Head to head: Z.ai: GLM 5.3 vs grok-4.7

Z.ai: GLM 5.3: # Cloudnote Says Customer Prompt Excerpts Were Exposed for Four Hours *A debugging page showing excerpts from 1,842 user sessions was accessible without login; Cloudnote says passwords and payment information were not involved.* Cloudnote, an AI note-taking app with 3.2 million registered users, said a debugging page containing excerpts of customer prompts was accessible online without authentication for four hours on Monday. The company removed the page 19 minutes after a security researcher notified it. Server logs show the page displayed excerpts from 1,842 user sessions, each of which could include up to 200 characters from a prompt. Passwords, payment details and full documents were stored in separate systems and were not present on the page, the company said. Even those fragments could be sensitive: people routinely type health questions, drafts and personal details into AI tools, and a debugging page that requires no login turns private input into something reachable by anyone who finds the address. Cloudnote says it has found no evidence that anyone besides the researcher accessed the page. But the company has not completed its investigation, and it cannot yet say for certain who — if anyone else — viewed the excerpts. For now, the known facts point to a contained incident: a single tip, a 19-minute takedown, no evidence of other visitors. Whether that picture holds — and what, if anything, Cloudnote tells the 1,842 users whose prompts appeared — depends on an investigation that is still underway.

grok-4.7: # Cloudnote Debug Page Exposed Prompt Excerpts From 1,842 Sessions *The unauthenticated page was open for four hours on Monday. Cloudnote says only a researcher is known to have viewed it, and the company has not finished its review.* Cloudnote said a debugging page containing customer prompt excerpts was accessible without authentication for four hours on Monday. Logs show excerpts from 1,842 user sessions. Each excerpt could include up to 200 characters from a prompt. The exposure is narrow in one sense and unsettled in another. Cloudnote has 3.2 million registered users, and the company has not said how many distinct accounts the 1,842 sessions represent. It also has not said whether any excerpt included personal or confidential text. What is known is that a prompt fragment of that length can still show part of what a user asked the product to do. Cloudnote said it removed the page 19 minutes after a researcher notified it. The company says it has found no evidence that anyone besides the researcher accessed the page. That statement is not a final conclusion. Cloudnote said it has not completed its investigation, leaving unconfirmed whether anyone else viewed the page. The company was more definite about what the page did not contain. Full documents, passwords and payment information were stored elsewhere and were not present, Cloudnote said. Left open are the contents of the excerpts themselves and the question of who, if anyone beyond the researcher, saw them.

Z.ai: GLM 5.3: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

grok-4.7: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

Z.ai: GLM 5.3: # QueueZero's AI Meeting Assistant Takes Its Boss Literally at Launch Event QueueZero introduced Proxy, an artificial-intelligence meeting assistant, on Wednesday, and the product immediately demonstrated both its promise and its pitfalls by assigning its own chief executive a pair of tasks mid-presentation. Proxy is designed to attend meetings, take notes, and delegate follow-up work to software agents. Its debut took an unexpected turn when a customer in the audience remarked, "Someone should tell Evan to stop talking and send us the pricing sheet." The assistant interpreted the comment as an instruction. Moments later, two tasks appeared on CEO Evan Wu's task list: "Stop talking" and "Send pricing sheet." Wu laughed, deleted the first task, and completed the second — providing attendees with the pricing information they had requested. He then continued his presentation. ## How Proxy Handles Instructions The incident highlights how Proxy's delegation system works — and the questions it raises for workplace automation. The assistant, like many AI tools, does not always distinguish between rhetorical commentary from meeting attendees and direct commands. QueueZero says the risk of uncontrolled messaging is limited. Proxy requires confirmation before it can send any message to an external party, such as a customer or partner. Internal tasks, however, can be created without human sign-off. Administrators concerned about unwanted or frivolous task creation can disable automatic task creation entirely, requiring a human to approve each item before it is assigned. ## Pricing and Availability Proxy will cost $30 per user per month. The product enters public beta next week, giving customers a chance to test the assistant in real meetings before general release. ## The Road Ahead The demonstration was, in effect, a real-world test of the product's core functionality. Proxy successfully parsed conversational speech, identified actionable items, and assigned tasks — all within moments of hearing the customer's remark. That speed and initiative are the product's selling points. Still, the launch also served as a preview of the kind of confusion that can arise when AI assistants act on ambiguity. A comment directed at a presenter, meant as a joke about pricing, became a work assignment that was taken literally by software with no ear for context. QueueZero declined to say whether the demonstration would prompt changes to how Proxy handles audience remarks. For now, the company appears content with the free publicity — and with the fact that its CEO's task list contained only one item he did not intend to complete. Proxy's public beta begins next week. The pricing sheet, as requested, is available.

