cascade into thousands of duplicate payments — and of how the gap between a marketing promise and a fallback path can quietly become a $2.8 million lesson.

Meta: Muse Spark Open: # The Doors Are Closing at Foundry Hall *After seven years and more than 200 startups, the west side incubator is preparing for its final cohort.* Foundry Hall, the nonprofit startup accelerator on Milwaukee's near west side, will close on October 31, ending seven years as a low-cost launchpad for early-stage companies. The organization, which opened in 2018, has enough money to stay open through October but not to run another full program, its leaders said. The closure will eliminate six jobs and force the 12 startups currently working from its building to find new space. The hall never charged market rent, keeping its doors open to founders who could not afford a traditional office. The immediate cause was the withdrawal of its largest sponsor. Blank River Bank, which supplied 38% of Foundry Hall's annual budget, declined to renew its sponsorship this year. The bank said the decision reflected a shift in its community investment priorities rather than any dissatisfaction with Foundry Hall's performance. Executive Director Camille Ross said the team spent months trying to replace the funding, exploring grants, smaller donors and cost cuts before concluding another cycle was unsustainable. “We kept waiting for the replacement check that never came,” she said. In its seven years, Foundry Hall supported 212 startups, companies that now collectively employ 1,140 people. Many arrived with little more than an idea and left with incorporation papers, prototypes, or, in some cases, first checks from investors. For Dev Malik, whose logistics-software company was part of the 2021 cohort, the hall was the first place that felt like a proper office. “My first investor meeting happened in that building. Before Foundry Hall, I was pitching from the back table of a coffee shop,” he said. He now employs eight people. Stories like his are part of why the news has landed heavily in Milwaukee's small-business community. The accelerator did not promise unicorns, but it offered desks, mentorship, and a shared sense of momentum that is hard to replicate. The quiet impact is visible in the building itself, where whiteboards still carry half-erased pitch decks and calendar invites stretch to the end of October. City officials said they are talking with building owners and local universities about temporary space for the displaced startups, though no specific plan has been announced. For now, the hall's calendar runs through October: final workshops, final demo night, final leases. What remains is less a dramatic ending than a steady winding down, founders packing boxes while trying to plan what comes next. Ross said the staff is focused on helping founders transition, even as they prepare for their own exits.

MiniMax: MiniMax M3: # Foundry Hall to Close Its Doors After Seven Years The doors will close on October 31. Foundry Hall, a nonprofit startup accelerator that has nurtured 212 early-stage companies since opening in 2018, will wind down operations at the end of next month, executive director Camille Ross confirmed this week. The organization plans to spend its remaining weeks helping the founders inside its walls land somewhere soft. The closure comes down, in the end, to a single sponsor. Blank River Bank, which supplied 38 percent of the accelerator's annual budget, declined to renew its partnership earlier this year. According to a spokesperson for the bank, the decision reflects a shift in the institution's community investment priorities and is unrelated to how Foundry Hall has performed. Without that anchor funding, the organization has enough cash on hand to honor its commitments through October but cannot sustain another full program cycle. For Ross, the months since the bank's decision have unfolded like a slow, quiet kind of waiting. "We kept waiting for the replacement check that never came," she said. In the meantime, the staff has tried to stretch each program, each cohort, and each dollar as far as it would go. The ripple effects are concrete and immediate. Six employees will lose their jobs when the accelerator closes. Twelve startups currently sharing workspace inside Foundry Hall will need to find somewhere else to operate by November. Across the program's lifetime, the 212 companies that passed through its doors have collectively grown to employ roughly 1,140 people in the Milwaukee area—a track record that supporters say makes the loss feel disproportionate to a single line item on a bank's annual community report. For some founders, the building represented the first time an idea felt like it might actually become a company. Dev Malik, whose logistics software startup was among the early cohorts, recalled what the space meant when he was still pitching alone. "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 now trying to absorb the shock of the closing. Deputy mayor Susan Cole said her office is in conversation with building owners and nearby universities about transitional space for the companies still working inside Foundry Hall. What becomes of the accelerator itself—an organization that, by its own accounting, helped turn kitchen-table ideas into payroll-generating businesses—remains an open question. A handful of national foundations have been approached about stepping in, Ross said, but none have committed. For now, the calendar moves toward Halloween, and a building that once felt like the beginning of something will quietly go dark.

