Head to head: MoonshotAI: Kimi K3 vs Anthropic: Claude Fable 5.1

MoonshotAI: Kimi K3: 6

Anthropic: Claude Fable 5.1: **6**

MoonshotAI: Kimi K3: # LedgerLark Sends AI Agents Into the Back Office, Under Supervision *At small investment firms, resolving a single bad trade can still eat up days. LedgerLark's founders think software agents can shorten that — if a human approves every fix.* Mina Batra spent seven years in operations at a regional brokerage, so she knows what a trade mismatch actually costs. In one instance she recalls clearly, a six-person team spent three days untangling an error that traced back to a single account number copied into the wrong field. That kind of episode is familiar in post-trade operations — the unglamorous work of confirming that a trade recorded in one system matches what custodians and counterparties show in theirs. Large institutions automate much of it. Smaller investment firms, with leaner budgets and older software, often resolve discrepancies by hand: emails, spreadsheets, phone calls. Batra's experience became the seed for LedgerLark, which she co-founded with Owen Carr, who previously built reconciliation systems at a large bank. The pairing is notable mainly because it covers both sides of the problem: Batra has lived the workflow, and Carr has built the plumbing. ## What the agents do — and don't do LedgerLark's software agents compare trade confirmations, custodian records, and a firm's internal ledgers, flag inconsistencies, and propose corrections. A human reviewer approves or rejects each proposal before anything changes. The product does not execute trades or move money — a boundary that keeps the company clear of the regulatory weight that comes with touching client funds, and that gives operations staff the final say. Whether that human-in-the-loop design is a feature or a ceiling is an open question. It makes the product easier to trust today; it may also limit how much labor it can remove tomorrow. ## Early traction, small sample LedgerLark says five firms are paying customers, on annual contracts ranging from $36,000 to $110,000. The company also reports that one customer cut its unresolved daily exceptions from 70 to 18 — a striking figure, though it comes from LedgerLark itself, reflects a single customer, and says little about how typical that result is. The company has raised a $3.5 million seed round led by Motive Row. For now, the funding and the five contracts are best read as evidence that a handful of firms consider the problem worth paying to solve — not yet proof that the approach scales. ## A crowded, slow-moving market LedgerLark is not entering an empty field. Established reconciliation vendors already serve financial institutions, and some banks run internal software that performs similar matching. The company's implicit argument is that smaller firms are underserved by tools built for bigger clients, and that agents able to read messy documents and reason about discrepancies can go further than rules-based matching. That argument will be tested by procurement as much as by technology. Financial institutions tend to require lengthy security and compliance reviews before adopting new software, which can stretch sales cycles beyond what a young company comfortably absorbs. Accuracy is another open question: an agent that proposes wrong corrections, even occasionally, could erode the trust that makes the approval step meaningful in the first place. ## What to watch The next year should clarify whether LedgerLark's early traction reflects a genuine gap or a few accommodating pilots. Useful signals include whether existing customers renew and expand, whether the exception-reduction results repeat beyond one firm, and how quickly new prospects clear security review. The underlying insight — that small firms carry a disproportionate share of manual reconciliation work — is credible precisely because it comes from someone who did that work. Whether AI agents are the right tool to relieve it is exactly the kind of question a seed round exists to answer. --- **Quick facts** — Founders: Mina Batra and Owen Carr · Customers: 5 · Pricing: $36K–$110K/year · Funding: $3.5M seed, led by Motive Row

Anthropic: Claude Fable 5.1: # Startup Spotlight: LedgerLark **What it does:** AI agents that help small investment firms find and fix post-trade discrepancies. ## The problem that prompted it Most trades settle without incident. The ones that don't can consume a surprising amount of human attention. Mina Batra, who spent seven years in operations at a regional brokerage before co-founding LedgerLark, describes watching a six-person team spend three days resolving a single trade mismatch. The cause, once found, was mundane: an account number copied into the wrong field. That episode is the company's origin story, and it is a modest one. There was no sweeping thesis about the future of finance, just a familiar frustration: at smaller firms, reconciliation work is often spread across spreadsheets, email threads and a handful of people who know where to look. Batra's co-founder, Owen Carr, approached the same problem from the opposite end of the industry, having built reconciliation systems at a large bank. ## How the product works LedgerLark's agents compare three sources that should agree but frequently don't: trade confirmations, custodian records and a firm's internal ledger. When the agents detect a mismatch, they attempt to identify the likely cause and propose a correction. A human operations staffer then reviews and approves, rejects or edits the suggestion. The boundaries here are worth noting. The product does not execute trades, and it does not move money. That design choice narrows the scope of what the software can get wrong, which matters in a setting where a bad automated action could be costly and hard to reverse. It also sidesteps some regulatory questions, though not all of them, since a system that influences which corrections get made still shapes a firm's books. A reasonable open question is how much the "agent" framing changes the underlying work. Rule-based reconciliation tools have flagged exceptions for decades. LedgerLark's bet is that language models can handle the messier parts, such as inferring that a mismatch stems from a transposed digit or an inconsistent security identifier, and can explain that reasoning to a reviewer in plain terms. Whether that holds up across the variety of errors real firms encounter is something only more customers and more time will show. ## Early traction Five firms currently pay for the product, with annual contracts ranging from $36,000 to $110,000. The company says one customer reduced its count of unresolved daily exceptions from 70 to 18. That figure comes from LedgerLark rather than an independent audit, and a single customer's result is a data point, not a pattern. Still, if the reduction is durable, it suggests the kind of outcome operations managers would notice in their day-to-day work. LedgerLark has raised a $3.5 million seed round led by Motive Row. That is enough to fund a small team for a while, not enough to outspend incumbents. ## Competition and headwinds The company is entering a crowded field. Established reconciliation vendors have long relationships with the firms LedgerLark is courting, and some larger institutions maintain internal software that does similar work. The pitch to a small firm is that existing tools are either too expensive, too rigid or too dependent on configuration work that small operations teams can't spare. Incumbents, for their part, are adding their own machine learning features. The more predictable obstacle is procurement. Financial institutions, even small ones, often require lengthy security and compliance reviews before adopting software that touches trade data. A five-person startup can find itself waiting months for a yes. That slows revenue and tests patience on both sides. ## What to watch The near-term signals are fairly concrete: whether the customer count grows beyond the current five, whether the exception-reduction result repeats elsewhere, and how long sales cycles actually run. LedgerLark's founders know the problem from the inside. The harder part, as with most operations software, is proving the solution generalizes.