grok-4.7: # QueueZero's Meeting Assistant Assigns Its CEO Tasks From an Audience Remark QueueZero launched Proxy, an artificial intelligence meeting assistant, on Wednesday. The product can attend a meeting, produce notes and assign follow-up work to software agents. The debut also supplied an unplanned demonstration of how literally the system can read a room. During the launch demonstration, a customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy interpreted the remark as an instruction. It created two tasks and assigned both to QueueZero chief executive Evan Wu. One task read "Stop talking." The other read "Send pricing sheet." Wu laughed, completed the second task and deleted the first. The sequence showed a function QueueZero is selling as part of the product: speech in a meeting becoming assigned work, without a separate person restating the request as a command. That function sits beside limits the company has stated. Proxy requires confirmation before an external message is sent. Internal tasks can be created automatically. Administrators can disable automatic task creation. An organization can therefore keep the assistant from opening internal work on its own, but that restraint is a setting, not the behavior shown on stage. Confirmation before an external message restricts what the assistant can send outside the company without a person agreeing to the send. Automatic internal tasks follow another rule. If Proxy hears language it treats as an instruction, it can file the work inside the organization immediately. Wednesday's demonstration took that route. The tasks existed before Wu decided which of them deserved to remain. He kept the pricing request. He removed the instruction that he stop talking. The company has not said that completing the pricing task meant an external message was sent. Its stated rule is that Proxy requires confirmation before an external message goes out. What the demonstration established is narrower, and it is enough to matter: an internal task can be created from a remark in the room, completed by the person it names, and deleted by that person if it should not have been created. Proxy costs $30 per user per month. It enters public beta next week. The beta will extend the product beyond a launch demonstration and into meetings where the person named in a task may not be present to complete it or delete it. Customers will be able to judge the notes, the handoff to software agents, and the point at which a sentence becomes an assignment. The launch left that separation to Wu, who was present and deleted the task he did not accept.

Z.ai: GLM 5.3: # KiteBridge Software Update Triggered $2.8 Million in Duplicate Reimbursements at 91 Companies A botched software update at expense-management firm KiteBridge generated 13,400 duplicate reimbursements totaling $2.8 million, paying employees of 91 corporate customers for expenses that had already been covered before the company could halt its automated payment system Tuesday morning. The failure was triggered by a new release of KiteBridge's software and halted within roughly 80 minutes. But it exposed both a flaw in the company's payment safeguards and a lag in how candidly it described the problem to customers. KiteBridge makes software that corporations use to approve employee expense claims and pay them back — a business that depends on moving employer money quickly and, above all, exactly once. Version 4.2 of that software went live at 8:45 a.m. Tuesday. The first customer complaint arrived 27 minutes later, and by 10:03 a.m. KiteBridge had suspended automated payments. By then, thousands of reimbursements had already been duplicated. The technical trigger was mundane: a payment-processor timeout. When an expense claim is approved, KiteBridge hands a payment instruction to an outside payment processor, which executes the transfer and confirms that it succeeded. On Tuesday, that processor began timing out — taking so long to respond that KiteBridge's systems gave up waiting and treated the transactions as unresolved. A timeout leaves software in an awkward position: the payment may have gone through, or it may not have, and the waiting system cannot tell which. That is what fallback systems are for. KiteBridge's fallback, built to keep payments moving when the primary channel fails, responded by resubmitting the stalled payments. But it made a consequential error: it treated each resubmission not as a retry of an existing payment but as a brand-new reimbursement request. Unable to tell "try this again" from "pay this again," the system multiplied every stalled transaction — 13,400 duplicate payments worth $2.8 million. The cost fell on employers rather than employees. No worker had money withdrawn from an account. Instead, the affected companies funded the duplicates from their own expense budgets — more than $30,000 per company on average — and now face the work of recouping the money. The appeal of expense software is that it automates a high-volume, error-prone process; when the automation errs, it errs at machine speed. The failure also undercut a safeguard KiteBridge had advertised. An April changelog entry said version 4.0 had introduced "exactly-once payment protection" — engineering shorthand for a guarantee that a payment is processed one time, no more. Tuesday's cascade was precisely the kind of duplication that feature was meant to prevent. The company's disclosure has drawn scrutiny as well. KiteBridge's public incident page, the channel customers rely on during outages, initially described the problem as delayed reimbursements — a description that understated matters considerably, since money had not merely arrived late but multiplied. The page did not disclose the duplicate payments until six hours after automation was suspended, around 4 p.m. Tuesday. Chief Executive Lena Ortiz apologized for the incident and said KiteBridge would publish a technical report explaining what went wrong. By Wednesday evening, $2.3 million of the $2.8 million had been recovered, leaving roughly $500,000 still pending — along with the promised explanation.

