Head to head: Anthropic: Claude Fable 5.2 vs Meta: Muse Spark Open

Anthropic: Claude Fable 5.2: # Arcwell's Relay puts AI-suggested answers in front of support agents *The company says testers closed tickets 14% faster, but it has not shared data on customer retention.* Arcwell has launched Relay, an AI platform that reads customer support conversations and suggests answers to the agents handling them. The company is pitching it as a remedy for the time support staff lose hunting for information. "Support agents spend too much time searching for information that their companies already have," chief executive Priya Shah said. Answer suggestions are only part of the package. Relay also identifies customers who may cancel and creates summaries, and it is priced at $65 per agent each month. Arcwell has early numbers to support the pitch. Twelve companies tested Relay over six months, and the company says agents using it closed tickets 14% faster. That figure comes from Arcwell itself, and it measures speed alone. The company did not release customer-retention results, a notable omission for a product that promises to spot customers at risk of leaving. Faster ticket closures suggest agents are moving through their queues more quickly; they say nothing about whether flagging a likely cancellation helps a company keep that customer. Relay comes with a trade-off. To do its work, the platform requires access to support transcripts and internal help documents, so any company considering it will have to decide how comfortable it is opening that material to an outside vendor. Arcwell's answer is a set of safeguards. It says customer data is encrypted and is not used to train shared models. Companies can also choose between a 30-day and a 90-day retention period. For now, the case for Relay rests on a single figure reported by its maker: tickets closed 14% faster across a dozen test customers. Buyers will have to weigh that against the $65 monthly cost per agent and the data access the platform demands. Whether the platform can help companies hold on to the customers it flags is a question Arcwell has yet to answer publicly.

Meta: Muse Spark Open: # Arcwell Launches Relay to Put AI Inside the Support Queue Arcwell has launched Relay, a new platform that uses artificial intelligence to examine customer support conversations and suggest answers to agents. The tool is aimed at reducing the time agents spend searching through internal help documents and past tickets for information their companies already hold. In addition to surfacing suggested replies, Relay is built to identify customers who may be about to cancel and to automatically create summaries of support interactions. The pitch comes as support teams face rising ticket volumes and growing pressure to resolve issues without adding staff, a squeeze that has made customer service a crowded testing ground for AI vendors. Relay was tested over six months by 12 companies, Arcwell said. During that trial, agents using the system closed tickets 14% faster. The company did not release customer-retention results tied to its ability to flag at-risk customers. The platform will cost $65 per agent each month. The company is betting that shaving minutes off each interaction will matter more at scale than any single dramatic capability. “Support agents spend too much time searching for information that their companies already have,” CEO Priya Shah said. To work, Relay requires broad access to sensitive internal material, including past support transcripts and help documentation that it uses to formulate its suggestions. That dependency is likely to prompt questions from security and compliance teams weighing the productivity gains against data exposure. Arcwell said customer data processed by Relay is encrypted and is not used to train shared models. Companies can choose whether that data is retained for 30 days or 90 days, a window that determines how long transcripts and related material remain in the system. For Arcwell, Relay represents a push to bring its AI capabilities directly into day-to-day support work — from resolving tickets to preventing churn and documenting conversations — at a moment when enterprises are deciding how deeply to let AI into customer-facing operations.

Anthropic: Claude Fable 5.2: # In a Support-Ticket Benchmark, the Top Two Models Tie, and One Costs 61% Less A new benchmark of five language models on customer-support tickets has a nominal winner, Model North, with an average score of 84.2. The more useful finding is that the runner-up, Model Vale, scored 83.8, a difference the researchers say is not statistically significant. Vale also costs about 61% less to run and answers 18% faster. ## How the test worked Researchers drew 200 anonymized tickets from a software company's support queue and had each model write one proposed answer per ticket. Three experienced support managers graded the answers without knowing which model had written which. Each score combined factual correctness, usefulness, and tone. The blind grading by practitioners is a strength of the design, since the scores reflect what support managers consider a good answer. But a combined score can hide trade-offs. Two answers could earn similar marks for different reasons, such as one being accurate but curt and the other warm but less precise. And because each model answered each ticket only once, the results cannot show how much a model's quality varies from one attempt to the next. ## A tie at the top, a gap in price | Model | Average score | Est. cost per 1,000 tickets | |---|---|---| | North | 84.2 | $1.90 | | Vale | 83.8 | $0.74 | | Ember | 79.1 | $0.31 | | Kite | 76.4 | $0.22 | | Harbor | 71.0 | $0.08 | The 0.4-point gap between North and Vale is within the range chance could produce on a 200-ticket sample. On overall quality, the benchmark cannot separate them. It can separate them on cost and speed. North's estimated cost of $1.90 per 1,000 tickets is roughly 2.6 times Vale's $0.74. In absolute terms the sums are small: at a million tickets, the difference is about $1,160. For most support teams, that makes the price gap a tiebreaker, and Vale's 18% speed advantage may matter more wherever customers are waiting on a reply. The results do contain one reason to pay more for North, which is ticket type. North performed best on billing disputes, while Vale performed best on technical troubleshooting. A team whose queue is dominated by billing problems could reasonably choose North. The summary results don't say how large those category advantages were, though, and each category contains only a fraction of the 200 tickets, so they are best treated as leads to check against a team's own tickets. ## Lower cost, lower scores Below the top pair, scores fall along with price. Ember trails Vale by nearly five points at less than half the cost, and Harbor, at $0.08 per 1,000 tickets, sits 13 points below the leaders. These gaps are much larger than the one between North and Vale, though the reported significance test covers only the top pair. Whether the savings justify the drop depends on what a weak answer costs. When even the most expensive model runs under $2 per 1,000 tickets, one mishandled ticket that escalates to a human agent or loses a customer could cancel out the savings from a cheaper model. The cheaper models might still suit high-volume, low-stakes queues, but the benchmark doesn't test that directly. ## A shared weakness All five models performed poorly when a ticket depended on an image attachment, such as a screenshot of an error. Because the weakness appeared at every price point, paying for a better model would not fix it. Teams deploying any of these models should plan to detect such tickets and route them to people. ## What the benchmark doesn't show The study has limits that matter for anyone applying its results: - **One company, one language.** All tickets were in English and came from a single software company. Results may differ in other industries, products, and languages, and multilingual support was not tested. - **Drafting only.** The models wrote proposed answers. The researchers did not test agentic actions or tool use, such as looking up an account or issuing a refund, so the scores don't show how the models would perform in an automated workflow. - **Estimated costs.** The cost figures are estimates, and actual costs will depend on ticket length and pricing terms. ## The takeaway In this benchmark, the highest score did not identify a clear best model. North and Vale are indistinguishable on overall quality, Vale is cheaper and faster, and each leads in a different ticket category. Teams choosing between them should test both on their own tickets, and whichever they pick, they should keep people on tickets that depend on images.