MoonshotAI: Kimi K3: # Cloudnote Debugging Page Exposed Excerpts of Customer Prompts for Four Hours *Logs show the page contained excerpts from 1,842 user sessions. The company says it has no evidence anyone beyond the researcher who flagged the problem saw them — but its investigation is not finished.* Cloudnote, which counts 3.2 million registered users, said a debugging page containing excerpts of customer prompts was accessible on the public internet without a password for roughly four hours on Monday. The company removed the page 19 minutes after a security researcher alerted it to the problem. According to the company's logs, the page held excerpts from 1,842 user sessions, each running up to 200 characters — a fraction of a full prompt, but potentially enough to capture whatever a user happened to type: names, drafts, business plans or personal details. Cloudnote said full documents, passwords and payment information were stored elsewhere and were not present on the page. It also said it has found no evidence that anyone besides the researcher accessed the page during the four-hour window. The investigation is not complete, and key facts remain unknown: how the page became publicly accessible in the first place, and whether Cloudnote will notify the users whose sessions appear in its logs. For now, nearly everything confirmed about the incident — 1,842 sessions, four hours, no evidence of access by anyone other than the researcher — comes from the company's own accounting.

Anthropic: Claude Fable 5.1: # Cloudnote Debugging Page Exposed Snippets of Customer Prompts for Four Hours *Excerpts from 1,842 user sessions were reachable without a login on Monday. The company says it has found no sign anyone but the researcher who flagged the problem viewed them, but its review is ongoing.* Cloudnote said an internal debugging page that displayed excerpts of customer prompts was accessible without authentication for about four hours on Monday. Server logs show the page contained fragments from 1,842 user sessions. Each excerpt could include up to 200 characters of text from a prompt, according to the company. Full documents, passwords and payment information were stored on separate systems and were not present on the page. Cloudnote has about 3.2 million registered users; the affected sessions represent a small fraction of that base, though the company has not said how many distinct accounts they belong to. A security researcher notified Cloudnote of the exposure, and the company said it removed the page 19 minutes later. Cloudnote said it has found no evidence that anyone other than the researcher accessed the page, while cautioning that it has not completed its investigation. That caveat matters. Two hundred characters is enough to capture a name, a draft of a sensitive message or the outline of a confidential project, depending on what users typed. What the exposed excerpts actually contained, and whether the logs reviewed so far give a complete picture of who reached the page, remain open questions.