grok-4.7: # KiteBridge Sent $2.8 Million in Duplicate Reimbursements After a Retry Was Treated as a New Payment KiteBridge’s expense-management software paid the same reimbursements twice on Tuesday, generating 13,400 duplicate transactions worth $2.8 million and leaving 91 corporate customers to sort out money their employers had already funded once. By Wednesday evening the company had recovered $2.3 million. The remaining $500,000 was still pending, and no employee had money withdrawn from a personal account. The failure began with a routine release. KiteBridge, which builds software that tracks employee expenses and sends reimbursement payments, deployed version 4.2 at 8:45 a.m. Tuesday. Within half an hour, at 9:12 a.m., the first customer complaint arrived. What looked at first like a delay was a payment logic error: when a payment processor timed out, KiteBridge’s fallback system did not recognize a retry as a second attempt at the same transfer. It treated the retry as a new reimbursement request and paid it again. In ordinary online payments, a timeout is ambiguous. The sender may not know whether the processor received the instruction, completed it, or dropped it. Systems are supposed to answer that uncertainty with an idempotency check—a record that says this particular payment has already been attempted, so a repeat must not create a second one. KiteBridge had publicly claimed that protection. An April changelog said version 4.0 introduced “exactly-once payment protection,” the industry phrase for ensuring a payment is executed once even if the request is sent more than once. Tuesday’s release showed that the safeguard did not hold when the processor timed out and the fallback path ran. Automated payments continued until 10:03 a.m., when KiteBridge suspended them. In that window the software issued the 13,400 duplicates. The money did not come out of employees’ accounts. Employers fund reimbursements through the platform, so the duplicate charges landed on the companies that use KiteBridge, not on the workers those payments were meant to repay. That limited the immediate harm to individuals, but it did not shrink the accounting problem: 91 customers had to reconcile a sudden second wave of outbound payments against expenses they had already settled. Public disclosure lagged the internal response. KiteBridge’s incident page initially described the problem as delayed reimbursements, a framing that matched the early complaints but not the duplicate charges already on the books. The page first disclosed the duplicate payments six hours after automation was suspended, well after finance teams at affected companies would have begun seeing unexpected outflows. The company has not said whether that delay reflected incomplete diagnosis or a choice to wait for a fuller picture. Recovery moved faster than the initial notice. By Wednesday evening, $2.3 million of the $2.8 million had been recovered, with the rest still pending. Chief executive Lena Ortiz apologized and said KiteBridge would publish a technical report. That report will have to explain how a protection advertised since version 4.0 failed in 4.2, why a timeout on one processor was enough to mint thousands of new payment records, and why customers were first told the issue was delay rather than duplication. For expense platforms, the stakes of that explanation are practical. Companies route payroll-adjacent money through these systems and treat “exactly once” as a promise, not a slogan. A retry that becomes a second payment is not a cosmetic glitch. It is a direct charge against a customer’s funds, and on Tuesday it was a $2.8 million one.

Z.ai: GLM 5.3: # Foundry Hall to Close After Six Years MILWAUKEE — Foundry Hall, the nonprofit startup accelerator that has anchored Milwaukee's entrepreneurial scene since 2018, will close its doors on October 31. The shutdown follows Blank River Bank's decision not to renew its sponsorship. The bank, Foundry Hall's largest corporate backer, supplied 38 percent of the accelerator's annual budget. A spokesperson said the decision reflects a shift in the bank's community investment priorities and is unrelated to Foundry Hall's performance. For Executive Director Camille Ross, the answer arrived not as a verdict but as a silence. "We kept waiting for the replacement check that never came," she said. The organization has enough money to operate through October, she confirmed, but not enough to fund another full program year. The numbers behind the closing tell a story of what will be lost. Foundry Hall has supported 212 startups since opening, companies that now collectively employ 1,140 people. Six employees will lose their jobs when the accelerator shuts down, and the twelve startups currently working out of its shared space must find new homes by month's end. For founders like Dev Malik, the building was more than an address. His first investor meeting took place there, a moment he says would have been impossible before Foundry Hall opened. "Before Foundry Hall, I was pitching from the back table of a coffee shop," Malik said. That sentiment — of a place turning improvised ambition into something presentable, and fundable — recurred among the founders who passed through the accelerator's doors. Milwaukee has no shortage of coffee shops. What it had in Foundry Hall, founders say, was a room where their ideas were taken seriously. The timing leaves the current cohort in a difficult position. Twelve startups must relocate with little more than six weeks' notice, in a market where affordable flexible workspace is scarce. Deputy Mayor Susan Cole said the city is aware of the gap and is moving to fill it, speaking with building owners and local universities about space for the displaced companies. She offered no timeline for those arrangements. The bank's withdrawal raises familiar questions about the fragility of nonprofit models built heavily on a single corporate patron. Nearly four in ten of Foundry Hall's dollars came from one source, and when that source changed direction, there was no second checkbook large enough to absorb the loss. Ross declined to speculate about whether diversification could have saved the organization, saying only that the board explored "every path" before voting to close. Foundry Hall will end where it began: with Milwaukee entrepreneurs in need of a room of their own. Whether the city can provide one before the end of October — and whether the founders who built 1,140 jobs inside those walls can keep building elsewhere — remains unanswered. The doors close on the thirty-first.