Meta: Muse Spark Open: # The Price of "Good Enough" Is Dropping Fast For customer-support teams shopping for AI, the most expensive model is no longer the obvious choice. In a new blind benchmark of five language models, the highest-priced system failed to outperform a rival that costs less than half as much — and answers faster. That tie at the top is the real story. Not who scored 84.2 versus 83.8, but what that deadlock means for buyers: top-tier quality is now available without paying a premium, while the cheapest options still lag far enough behind to matter. ### How the Test Worked Researchers tested the models on 200 anonymized, real customer-support tickets from a single software company. Each model generated one proposed answer per ticket, with no access to tools, account systems, or follow-up actions. Three experienced support managers graded every answer blindly, without knowing which model wrote it. Scores combined three things customers actually care about: factual correctness, usefulness in resolving the issue, and tone. This is a focused, human-judged test, not an automated leaderboard. That makes the results more relevant to support leaders, but also narrower. With 200 English-only tickets from one company, the benchmark captures a realistic slice of work, not the full range of support. The researchers also deliberately excluded agentic tasks like issuing refunds, looking up accounts, or handling other languages. ### A Tie at the Top Changes the Math Two models, North and Vale, separated themselves from the rest with average scores in the mid-80s. Statistically, there was no winner between them; the gap between North's 84.2 and Vale's 83.8 was not significant given the sample size and blind grading. The business difference, however, was stark. Vale is estimated to cost $0.74 per 1,000 tickets, compared to $1.90 for North — roughly a 60% savings for effectively identical quality. Vale was also 18% faster in response time, a margin that compounds quickly in high-volume queues. That reframes the buying question. If quality is a tie, cost and latency become the tiebreakers, and the cheaper, quicker model wins on practicality. The remaining models help illustrate why that matters. Two mid-cost options landed meaningfully lower, in the upper to mid-70s, while the budget model, at just $0.08 per 1,000 tickets, averaged 71.0 — more than a full tier below the leaders. In other words, going cheap still has a cost in answer quality, but paying the most no longer guarantees the best result. The price-performance curve has bent. ### Special Strengths, Shared Weakness The averages hide some useful nuance. North edged ahead on billing disputes, the kind of tickets that hinge on policy precision and careful phrasing. Vale led on technical troubleshooting, where step-by-step clarity matters more than diplomacy. For a team drowning in one type of ticket, that specialty could nudge a decision. For most generalist support desks, though, the difference reinforces the main finding: choose based on your ticket mix and workflow, not a fractions-of-a-point overall lead. Every model shared one clear failure mode. When a ticket depended on an image attachment — a screenshot of an error, a photo of an invoice — performance dropped across the board. Text-only models, however fluent, cannot reliably interpret visual context, and this benchmark confirmed that limitation in a support setting. ### What We Still Don't Know The benchmark's constraints should shape how readers use it. First, language and domain matter. All tickets were in English and came from one software company. A model that excels at SaaS troubleshooting may not handle retail returns, medical queries, or multilingual customers with the same poise. Second, the test was intentionally non-agentic. Modern support AI is often asked to *do* things, not just draft answers: check order status, reset a password, escalate with context. None of that was evaluated here. A model that writes the best draft is not automatically the best at taking action. Finally, human grading, while valuable, is subjective. Combining correctness, usefulness, and tone into a single score is sensible for customer experience, but it can mask trade-offs. A perfectly accurate answer with a cold tone might score similarly to a friendly answer with a minor factual slip. Leaders should weight those dimensions for their own brand. ### The Takeaway for Buyers This benchmark is not a coronation of one model. It is evidence that the support-AI market has matured into distinct tiers. At the top, premium pricing no longer equals premium performance. The two leading models were indistinguishable on quality, but one was dramatically cheaper and faster — a combination that should push procurement teams to demand proof, not just reputation. In the middle and bottom tiers, lower cost still correlates with lower scores, but the gap is narrowing enough that a cost-conscious team with strong human review might find acceptable trade-offs. The smart next step is not to chase the highest number in a table. It is to run a similar blind test on your own tickets, including your hardest cases — especially those with images — and decide whether the draft needs to be perfect, or just good enough to help your agents move faster.