MoonshotAI: Kimi K3: # The Best Model in This Support-Ticket Benchmark Is a Statistical Tie — and One Contender Costs 60% Less A new benchmark of five language models on customer-support work has a headline finding that matters more than any leaderboard position: the two best models finished in a statistical dead heat, and the cheaper of them costs roughly 60% less while answering faster. Researchers gave each model 200 anonymized customer-support tickets and collected one proposed answer per ticket. Three experienced support managers graded every response blindly — without knowing which model wrote it — using a composite score that combined factual correctness, usefulness and tone. Model North averaged 84.2. Model Vale averaged 83.8. That 0.4-point gap was not statistically significant, which means the honest reading is that the two models performed equally well on this test. The meaningful difference is economic: North's estimated cost is $1.90 per 1,000 tickets, versus $0.74 for Vale. Vale also produced its answers 18% faster. ## The rest of the field follows a normal price curve Below the top two, quality falls off in a way that tracks price closely. Model Ember scored 79.1 at $0.31 per 1,000 tickets, Model Kite scored 76.4 at $0.22, and Model Harbor scored 71.0 at $0.08. Two things stand out. First, the drop from Vale to Ember — nearly five points — dwarfs anything separating the leaders. The researchers only reported significance testing for the top pairing, but a gap of that size is unlikely to be noise, so buyers who step down a tier are making a real quality trade-off, not a rounding-error one. Second, Harbor's result deserves context: a score of 71 at barely a tenth of Vale's price might be entirely adequate for drafting replies that human agents edit, even if it would disappoint as a fully autonomous responder. Cost per ticket only means something relative to how the output gets used. ## "Best model" depends on the ticket Treating North and Vale as interchangeable would also miss something. North performed best on billing disputes; Vale performed best on technical troubleshooting. For teams whose queues skew heavily toward one category, the tie at the top may effectively resolve itself. For teams with mixed queues, the results hint at a routing strategy — sending different ticket types to different models — rather than a single-model deployment. The benchmark doesn't test that approach directly, but the category-level variation makes it worth piloting. ## Every model shared one blind spot The most consistent finding in the study had nothing to do with price: all five models performed poorly when a ticket depended on an image attachment. Screenshots are routine in software support, so this is a practical limitation, not an edge case. Until multimodal performance improves, image-heavy tickets are an argument for keeping human review in the loop regardless of which model a team chooses. ## What the test didn't measure The methodology has real strengths — blind grading by three raters reduces brand bias and softens any single grader's quirks — but the scope is narrow in ways readers should weigh: - **One company, one language.** All 200 tickets were English-language tickets from a single software company. Results may not transfer to other industries, support styles or languages. - **No agentic behavior.** The researchers did not test tool use or agentic actions — looking up an account, issuing a refund, escalating a case. Modern support automation increasingly depends on exactly those capabilities. - **Composite scores.** Bundling correctness, usefulness and tone into one number hides which dimension drove any difference. A model could be winning on tone while quietly lagging on factual accuracy. - **One answer per ticket.** With a single response per model per ticket, the test says nothing about consistency — how much a model's quality swings from one run to the next. - **Estimated costs.** Actual per-ticket spend will vary with prompt length, caching and vendor pricing changes. ## The takeaway The benchmark's clearest message is that the most expensive option bought no measurable quality advantage. Teams evaluating these models should resist defaulting to the top-priced name, run pilots on their own ticket mix, and pay attention to category strengths — the gap that matters most in this study isn't between models, it's between ticket types. And whatever the deployment, tickets with images still belong in human hands.

Anthropic: Claude Fable 5.1: # The Best Support Model Isn't the Most Expensive One If you only read one number from this benchmark, make it this: the second-highest-scoring model cost 61% less than the top one, and the quality gap between them was too small to measure. Researchers ran five language models against 200 anonymized customer-support tickets from a single software company. Each model drafted one answer per ticket. Three experienced support managers graded every answer blind, combining factual correctness, usefulness and tone into one composite score. The output is a ranking, but the ranking hides the real story. ## Two models at the top, one price that makes sense Model North finished first with an average of 84.2. Model Vale finished second at 83.8. That 0.4-point difference did not reach statistical significance; with 200 tickets and three graders, it sits inside the noise you'd expect from re-running the test on a different sample. In practical terms, the graders could not tell the two apart. The costs are easy to tell apart. North was estimated at $1.90 per 1,000 tickets, Vale at $0.74. Vale also returned answers 18% faster. For a team handling 50,000 tickets a month, that's roughly $95 versus $37 — small in absolute terms, but it's a 2.5x cost multiple for a quality difference nobody could detect. Below the top two, the picture is more conventional: quality declines as price does. | Model | Avg. score | Cost per 1,000 tickets | |---|---|---| | North | 84.2 | $1.90 | | Vale | 83.8 | $0.74 | | Ember | 79.1 | $0.31 | | Kite | 76.4 | $0.22 | | Harbor | 71.0 | $0.08 | Harbor at $0.08 is remarkably cheap, but a 71.0 average is 13 points behind the leaders, and a 13-point gap on a composite that includes factual correctness is not something to wave away. The budget tier involves a real tradeoff. The choice between North and Vale doesn't. ## Where the models differed in kind, not degree Composite averages flatten some texture worth keeping. North did best on billing disputes — tickets where a customer is contesting a charge, which tend to demand careful tone and precise policy wording. Vale did best on technical troubleshooting, where usefulness depends on diagnosing a problem from an imperfect description. That split may matter more than the headline numbers for anyone who routes tickets by category. A billing-heavy operation might see a genuine advantage from North that the overall average conceals; one dominated by "it doesn't work" tickets might find Vale both cheaper and better. The benchmark doesn't report per-category scores or significance, so treat this as a lead to investigate rather than a settled fact. ## The failure every model shared One result cut across all five: when a ticket depended on an image attachment — a screenshot of an error, a photo of a receipt — every model performed poorly. No amount of money bought a model that handled these tickets well. For a deployment, that argues for triage. Detect tickets with attachments and route them to a human or a separate pipeline before any model drafts a reply. It's also the most useful question to put to future benchmarks: image-dependent tickets are common in software support, and a model that handled them even adequately would change the calculation. ## What this test can't tell you The tickets came from one software company, all in English. Support vocabulary, product complexity and customer expectations vary widely, and a model that handled this company's tickets well may not handle a bank's or a hardware vendor's the same way. The test also measured a single, narrow task: drafting one reply to one ticket. It did not test agentic actions, tool use or multilingual support — no looking up an order, issuing a refund, or answering in Spanish. Many production systems depend on exactly those capabilities, and models that look alike on drafting can diverge sharply once they have to act. Finally, the cost figures are estimates, and pricing moves. The 2.5x ratio between North and Vale could be stale within a quarter. ## The takeaway None of that undercuts the central result. On a blind, human-graded test of real support tickets, the most expensive model and the second-most expensive were indistinguishable in quality, and the cheaper one was also faster. For English-language drafting among these five, the burden of proof sits with the $1.90 option, not the $0.74 one.