grok-4.7: # The Last Month at Foundry Hall The glass doors of Foundry Hall still open onto the lobby they have opened onto since 2018. People still come through. The coffee maker still runs. In the main room, whiteboards hold the residue of pitches that will not be finished there. The nonprofit startup accelerator in Milwaukee will close on October 31, and the building has already begun to feel like a place that knows its date. The record is not small. Foundry Hall has supported 212 startups. Those companies collectively employ 1,140 people, a payroll built one hire at a time in a city that has learned to count such things. None of that replaced the money that stopped arriving. Blank River Bank, the hall’s largest corporate sponsor, had supplied 38 percent of its annual budget. The bank declined to renew. A spokesperson said the decision reflected a change in the bank’s community investment priorities and was unrelated to Foundry Hall’s performance. The statement was precise about what the choice was not. It was silent about what would fill the hole. Executive director Camille Ross did not reach for a softer phrase. “We kept waiting for the replacement check that never came,” she said. What remains will carry the organization through October. It will not carry another full program. Six employees will lose their jobs. Twelve startups still using the workspace will have to leave. Founder Dev Malik measured the loss in rooms. His first investor meeting happened in that building. Before Foundry Hall, he was pitching from the back table of a coffee shop. He spoke of the hall less as a brand than as the first place that treated his company as if it might last. Deputy mayor Susan Cole said the city is speaking with building owners and local universities about space for the current companies. The talks are underway. They are not yet a door with a key. No one has promised that the twelve will land together, or that the six people who kept the program running will be hired into whatever follows. What ends on October 31 is not the 212 companies. Most of them have already gone on, employing people in rooms Foundry Hall will never see. What ends is the bargain the hall made with founders who did not yet have an office: a desk, a deadline, and someone who would read the pitch twice. Milwaukee has other programs. After this month, it will not have this one. The lights will stay on through the last day. Then the address will be only an address, and the work that started there will have to continue without the room that taught it how to begin.

Z.ai: GLM 5.3: C

grok-4.7: C

Z.ai: GLM 5.3: # Raincheck Raises $8 Million to Help Cities Prepare for Floods *The Austin-based company maps drainage systems and pairs them with rainfall forecasts so cities can send crews where they're needed before a storm hits.* Raincheck, an Austin company that makes flood-risk software for local governments, has raised $8 million in seed funding. The company maps drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm. The round was led by Basin Ventures, with participation from existing investor Long Mile. Raincheck has raised $10.5 million to date. 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." The company describes its approach in detail in [its published methodology](https://example.com/raincheck-methodology). Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations. Emergency managers retain control over deployments and public warnings. Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida.

grok-4.7: # Raincheck raises $8 million for flood-risk software *The Austin company maps drainage systems and pairs them with rainfall forecasts so cities can send crews out before a storm.* Raincheck, which makes flood-risk software for local governments, said today it has raised $8 million. The Austin 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. Further detail is available in its [methodology overview](https://example.com/raincheck-methodology). The seed round was led by Basin Ventures. Existing investor Long Mile also participated. Raincheck has 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. Its forecasts are recommendations. Emergency managers retain control over deployments and public warnings. Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida.

Z.ai: GLM 5.3: # The Best Model Isn't the Best Value: What 200 Support Tickets Revealed When a team of researchers handed the same 200 anonymized customer-support tickets to five language models and asked three experienced support managers to grade the results blindly, the top of the leaderboard looked familiar: the most expensive model won. But the more interesting story sat just below it — where a model costing less than half as much scored statistically the same. ## The Setup The benchmark was straightforward. Each model received one ticket and produced one proposed reply. The tickets were real, anonymized support requests written in English, drawn from a single software company's queue. Three support managers — unaware of which model had produced which answer — graded every response on a combined score reflecting factual correctness, usefulness, and tone. The five models spanned a wide range of price and capability: | Model | Average 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 | ## The Finding That Matters Model North posted the highest score, at 84.2. Model Vale came in second at 83.8 — a gap of less than half a point that the researchers confirmed was not statistically significant. In plain terms: the graders could not reliably tell the two models apart on overall quality. But the price gap between them was anything but subtle. North cost an estimated $1.90 per 1,000 tickets. Vale cost $0.74 — about 39% of North's price. Vale also answered 18% faster. For any team triaging support volume at scale, that's the headline. If two models are indistinguishable on quality but one costs 2.6 times more, the cheaper one isn't a compromise — it's the rational default. The premium model has to justify its premium somewhere other than the average score. There is a nuance worth keeping: it did, in one place. North performed best on billing disputes, while Vale led on technical troubleshooting. If a support operation is dominated by one category of ticket, the category-specific results may matter more than the aggregate. A payments-heavy company might still find North worth the premium; a troubleshooting-heavy one probably won't. ## The Cost Curve Flattens Fast Below the top two, the picture is less flattering. Ember scored 79.1 at $0.31 per 1,000 tickets — a meaningful quality drop from Vale for meaningful savings. Kite and Harbor fell further, and Harbor's 71.0 average suggests it would need human review on enough tickets to erode its cost advantage. The pattern here is common in model benchmarks: quality rises steeply from the cheapest tier to the middle, then plateaus near the top. The last few points of quality are disproportionately expensive — and in this test, the last 0.4 points vanished entirely into statistical noise. ## Where Every Model Failed The most important weakness wasn't a differentiator between models — it was shared across all five. Every model performed poorly when a ticket's answer depended on an image attachment, such as a screenshot of an error message. Any organization whose tickets routinely include images should treat these benchmark numbers as upper bounds, not expectations. ## Read the Limits Before You Copy the Conclusion This benchmark is useful precisely because it's narrow, and it should be read that way. All 200 tickets were in English and came from one software company — meaning one product line, one customer profile, and one internal knowledge base. Scores from a different industry, a chattier customer base, or a multilingual queue could look quite different. The evaluation also covered only the generation of a proposed answer. It did not test agentic actions (refunding a charge, updating an account), tool use (querying a CRM, escalating to a human), or multilingual support. A model that drafts excellent replies is not the same as a model that operates well inside a real support workflow, where integration, escalation paths, and correctness under tool-use pressure matter as much as prose quality. And the scores themselves rest on three graders. Blind grading by experienced managers is a solid methodology, but human grading has its own variance, and "usefulness" in a draft answer only loosely predicts usefulness once that answer is edited, approved, and sent by a real agent. ## The Takeaway For teams whose needs resemble this test — English-language tickets for a software product, drafts reviewed by humans — the benchmark points to Vale as the obvious starting point: near-top quality at roughly a third of the flagship price, with faster responses to boot. North earns its premium only for billing-heavy workloads, and the budget models trade away more quality than their savings justify. The broader lesson generalizes better than any single number: before paying for the top of the leaderboard, check whether you can tell it apart from second place. Often, you can't.

