Head to head: MoonshotAI: Kimi K3.1 Pro vs grok-4.7

MoonshotAI: Kimi K3.1 Pro: # KiteBridge failure creates $2.8 million in duplicate reimbursements Employers using KiteBridge were left funding $2.8 million in duplicate reimbursements on Tuesday after the company’s expense-management software treated repeated payment attempts as new requests. The failure generated 13,400 duplicate reimbursements across 91 corporate customers and led KiteBridge to suspend automated payments. No employees had money withdrawn from their accounts; employers funded the duplicate payments. The sequence began at 8:45 a.m. Tuesday, when KiteBridge deployed version 4.2. The first customer complaint arrived at 9:12 a.m., and the company suspended automated payments at 10:03 a.m. The halt therefore came less than 90 minutes after the release, but not before the software had produced thousands of additional reimbursements. The technical problem involved a payment-processor timeout and KiteBridge’s fallback system. A timeout generally means that one system did not receive an expected response within a set period. It does not necessarily establish that the underlying payment failed: the system requesting the payment may simply lack confirmation of what happened. A fallback system provides an alternative path when the normal process cannot proceed as expected. In this incident, its handling of the timeout was central. Retries are a common response when a payment’s outcome is uncertain, but KiteBridge’s fallback system treated those retries as new reimbursement requests rather than repeated attempts associated with an earlier request. In practical terms, trying again could create another reimbursement instead of merely continuing the effort to complete the original one. That distinction helps explain how a response failure could become a much larger financial error: repeating an attempt did not just repeat communication with the processor, but generated another employer-funded payment. KiteBridge’s public explanation developed more slowly than its payment suspension. The company’s incident page initially described the problem as delayed reimbursements. It first disclosed the duplicate payments six hours after automation had been suspended, shifting the public account from a question of when reimbursements would arrive to one involving extra payments that should not have been created. The incident also cast a spotlight on KiteBridge’s April changelog, which said version 4.0 had introduced “exactly-once payment protection.” In payment software, “exactly once” generally describes safeguards intended to ensure that a transaction has only one financial effect, even if messages are retried or responses arrive late. It is more than sending a request once: it concerns preventing repeated communication from producing repeated payments. The difference between delayed and duplicate reimbursements is significant for affected customers. A delay changes when money moves, while a duplicate changes how much employers pay out. Recovering the funds was therefore only part of the incident; customers also needed an explanation of how the duplication occurred despite the previously announced protection. CEO Lena Ortiz apologized and said KiteBridge would publish a technical report. By Wednesday evening, $2.3 million had been recovered, while the rest remained pending. Most of the duplicate funds had thus been retrieved, but the financial recovery was not yet complete, and the promised report remained the company’s next opportunity to explain how its payment safeguards and fallback behavior interacted.

grok-4.7: # $2.8 Million in Duplicate Reimbursements Follows KiteBridge Update A flawed software update at expense-management company KiteBridge sent employers’ money out twice on Tuesday, generating 13,400 duplicate reimbursements worth $2.8 million before the company shut off automated payments. KiteBridge’s software lets companies approve worker expenses and pay them back. At 8:45 a.m. Tuesday it deployed version 4.2. The first customer complaint arrived at 9:12 a.m. By 10:03 a.m. the company had suspended automated payments. In that stretch, 91 corporate customers were hit. The break was not a hack and not a bank error on the employee side. A payment processor—the outside service that moves the money and confirms the transfer—timed out. A timeout means the processor did not answer in time, so KiteBridge could not tell whether a reimbursement had gone through. Its fallback is meant to retry uncertain payments so employees are not left waiting. On Tuesday that fallback treated retries as new reimbursement requests. Each uncertain transfer could therefore be paid again, and the software did so at scale: 13,400 extra payments, $2.8 million in all. No employee had money withdrawn from a personal account. Employers fund reimbursements through the platform, and employers paid for the duplicates. Workers may have received money their companies never intended to send, but the cash left corporate accounts, not paychecks in reverse. Suspending automation at 10:03 a.m. stopped further duplicates. It did not pull back funds banks had already accepted. Recovery ran through the next day. By Wednesday evening KiteBridge had recovered $2.3 million. The rest, about $500,000, remained pending—reversals unfinished or repayments not yet complete. The miss was sharper because of a claim the company had already made. An April changelog, the public note that accompanies a software release, said version 4.0 had introduced “exactly-once payment protection.” In plain terms, that is a promise that one approved expense produces one payment, even if a glitch forces the system to try again. Version 4.2 did not keep the promise when a processor timeout arrived. KiteBridge has not yet said whether the protection never covered this path or whether later changes weakened it. What the company told the public also lagged what it knew. Its incident page initially described the problem as delayed reimbursements, which suggested employees were waiting on money rather than receiving it twice. The page first disclosed the duplicate payments six hours after automation was suspended. A delay and a double payment are different problems. Customers who saw only the early notice had no clear reason to alert finance teams or hold related transfers. Chief executive Lena Ortiz apologized and said the company would publish a technical report. The report is the next public test of the incident. It will have to show how a retry became a new reimbursement, and what the company will change so a processor timeout cannot do that again. For those customers the open question is the unrecovered money. Most of the $2.8 million was back by Wednesday evening. Until the remainder is recovered or written off, Tuesday’s update is still an open bill, however quickly KiteBridge cut the automation that ran it up.