MoonshotAI: Kimi K3: # Foundry Hall, Milwaukee's Startup Launchpad, Will Close October 31 *Since 2018, the nonprofit accelerator has supported 212 startups that collectively employ 1,140 people. This year, its largest sponsor declined to renew.* Foundry Hall, the nonprofit accelerator that has backed Milwaukee startups since 2018, will close its doors on October 31, after its largest corporate sponsor declined to renew its funding and no replacement stepped forward. The organization has enough money to operate through October but not enough to run another full program. The program currently underway will be its last. Six employees will lose their jobs, and the twelve startups working out of its space must find somewhere else to go. The decisive loss was Blank River Bank, which supplied 38 percent of Foundry Hall's annual budget. It was an arrangement that sustained the accelerator for years — and one that left it exposed the moment the bank walked away. Replacing more than a third of an annual budget on short notice proved impossible. Executive director Camille Ross said the organization held out hope that another funder would step in. "We kept waiting for the replacement check that never came," she said. A spokesperson for the bank said the decision reflected a change in its community investment priorities and was unrelated to Foundry Hall's performance. The closure will end an organization with a sizable imprint on the city's economy. Since opening in 2018, Foundry Hall has supported 212 startups, which collectively employ 1,140 people. In Milwaukee's startup community, it became a fixture — part office, part meeting ground, part first rung on the ladder. The companies it supported will go on without it. But founders who came through the program describe its value in smaller, more personal terms. For Dev Malik, the building was the difference between having an idea and having a company. "My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." City officials are now trying to limit the fallout for the companies still inside. Deputy mayor Susan Cole said the city is speaking with building owners and local universities about space for the twelve displaced startups. For those companies, the timeline is short: they must be out by the end of the month, and where they land next is still unclear. In the weeks that remain, Foundry Hall will keep running its program while its final startups pack up around it. On October 31, the doors will close for good on the building where Malik took his first real meeting — and Milwaukee will lose the place where, since 2018, its newest companies found their footing.

Anthropic: Claude Fable 5.1: # Foundry Hall Will Close Its Doors October 31 Foundry Hall, the nonprofit accelerator that has helped launch 212 Milwaukee companies since 2018, will close on October 31. The organization's board made the decision after Blank River Bank, its largest corporate sponsor, declined to renew a sponsorship that had covered 38 percent of Foundry Hall's annual budget. In a statement, the bank described the move as a shift in its community investment priorities and said it had nothing to do with Foundry Hall's performance. That distinction offered little practical help to executive director Camille Ross, who spent months trying to find other donors to fill the gap. None came through at the scale required. "We kept waiting for the replacement check that never came," she said. Foundry Hall has enough cash to keep operating through the end of October. It does not have enough to run another full program, which carries a group of founders through months of mentorship, workspace, and introductions to investors. Rather than start something it could not finish, the board chose to wind down. Six staff members will lose their jobs. Twelve startups currently working out of the building have been told they must find new space. The companies that have passed through Foundry Hall now employ 1,140 people. For many of their founders, the building was the first place anyone treated their idea as a business rather than a hobby. Dev Malik was one of them. His company grew out of a Foundry Hall cohort, and he remembers what changed when he walked through the door. "My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." Alumni tend to describe the accelerator less as a program than as a place — somewhere to show up every morning when a company was still a spreadsheet and a hunch. The loss of that address, more than any single service, is what founders say they will feel. City Hall has taken notice. Deputy mayor Susan Cole said the city is talking with building owners and local universities about space for the twelve displaced companies. She did not offer a timeline, and it remains unclear whether any arrangement can replicate the concentrated mix of desks, advisers, and chance encounters that Foundry Hall built over seven years. For now, the numbers Foundry Hall leaves behind — 212 companies, 1,140 jobs — are the ones its supporters will point to when they argue the experiment worked. The numbers that ended it are smaller and simpler: one sponsor, 38 percent, no replacement. On October 31, the last of the staff will turn in their keys and the building will go quiet. Somewhere across the city, a founder who has never heard of Foundry Hall will be pitching from the back table of a coffee shop, waiting for a room like this one to open up.