Meta: Muse Spark Open: # Raincheck Raises $8 Million to Help Cities Fix Flood Risks Before Storms _Austin startup's software maps drainage systems and forecasts rainfall to guide maintenance crews, with plans to expand into Florida_ Raincheck, an Austin-based company that makes flood-risk software for local governments, has raised $8 million in a seed funding round announced today. The company maps municipal drainage systems and combines those maps with rainfall forecasts to predict where flooding is most likely. Cities use the software to decide where to send maintenance crews to clear blocked drains, repair pumps and address unfinished work before a storm arrives. The company has published [details on its forecasting methodology](https://example.com/raincheck-methodology). The round was led by Basin Ventures, with participation from existing investor Long Mile. The new investment brings Raincheck's total raised to $10.5 million. CEO Ana Flores said the platform is designed to prioritize pre-storm maintenance: > “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, saying some contracts have not been publicly announced. The company's forecasts are provided as recommendations, with emergency managers retaining control over deployments and public warnings. Raincheck competes with other flood-planning vendors including HydroMap and municipal engineering contractors. The company said the new funding will allow it to hire eight engineers and expand into Florida.

MiniMax: MiniMax M3: # Raincheck raises $8 million to expand flood-risk software for cities *The Austin-based startup builds drainage maps and rainfall forecasts that help municipalities decide where to send maintenance crews before a storm.* Raincheck, an Austin-based company that builds flood-risk software for local governments, announced today that it has raised $8 million in seed funding. The company maps municipal drainage systems and combines those maps with rainfall forecasts, allowing cities to decide where to send maintenance crews before a storm arrives. The seed round was led by Basin Ventures, with existing investor Long Mile also participating. Raincheck has raised $10.5 million to date. > "Our customers usually know which streets flood," CEO Ana Flores said. "The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives." [Read more about Raincheck's forecasting methodology.](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. ## Competition 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 Open: 6

MiniMax: MiniMax M3: 6

Meta: Muse Spark Open: # Arcwell Launches Relay, an AI Assistant for Customer Support Teams *The platform drafts replies, flags potential cancellations and summarizes interactions — promising faster resolutions without replacing agents.* Arcwell has introduced Relay, a platform that uses artificial intelligence to examine customer support conversations and propose answers its human agents can send. For support managers, the promise is less toggling between systems and more consistent answers drawn from material companies have already produced. The system surfaces suggested replies in the flow of work, flags customers who appear likely to cancel, and creates concise summaries of each interaction. That combination is central to Arcwell's pitch. Suggested answers aim to shorten response times. Churn signals are meant to give teams earlier warning before a customer leaves. Summaries are intended to cut the administrative work that follows a resolved ticket. Arcwell said it tested Relay with 12 companies over six months. During that period, agents using Relay closed tickets 14% faster, according to the company. It did not release data on whether the tool helped those companies retain more customers. The platform will be sold at $65 per agent per month. The pricing puts Relay in line with other seat-based support tools, leaving managers to weigh time saved against per-agent cost. Chief executive Priya Shah said the product was built around a familiar bottleneck: “Support agents spend too much time searching for information that their companies already have.” Using Relay comes with data trade-offs common to AI support tools. To function, it requires access to support transcripts and internal help documents. Arcwell says customer data is encrypted and is not used to train shared models. Companies deploying Relay can choose a 30-day or 90-day retention period for that data. How teams balance those efficiency gains against data access will likely determine whether Relay moves beyond pilot into everyday use.

MiniMax: MiniMax M3: # Arcwell's Relay Uses AI to Coach Customer Support Agents in Real Time Arcwell is rolling out a new artificial-intelligence platform called Relay that listens to live customer support conversations and recommends responses to agents handling them. The platform does more than suggest replies. It can flag subscribers who appear likely to cancel and generate post-conversation summaries, two capabilities aimed at lightening the administrative load that has piled up on support teams as ticket volumes have grown. Relay was tested over six months by 12 companies, according to Arcwell. Agents using the tool closed tickets 14 percent faster than those who were not, the company says. Arcwell did not share figures on whether customers flagged as likely to leave were retained at higher rates. The software lists at $65 per agent per month. "Support agents spend too much time searching for information that their companies already have," CEO Priya Shah said in a statement. The pitch is straightforward: pull answers from a company's existing knowledge base and surface them at the moment an agent needs them. Relay enters a market that has become one of the more crowded corners of enterprise software. A growing number of vendors have spent the last two years weaving similar agent-assist features into their helpdesk products, and Arcwell is positioning Relay as a more focused alternative built specifically around real-time conversation analysis rather than as an add-on to a broader customer relationship platform. That focus comes with privacy trade-offs. The platform needs access to full support transcripts and internal help documents to function, raising familiar questions about how AI vendors handle sensitive customer data. Arcwell says the information is encrypted in transit and storage and is not used to train shared models. Customers can choose to retain transcripts for either 30 days or 90 days, a setting that companies in regulated industries, such as healthcare or financial services, will want to examine closely before signing on. The pricing puts Relay in the middle of the market, more expensive than the basic AI features bundled into larger helpdesk subscriptions but cheaper than a custom enterprise rollout.