MoonshotAI: Kimi K3.1 Pro: C

grok-4.7: C

MoonshotAI: Kimi K3.1 Pro: # Cloudnote debugging page exposed customer prompt excerpts for four hours *Logs show excerpts from 1,842 user sessions. The company says it has found no evidence of access beyond the researcher who reported the issue, but its investigation is unfinished.* Cloudnote discovered that a debugging page containing excerpts of customer prompts was accessible without authentication for four hours on Monday. The company removed the page 19 minutes after a researcher notified it. Logs show excerpts from 1,842 user sessions, and each excerpt could include up to 200 characters from a prompt. The session count is not a confirmed count of individual users. Cloudnote has 3.2 million registered users overall; the records described so far do not establish how many of those customers had excerpts on the page. The incident matters because customer material was reachable without authentication, even if only in excerpts. Prompts can contain information users expect to remain private, but the content of these excerpts—and therefore the potential privacy impact—has not been established. Full documents, passwords and payment information were stored elsewhere and were not present on the page. Cloudnote says it has found no evidence that anyone besides the researcher accessed the page. Its investigation is not complete, so that is a current assessment rather than a final finding that no other access occurred. That leaves open whether further review will identify other access or alter the scope now described.

grok-4.7: # Cloudnote Debug Page Exposed Prompt Excerpts From 1,842 Sessions *The unauthenticated page was reachable for four hours Monday. Cloudnote says it has found no other access, but its investigation is not finished.* Cloudnote discovered that a debugging page containing customer prompt excerpts was accessible without authentication. The page was exposed for four hours on Monday. Logs show excerpts from 1,842 user sessions, and those excerpts could include up to 200 characters from a prompt. The company has 3.2 million registered users. Against that figure, 1,842 sessions is a narrow count, though a session is not the same as an account. The excerpts still matter: even 200 characters can capture a fragment of what a customer submitted. Cloudnote has not said how many accounts were involved or what any excerpt contained. Full documents, passwords and payment information were stored elsewhere and were not present on the page, the company said. That is a confirmed boundary on what the page could display. It does not answer whether the short excerpts themselves included sensitive details. Cloudnote removed the page 19 minutes after a researcher notified it. The company says it has found no evidence that anyone besides the researcher accessed the page. It has not completed its investigation, so that statement remains preliminary. How the page became reachable without authentication, and whether every visit would appear in the logs, are still unknown. Until the review is finished, the number of people who saw the excerpts cannot be treated as settled.

MoonshotAI: Kimi K3.1 Pro: # Arcwell launches Relay, an AI platform for customer support Arcwell says agents using its new artificial intelligence platform closed support tickets 14% faster during a six-month test with 12 companies. The platform, Relay, examines customer support conversations and suggests answers to agents, drawing on information that can otherwise be difficult to find during a customer interaction. “Support agents spend too much time searching for information that their companies already have,” CEO Priya Shah said. Relay’s capabilities extend beyond drafting responses. It can also identify customers who may cancel and create summaries. That puts three functions in one product: answer suggestions, conversation summaries and warnings about potential customer churn. Arcwell is presenting the combination as a way to improve customer service, but the performance figures released with the launch address only part of that promise. The 14% improvement concerns ticket-closing speed. Arcwell did not release customer-retention results, leaving open whether Relay’s ability to flag customers who may cancel leads to fewer cancellations. The distinction matters because faster ticket handling and stronger retention are different measures of the platform’s impact. Relay costs $65 per agent each month. Using it also requires giving the platform access to support transcripts and internal help documents. Those sources connect Relay to both sides of a support interaction: the customer’s conversation with the company and the internal material available to help resolve the issue. As a result, data access and retention are central to the product, not just administrative details. Arcwell says customer data is encrypted and is not used to train shared models. Companies can choose either a 30-day or a 90-day retention period. Relay therefore arrives with a clearer case for speed than for retention. Its test results point to faster ticket resolution, while its security and data-retention policies address the access the platform needs. Whether it can also reduce cancellations remains unanswered by the results Arcwell has released.