grok-4.7: # The Expensive Model Did Not Win Five language models were scored on the same stack of customer-support tickets, and the result that should shape a buying decision is not a new champion. The most expensive model and a much cheaper one were, for practical purposes, tied. What separated them was price and speed, not a quality gap the test could detect. Researchers gave each model 200 anonymized tickets from a single software company and asked for one proposed answer per ticket. Three experienced support managers graded the drafts blindly, folding factual correctness, usefulness, and tone into a single score. Model North averaged 84.2. Model Vale averaged 83.8. That difference was not statistically significant. Treating 84.2 as a win over 83.8 would invent a ranking the data do not support. The bill does support a ranking. North was estimated at $1.90 per 1,000 tickets. Vale was estimated at $0.74, and it answered 18 percent faster. A team generating drafts at volume would pay roughly two and a half times as much for North without a measurable gain in the qualities the managers were asked to judge. If the job is a first-pass reply a human still reviews, the rational default is Vale, not North. Further down the price list, the tradeoff becomes real. Ember scored 79.1 at $0.31 per 1,000 tickets. Kite scored 76.4 at $0.22. Harbor, at $0.08, scored 71.0. Those gaps are several times larger than the 0.4-point difference at the top. The study did not report significance tests for them, so they should not be treated as a formal ranking. They are, though, the kind of drop a support queue would feel: extra edits, missed steps, a reply a customer might not trust. The curve flattens at the top. Moving from Vale to North bought nothing the graders could reliably distinguish. Category results argue for fitting the model to the queue rather than crowning one system. North did its best work on billing disputes. Vale did its best work on technical troubleshooting. The study does not quantify how wide those leads were, so they should not be used to overturn the overall tie. They do suggest that a billing-heavy desk and a troubleshooting-heavy desk might pilot different defaults, and that an average score can hide a mismatch with the tickets that actually arrive. Every model shared a blind spot. When a ticket depended on an image attachment, all five performed poorly. Software support is full of screenshots, photos, and marked-up interfaces. A score earned on the words in the ticket does not transfer to a problem whose evidence sits in a file the model handled badly. Routing those tickets to any of these systems, on the strength of this benchmark, would be an extrapolation the results contradict. The method is clear enough to trust the tie, and narrow enough to limit what else it proves. Tickets were in English and came from one software company. The researchers did not test agentic actions, tool use, or multilingual support. One draft per ticket does not show whether a model recovers when asked to revise, and a manager’s grade is not the same as a customer marking the issue resolved. Cost figures are estimates, not production invoices. Prompt length, retries, and human rewrite rates will move the real bill. The speed comparison covers only Vale against North, so it cannot rank the other three on latency. None of that weakens the central comparison. On this set of tickets, judged this way, North’s higher price did not purchase a statistically detectable improvement over Vale. The cheaper models below Vale did give something up. And image-dependent tickets remain a problem none of the five solved. The useful response is a short pilot on the live queue, split by billing, troubleshooting, and anything that arrives with an attachment, with a person still responsible for the reply that goes out.