Meta: Muse Spark Open: # 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.

MiniMax: MiniMax M3: # 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 Open: # QueueZero Launches AI Meeting Assistant Proxy **San Francisco —** QueueZero on Wednesday introduced Proxy, an AI meeting assistant designed to attend meetings on behalf of users, generate notes and delegate follow-up work to connected software agents. The announcement, made at the company's product showcase in San Francisco, was intended to demonstrate how Proxy handles live conversation. During the demonstration, Chief Executive Evan Wu was presenting the assistant's capabilities when a customer in the audience said, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy, which was transcribing the session, interpreted the comment as an instruction. The system subsequently generated two internal tasks and assigned both to Wu: “Stop talking” and “Send pricing sheet.” Wu acknowledged the tasks on stage, laughed, and continued the presentation. He later completed the second task by distributing the pricing information to attendees and deleted the first. QueueZero said the incident illustrates both the responsiveness of the system and the safeguards built around it. According to the company, Proxy is designed to require explicit confirmation from a user before it sends any external message on their behalf. Internal tasks, such as to-do items, calendar holds, or assignments to integrated software agents, can be created automatically without confirmation to reduce manual coordination. QueueZero said administrators have the option to disable automatic task creation entirely or to require approval for all actions. QueueZero, founded in 2022, has focused on workplace productivity tools that connect large language models to enterprise software. Proxy represents its first product designed to participate actively in meetings rather than simply record them. The company said early enterprise testers have used the assistant to draft recaps, update customer records and initiate internal workflows without manual input. The assistant integrates with calendar platforms and workplace suites including Slack, Microsoft Teams and Google Workspace, the company said. Access during the beta will be available to organizations on QueueZero's waitlist, with broader availability expected later this year. The company positioned Proxy as part of a broader effort to automate routine meeting follow-up. In addition to transcription and summarization, the assistant can assign action items to downstream agents that handle scheduling, documentation and project updates. Proxy will be priced at $30 per user per month and is scheduled to enter public beta next week.

MiniMax: MiniMax M3: # QueueZero's AI Assistant Turns Audience Quip Into Real Task for Its Own CEO QueueZero unveiled its artificial intelligence meeting assistant, Proxy, on Wednesday, pitching the product as a way to offload note-taking and follow-up work to autonomous software agents. During the live demonstration at the company's product launch event, an audience remark aimed at chief executive Evan Wu was treated by the system as a legitimate command, producing two tasks assigned to the CEO himself. Proxy is designed to listen in on meetings, generate structured notes, and convert spoken requests into follow-up tasks routed to either human teammates or connected software agents. The product is one of several recently launched "AI chief of staff" tools that promise to translate workplace conversation into action without manual input. The unexpected moment came when a customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy, parsing the remark as an actionable instruction, created two internal tasks assigned to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first before continuing the presentation. QueueZero said the behavior reflects how internal task creation is intended to work. Unlike outbound communications, which the company says require explicit user confirmation before any external message is sent, internal tasks can be generated automatically based on meeting context. Administrators can disable automatic task creation entirely, and Wu said the feature is configurable per workspace. The incident drew laughter from the audience and quickly circulated on social media, where several attendees posted short clips. Wu acknowledged the moment on stage and used it to illustrate the difference between automated internal workflows and externally facing actions. He also noted that Proxy is designed to flag ambiguous instructions for human review, though internal task creation can happen without that step. Proxy will enter public beta next week and is priced at $30 per user per month. QueueZero, founded in 2022, has positioned the product for engineering, sales and operations teams that already rely on connected software agents to move work forward. Whether customers treat automatic task creation as a feature or a hazard may shape how aggressively the company markets the capability.