grok-4.7: # Arcwell Launches Relay, Citing Faster Ticket Closures in a Six-Month Test Arcwell has announced Relay, a platform that uses artificial intelligence to examine customer support conversations and suggest answers to agents. It also identifies customers who may cancel and creates summaries. Twelve companies tested it over six months. Arcwell says agents using Relay closed tickets 14% faster. The company did not release customer-retention results. Relay costs $65 per agent each month. That is an uneven set of disclosures. A faster close is a measurable change in how agents work. It is not evidence that customers stayed. Relay can identify people who may cancel, but Arcwell has not said whether fewer of them left. Anyone weighing the six-month test will find a number on speed and nothing on retention. Chief executive Priya Shah described the problem Relay is meant to reduce. "Support agents spend too much time searching for information that their companies already have," she said. Suggested answers follow from that complaint more directly than the other two features do. The platform examines the conversation and proposes a reply, rather than leaving the agent to search out material the company already holds. Summaries turn the exchange into a shorter account of what was said. A third function marks customers who may cancel. Arcwell has not released a separate result for summaries or cancellation flags. The 14% figure stands as the result on offer. Relay requires access to support transcripts and internal help documents before it can do that work. Arcwell says customer data is encrypted and is not used to train shared models. Companies can choose a 30-day or 90-day retention period, deciding how long those records remain once a conversation is over. The practical reading is simpler than the feature list. Relay is a $65-a-month seat that needs a company's own transcripts and help documents, handles that data under the limits Arcwell describes, and, in the company's telling, helps agents finish tickets faster. Whether it helps them keep customers is not a result Arcwell has chosen to share.

MoonshotAI: Kimi K3.1 Pro: # HelioDock launches Night Shift for remote warehouse robot monitoring *The $90-per-robot

grok-4.7: # HelioDock Launches Night Shift to Let Operators Watch Warehouse Robots From Afar *The robotics company paired the remote-monitoring software with a $24 million Series B after a 14-month private beta, citing a 17 percent drop in unplanned stoppages that it has not had independently verified.* HelioDock, which makes autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely. The announcement arrived together with word of a $24 million Series B led by North Coast Ventures, bringing the company’s total funding to $39 million. The software is leaving a closed trial that ran long enough to generate the only performance number the company has released. Night Shift spent 14 months in a private beta involving nine logistics companies and 310 robots. HelioDock says beta customers experienced 17 percent fewer unplanned robot stoppages. The figure has not been independently verified. That company-supplied percentage is the sole quantitative result attached to the beta, so any operator weighing the product must treat the reduction as HelioDock’s own measurement from those nine sites and those 310 machines rather than a result confirmed by an outside party. Chief executive Mara Velez described the problem the software is meant to shrink. “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. The sentence fixes the use case in a late-night interruption: one robot stopped by confusion over a pallet, and a manager pulled onto the road to deal with it in person. Night Shift, as HelioDock presents it, would give that manager a remote view of the fleet so the same confusion could be observed without the drive. The company has not elaborated further on how an operator would clear the pallet confusion once the remote session is open, leaving the quote itself as the clearest statement of intent. A customer who used the system during the beta supplied a practical limit that sits beside the stoppage claim. Aaron Pike of Fulcrum Logistics said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. His account separates two operating conditions inside the same product. Where the network held, remote monitoring functioned; where rural LTE dropped, the usefulness fell. HelioDock’s description of the failure mode matches that report. When connectivity fails, robots return to their locally configured behavior. Remote control becomes unavailable. The machines keep following the instructions already stored on site, yet the distant operator loses the ability to direct them until the link is restored. The fallback therefore preserves local motion at the exact moment it removes the remote oversight Night Shift is built to provide. Pike’s experience at Fulcrum Logistics shows that the difference is not theoretical: strong connectivity kept the system useful, and rural outages made it less so. The commercial details are now fixed. Night Shift will cost $90 per robot per month and become generally available September 15. Applied to a fleet the size of the beta group, the monthly fee would multiply directly by the number of robots placed under monitoring. HelioDock has not indicated whether any of the nine logistics companies have committed to the paid service or what share of a working warehouse fleet is expected to carry the subscription. The price and the date simply mark the end of the private period and the start of open availability. The category already contains other options. Competing fleet-management products are available from Relay Robotics and YardOS. Night Shift therefore enters alongside two named alternatives rather than into an empty field. HelioDock has not compared features, uptime, or price against either product, so the points of contrast available from the launch are the ones the company and its beta customer have already stated: fourteen months of private use, 310 robots, a 17 percent reduction in unplanned stoppages that remains unverified by any independent source, the chief executive’s focus on eliminating a 2 a.m. drive, and the reversion to local behavior whenever connectivity fails. The funding round supplies the capital context for that September date. North Coast Ventures led the $24 million Series B that raises HelioDock’s total to $39 million. No allocation of the new money—between additional robots, further work on Night Shift, or other purposes—has been disclosed, nor have any other investors been named. What is on record is the joint timing: the capital and the software were both announced after the beta had already run its 14 months with the nine logistics companies and their 310 robots. Operators looking at the September 15 release therefore have a compact set of facts. HelioDock says the beta produced 17 percent fewer unplanned stoppages; that number is unverified outside the company. Velez locates the value in sparing a manager an overnight drive caused by one robot and one pallet. Pike reports that the value held where connectivity was strong and receded during rural LTE outages. In every outage the robots continue under their locally configured behavior while remote control stays unavailable. The software will be offered at $90 per robot per month, and fleet-management tools from Relay Robotics and YardOS are already on the market. The $39 million funding total is the sum HelioDock has raised by the week it set that price and that date.