Anthropic: Claude Fable 5.2: **C** It is the only option that sticks to the stated facts and leads with the specifics: what happened, when, how many accounts, and for how long. - **A** is wrong because message contents were not sent, and "massive privacy disaster" is editorializing. - **B** buries the news, and "innovative" and "difficult decision" are unsupported promotional language. - **D** is false because only workspace titles were exposed, and the company found no evidence of malicious access.

Meta: Muse Spark Open: C

Anthropic: Claude Fable 5.2: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

Meta: Muse Spark Open: # 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 5.2: # Startup Spotlight: LedgerLark and the Work That Starts After the Trade *The seed-stage company builds AI agents that propose fixes for trade mismatches at small investment firms. Five customers are paying, and several important questions remain open.* A copied account number is a small mistake. In the case Mina Batra saw, it caused a trade mismatch that took a six-person team three days to resolve. The episode prompted her to start exploring the idea behind LedgerLark, a startup building AI agents for post-trade operations at small investment firms. ## The insight Batra spent seven years in operations at a regional brokerage. Her co-founder, Owen Carr, built reconciliation systems at a large bank. Between them, they have seen the problem from two angles: the people who chase discrepancies and the software meant to catch them. Much of post-trade work involves comparing records that should agree and working out why they don't. LedgerLark is betting that small firms feel this burden acutely, and that software able to suggest a fix, not just flag a break, could take much of the routine work off their operations teams. ## What the product does LedgerLark's agents compare trade confirmations, custodian records, and a firm's internal ledgers. When the records disagree, the agent proposes a correction, and a person decides whether to approve it. The product does not execute trades or move money. That keeps the software in an advisory role, which limits the damage a wrong suggestion can do and may make cautious buyers more comfortable. It also means the product's value depends on the quality of its proposals. If staff have to re-verify every suggestion from scratch, the time savings shrink. ## Early traction Five firms are paying customers, with annual contracts ranging from $36,000 to $110,000. The company has raised a $3.5 million seed round led by Motive Row. LedgerLark says one customer reduced its unresolved daily exceptions from 70 to 18. Unresolved exceptions are a number operations teams track closely, so the metric is a relevant one. But this is a single, company-reported data point. The reporting doesn't show how long the reduction took, whether other customers have seen similar results, or how much of the change came from the software rather than from process changes made alongside it. ## The open questions **Competition.** Established reconciliation vendors already sell into this market, and large banks build their own internal software. LedgerLark will need to show that agent-proposed corrections are a meaningful improvement over what those tools offer, and that incumbents can't easily add something similar. **Sales cycles.** Financial institutions may require lengthy security and procurement reviews before adopting new software, particularly software that touches trade and account records. For a seed-stage company, long reviews mean slower revenue and more pressure on a finite runway. Smaller firms may move faster than large ones, but that has yet to be demonstrated. **Reliability at scale.** Five customers is a start, not a pattern. It remains to be seen how the agents perform across different custodians, data formats, and exception types, and how firms will audit the agents' suggestions over time. ## What to watch - **Renewals and expansion.** Whether the first five customers stay and broaden their use will say more than the initial contracts do. - **Corroborated results.** More customers reporting outcomes, ideally in their own words, would strengthen the 70-to-18 claim. - **Time to close.** The length of procurement reviews will shape how quickly the company can grow. LedgerLark's founders are working on a problem they have encountered firsthand, and they have chosen a deliberately narrow role for their software: suggest, don't act. The next year or two should show whether that restraint, along with five early customers, is enough to win a place in a market with entrenched alternatives.