Meta: Muse Spark Open: # The $1.16 Question: Why the Priciest Support AI Didn't Win If you buy AI for customer support, the default assumption is that a higher price buys a better answer. A new benchmark built on real tickets suggests that assumption no longer holds at the top — and that buyers should be looking at speed and specialty, not just a single score. ## The finding that matters Researchers had five leading language models each draft answers to the same 200 anonymized customer-support tickets. Three experienced support managers then graded those answers blindly, judging them on factual correctness, usefulness, and tone. The result at the top was a statistical tie. Model North averaged 84.2 and Model Vale 83.8 on the combined 100-point scale. The researchers report that difference is not statistically significant, meaning we cannot confidently say North is actually better. In practice, they performed indistinguishably on overall quality. What was distinguishable was everything else. Vale costs an estimated $0.74 to handle 1,000 tickets; North costs about $1.90 for the same volume. Vale was also about 18% faster to respond. For a team handling tens of thousands of tickets a month, that is not a rounding error. It is the difference between paying a premium for no measurable gain and getting the same overall quality cheaper and quicker. That is the clear story in an otherwise noisy table: at the high end, price and performance have decoupled. ## What that means if you're buying The temptation is to crown an overall winner and move on. The benchmark argues for a more practical question: where do you need strength? North edged ahead on billing disputes, the kind of tickets that hinge on policy language, proration, and refunds. Vale led on technical troubleshooting, where step-by-step accuracy matters. Neither lead should be read as dominance — these are subcategory trends inside a tied overall result — but they point to fit. A software company drowning in “why won’t this integrate?” tickets might value Vale’s pattern; a subscription business living in chargeback threads might lean North. Below that top tier, the trade-off becomes real. The three more affordable models scored noticeably lower overall, clustering in the 70s rather than the mid-80s, while pushing costs down to well under $0.35 per 1,000 tickets — and in one case to just $0.08. The gap between those models and the leaders *is* meaningful, unlike the gap between the leaders themselves. You save a lot, but customers would likely notice the drop in correctness and tone. For triage, internal drafts, or low-risk FAQs that might be acceptable; for final customer-facing answers without human review, it is a gamble. ## How the test was built — and where it breaks The methodology is refreshingly close to how support teams actually work. Using 200 real tickets from one software company, rather than synthetic questions, forces models to deal with vague phrasing, incomplete context, and frustrated tone. Having three veteran managers grade blindly and weighting correctness, usefulness, and tone together is a better proxy for customer experience than a pure factuality quiz. But that realism comes with boundaries you should not ignore. First, all five models struggled when the ticket depended on an image attachment — a screenshot of an error, a photo of a bug. Text-only answers in those cases were consistently rated poor. If your queue is heavy on “see attached,” this benchmark does not prove any of these models are ready to work solo. Second, the tickets were in English and came from a single company’s product and customer base. Writing quality, terminology, and policy questions will not generalize perfectly to, say, multilingual retail support or regulated healthcare inquiries. Finally, the study measured one-shot answers. It did not test agentic actions, tool use, or multilingual support — no refunding an order, looking up an account, or continuing a multi-turn diagnosis. Modern support stacks often live or die on those abilities, and this data says nothing about them. ## The bottom line For leaders deciding where to invest, the lesson is not “Vale is the best model.” It is that paying more than double for an undetectable 0.4-point gain is hard to justify when the cheaper, faster option matches overall quality and even shows a relative strength in technical help. Start with that parity, then choose by ticket type, latency needs, and budget for human review. And if images or autonomous actions are core to your workflow, treat all scores as incomplete — because for those jobs, the benchmark found no winner at all.