MoonshotAI: Kimi K3: # Raincheck Raises $8 Million to Help Cities Prepare for Floods Before the Rain Falls *The Austin startup's software maps drainage systems and pairs them with rainfall forecasts, helping local governments decide where to send maintenance crews ahead of a storm.* Raincheck, a startup that builds flood-risk software for local governments, has raised $8 million in seed funding, the company announced today. The Austin-based company maps drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm. Raincheck has published a [detailed description of its methodology](https://example.com/raincheck-methodology). The seed round was led by Basin Ventures, with participation from existing investor Long Mile. Raincheck has now raised $10.5 million in total. CEO Ana Flores said: > "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. The company's forecasts are recommendations only. Emergency managers retain control over deployments and public warnings. ## Competition and Expansion 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.

Anthropic: Claude Fable 5.1: # Raincheck Raises $8 Million to Help Cities Clear Drains Before the Storm *The Austin startup's flood-risk software tells local governments where to send maintenance crews ahead of heavy rain. Basin Ventures led the seed round.* Raincheck, an Austin company that makes flood-risk software for local governments, has raised $8 million in seed funding, the company announced today. The round was led by Basin Ventures, with participation from existing investor Long Mile. It brings Raincheck's total funding to $10.5 million. Raincheck's software maps a city's drainage systems and combines those maps with rainfall forecasts. Cities use the results to decide where to send maintenance crews before a storm arrives. Chief executive Ana Flores said the hard part for cities is not knowing where flooding happens but deciding what to fix first. > "Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives." The company has published [an explanation of its forecasting methodology](https://example.com/raincheck-methodology). Raincheck currently works with seven cities in Texas and Louisiana. It declined to identify them, saying some contracts have not been publicly announced. Its forecasts are recommendations; emergency managers retain control over deployments and public warnings. Raincheck competes with other flood-planning vendors, including HydroMap, as well as municipal engineering contractors. The company says it will use the new funding to hire eight engineers and expand into Florida.

MoonshotAI: Kimi K3: # Arcwell's Relay Reads Support Chats and Drafts the Reply Arcwell wants customer support agents to spend less time searching for answers and more time approving them. Its new platform, Relay, uses artificial intelligence to follow support conversations as they unfold and suggest responses agents can send to customers. Relay's scope extends beyond drafting replies. The platform watches for signs that a customer may be preparing to cancel and flags those conversations, giving companies an opening to intervene before the account is lost. It also generates summaries of support interactions, so an agent picking up a long-running case — or a manager reviewing one — doesn't have to work through the entire thread. "Support agents spend too much time searching for information that their companies already have," said Priya Shah, Arcwell's chief executive. Relay is built around that diagnosis: rather than asking agents to dig through internal help documents, the platform reads that material itself and brings relevant answers into the conversation. Arcwell tested Relay with 12 companies over six months. Agents using the platform closed support tickets 14% faster, according to the company. Arcwell did not release customer-retention results from the trial — a conspicuous omission, given that spotting at-risk customers is one of the features it is promoting most heavily. Relay costs $65 per agent per month. The platform's design also means customers must hand over a great deal of raw material. Relay requires access to support transcripts and internal help documents, data many businesses treat as sensitive. Arcwell says customer data is encrypted and is not used to train models shared across its customer base, and companies can choose how long their information is retained: 30 days or 90. Those safeguards answer the obvious objections on paper. The harder question is the one the trial left open. Closing tickets faster is a straightforward efficiency gain; whether Relay's early-warning flags actually keep customers from leaving is something Arcwell has yet to put a number on.

Anthropic: Claude Fable 5.1: # Arcwell's Relay Puts AI Over Support Agents' Shoulders Arcwell has unveiled Relay, an artificial intelligence platform that reads customer support conversations as they unfold and feeds agents suggested replies, in a bid to cut the time staff spend hunting for answers. The pitch is straightforward. "Support agents spend too much time searching for information that their companies already have," said CEO Priya Shah. Relay is designed to surface that information when it is needed, drawing on a company's own support transcripts and internal help documents. Suggested answers are the headline feature, but not the only one. Relay also flags customers who appear likely to cancel, giving agents and their managers a chance to intervene before a frustrated customer walks away. And once a conversation ends, the system generates a summary, sparing agents the chore of writing up notes by hand. Arcwell has some early numbers to back its claims. Twelve companies piloted Relay over six months, and according to the company, agents using the tool closed tickets 14% faster than before. That figure speaks to efficiency, which is the metric most support teams are measured on. What Arcwell has not disclosed is whether the churn-detection feature actually kept more customers from leaving. The company released no retention results from the pilot, leaving the value of one of Relay's three core capabilities an open question for now. Pricing is set at $65 per agent per month. Because Relay works by digesting a company's support history, it necessarily sits close to sensitive material: customer complaints, account details, internal troubleshooting guides. Arcwell says that data is encrypted and is not used to train models shared across its customer base, an assurance aimed at businesses wary of seeing their support conversations inform a competitor's tool. Customers can also choose how long Relay keeps their data, with retention periods of either 30 or 90 days. The platform arrives as support teams face mounting pressure to resolve more issues with fewer people, and as vendors race to embed generative AI into the tools agents already use. Arcwell is betting that a system which stays in the background, suggesting rather than replacing, will find a readier audience than fully automated alternatives. The ticket-closure numbers will help make that case. The missing retention data means part of it remains to be proven.