Z.ai: GLM 5.3: 6

grok-4.7: 6

Z.ai: GLM 5.3: # Startup Spotlight: LedgerLark Bets Small Firms Want AI Agents That Catch Trade Errors — Not Make Them Most post-trade automation gets built for the biggest banks. Mina Batra and Owen Carr think there's an opening one or two tiers down. LedgerLark, a startup building AI agents for post-trade operations at small investment firms, emerged from stealth this month with a $3.5 million seed round led by Motive Row. The company's pitch is narrow and specific: its agents compare trade confirmations, custodian records and internal ledgers, flag mismatches, and propose corrections — which humans then approve. The system doesn't execute trades or move money. That constraint is deliberate, and it traces back to how the founders came to the problem. ## An insight from the trenches Batra spent seven years in operations at a regional brokerage before co-founding LedgerLark. The germ of the idea, she says, came from watching a six-person team spend three days resolving a single trade mismatch — one caused by a copied account number. That kind of story will sound familiar to anyone who has worked in middle- or back-office operations. Trade breaks are rarely exotic; they're usually mundane data problems that nonetheless eat enormous amounts of skilled people's time. What made Batra's observation commercially interesting was who was stuck doing the work. Small investment firms — regional brokerages, boutique asset managers — face the same reconciliation burden as larger institutions, but without the reconciliation staff. Enterprise vendors historically haven't chased them aggressively, because the contracts are small and the sales cycles are long. That's left a lot of firms running manual processes or patchwork spreadsheets. Carr's background is the technical complement. He built reconciliation systems at a large bank, giving him firsthand knowledge of how painful it is to retrofit large-firm tooling for smaller shops. ## What the product actually does LedgerLark's agents sit across three data sources — confirmations, custodian records and internal ledgers — and identify discrepancies among them. When the agents find a mismatch, they don't fix it autonomously. They propose a correction, and a human decides. This is a notably conservative design for an AI product in 2025, when plenty of startups are marketing agents that take actions on their own. The choice reflects the regulatory and reputational stakes in post-trade operations: an incorrect "fix" applied automatically to a ledger is a far worse failure mode than a flagged exception that sits unresolved. Keeping humans in the approval loop is also likely a prerequisite for getting past compliance teams at financial institutions, which have their own opinions about autonomous agents touching books and records. ## Early traction, with caveats Five firms are paying customers, on annual contracts ranging from $36,000 to $110,000. That's a real business — enough to suggest the problem is one firms will pay to solve — but it's also early. Five customers is a thin base from which to conclude much about product-market fit, and the company hasn't said how long those customers have been on the platform or what churn looks like. The traction figure LedgerLark highlights is more specific: the company says one customer reduced unresolved daily exceptions from 70 to 18. That's a meaningful improvement if it holds, but it's a single customer, self-reported, over an unspecified period. It's the kind of number that deserves follow-up more than celebration. ## The competition question The bigger uncertainty is competitive. LedgerLark isn't operating in empty space. Established reconciliation vendors already serve the broader market, and large banks have spent years building internal reconciliation software. Some of that tooling eventually trickles down-market or gets productized. If a major vendor decides small investment firms are worth pursuing seriously, LedgerLark would face competitors with deeper balance sheets, longer reference lists and existing procurement relationships. There's also a sales-cycle problem endemic to the category: financial institutions often require lengthy security and procurement reviews before adopting new software. Even a clearly superior product can spend a year or more in review. For a seed-stage company with $3.5 million in the bank, that pace puts real pressure on runway and forces discipline about which deals to pursue first. ## What to watch LedgerLark's bet has two parts. The first is that AI agents can do reconciliation work well enough to justify their cost at small firms. The early evidence is suggestive but not conclusive. The second is that the small-firm segment — historically underserved — will grow into a viable market before larger vendors move down-market to claim it. The company has chosen its constraints sensibly: a narrow problem, a conservative product design, and pricing that implies real operational value rather than experimental budgets. Whether that's enough to build a durable company is exactly what the next 12 to 18 months will reveal. The number to watch isn't the exceptions reduced at any single customer — it's whether customer count five becomes fifteen, and how much friction each new sale takes to close.