MiniMax: MiniMax M3: # The Hidden Winner in Customer-Support AI Benchmarks When five language models were tested head-to-head on 200 anonymized customer-support tickets, the most expensive model didn't win. In fact, it barely beat the cheapest serious contender — and the gap was small enough that statisticians called it a draw. ## The Headline Finding Model **North** finished first with an average score of 84.2, but **Vale** was right behind at 83.8. That 0.4-point difference was not statistically significant, meaning a reasonable observer should treat them as tied for the top spot. Where Vale pulls dramatically ahead is on price: $0.74 per 1,000 tickets, versus $1.90 for North. Vale is also 18% faster. For most support teams, that isn't a close call. The pattern continues down the leaderboard. **Ember**, **Kite**, and **Harbor** trailed with scores of 79.1, 76.4, and 71.0 at $0.31, $0.22, and $0.08 per 1,000 tickets respectively. Quality scales loosely with cost at the top of the market, then collapses well before the cheapest tier — Harbor is roughly 13 points behind North for less than 5% of the price. There is a real quality ceiling somewhere around the Ember mark, and paying less than that gets you a noticeably worse answer. ## How the Benchmark Worked Each model produced a single proposed answer per ticket. Three experienced support managers then graded those answers blindly, scoring them on factual correctness, usefulness, and tone. Blinded grading matters: it removes the temptation to favor a familiar brand or to assume an expensive model must be better. That methodology has real strengths. It measured what customers actually receive, not what models claim they can do. It combined three qualities that matter in real support work, rather than rewarding a single dimension like fluency. ## Where Each Model Actually Wins The headline tie hides an interesting split. North performed best on billing disputes, while Vale performed best on technical troubleshooting. Teams that route work might consider sending billing questions to North and technical ones to Vale — though the gain on either side would be modest given how small the overall gap is. ## The Image Problem The benchmark's most worrying finding isn't about which model won. It's that *all five* models performed poorly when a ticket depended on an image attachment. Screenshots of error messages, photos of damaged products, scanned receipts — these are common in customer support, and none of the systems in this test handled them well. Any team planning to deploy these models needs a fallback path for image-bearing tickets, because the proposed answer will often be wrong or irrelevant. ## What This Study Doesn't Tell You The test used 200 English-language tickets from a single software company. That's a reasonable starting point but a thin slice of customer support as a whole. The researchers also did not test agentic actions, tool use, or multilingual support. A model that can call a refund API, look up an order, or reply in Spanish wasn't being measured here — and those capabilities could easily flip the rankings in a real deployment. It's also worth noting that "score" here is a composite of three judgments. A team that cares disproportionately about tone (say, a luxury brand) might weight the results differently than a team optimizing for raw factual accuracy. ## The Bottom Line If you're choosing between these five models based on this benchmark, the answer is almost certainly Vale. It matches the leader on quality, costs less than half as much, and answers faster. North only earns its premium if you can show that it handles your specific billing workflows noticeably better than Vale — which is a question this benchmark raises but doesn't settle. The more durable lesson is methodological: when cost differences are large and quality differences are small, "best" usually means "best value." And when a clear weakness shows up across every model tested — as image handling did here — that's the area worth investing in next.

Meta: Muse Spark Open: # Startup Spotlight: LedgerLark Small investment firms face the same post-trade complexity as large institutions, but with far fewer people to manage it. LedgerLark is building AI agents to take on that operational load — not by executing trades, but by sorting out what happens after. ## An operations problem Co-founder Mina Batra spent seven years in operations at a regional brokerage, a role that often meant tracking down small data errors with outsized consequences. The early idea for LedgerLark came after she saw a six-person team spend three days resolving a single trade mismatch caused by a copied account number. It was, in her view, a representative example of how post-trade work tends to unfold: confirmations, custodian records, and internal ledgers need to line up, and when they don’t, people chase the discrepancy across systems, emails, and spreadsheets. Co-founder Owen Carr approaches the same problem from the engineering side. He previously built reconciliation systems at a large bank, where such matching and exception workflows were handled by larger teams and more customized internal software. Their shared insight is that smaller firms experience much of the same reconciliation burden without the same resources to automate it. Rather than replace the operations team, the founders want to reduce the time spent on routine comparisons. Whether that focus is narrow enough to build a sustainable product, or too narrow to support broader adoption, remains an open question. ## What the product does LedgerLark’s agents compare three sources — trade confirmations, custodian records, and a firm’s internal ledgers — to identify mismatches and propose corrections. Crucially, the system stops at the proposal. A human still reviews and approves any fix. The product does not execute trades or move money, a design choice that keeps it out of the direct flow of funds but still inside sensitive operational data. For firms that handle this work manually, the value proposition is straightforward: fewer exceptions to investigate by hand, and a clearer audit trail for those that remain. The practical test is whether the agents can handle the variety of formats, naming conventions, and edge cases that make reconciliation difficult in the first place, especially without creating new work to review inaccurate suggestions. ## Early customers LedgerLark says it now has five paying customers, all small investment firms. Annual contracts range from $36,000 to $110,000, suggesting some variation in firm size or scope of deployment. As an example of early results, 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 case rather than an average outcome. It also does not, on its own, explain how much human time was saved or how often staff override the agent’s proposals. The company has raised a $3.5 million seed round led by Motive Row, capital that will likely need to cover both product development and the longer sales cycles common in financial services. ## A crowded and cautious market LedgerLark is not alone in working on reconciliation. Established vendors already sell reconciliation software to investment firms, and many larger institutions rely on internal systems built over years. For small firms, the alternative is often a combination of vendor tools and manual processes that are familiar, if inefficient. That creates two related challenges. The first is differentiation: persuading firms that an agent-based approach is more accurate or adaptable than existing rule-based systems. The second is trust and procurement. Financial institutions, even small ones, may require lengthy security and procurement reviews before granting access to the kinds of records LedgerLark needs to function. Those reviews can slow adoption and raise the cost of acquiring each customer, particularly for a seed-stage company selling annual contracts in the five-figure range. LedgerLark’s bet is that a focused product for a historically underserved segment — small firms with big-firm reconciliation headaches — can earn its place by staying narrow and human-in-the-loop. The next evidence to watch will be whether its initial customers renew, whether the proposed corrections hold up across different custodians and asset types, and whether more firms are willing to let agents into their post-trade workflow at all.