MoonshotAI: Kimi K3.1 Pro: # QueueZero’s Proxy AI assistant turns an audience remark into CEO tasks QueueZero launched an AI meeting assistant called Proxy on Wednesday, pitching software that can attend meetings, produce notes and assign follow-up work to software agents. The launch demonstration also showed how easily an offhand remark can become an action item when a meeting tool is designed to listen for instructions. During the demonstration, a customer in the audience said, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy interpreted the comment as an instruction rather than an observation. 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 pricing sheet survived the workflow; the instruction to stop talking did not. The exchange offered a concise illustration of both the appeal and the difficulty of agentic meeting software. A system expected to turn conversation into useful work must determine not only what was said, but also whether a speaker was addressing the software, whether a request was serious and who has the authority to act on it. QueueZero says Proxy requires confirmation before an external message is sent. Internal tasks, however, can be created automatically, allowing a comment in a meeting to become an assignment without a separate approval step. Administrators can disable automatic task creation, providing a direct control for organizations that do not want that behavior. That distinction is significant because the consequences of an unwanted action can differ. An erroneous internal task may be deleted, as Wu demonstrated. An external message can affect customers or partners outside the meeting, making a pre-send checkpoint especially important. Confirmation at the point of sending does not by itself settle the earlier question of what a speaker intended. The demonstration also underlines the governance questions facing companies adopting AI agents in routine work. Meeting notes are generally a record; assignments and messages are actions. Products that combine them need clear boundaries around permissions, review and automation, particularly when casual speech can be ambiguous even to people in the room. Proxy costs $30 per user per month and enters public beta next week. Its debut suggests that the practical test for meeting assistants is no longer limited to transcription accuracy. It also includes whether they can reliably distinguish a useful follow-up from a remark that should have remained commentary.

grok-4.7: # QueueZero Launches Proxy, an AI Meeting Assistant That Assigns Follow-Up Work QueueZero on Wednesday launched Proxy, an artificial intelligence meeting assistant that can attend a meeting, produce notes and assign follow-up work to software agents. The product costs $30 per user per month and enters public beta next week. The launch demonstration showed how quickly Proxy can turn speech into assigned work. 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 chief executive Evan Wu: “Stop talking” and “Send pricing sheet.” Wu laughed, completed the second task and deleted the first. For a moment, the company’s own product had given its chief executive a task to be quiet. The incident illustrated a product built to do more than take notes. Proxy is meant to attend, take notes and pass follow-up work to software agents, so that a meeting can end with tasks already created rather than with a list someone still has to write. In the demonstration, that handoff ran on a sentence a customer spoke to the room. QueueZero says Proxy requires confirmation before an external message is sent. Internal tasks can be created automatically. Administrators can disable automatic task creation. A message that leaves the organization therefore waits on a person, while a task that stays inside can be created without that step. The two items Proxy assigned to Wu were internal tasks, not external messages. The two tasks were not equal. “Send pricing sheet” matched a routine request for commercial information, and Wu completed it. “Stop talking” was an instruction about his conduct during the demonstration. Deleting it kept that assignment from remaining as open work. It did not change the fact that Proxy had created the task from an audience remark. Buyers looking at the public beta will have to decide whether automatic internal task creation is a feature they want left on. The company is offering an administrator control that disables it. Organizations that leave automatic creation on are accepting that Proxy may treat indirect language as something to assign. Organizations that turn it off are choosing a narrower assistant, one that can still attend a meeting and produce notes but will not create internal tasks on its own. At $30 per user per month, Proxy’s public beta next week will let organizations test that tradeoff in their own meetings.