Meta: Muse Spark Open: # Startup Spotlight: LedgerLark Small investment firms face the same post-trade tangle as larger shops — confirmations that don't match, custodian records that drift from internal books, exceptions that linger until someone tracks down the typo. LedgerLark is building AI agents intended to take on that reconciliation work. The early pitch is narrow on purpose: help small operations teams find and fix mismatches faster, without touching execution or the movement of money. ## The insight from the back office Co-founder Mina Batra spent seven years in operations at a regional brokerage before starting LedgerLark. Co-founder Owen Carr previously built reconciliation systems inside a large bank. The idea for LedgerLark, according to Batra, took shape after a single account-number error — a copied number on a trade — led a six-person team to spend three days resolving the mismatch. It was not an unusual failure mode, but the time it consumed raised a question the founders kept returning to: how much operational capacity goes to hunting discrepancies that are, in principle, detectable? That question points to a familiar tension in operations. Exceptions are often simple at root — a transposed digit, a missed corporate action, a timing difference — but confirming the correct record across systems still requires pulling data from multiple places and building confidence in a fix. ## What the agents do LedgerLark's agents compare three sources: trade confirmations, custodian records, and a firm's internal ledger. When the sources disagree, the system surfaces the difference and proposes a correction, leaving approval to a human. Importantly, the product does not execute trades or move money. The human-in-the-loop design is central both for control and for adoption; firms can review proposed corrections before any ledger is updated. In practice, that means the work LedgerLark is automating is not judgment about investment strategy but assembly and comparison of records. The construct is less autonomous agent than assistant that drafts an exception note, pulls supporting evidence, and queues a suggested edit. Whether firms will trust an AI-generated proposal as much as a junior analyst's will likely depend on how clearly the system shows its reasoning and audit trail — a detail that remains to be seen outside customer demos. ## Early traction LedgerLark says it has five paying customers, all small investment firms, on annual contracts that the company says range from $36,000 to $110,000. The company also says one customer reduced its unresolved daily exceptions from about 70 to 18 after adopting the product. LedgerLark has not shared the time frame for that change or what share can be attributed to its software versus process changes, and the figure has not been independently verified. The company has raised a $3.5 million seed round led by Motive Row. For a seed-stage operations tool, five paying annual contracts suggests some willingness to pay for relief on a painful, if unglamorous, problem. It is not yet evidence of repeatable sales, especially in a buyer group that can be hard to reach and slow to purchase. ## Competition and constraints LedgerLark is entering a market with entrenched options. Established reconciliation vendors already serve many firms, and larger institutions often rely on internal software built over years to match local workflows. Newer entrants also face procurement realities that have little to do with product quality. Financial institutions, even small ones, can require lengthy security reviews, vendor diligence, and operational sign-off before a tool that touches ledgers is trusted in production. Those cycles can extend sales timelines and strain a small startup's resources. There is also the technical threshold: firms will need to connect sensitive data sources and be satisfied that proposed corrections are accurate and explainable. Mistakes in this context, even if still routed through humans, carry risk. ## What to watch The next test for LedgerLark will be whether it can move from a handful of early customers — likely more tolerant of rough edges — to a broader set of firms with varied custodians, data quality, and exception types. Constructively, the opportunity is clear: small teams often lack the off-hours coverage that larger operations have, so a tool that reliably shortens the exception queue could free meaningful time. The open questions are adoption-driven: how quickly can LedgerLark complete security reviews, integrate without heavy services work, and build enough trust that human reviewers approve proposals quickly? If it can, the mundane nature of the problem may be an advantage. Unlike a trading signal, reconciliation is easy to define and measure — exceptions resolved, time to close — which gives both the startup and its customers a concrete way to judge value.

Anthropic: Claude Fable 5.2: **6** Every non-blank line in the draft breaks a rule: 1. `#Flood software moves upstream` has no space after `#`, so it is not in `# Headline` form. 2. `The company is expanding its pilot.` is the dek but is not italicized. 3. `##Details` has no space after `##`, so it is not in `## Heading` form. 4. `Read [the announcement](www.example.com/news).` links to a URL that is not absolute HTTPS. 5. `* The first pilot begins Monday.` starts with an asterisk instead of a hyphen. 6. `<div>The second pilot begins Friday.</div>` is raw HTML.

Meta: Muse Spark Open: 6

Anthropic: Claude Fable 5.2: # Raincheck raises \$8 million to help cities prepare for floods *The Austin startup, whose software helps local governments decide where to send maintenance crews before a storm, plans to hire eight engineers and expand into Florida.* Raincheck, an Austin company that makes flood-risk software for local governments, has raised \$8 million in a seed round led by Basin Ventures. The funding was announced today. Existing investor Long Mile also participated. The round brings Raincheck's total funding to \$10.5 million. The company says the new money will help it hire eight engineers and expand into Florida. Raincheck maps drainage systems and [combines those maps with rainfall forecasts](https://example.com/raincheck-methodology). Cities use the software to decide where to send maintenance crews before a storm. CEO Ana Flores said knowing where flooding happens is not the hard part for cities: > 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's forecasts are recommendations. Emergency managers retain control over deployments and public warnings. Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Other flood-planning vendors include HydroMap and municipal engineering contractors.

Meta: Muse Spark Open: # Raincheck Raises $8 Million to Help Cities Stay Ahead of Floods *The Austin startup's software maps drainage systems and rainfall forecasts to guide storm preparations as it plans Florida expansion* Raincheck, an Austin-based company that makes flood-risk software for local governments, has raised $8 million in a seed round announced today. Its platform maps municipal drainage systems and combines those maps with rainfall forecasts to help cities decide where to send maintenance crews before a storm. More details on [how Raincheck's forecasting system works](https://example.com/raincheck-methodology) are available from the company. The seed round was led by Basin Ventures, with participation from existing investor Long Mile. The new funding brings Raincheck's total raised to $10.5 million. CEO Ana Flores said the challenge for cities is not identifying where flooding occurs, but prioritizing how to respond: > “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 provided as recommendations, with emergency managers retaining control over deployments and public warnings. The company competes with other flood-planning vendors including HydroMap and municipal engineering contractors. Raincheck said the new funding will allow it to hire eight engineers and expand into Florida.