grok-4.7: # LedgerLark Wants Software to Read the Records, Not Run the Trade Post-trade work rarely appears in the story of a market day. A trade is agreed, a price is set, and then someone has to make the paperwork agree. For a small investment firm, that task often falls to a few people comparing confirmations, custodian records and an internal ledger. LedgerLark is building AI agents for that narrow job. The premise is practical rather than grand: much of the delay sits in finding where the records diverge, and software might do more of the searching without being allowed to trade or move money. The company’s starting point is a specific operational failure. Co-founder Mina Batra spent seven years in operations at a regional brokerage. She began exploring the idea after seeing a six-person team spend three days resolving a trade mismatch caused by a copied account number. The cause was ordinary. The cost was attention. Three days is a long interruption for a team that size, and it suggests how little spare capacity some back offices have when a routine check fails. Batra’s co-founder, Owen Carr, built reconciliation systems at a large bank, where the staff, vendors and internal tools are usually deeper. Their combined experience explains why the problem is legible to them. It does not establish that a product shaped by those settings will fit every smaller firm. A regional brokerage, a large bank and a small investment adviser can share a headache without sharing the same data, permissions or tolerance for a new system. LedgerLark’s agents compare confirmations, custodian records and internal ledgers, then propose corrections for human approval. The product does not execute trades or move money. That limit is central to the offer. In a workflow where one copied account number can already consume days, a suggestion is a different kind of risk from an instruction that reaches a market or a payment rail. The remaining risk is still real. A proposed correction can be wrong, and an approval step helps only if the reviewer has enough context, and enough time, to reject a plausible mistake. Whether the agent shortens the search or simply relocates the judgment is a question the product will have to answer in use. The commercial record is early and concrete. Five firms are paying customers, on annual contracts ranging from $36,000 to $110,000. The spread is wide enough to imply different scopes or firm sizes, though the public account does not say what each price includes. LedgerLark says one customer reduced its unresolved daily exceptions from 70 to 18. That is a large reported change. It is also one customer’s result, as described by the company, not a result measured across the base. Five firms cannot yet show whether the same drop appears with a different custodian, messier files or a team that is already short of reviewers. The company has raised a $3.5 million seed round led by Motive Row. The round gives LedgerLark time. It does not answer the competitive question. Established reconciliation vendors already sell into this work, and institutions often rely on internal software built around their own books and custodians. An agent that proposes fixes has to be accurate enough, and adoptable enough, to justify standing beside tools a firm already trusts. For some buyers, that decision will depend less on a demonstration than on audit trails, security review and the awkward variety of custodian files. Those reviews can be slow. Financial institutions may require lengthy security and procurement processes before a new system touches operational data, even when it never touches cash. For a company with five customers, that pace is part of the business model rather than an external inconvenience. A long sales cycle can fill a pipeline while leaving revenue thin if implementations stall or renewals disappoint. LedgerLark is making a limited claim with demanding conditions attached. It is not asking to run the trade. It is asking firms to let software read three versions of the same event and suggest where they part. If the proposals are accurate and inspectable, small teams may recover hours now spent on mismatched fields. If they are not, the agent becomes another record that operations staff must check. The useful evidence ahead is modest: whether more firms keep paying, and whether exception counts fall for reasons a buyer can examine.

Z.ai: GLM 5.3: # HelioDock Launches Night Shift Fleet-Monitoring Software, Announces $24M Series B *The company says a 14-month beta with nine logistics firms and 310 robots cut unplanned robot stoppages by 17%. The software ships September 15 at $90 per robot per month.* HelioDock, a maker of autonomous warehouse robots, launched Night Shift on Thursday, a software product that lets operators monitor their robot fleets remotely rather than walking the warehouse floor or driving to a site to intervene when something goes wrong. Alongside the product launch, the company announced a $24 million Series B round led by North Coast Ventures, bringing its total funding to $39 million. The pitch behind Night Shift is simple: warehouses run around the clock, but the people who manage robot fleets historically haven't been able to. CEO Mara Velez framed the problem in personal terms. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," she said in a statement. The software is the result of a 14-month private beta involving nine logistics companies and a fleet of 310 robots. HelioDock says beta customers experienced 17% fewer unplanned robot stoppages while using the system. That figure comes from the company itself and has not been independently verified, and HelioDock has not detailed how the reduction was measured across the nine participating firms. The beta also surfaced a structural limitation that HelioDock's software alone can't solve: connectivity. Aaron Pike of Fulcrum Logistics, one of the beta customers, said Night Shift worked well in facilities with strong network connections but became less useful during rural LTE outages. When connectivity fails, the robots fall back to locally configured behavior and remote control becomes unavailable — meaning the same late-night trips the Night Shift is meant to prevent can still be on the table at sites with spotty coverage. The pricing reflects a per-robot subscription model rather than a site license: Night Shift costs $90 per robot per month, a structure that scales costs with fleet size but adds up quickly for operators running hundreds of machines. General availability begins September 15. HelioDock is not entering empty territory. Competing fleet-management products are already on the market from Relay Robotics and YardOS, meaning warehouse operators shopping for remote oversight tools have alternatives to evaluate on price, features and integration with their existing fleets. The Series B round gives the company capital to push beyond the beta phase. The new $24 million brings HelioDock's total raised to $39 million, with North Coast Ventures leading the latest round.