MoonshotAI: Kimi K3.1 Pro: # Startup Spotlight: LedgerLark Focuses on Post-Trade Exceptions LedgerLark is building AI agents for a less visible part of investing: the work required to check, reconcile and correct records after trades are placed. It is focused on small investment firms, where a mismatch can draw operations staff away from the rest of the day’s work. ## A copied number and a three-day search Co-founder Mina Batra spent seven years in operations at a regional brokerage. She began exploring the idea after seeing a six-person team spend three days resolving a trade mismatch caused by a copied account number. The episode shaped the company’s focus. The problem was not necessarily a lack of trade data; it was the effort required to compare records, locate a discrepancy and decide what should change. Co-founder Owen Carr previously built reconciliation systems at a large bank, giving the pair experience with both day-to-day operations and the software used to manage them. That background does not guarantee a better product, but it helps explain LedgerLark’s design choice: concentrate on investigation and correction rather than trade execution. ## Agents with a narrow remit LedgerLark’s agents compare trade confirmations, custodian records and internal ledgers. When they find inconsistencies, they propose corrections for human approval. The product does not execute trades or move money. That boundary matters. Post-trade work involves controls as well as efficiency, and a suggested correction still needs to be assessed by the people responsible for the firm’s books and processes. LedgerLark’s approach therefore puts its agents in an advisory role: they assemble and interpret evidence, while approval remains with a person. One technical question is how consistently the system handles exceptions that are less straightforward than the founders’ original example. Firms will also need to understand how recommendations are produced, reviewed and audited. Those requirements could matter as much as time saved. ## Early revenue, limited evidence Five firms are paying customers, with annual contracts ranging from $36,000 to $110,000. That is useful evidence that some firms will pay for the product, though the range leaves open how deployments differ in scope and size. LedgerLark says one customer reduced its unresolved daily exceptions from 70 to 18. If sustained, that would represent a substantial operational improvement. But it is a company-reported result from one customer, not evidence of a typical outcome across the customer base. Useful next evidence would include results over longer periods, at other firms, and with a clear account of implementation effort and exception types. The company has raised a $3.5 million seed round led by Motive Row. The funding gives it resources to develop the product and pursue more deployments; it does not resolve questions about adoption or performance. ## Competition and the pace of procurement LedgerLark is not entering an empty category. It competes with established reconciliation vendors and with internal bank software. Those alternatives may already be embedded in a customer’s workflows, while a newer product must demonstrate that it adds enough value to justify adoption and review. Financial institutions may also require lengthy security and procurement reviews. That can slow sales even when operations teams see a clear use case, consuming time for both LedgerLark and prospective customers. For now, LedgerLark has a specific operational thesis, a constrained product boundary and a small group of paying customers. The constructive test is whether its agents can produce reliable, reviewable work across varied environments—and whether the resulting savings justify the effort of buying and supervising them.

grok-4.7: # Startup Spotlight: LedgerLark Looks for the Error After the Trade Post-trade operations sit in the interval after a buy or sell is agreed, when confirmations, custodian records and a firm's own ledgers are supposed to describe the same event. At small investment firms, that work often falls to a small team comparing files and chasing one mismatched field. LedgerLark is building AI agents for that job, on the view that the problem they have chosen is not the trade itself but the break that follows a bad entry. Mina Batra, a co-founder, spent seven years in operations at a regional brokerage. She began exploring the idea after seeing a six-person team spend three days resolving a trade mismatch caused by a copied account number. The episode is small in origin and large in staff time: one transcription error can hold up settlement work long after the trade is done. Owen Carr, the other co-founder, built reconciliation systems at a large bank. Their product is aimed at smaller investment firms, which may lack a bank's internal engineering staff and still have to reconcile custodian breaks and confirmation gaps. LedgerLark's agents compare confirmations, custodian records and internal ledgers, then propose corrections for human approval. The product does not execute trades or move money. That boundary is worth stating plainly. In this part of the market, a proposed correction can become an instruction, and an instruction can move cash if someone acts on it. Keeping a person in the approval path limits one kind of risk. It does not settle whether the comparison is right often enough to save time. A firm still has to decide which records the agent may read, how suggestions are logged, and who is accountable when a proposal is wrong. Five firms are paying customers. Annual contracts range from $36,000 to $110,000, a wide band this early, and one that may reflect different scopes as much as different prices. The company says one customer reduced its unresolved daily exceptions from 70 to 18. The drop is large, and it is a single case reported by the vendor. Nothing public establishes how exceptions were defined, how long the improvement held, or what else changed in that customer's process. It is a result LedgerLark is putting forward. It is not yet a result others can assume. The company has raised a $3.5 million seed round led by Motive Row. The capital can fund product work and the slow parts of selling; it does not show that customers will remain. Competitors include established reconciliation vendors and internal bank software. A concentration on small investment firms may be a way to avoid competing first for the largest buyers. It also caps each contract. A fee of $36,000 a year has to justify itself against staff time, an incumbent tool, and the friction of switching. Financial institutions may require lengthy security and procurement reviews before an agent is allowed near client records, even though LedgerLark does not move money. Those reviews can stretch a sale and tend to favor vendors already known to a buyer's security and procurement teams. Five paying customers mean some firms have bought the product. They do not show how often the next review will end the same way, or how long it will take. The founders' insight is easy to grant: a copied account number should not take three days from a six-person team, and much of the repair is comparison rather than judgment. The unsettled part is whether the agents can tell a true break from a formatting difference reliably enough that operators trust the proposal, and whether the human approval step stays intact once the queue is long. A seed round and five contracts make that test possible. They do not complete it.