Anthropic: Claude Fable 5.2: # HelioDock Launches Remote Monitoring Software for Warehouse Robots, Raises $24 Million *The company says its Night Shift software cut unplanned stoppages in a 14-month beta. One tester found its usefulness depended on the network.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely. It also announced a $24 million Series B funding round led by North Coast Ventures. The release adds a per-robot monthly software fee to HelioDock's hardware business and puts the company in a category where Relay Robotics and YardOS already sell fleet-management products. Its pitch centers on what happens when a robot stalls and nobody is nearby to deal with it. The pitch also leans on a performance claim that so far comes only from the company. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," CEO Mara Velez said. ## A 14-month beta Night Shift spent 14 months in a private beta before Thursday's launch. Nine logistics companies took part, running the software across 310 robots, an average of roughly 34 machines per company. HelioDock says those customers experienced 17% fewer unplanned robot stoppages while using the software. The figure has not been independently verified. ## Where the connection matters The experience of one beta customer shows where the product works best. Aaron Pike of Fulcrum Logistics said the system worked well in facilities with strong connectivity. It became less useful, he said, during LTE outages at rural sites. That limitation follows from how the product is built. Night Shift depends on a live link between the operator and the warehouse floor. When connectivity fails, remote control becomes unavailable and the robots return to their locally configured behavior. In practice, a dropped connection does not leave the machines without instructions, because they fall back on the settings already configured on site. But for as long as the outage lasts, the remote operator has no way to step in. The off-site manager in Velez's example would be back to choosing between waiting and making the drive. For operators with well-connected buildings, that may rarely come up. For those running facilities in areas with unreliable coverage, Pike's account suggests the software's value will vary from site to site. ## Pricing and competition Night Shift will cost $90 per robot per month and becomes generally available on September 15. At that rate, a warehouse running 50 robots would pay $4,500 a month, or $54,000 a year. Covering a fleet the size of the entire beta, all 310 robots, would cost $27,900 a month. Prospective buyers can weigh that price against existing alternatives. Relay Robotics and YardOS both offer competing fleet-management products, so HelioDock is entering a market with established options. The 17% stoppage figure is the main evidence the company has offered for its own product so far. ## The funding The Series B brings HelioDock's total funding to $39 million. The $24 million round, led by North Coast Ventures, accounts for more than 60% of all the money the company has raised. The round was announced alongside the product, about a month and a half before general availability on September 15. From that date, the software will be open to customers beyond the nine companies that tested it.

Meta: Muse Spark Open: # HelioDock Launches Remote Monitoring Software in Push Beyond Warehouse Floors *Fourteen-month beta preceded Thursday's debut of Night Shift, which HelioDock will sell alongside fresh Series B funding* HelioDock on Thursday introduced software that lets a single operator watch over fleets of its autonomous warehouse robots from outside the building, a shift the company is pairing with a $24 million funding round to expand its reach among logistics providers. The launch of Night Shift moves HelioDock beyond selling on-site automation and into the market for remote fleet oversight, where logistics firms have been looking for ways to keep robots moving through the night without stationing managers on the floor. The company’s claims about reduced stoppages have not been independently verified and the system depends on network availability, two caveats that frame an otherwise routine product expansion in warehouse automation. Night Shift will become generally available September 15 and cost $90 per robot per month. Built for HelioDock’s autonomous warehouse robots, the software lets off-site operators monitor fleet activity without being physically present during off-hours. Operators using Night Shift can monitor fleets from a remote location rather than walking the warehouse floor. The goal, according to the company, is to avoid the late-night trips to a facility that have become a familiar pain point for managers overseeing automated sites. “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet,” CEO Mara Velez said. Development of Night Shift stretched across 14 months in a private beta that HelioDock ran with nine logistics companies and 310 robots. The nine logistics companies in the beta collectively operated 310 robots. The extended trial put the software inside live warehouse environments where operators used remote monitoring during regular and off-hours operations. HelioDock says beta customers in that group experienced 17% fewer unplanned robot stoppages. The figure has not been independently verified. Performance in those live settings tracked closely with connectivity. Aaron Pike of Fulcrum Logistics, a beta customer, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. HelioDock acknowledged that limitation in how the software handles signal loss. When connectivity fails, robots return to their locally configured behavior. Remote control becomes unavailable until a connection is restored, leaving machines to operate on their on-board routines. That dependency places HelioDock alongside other vendors navigating the same infrastructure constraint. Relay Robotics and YardOS sell competing fleet-management products that provide centralized oversight of warehouse robots, and all such systems require a stable link between the floor and the remote operator to function as designed. HelioDock has not disclosed how Night Shift manages handoffs between local and remote control beyond the fallback to local behavior. HelioDock makes autonomous warehouse robots. Night Shift is intended to extend the utility of those machines into hours when fewer staff are on site. The beta period covered 14 months, a timeframe the company used to test remote monitoring with nine logistics companies and 310 robots. Funding for the rollout comes from a $24 million Series B led by North Coast Ventures, the company announced Thursday alongside the product. The round brings HelioDock’s total funding to $39 million. HelioDock has positioned Night Shift as a companion to its hardware, rather than a standalone platform. The product will be sold on a per-robot monthly subscription basis.