MoonshotAI: Kimi K3: # Software Bug at KiteBridge Sends $2.8 Million in Duplicate Reimbursements to Employees A routine Tuesday-morning software update at KiteBridge turned into a costly failure when the expense-management company generated roughly 13,400 duplicate reimbursement payments totaling $2.8 million. Ninety-one corporate customers were affected before KiteBridge suspended automated payments later that morning, and by Wednesday evening the company was still working to recover about half a million dollars. The trouble began at 8:45 a.m. Tuesday, when KiteBridge deployed version 4.2 of its platform. The first customer complaint arrived at 9:12 a.m., less than half an hour later. By 10:03 a.m., the company had suspended automated payments entirely — a window of just 78 minutes between deployment and shutdown, but long enough for the error to cascade. ## A safety net that caught nothing The technical failure hinged on how payment systems handle uncertainty. When KiteBridge's software sends a reimbursement to its payment processor and doesn't receive confirmation in time — a "timeout" — it can't be sure whether the payment actually went through. Standard practice is to retry the request, and payment systems normally carry safeguards to recognize a retry as a repeat of an earlier instruction rather than a new one. KiteBridge's fallback system — the backup logic meant to keep money moving during exactly this kind of disruption — failed to make that distinction. Each retry was treated as a fresh reimbursement request, so a single legitimate expense could be paid out multiple times. In that short window, the behavior multiplied into 13,400 duplicate payments. The failure is particularly awkward for KiteBridge because of its own marketing. An April changelog for version 4.0 touted the introduction of "exactly-once payment protection," the industry term for the very guarantee that failed on Tuesday. The company has not yet explained how a safeguard it announced just months earlier was defeated by a subsequent update. ## Who bore the cost, and how it was communicated The duplicates drew on employer funds, not employee accounts. Workers who received extra reimbursements did not have money pulled from their bank accounts; it was their employers' accounts that funded the overpayments. Recovering duplicate payments is generally possible — the money trail is well documented — but the process takes time and cooperation from recipients' banks. As of Wednesday evening, some $2.3 million had been recovered, leaving roughly $500,000 still pending. Customers may also question how KiteBridge communicated during the incident. The company's public status page initially described the problem as delayed reimbursements — a framing that understated the situation while duplicate payments were still being tallied. The page did not disclose the duplicate payments until six hours after automated payments were suspended, meaning affected companies spent much of Tuesday without a full picture of what had happened to their funds. CEO Lena Ortiz apologized for the incident and said the company would publish a technical report on the failure. Such postmortems have become standard practice after major software outages, and customers will likely look to it for answers on two questions: how the fallback system came to treat retries as new payments, and why the company's public accounting lagged hours behind what it knew internally.

Anthropic: Claude Fable 5.1: # Software Update Sends $2.8 Million in Duplicate Reimbursements to Employees at 91 Companies A routine software update at expense-management provider KiteBridge triggered 13,400 duplicate reimbursement payments on Tuesday morning, pushing roughly $2.8 million out of corporate accounts before the company shut down its automated payment system. By Wednesday evening, $2.3 million had been recovered, with the remaining half-million still pending. The error struck 91 of KiteBridge's corporate customers. Employees who had filed expense reports were not charged; in every case the duplicate money came from employers, who fund reimbursements through the platform. ## How a timeout became a double payment KiteBridge released version 4.2 of its software at 8:45 a.m. Tuesday. The first customer complaint arrived 27 minutes later, at 9:12 a.m. According to the company, the trouble began when its payment processor — the outside service that actually moves money from an employer's account to an employee's — stopped responding promptly. That kind of delay, known as a timeout, is routine in payment systems, which are built to retry a request when no confirmation comes back. The failure was in what happened next. KiteBridge's fallback system, meant to keep payments flowing when the primary path stalls, did not recognize the retries as repeats. It treated each one as a brand-new reimbursement request and paid it again. Because the retries fired automatically and at scale, thousands of payments doubled within about an hour. That outcome is precisely what safeguards introduced earlier this year were supposed to prevent. An April changelog for version 4.0 advertised "exactly-once payment protection" — a software guarantee that a payment is executed a single time, even if the instruction to make it is sent repeatedly. KiteBridge has not yet said whether version 4.2 altered that protection or bypassed it, and the gap between the April promise and Tuesday's outcome is likely to be a central question in the company's post-mortem. KiteBridge suspended automated payments at 10:03 a.m., about 78 minutes after the deployment and 51 minutes after the first complaint. ## A slow public account The company's public communications lagged behind its internal response. KiteBridge's incident status page initially described the problem as *delayed* reimbursements — the opposite of what affected finance teams were seeing, as money went out twice rather than not at all. The page did not acknowledge duplicate payments until around 4 p.m. Tuesday, six hours after automation had been switched off. During that window, some customers were trying to reconcile unexpected outflows against a status page that gave no indication they had been overpaid. ## Recovery and what comes next Of the $2.8 million sent in error, $2.3 million had been recovered by Wednesday evening. The remaining roughly $500,000 was listed as pending, which typically means reversals have been initiated but have not yet cleared the banking system. Because employees' accounts were never debited, recovery involves pulling back the duplicate deposits rather than correcting charges. Chief executive Lena Ortiz apologized for the incident and said KiteBridge would publish a technical report. She did not give a timeline for the document or say whether the company would compensate affected customers. For the 91 companies involved, the immediate cost is largely administrative: reconciling ledgers, confirming which payments were legitimate and waiting for the last reversals to land. The longer-term question is one of trust in a product whose core job is to move the right amount of money exactly once.