MoonshotAI: Kimi K3.1 Pro: # Foundry Hall to Close After Losing Its Largest Sponsor Foundry Hall, a Milwaukee nonprofit startup accelerator, will close Oct. 31 after failing to replace its largest source of funding. Since opening in 2018, the organization has supported 212 startups, which collectively employ 1,140 people. Its closure will also cost six employees their jobs and force the 12 startups currently using its workspace to find somewhere else to work. The decisive loss came when Blank River Bank, its largest corporate sponsor, declined to renew its sponsorship. The bank supplied 38% of Foundry Hall’s annual budget. The accelerator has enough money to operate through October, but not enough to run another full program. Executive director Camille Ross described an extended wait for new backing as the organization tried to close the gap. “We kept waiting for the replacement check that never came,” she said. Blank River Bank said the decision reflected a change in its community investment priorities and was unrelated to Foundry Hall’s performance. That distinction separates the choice from the accelerator’s record, but it does not change the financial reality of its final months. For founders, the closing is more than a budget story. Dev Malik’s connection to Foundry Hall rests on a specific contrast: where he pitched before, and where he first met an investor. “My first investor meeting happened in that building,” Malik said. “Before Foundry Hall, I was pitching from the back table of a coffee shop.” His account suggests what the workspace offered beyond desks and meeting rooms: a setting where an early-stage company could feel less provisional. That value is harder to quantify than a sponsorship, and harder to replace on short notice. Deputy mayor Susan Cole said the city is speaking with building owners and local universities about space for the current companies. Those discussions could give the 12 startups somewhere to go, even as the organization that brought them together prepares to shut down. For now, the city’s effort concerns where the companies will work, not whether Foundry Hall itself can continue. Foundry Hall’s legacy will remain in the companies it assisted and the people they employ. But its final count is more immediate: six jobs ending, 12 teams preparing to leave, and a staff closing down work it had hoped to continue. When the accelerator closes, it will leave behind a substantial record and a quieter absence. For those who began there, Foundry Hall was not only a program or a line in a budget. It was the room where a first meeting could happen.

grok-4.7: # The Doors on October 31 The glass doors of Foundry Hall still opened onto the concrete floor where founders had once taped pitch decks to the walls. On an ordinary October afternoon the nonprofit startup accelerator in Milwaukee looked much as it had since 2018: long tables, a coffee urn that never quite kept up, whiteboards crowded with names that had since become companies. The doors would lock for the last time on October 31. In seven years the hall had supported 212 startups. Those companies now employ 1,140 people, a figure the staff sometimes recited the way other institutions recite endowments. The building itself had never been grand. What it offered was a room, a schedule, and the ordinary dignity of being taken seriously. That dignity depended, more than anyone liked to say aloud, on a single sponsor. Blank River Bank, its largest corporate sponsor, had supplied 38 percent of the annual budget. When the bank declined to renew, the arithmetic was immediate. Foundry Hall could pay its bills through October. It could not fund another full program. Executive director Camille Ross did not soften the sequence. “We kept waiting for the replacement check that never came,” she said. A spokesperson for the bank described the decision as a shift in community investment priorities, unrelated to the hall’s performance. The distinction was real. It did not restore the 38 percent that had walked out the door. Six employees would lose their jobs when the month ended. Twelve startups still working at the long tables would have to find somewhere else to open their laptops. Deputy Mayor Susan Cole said the city had begun talking with building owners and local universities about space for those companies. Talk was not a lease. It was, for the moment, what existed. No one confused that with a solution. The twelve companies needed desks before November, and the six staff members needed the month to end without pretending otherwise. Dev Malik did not talk about metrics. “My first investor meeting happened in that building,” he said. “Before Foundry Hall, I was pitching from the back table of a coffee shop.” He did not claim the hall had made his company. He said only that it had given the meeting a door that closed, and a table that was not borrowed. By late afternoon someone was labeling boxes in the corridor. Outside, the river moved under the bridges as it always had, indifferent to sponsorship cycles and cohort calendars. The work of seven years would not vanish with the keys. It would have to continue, for a while, in rooms that had not yet been found, without the building that had made it visible.