Anthropic: Claude Fable 5.2: # Cloudnote Left Customer Prompt Excerpts Exposed on Unsecured Debugging Page *Snippets from 1,842 user sessions were viewable without a login for four hours Monday. The company says only the researcher who reported it is known to have seen them, but its investigation is unfinished.* Cloudnote said a debugging page containing excerpts of customer prompts was accessible without authentication for four hours on Monday, leaving fragments of what users typed into the service open to anyone who found the page. The company removed it 19 minutes after a researcher reported the problem. Logs show the page displayed excerpts from 1,842 user sessions, each limited to 200 characters of a prompt. That figure counts sessions, not people, and the number of individual customers involved has not been established. Cloudnote has 3.2 million registered users. The exposure was narrower than a breach of stored accounts. Full documents, passwords and payment information were kept elsewhere and were not on the page, according to the company. But 200 characters is enough room for a name, an address or a line of confidential work, and what the exposed excerpts actually contained has not been disclosed. The larger open question is whether anyone else saw them. Cloudnote says it has found no evidence that anyone besides the researcher accessed the page. That is a preliminary finding, not a conclusion: the company's investigation is still under way, and an absence of evidence so far does not settle what happened during those four hours.

Meta: Muse Spark Open: # Cloudnote Says Debugging Page Briefly Exposed Excerpts From Nearly 2,000 Customer Prompts *An unauthenticated page was accessible for four hours Monday; the company says it acted quickly after a researcher's report but has not finished investigating potential access* Cloudnote exposed excerpts from customer prompts on an unauthenticated debugging page for four hours on Monday before removing it 19 minutes after being alerted by an outside security researcher, the company said. According to Cloudnote's initial review of logs, the page displayed snippets of up to 200 characters from 1,842 user sessions — less than one-tenth of one percent of its 3.2 million registered users. Full documents, passwords and payment information were not present on the page, as they are stored separately, the company said. Even truncated prompts can contain sensitive personal, business or legal details typed by users who assume their inputs are private, a risk that makes any exposure of prompt data potentially consequential even when passwords, payment information or full files remain secured in separate systems. What remains unclear is whether anyone beyond the researcher saw or captured the data. Cloudnote said it has found no evidence of additional access so far, but it has not yet completed its investigation and did not say when it would notify affected users or disclose how the debugging page became publicly accessible. The pending findings will determine how many users need to take protective steps.

Anthropic: Claude Fable 5.2: # QueueZero's AI Meeting Assistant Tells Its CEO to "Stop Talking" During Launch Demo *Proxy, which enters public beta next week, turned an audience member's remark into two tasks for the chief executive.* QueueZero launched an AI meeting assistant called Proxy on Wednesday, and the launch demonstration included an unscripted test of one of its core features. The assistant assigned the company's chief executive a task to stop talking. Proxy is designed to attend a meeting, produce notes and assign follow-up work to software agents. 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 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. As demonstrations of automatic task creation go, it was a thorough one. ## Where the safeguards sit The episode drew attention to how Proxy handles the actions it generates. QueueZero says the assistant requires confirmation before an external message is sent. Internal tasks can be created automatically, which is what happened on Wednesday. Administrators can disable automatic task creation, according to the company. The consequences in this case were minor. Proxy added two items to a task list, and Wu treated one as a legitimate request and fulfilled it. But the moment illustrated a question facing any software that listens to workplace conversation: how to tell a directive from an aside, and whose remarks count as instructions. Proxy acted on a comment from an audience member and split a single sentence into two separate assignments. QueueZero's design puts the checkpoint at the edge of the organization. Anything sent outside the company needs human sign-off. Work created inside it does not, unless an administrator changes the setting. That default is likely to matter to organizations evaluating the product. A mistaken internal task is easy to delete, as Wu showed. The calculation may differ in a busy workplace where tasks are generated across many meetings, or where follow-up work is handed to software agents and not to a person who can laugh it off. ## Pricing and availability Proxy costs $30 per user per month. It enters public beta next week.