MoonshotAI: Kimi K3: # HelioDock Launches Night Shift Remote Fleet Software, Raises $24 Million Series B *The company says a 14-month beta with nine logistics firms cut unplanned robot stoppages by 17 percent, a figure it has not independently verified.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely instead of keeping staff on-site around the clock. The company also announced a $24 million Series B round led by North Coast Ventures, bringing its total funding to $39 million. The launch is a bet that the constraint on warehouse automation is shifting from the robots themselves to the people paid to watch them. HelioDock's pitch is that a single operator, working from anywhere, can supervise machines spread across multiple facilities and respond when one runs into trouble — no pre-dawn drive required. The company is tying that pitch to a subscription model, charging a recurring per-robot fee rather than treating remote oversight as a built-in feature, and the paired funding announcement gives it new capital as Night Shift moves from a closed beta of nine customers to open sale. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," said Mara Velez, HelioDock's chief executive. The scenario she describes — one malfunctioning machine demanding an in-person trip — is the use case the product's name invokes directly. Night Shift arrives after a 14-month private beta involving nine logistics companies and 310 robots, an average of roughly 34 machines per participating company. That extended run gave HelioDock more than a year of real-world operation ahead of Thursday's release. The company says beta customers experienced 17 percent fewer unplanned robot stoppages while using the software. The figure comes from HelioDock and has not been independently verified, so it is best read as the company's own accounting of the program rather than an audited result. The software also carries a structural dependency: network connectivity. Aaron Pike of Fulcrum Logistics, one of the beta customers, said Night Shift worked well in facilities with strong connectivity but became less useful during rural LTE outages. His account suggests the product's usefulness tracks the quality of a facility's network connection more than the fleet itself. When a connection drops, the robots do not shut down. They return to their locally configured behavior and continue operating; what disappears is the remote layer. Operators lose the ability to control the fleet until service is restored. That creates a wrinkle in the pitch. Distribution centers in rural areas, where cellular coverage can be inconsistent, are in some ways the natural market for remote oversight — and also the sites where Night Shift's core function is most likely to cut out. The software sits on top of the robots' local programming rather than replacing it, so an outage does not idle the fleet. But the value of a per-robot subscription depends on the connection holding, and prospective buyers will need to weigh whether their facilities' coverage is reliable enough to justify the cost. HelioDock will not have the category to itself. Relay Robotics and YardOS both offer competing fleet-management products, so buyers weighing Night Shift will have alternatives — and a basis for comparison when assessing the performance claims HelioDock is bringing out of its beta. Night Shift will cost $90 per robot per month and becomes generally available September 15. At that price, a fleet matching the beta's full 310-robot population would run about $27,900 a month.

Anthropic: Claude Fable 5.1: # HelioDock Launches Remote Fleet Monitoring Software, Raises $24 Million *Night Shift lets warehouse operators oversee robot fleets from afar after a 14-month private beta, though its usefulness depends on the network staying up.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that allows operators to monitor and manage their robot fleets remotely, and announced a $24 million Series B round led by North Coast Ventures. The software targets a persistent headache for warehouses that run robots around the clock: when a machine stalls in the middle of the night, a person usually has to go deal with it. Night Shift, which will cost $90 per robot per month and become generally available September 15, is HelioDock's attempt to let that person handle the problem from a screen instead of a car seat. The product enters a field where Relay Robotics and YardOS already sell fleet-management tools. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," HelioDock CEO Mara Velez said. ## Fourteen months in beta Before Thursday's launch, Night Shift spent 14 months in a private beta with nine logistics companies, covering 310 robots in total. HelioDock says customers in that program experienced 17% fewer unplanned robot stoppages. The company has not released the data behind that figure, and it has not been independently verified. One beta participant described a more qualified result. Aaron Pike of Fulcrum Logistics said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. That shortcoming traces directly to how the product is built. Night Shift depends on a live connection between the warehouse floor and whoever is watching it. When that connection fails, remote control becomes unavailable and the robots fall back to whatever behavior they were configured to follow locally. For a site with reliable broadband, that fallback may rarely come into play. For a facility relying on cellular service in a remote area, the software's central promise — fixing problems without being on site — holds only as long as the signal does. The distinction matters because the scenario Velez describes, a lone robot confused by a pallet at two in the morning, is exactly the moment when a manager would want to intervene remotely. If the outage that strands the robot also severs the monitoring link, Night Shift cannot help until service returns. HelioDock has not said whether the 17% figure reflects beta sites with uniformly strong connectivity or a mix of conditions like those Pike described. ## Pricing and competition At $90 per robot per month, a facility running 100 robots would pay $9,000 monthly, or $108,000 a year, on top of whatever it already spends on the robots themselves. HelioDock has not said whether volume discounts are available or whether the price differs for customers who use robots from other manufacturers. Relay Robotics and YardOS offer competing fleet-management products. HelioDock did not address how Night Shift differs from those offerings, and the company's beta results, as reported, do not include comparisons against them. ## The funding round The $24 million Series B brings HelioDock's total funding to $39 million. North Coast Ventures led the round; HelioDock did not name other participants or disclose a valuation. The raise arrives as HelioDock moves Night Shift from a controlled group of nine customers into general release, where the software will face a wider range of warehouse environments, network conditions and operator expectations than the beta covered. The 310 robots in the trial represent a fraction of what a broad commercial rollout would encompass, and the connectivity gaps Pike flagged suggest that the product's performance will vary considerably by site. General availability begins September 15. Until then, Night Shift remains limited to the nine logistics companies that have been running it since the beta opened.