MoonshotAI: Kimi K3.1 Pro: # 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.

grok-4.7: # 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.

MoonshotAI: Kimi K3.1 Pro: 6

grok-4.7: 6

MoonshotAI: Kimi K3.1 Pro: # Vale stands out in a support-ticket benchmark—on value, not a decisive quality win The clearest finding from a five-model customer-support benchmark is not that one system produced decisively better answers. It is that Vale’s average quality score was close to North’s, and the difference was not statistically significant, while Vale responded faster and cost substantially less. For support leaders choosing a model for a queue like the one tested, that makes Vale the strongest default candidate—but not a universal winner. ## How the benchmark worked Researchers gave each model the same 200 anonymized customer-support tickets and collected one proposed answer per ticket. Three experienced support managers graded the responses blindly, using a composite score covering factual correctness, usefulness and tone. Blind grading reduces the chance that a model’s name or reputation influenced the scores. The benchmark therefore measures the quality of draft answers under controlled conditions; it does not show that tickets were actually resolved or that customers were satisfied. ## Quality was close; operating differences were not North averaged 84.2 and Vale 83.8. That 0.4-point gap was not statistically significant, so describing North as the quality champion would overstate the evidence. The operational differences were much clearer. Vale’s estimated cost was $0.74 per 1,000 tickets, compared with $1.90 for North—about 61% lower. Vale also answered 18% faster. When measured quality is this close, those differences matter. A cheaper, faster model can free budget and reduce response latency, provided the savings are not erased by extra corrections or escalations. Still, a non-significant result is not proof that the models are identical; it means this test did not establish a reliable quality difference. The averages also hide useful specialization. North performed best on billing disputes, while Vale led on technical troubleshooting. A company with a billing-heavy queue should therefore test North on that workload rather than assume Vale’s overall value settles the choice. The reported results do not include category-level sample sizes or significance tests, so these strengths are better treated as prompts for follow-up testing than as decisive rankings. The other models sharpen the trade-off rather than overturning it. Ember and Kite occupied the middle ground on price and quality, while Harbor was cheapest at $0.08 per 1,000 tickets but had the lowest average score, 71.0. That may be acceptable for low-risk drafts with review, but the benchmark alone cannot determine whether the savings compensate for lower answer quality. ## The shared weakness may matter more than the leaderboard All five models performed poorly when a ticket depended on an image attachment. For real support operations, that is a substantial warning: screenshots and error dialogs can contain the information needed to respond accurately. The result argues for human review, an attachment-based routing rule, or a separate multimodal evaluation before automating such cases. A strong overall average should not be read as evidence of competence on these tickets. ## What the test cannot establish The study used English-language tickets from one software company, and only 200 tickets in total. Its results may not transfer to other products, customer bases, languages or ticket mixes. Researchers also did not test agentic actions, tool use or multilingual support, so the scores say nothing about issuing refunds, querying account systems, following runbooks or handling a global queue. The composite score further obscures whether a model’s average came from correctness, usefulness, tone, or some combination. The cost figures are estimates, not a demonstrated total cost of ownership. Integration, review and failure-handling costs could change the business case. ## A practical next step For a deployment resembling the test, the sensible next step is a focused pilot of Vale against North, stratified by ticket type, alongside a separate evaluation of image-dependent cases. Teams should track factual errors, reviewer edits, escalation rates, latency and full operating cost. On the evidence available, Vale is the leading value candidate, while North remains worth testing for billing-heavy work. Neither result justifies unattended support across the board.