Meta: Muse Spark Open: # QueueZero Debuts Proxy, an AI Assistant That Turns Meetings Into Tasks **San Francisco —** QueueZero on Wednesday launched Proxy, an AI meeting assistant designed to attend virtual meetings, generate notes and automatically delegate follow-up work to software agents. The launch, held at an invite-only event in San Francisco, marks QueueZero's first major product release since its founding two years ago. The company said Proxy is built to reduce post-meeting administrative work by connecting transcription directly to task execution. According to QueueZero, Proxy can join meetings as a participant, produce summaries and action items, and assign tasks to integrated software agents that handle functions such as scheduling, email drafting and project management updates. The system's ability to act autonomously was illustrated during the live demonstration. Midway through a presentation by QueueZero Chief Executive Evan Wu, an audience member — identified by the company as a customer — remarked, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy, which was actively transcribing the session, interpreted the comment as an instruction. Within seconds, it created two tasks assigned to Wu: “Stop talking” and “Send pricing sheet.” The tasks appeared on a shared screen visible to attendees. Wu laughed and acknowledged the assignments, completing the second task by distributing the pricing sheet and deleting the first. The moment prompted applause from the audience but did not disrupt the presentation. QueueZero said the incident highlights both the capabilities and configurable limits of Proxy's automation. The company said Proxy requires user confirmation before sending external messages, such as emails to customers or partners, to prevent unintended communications. Internal tasks, however, can be created automatically without confirmation, allowing teams to move quickly from discussion to execution. Administrators can disable automatic task creation entirely or require approval for all actions, the company said. Analysts note that automatic task generation raises questions about accuracy and oversight, particularly when AI systems interpret casual conversation as directive. QueueZero said its permission model is intended to give organizations control while preserving speed. Proxy will cost $30 per user per month and is set to enter public beta next week. The company said general availability is planned for later this year, with additional integrations to follow.

Anthropic: Claude Fable 5.2: # KiteBridge error sends out $2.8 million in duplicate reimbursements *Ninety-one corporate customers funded 13,400 payments they did not owe. About $500,000 had yet to be recovered by Wednesday evening.* Expense-management software maker KiteBridge issued 13,400 duplicate reimbursements worth $2.8 million on Tuesday morning, leaving 91 corporate customers to cover payments they never owed. For hours afterward, the company's public incident page described the problem only as delayed reimbursements. By Wednesday evening, $2.3 million had been recovered. The remaining sum, roughly $500,000, was still pending. The cost fell on employers, not their workers. The duplicate payments were funded by KiteBridge's corporate customers, and no employees had money withdrawn from their accounts. Chief executive Lena Ortiz apologized and said the company would publish a technical report on the failure. ## A morning release, then duplicate payments The trouble began after KiteBridge deployed version 4.2 of its software at 8:45 a.m. Tuesday. The first customer complaint arrived at 9:12 a.m., less than half an hour later. The duplicates trace to a timeout involving KiteBridge's payment processor, the outside service that moves the money. A timeout occurs when one system sends a request to another and does not get an answer within a set period. The sender cannot tell whether the request failed or whether it succeeded and the confirmation was lost. Software typically responds to a timeout by trying again. That is safe only if the system recognizes the second attempt as a repeat of the first. KiteBridge's fallback system, the backup process that takes over when the normal payment path fails, did not. It treated each retry as a new reimbursement request, so a single approved expense could be paid more than once. KiteBridge suspended automated payments at 10:03 a.m., 51 minutes after the first complaint and a little more than an hour and a quarter after the new version went live. ## A safeguard that was supposed to prevent this The failure is the kind KiteBridge had told customers its software was built to prevent. An April changelog said version 4.0 had introduced "exactly-once payment protection." In payment systems, "exactly-once" refers to a guarantee that a transaction will be carried out a single time no matter how often the underlying request is retried. Treating retries as new requests is the specific error such protection exists to catch. The reporting available so far does not establish why that protection failed to stop Tuesday's duplicates, or whether the fallback system was covered by it at all. ## Six hours before duplicates were disclosed KiteBridge's public account of the incident lagged behind events. Its incident page initially described the problem as delayed reimbursements, which suggested money was arriving late, not that it was being sent twice. The page first disclosed the duplicate payments six hours after automation was suspended, which puts the disclosure at around 4 p.m. Tuesday. In the meantime, customers relying on the page had no public indication that excess money had left their accounts. ## What remains open Several questions are unanswered, including: - when the outstanding $500,000 will be recovered - what role, if any, the version 4.2 release played in the failure - why the incident page's first description did not match the problem Ortiz did not give a date for the technical report.

Meta: Muse Spark Open: # Software Glitch Triggers $2.8 Million in Duplicate Employee Reimbursements **Ninety-one companies were charged twice after an update to KiteBridge's expense platform overrode a key safety check. Most of the money has been clawed back.** Ninety-one employers are working to recover millions of dollars in duplicate payments sent to their workers this week after a routine software update at expense-management provider KiteBridge backfired, issuing thousands of reimbursements a second time and pulling the funds directly from corporate accounts. The error, which generated 13,400 extra payments totaling $2.8 million, has renewed scrutiny of the company's promises around payment reliability. While KiteBridge says most funds have already been recovered and no employees lost money, the incident exposed a gap between the safeguards the company advertised and how its system behaved under stress. KiteBridge, which makes software that automates employee expense reimbursements for corporate clients, deployed version 4.2 of its platform on Tuesday morning. The update coincided with a wave of expense processing for the start of the work week. Shortly after the deployment went live, the company's payment processor began timing out — acknowledging payment requests but responding too slowly for KiteBridge's system to confirm they had succeeded. In response, KiteBridge's automated fallback system retried the payments. Instead of recognizing the retries as duplicates, the fallback treated each one as a brand-new reimbursement request and issued it again. For a general user, the distinction matters: modern payment systems are designed to ensure that even if a network hiccup forces a retry, a single approval from an employer can result in only one transfer to an employee. That concept is often called "exactly-once" processing. It was a protection KiteBridge had specifically highlighted. An April changelog for version 4.0 of its software said the update had introduced "exactly-once payment protection," assuring customers that retries would not create extra payouts. On Tuesday, that safeguard did not work as intended. Within about half an hour of the deployment, KiteBridge received its first complaint from a customer noticing double reimbursements landing in employee accounts. About an hour after that, the company suspended all automated payments, halting further duplicates. The scale, however, was already significant. Employers — not employees — funded the duplicate transfers, KiteBridge said, meaning workers received extra money but did not have funds pulled from their personal accounts. Communication around the failure lagged behind the technical response. KiteBridge's public incident page initially described the issue only as delayed reimbursements. It was not until roughly six hours after automated payments were paused that the page was updated to disclose that duplicate payments had been sent. By Wednesday evening, KiteBridge said it had recovered $2.3 million by reversing transfers in coordination with banks and employers. About $500,000 remained outstanding and pending recovery. Chief Executive Lena Ortiz apologized for the incident and said the company would publish a full technical post-mortem. She did not give a date for the report, which is expected to detail why the fallback system bypassed the exactly-once protection and what changes are now in place. For affected companies, the immediate disruption has largely been administrative — reconciling accounts, confirming reversals, and reassuring employees — rather than financial loss. But for KiteBridge, a company whose core pitch is that it makes reimbursements invisible and reliable, the failure has struck at its central promise of trust.