MoonshotAI: Kimi K3: C

Anthropic: Claude Fable 5.1: C

MoonshotAI: Kimi K3: # 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.

Anthropic: Claude Fable 5.1: # 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: # QueueZero's AI Assistant Debuts by Telling Its CEO to Stop Talking QueueZero on Wednesday launched Proxy, an AI meeting assistant that attends meetings, produces notes and assigns follow-up work to software agents. The product's first public demonstration delivered an unplanned illustration of both its capabilities and its limits. Midway through the event, a customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy, which was listening to the room, interpreted the remark as an instruction. It created two tasks, both assigned to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. He then resumed talking. The unscripted episode cut to the central question facing a growing category of workplace software: when a machine is authorized to act on what it hears, where does listening end and acting begin? Proxy is designed to sit inside meetings as a participant rather than a passive recorder. It transcribes and summarizes discussion, then routes follow-up work to software agents — automated systems that carry out routine tasks on an employee's behalf. QueueZero is pitching the product to teams with heavy meeting schedules, where commitments made on calls are easily lost once everyone hangs up. The company says guardrails are built in. Proxy requires human confirmation before any external message is sent, so a note to a client or partner cannot leave on the assistant's authority alone. Internal tasks, by contrast, can be created automatically — which is how an audience aside became an item on the chief executive's to-do list. Administrators who would rather not test that distinction can disable automatic task creation entirely. Proxy costs $30 per user per month, in line with other AI note-taking and productivity tools sold on a per-seat basis. It enters public beta next week. The market for meeting assistants is crowded, with transcription services and office suites adding similar features at a steady clip. QueueZero is betting that execution — doing the follow-up work, not just recording it — will set Proxy apart from tools that stop at summaries. Whether Wednesday's demonstration helps or hurts that case may depend on the observer. On one hand, the product heard a request, parsed it correctly and acted within seconds — a live demonstration of exactly what QueueZero is selling. On the other, it failed to recognize that the request was never addressed to it, and it assigned its own chief executive a task no employee would dare create. For prospective customers, the incident may serve as a useful preview of the configuration decisions ahead: which meetings Proxy should attend, which tasks it may create on its own, and who gets to tell it — explicitly, this time — when to stop.

Anthropic: Claude Fable 5.1: # QueueZero's New AI Meeting Assistant Assigns Its First Task to the CEO **Proxy, which converts meeting conversation into work items, created two tasks for Evan Wu after an audience comment during the launch demonstration.** QueueZero on Wednesday launched Proxy, an AI meeting assistant designed to attend meetings, produce notes and hand off follow-up work to software agents. The launch event also provided an unplanned illustration of how literally the system can interpret what it hears. Partway through the presentation, a customer in the audience remarked, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy, which was running during the demonstration, treated the comment as an instruction. It generated two tasks and assigned both to QueueZero chief executive Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. The exchange drew attention to a design choice at the center of the product. Proxy is built to shorten the gap between a decision made in a meeting and the work that follows from it. Rather than producing a transcript for someone to review later, it identifies action items as they arise and either assigns them to people or dispatches them to software agents that can carry out routine work. That capability depends on Proxy deciding what counts as an instruction, and Wednesday's demonstration showed that a passing remark from the audience can clear that bar. QueueZero says the system has limits on what it will do without human approval. Proxy requires confirmation before it sends any external message, so a task such as "Send pricing sheet" would not have resulted in a document leaving the company on its own. Internal tasks, however, can be created automatically, which is how the two items appeared on Wu's list without anyone signing off on them. Administrators can disable automatic task creation for their organizations, the company says. With that setting turned off, Proxy would propose tasks rather than create them. The incident is likely to be cited by both sides of an ongoing debate about AI assistants in the workplace. Supporters will note that the system did exactly what it was asked and that the safeguard on external messages held. Skeptics will observe that nobody, strictly speaking, asked it anything. For QueueZero, the demonstration ended with the pricing sheet delivered and the product's behavior on display in a way no scripted demo could have arranged. Whether prospective customers read that as reassurance or caution may depend on how often their own meetings include offhand comments about the people running them. Proxy costs $30 per user per month. It enters public beta next week.