grok-4.7: # The Pricier Model Did Not Win The clearest result in this benchmark is not a champion. It is a tie that changes the bill. On 200 anonymized customer-support tickets, Model North scored 84.2 and Model Vale 83.8. Researchers found that gap was not statistically significant. Vale still cost an estimated $0.74 per 1,000 tickets against North’s $1.90, and it answered 18 percent faster. Paying more than twice as much did not buy a measurable edge in graded answer quality. That is the decision the numbers can support. The other three models sat lower, and their discounts were large, but the drop in score was large enough that “cheaper” and “about as good” are not the same claim. Ember averaged 79.1 at $0.31 per 1,000 tickets, Kite 76.4 at $0.22, and Harbor 71.0 at $0.08. Quality did not fall in neat proportion to price. It held on a high plateau for two models, then stepped down. ## What was actually tested Each model produced one proposed answer per ticket. Three experienced support managers graded those answers blindly, without knowing which system wrote which reply. The score mixed factual correctness, usefulness, and tone, so a smooth voice could not carry a wrong or unhelpful draft by itself. That design is better than a demo and thinner than a production trial. Blind review by people who manage support is a real check on customer-facing judgment. A single draft per ticket does not show what happens when an agent pushes back, asks for a revision, or has to finish the case. The cost figures are estimates, not invoices. Speed was reported only for the North–Vale comparison, so the rest of the field cannot be ranked on latency from this study. The sample is also narrow in ways that matter. The tickets were in English and came from one software company. What counts as a correct billing answer, a useful troubleshooting step, or an acceptable tone will not travel unchanged to another product, another policy set, or another language. Researchers did not test agentic actions, tool use, or multilingual support. A model that drafts well may still fail if the job is to look up an order, issue a credit, or answer in a second language. ## How to read the spread North’s four-tenths of a point over Vale is the sort of difference a leaderboard treats as a win and a significance test does not. The honest summary is a quality tie, not a victory for the expensive model. Speed points the same direction: an 18 percent faster answer, on top of a score that did not separate the two, is an operational difference rather than a rounding footnote. Below that pair, the gaps are wide enough to discuss without pretending the decimals are precise. Harbor’s average of 71 is not a near miss on a blend of correctness, usefulness, and tone. Ember and Kite land between the leaders and that floor. A team that treats the reply as something a customer may act on should read the cheap end of this list as a different product: much lower estimated cost, and a visible concession in graded quality. The study does not show that those concessions shrink on the hardest tickets. Averages also hide a split the researchers did notice. North performed best on billing disputes. Vale performed best on technical troubleshooting. Those are rankings, not published subscores, so they are not a routing policy. They are a warning against treating one overall number as the only fact. A queue dominated by refunds and plan changes is not the same queue as one dominated by setup failures, even when the headline scores match. ## The failure they shared All five models performed poorly when a ticket depended on an image attachment. That result is more useful than another tenth of a point between North and Vale. The task sometimes required evidence the text did not contain — a screenshot, an error dialog, a photo — and none of the systems closed the gap. A higher overall score here does not mean broader competence. Until image handling is tested on its own, these rankings describe text replies to tickets that could be understood as text. ## What a buyer should not conclude Small differences elsewhere should be handled with the same restraint as the top-line gap. The study flags statistical significance only to say North and Vale were not separated. It does not establish that Ember’s lead over Kite, or Kite’s over Harbor, would hold in a larger sample. Direction is the safer claim: two models clustered at the top, and quality then declined as price fell sharply. Nor does a strong draft grade prove a strong support operation. The test stopped at a proposed answer. It did not measure reopen rates, handle time after a human edit, escalations, or whether customers accepted the reply. One company’s English queue is not the industry. Tool use was out of scope, and every model stumbled when the ticket needed an image. Within those limits, the finding is narrow and practical. If the job is a first-draft English reply on tickets like the ones tested, Vale matched the most expensive model closely enough that the difference did not clear a significance test, cost well under half as much on the researchers’ estimate, and was faster. North remains plausible if billing disputes dominate and a buyer wants the specialization this study only sketches. The cheaper three are a cost trade with a real quality gap, not a hidden tie. Spending more, on this test, did not buy a clearer answer.

MoonshotAI: Kimi K3.1 Pro: # Raincheck Raises $8 Million in Seed Funding for Flood-Risk Software *The Austin company’s latest round brings its total funding to $10.5 million and will support hiring and expansion into Florida.* Raincheck, an Austin-based company that makes flood-risk software for local governments, announced today that it has raised $8 million in seed funding. Basin Ventures led the round, with participation from existing investor Long Mile. Raincheck has raised $10.5 million to date. Raincheck maps drainage systems and combines those maps with rainfall forecasts, according to the company’s [flood-risk methodology](https://example.com/raincheck-methodology). Cities use its software to decide where to send maintenance crews before a storm. CEO Ana Flores said: > “Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives.” Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations, and 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.

grok-4.7: # Raincheck Raises $8 Million for Flood-Risk Software *The seed round, led by Basin Ventures, brings the Austin company’s total funding to $10.5 million.* Raincheck, an Austin company that makes flood-risk software for local governments, announced today that it has raised $8 million. It maps drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm. The seed round was led by Basin Ventures. Existing investor Long Mile also participated. Raincheck has raised $10.5 million total. CEO Ana Flores said: > 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. Further detail is available in Raincheck’s [flood-risk 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. 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.