Anthropic: Claude Fable 5.2: # Foundry Hall to Close Oct. 31 After Losing Its Largest Sponsor *Six employees will lose their jobs, and 12 startups must find somewhere else to work.* MILWAUKEE — Foundry Hall, the nonprofit accelerator that has supported 212 startups since it opened in 2018, will close on October 31 after losing the sponsor that covered more than a third of its budget. Blank River Bank, the organization's largest corporate backer, declined to renew a sponsorship that supplied 38% of Foundry Hall's annual funding. The accelerator has enough money to keep operating through October, but not enough to carry another full program. So it will not start one. "We kept waiting for the replacement check that never came," said Camille Ross, the executive director. The closing will cost six employees their jobs. Twelve startups that currently use Foundry Hall's workspace will have to leave. ## A decision about priorities The bank has not faulted the accelerator. A spokesperson said the decision reflects a change in Blank River's community investment priorities and is "unrelated to Foundry Hall's performance." That performance is easy enough to count. The startups Foundry Hall has supported now employ 1,140 people between them, a figure that grew one small company at a time. What is harder to count is what the place meant to people at the very beginning, when a company was mostly an idea and a person willing to explain it to anyone who would listen. Dev Malik was one of those people. "My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." The distance between that back table and a room where an investor would sit down is what the accelerator offered. It is also what the 12 companies still inside must now find elsewhere, on a deadline they did not choose. ## What happens next The city has begun looking for places to put them. Deputy Mayor Susan Cole said officials are speaking with building owners and local universities about space for the current companies. Those talks concern where a dozen existing businesses will work once the month ends. Foundry Hall itself has only a date. Through October, the doors stay open, the staff keeps working and the companies keep their desks. After the 31st, the program that supported 212 startups in Milwaukee will not take on another.

Meta: Muse Spark Open: # The Last Program at Foundry Hall Foundry Hall, the nonprofit startup accelerator in Milwaukee that opened in 2018, will close on Oct. 31 after losing its largest funder. Executive director Camille Ross announced the closure this week. The accelerator has enough money to operate through October, but not to sustain another full program cycle. Its largest corporate sponsor, Blank River Bank, which supplied 38% of its annual budget, declined to renew its sponsorship. Ross put it plainly: “We kept waiting for the replacement check that never came.” The loss is significant for a small organization with an outsized footprint. Since opening seven years ago, Foundry Hall has supported 212 startups, which collectively employ 1,140 people. For many founders, it provided the first professional space beyond coffee shops and kitchen tables. Founder Dev Malik said his early break came there. “My first investor meeting happened in that building,” he said. Before Foundry Hall, he had been pitching from the back table of a coffee shop. That kind of transition was the accelerator’s core work — offering workspace, workshops, and introductions that early-stage companies rarely access on their own. Now the 12 startups currently using its workspace must leave by the end of October, when operations cease. Six employees will lose their jobs, as well, underscoring that the nonprofit itself operated under the same precarious economics it tried to ease for others. The bank said the decision was not a reflection on Foundry Hall’s performance, attributing it instead to a change in its community investment priorities. At City Hall, officials acknowledged the disruption. Deputy Mayor Susan Cole said the city is talking with building owners and local universities about interim space for the companies now facing displacement, though no plan has been finalized. The closure will not come with a final showcase, just a quiet winding down of office hours and mentoring sessions through the fall. Desks will be cleared, key cards deactivated, and a familiar address for early meetings will disappear from founders’ calendars. For Milwaukee’s startup community, Foundry Hall was rarely about scale alone. It offered a kind of permission — to treat a fragile idea as worth a lease, a logo, a first hire. Its absence will be measured not just in square footage vacated, but in the conversations that will now need to find another room. Foundry Hall said it will spend its remaining weeks helping its current cohort transition before shutting its doors at the end of October.