grok-4.7: # HelioDock Launches Remote Warehouse-Robot Software and Raises $24 Million *After 14 months with nine logistics companies and 310 robots, Night Shift becomes generally available September 15 at $90 per robot a month, beside a $24 million round led by North Coast Ventures.* HelioDock, which makes autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely, and announced a $24 million Series B led by North Coast Ventures. The launch follows the private test and arrives with new capital. Night Shift spent 14 months in a private beta involving nine logistics companies and 310 robots. HelioDock says beta customers experienced 17 percent fewer unplanned robot stoppages. The figure has not been independently verified. The software will cost $90 per robot per month and become generally available on September 15. The Series B brings HelioDock's total funding to $39 million. Chief executive Mara Velez described the trip Night Shift is meant to make unnecessary. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," she said. The line names one robot, a pallet, a drive across town, and two in the morning. HelioDock is offering remote fleet monitoring, with remote control available while a connection holds, as the alternative to that drive. She offered the scene as the problem Night Shift is meant to remove, not as a statistic from the 14-month test. The 17 percent figure is the performance claim HelioDock has chosen to stand behind. It covers unplanned robot stoppages among beta customers, and it comes from the company alone. The figure has not been independently verified. A buyer looking at the launch has HelioDock's account of the beta and no outside check on the percentage. Aaron Pike of Fulcrum Logistics, a beta customer, described a result that turned on connectivity. The system worked well, he said, in facilities with strong connectivity, and became less useful during rural LTE outages. His comment separates facilities by the strength of their connection, a different measure from the stoppage percentage HelioDock supplied. The product's own failure mode lines up with what Pike reported. When connectivity fails, robots return to their locally configured behavior, and remote control becomes unavailable. A robot that loses its connection does not wait for a remote operator to reappear. It returns to the behavior configured locally, and the person monitoring from elsewhere can no longer control it. Remote monitoring and remote control stop with the link. Local operation continues under settings already on the robot. An outage takes away the remote layer Night Shift adds and leaves the machine on that local behavior. Velez's example assumes a manager can deal with that robot from somewhere other than the warehouse. Where connectivity is strong, Pike said the system worked well, which is the setting in which remote control would still be available. Where rural LTE fails, remote control becomes unavailable, and the robot returns to locally configured behavior. The distance Night Shift is meant to close is the same distance an outage puts back between the operator and the robot. Night Shift will cost $90 per robot per month once it is generally available on September 15. The price is attached to the robot rather than to a building or a company. September 15 moves the software from a private beta of nine logistics companies and 310 robots to general sale. Competing fleet-management products are available from Relay Robotics and YardOS. Night Shift is HelioDock's entry in that category: remote monitoring of robot fleets, sold by the robot, with remote control that disappears when connectivity fails. North Coast Ventures led the $24 million Series B announced the same day. HelioDock's total funding is now $39 million, of which this round supplies $24 million and earlier financing supplies the other $15 million.

Z.ai: GLM 5.3: # Arcwell Launches Relay, an AI Assistant for Customer Support Agents Arcwell has unveiled Relay, a platform that uses artificial intelligence to analyze customer support conversations and suggest answers to agents in real time. Beyond drafting replies, Relay identifies customers at risk of canceling and generates summaries of support interactions. Arcwell is pitching the platform, priced at $65 per agent per month, as a way to strip inefficiency out of the support workflow. “Support agents spend too much time searching for information that their companies already have,” CEO Priya Shah said in the announcement. The company tested Relay with 12 companies over six months before launch. Agents using the platform closed tickets 14% faster, according to Arcwell. The company did not release results on the platform's effect on customer retention. Relay's speed gains come with data requirements. The platform needs access to support transcripts and internal help documents to function, raising the kinds of privacy questions that have dogged enterprise AI tools. Arcwell says customer data is encrypted and is not used to train shared models, and companies can choose a 30-day or 90-day retention period. Whether those assurances will be enough to ease adoption concerns remains to be seen, but Arcwell is betting that the productivity gains will win out. For a company building its future on AI-assisted customer service, Relay is the most concrete evidence yet of that direction.

grok-4.7: # Arcwell Announces Relay, an AI Platform for Support Agents Arcwell has announced Relay, a platform that uses artificial intelligence to examine customer support conversations and suggest answers to the agents handling them. In a six-month test with 12 companies, Arcwell said, agents using Relay closed tickets 14% faster. The company did not release customer-retention results, so the faster ticket closures are the only measured result it has made public. Relay does more than propose a reply. It can identify customers who may cancel, and it can create summaries of support conversations. Arcwell is offering the three capabilities together as it seeks to help companies improve customer service using artificial intelligence. “Support agents spend too much time searching for information that their companies already have,” chief executive Priya Shah said. That sentence is the problem Relay is built to address. Her point is that an agent working a ticket spends that time searching material the company already holds. Relay is designed to examine the conversation and return a suggested answer, the company’s response to the search Shah described. The cancellation feature and the summaries sit alongside that suggestion: one identifies a customer who may cancel, and the other creates a summary of the conversation. None of it runs without source material. Relay requires access to support transcripts and internal help documents — the conversations it examines and the documents an agent would otherwise search. Arcwell said customer data is encrypted and is not used to train shared models. Companies can choose a 30-day or 90-day retention period, a choice that sets how long that material is retained. On price, Arcwell was equally specific. Relay costs $65 per agent each month. A buyer can multiply that rate by headcount and know the invoice. What a buyer cannot do, on the evidence Arcwell has published, is match that invoice to a change in customer retention. Whether customers identified as likely to cancel actually stayed is a question the company has left unanswered. For companies asked to supply transcripts and help documents, that unanswered question is part of the decision, not a footnote to it.