MiniMax: MiniMax M3: ```markdown # Startup Spotlight: LedgerLark Behind every stock or bond trade at an investment firm sits a trail of paperwork: broker confirmations, custodian statements, internal ledgers. When those records don't match, someone has to figure out why. A small Boston-area startup called LedgerLark is betting that AI agents can shoulder more of that drudgery — and that smaller firms, often left behind by enterprise tooling, are willing to pay for it. The company's origin story is deliberately unglamorous. Co-founder Mina Batra spent seven years in operations at a regional brokerage, where she watched a six-person team spend three days resolving a trade mismatch traced back to a single copied account number. "It was the kind of thing that happens constantly," Batra said in an interview. "The fix was easy. Finding it was the problem." Co-founder Owen Carr brings the technical counterweight. He previously built reconciliation systems inside a large bank, giving him firsthand familiarity with the rules-based software that has long dominated back-office work. The two met through industry contacts and began sketching what a more flexible, agent-based system might look like. LedgerLark's product works like this: AI agents compare broker confirmations, custodian records, and a firm's internal ledgers, then surface inconsistencies and propose corrections. A human operator reviews and approves any changes before they are written back to the system. The agents do not execute trades and cannot move money on their own — a boundary the founders describe as deliberate, given the regulatory sensitivity of post-trade workflows. The early traction is modest but concrete. Five firms are paying customers, with annual contracts ranging from $36,000 to $110,000 depending on trade volume. The company says one customer reduced its unresolved daily exceptions from 70 to 18 after deploying LedgerLark, though the figure is self-reported and the underlying conditions aren't publicly detailed. Whether that kind of result holds across other clients, and how durable the improvement proves to be, remains the open question. To fund growth, LedgerLark has raised a $3.5 million seed round led by Motive Row. According to Batra, the capital gives the team roughly 18 months of runway and will go toward hiring engineers and expanding beyond the post-trade workflows the product currently handles. Competition is real and well-established. Legacy reconciliation vendors — firms like Broadridge, FIS, and a handful of smaller specialists — already serve this market, often bundled with custody or clearing services. Larger banks typically build similar tools in-house, sometimes maintained by teams of dozens. LedgerLark's pitch is that its agents are more adaptable than rules engines and cheaper than enterprise platforms, but it has yet to prove that advantage at scale, and incumbents are hardly standing still. The regulatory and procurement environment adds another layer of friction. Financial institutions, particularly those handling client assets, can require six to twelve months of security review, vendor diligence, and compliance sign-off before a new tool goes live. That pace works against a young company trying to land new customers quickly, and it shapes the kind of sales motion the founders have had to design. For now, LedgerLark is a small team chasing a familiar but stubborn problem. The work it seeks to automate has resisted full digitization for decades, partly because the consequences of getting it wrong are serious, and partly because the people doing it are trained to be cautious. Whether AI agents can earn that caution — without ever crossing the line that separates suggestion from action — is the bet the company is asking its customers to make. ```