Head to head: Z.ai: GLM 5.3 vs DeepSeek-V4-Pro

Z.ai: GLM 5.3: # KiteBridge expense software update triggered $2.8 million in duplicate reimbursements *A version 4.2 rollout failed Tuesday morning, forcing the company to suspend automated payments less than 90 minutes after launch.* Companies using KiteBridge's expense-management software were hit Tuesday with 13,400 duplicate employee reimbursements totaling $2.8 million, after a software update caused the platform to treat payment retries as brand-new claims. Ninety-one corporate customers were affected. In practice, employees were repaid twice for the same expense, but no one had money withdrawn from their bank accounts: the duplicate payments were funded by employers, meaning the cost landed on corporate finance departments rather than individual staff. The trouble began at 8:45 a.m., when KiteBridge deployed version 4.2 of its platform, which companies use to approve and repay employees for out-of-pocket expenses such as travel and meals. The first customer complaint arrived at 9:12 a.m., less than half an hour after the rollout. By 10:03 a.m., the company had suspended all automated payments, a move that stopped new reimbursements from going out but could not claw back the duplicates already issued. ## How it happened The trigger was a payment-processor timeout. When an expense claim is approved, KiteBridge hands the payment to an outside processor — the intermediary that actually moves money between banks. On Tuesday, that processor repeatedly failed to confirm payments within the expected window, leaving KiteBridge's software unable to tell whether the transactions had gone through. Timeouts are a routine hazard of online payments: because the money may or may not have moved, well-built systems are designed to verify what happened before trying again. KiteBridge's fallback system — the safety net meant to keep operations running when a component fails — instead resent the stalled payments. And each retry was recorded not as a repeat of the original request but as an entirely new reimbursement. Every resend, in effect, minted a fresh payment. That behavior clashed with a guarantee the company had previously advertised. The April changelog for version 4.0 said that update introduced "exactly-once payment protection" — an engineering term for a system's ability to ensure a payment is executed once, and only once, no matter how many times it is retried. Version 4.2's fallback logic broke that promise: instead of resuming stalled payments, it duplicated them. ## Recovery and disclosure By Wednesday evening, $2.3 million of the $2.8 million had been recovered. The remaining roughly $500,000 was still pending. KiteBridge's public accounting of the incident got off to a slower start. The company's incident page initially described the problem as delayed reimbursements. Only six hours after automated payments were suspended — late Tuesday afternoon — did the page disclose that duplicate payments had been made. For affected customers, the difference was more than semantic: a delay holds up employees' money, while duplicates tie up company cash and create reconciliation work for finance teams. Chief Executive Lena Ortiz apologized for the failure and said KiteBridge would publish a technical report on the incident. That report will have to explain how software built to pay each claim exactly once ended up paying thousands of them twice. *(~510 words)*

DeepSeek-V4-Pro: # Expense-Platform Glitch Sends $2.8 Million in Duplicate Payments Before Safeguards Fail KiteBridge, a provider of expense-management software used by mid-sized and large companies, accidentally issued 13,400 duplicate reimbursements totaling $2.8 million on Tuesday after a routine deployment bypassed the very protection designed to prevent double payments. The failure left 91 corporate customers scrambling to track errant funds and reignited questions about whether financial-software vendors are adequately transparent during active incidents. The cascade began at 8:45 a.m., when KiteBridge engineers pushed version 4.2 of its platform into production. The release contained standard updates that the company has not yet detailed publicly. Almost immediately, a downstream payment processor began returning timeout errors—a momentary communication failure that, on its own, should have been unremarkable. But inside KiteBridge’s infrastructure, the timeouts interacted with a fallback mechanism that treated each failed confirmation as a brand-new reimbursement request. Instead of asking the processor “did this payment go through?” and waiting for an answer, the system repeatedly created fresh instructions to pay. Because the timeouts were transient, the processor eventually accepted and settled most of those instructions. Employees enrolled in customer programs saw reimbursements for the same meal, travel charge, or office-supply receipt land in their bank accounts two, three, or even more times. KiteBridge noticed the anomaly when customer complaints began arriving at 9:12 a.m. At 10:03 a.m., the company suspended automated payments across the platform to stop the bleeding. By then the duplicate disbursements were already substantial. Employers—not employees—funded the duplicate payments through their corporate accounts, so no funds were pulled out of workers’ personal balances without authorization. Still, the episode put finance teams in the awkward position of explaining to staff that extra money would need to be returned. By Wednesday evening, KiteBridge said it had recovered $2.3 million of the $2.8 million accidentally distributed, with the remainder still pending. The company did not clarify whether recovery came from transaction reversals, clawback requests to receiving banks, or direct employer-to-employee arrangements. The incident is especially bruising because KiteBridge promoted version 4.0 in an April changelog as introducing “exactly-once payment protection”—a term borrowed from distributed-systems engineering that signals a guarantee against duplicates. That claim was undermined by the fallback logic that treated retries as original requests, a classic failure mode when idempotency keys or deduplication layers are not enforced end-to-end. Compounding the operational failure was a communications gap. When KiteBridge first updated its public incident page, the posting described the problem merely as a delay in reimbursement processing. It took six hours after the 10:03 a.m. payment suspension for the page to disclose that duplicate payments had been issued. For finance managers racing to reconcile accounts, that lag meant critical context arrived only after they had already identified anomalies on their own. CEO Lena Ortiz issued an apology Tuesday afternoon and pledged the company would publish a detailed technical report. The promise mirrors postmortem practices at major cloud and payments firms, but KiteBridge has not yet set a date for the report’s release. Customers will be watching for specifics: why the exactly-once guarantee did not hold, whether the fallback behavior was newly introduced in version 4.2, and what architectural changes the company plans to make so that a processor timeout cannot again be mistaken for a fresh demand to pay. For the broader financial-technology sector, the event highlights a tension between rapid deployment cadences and the extreme reliability expected of software that moves money. Even a brief mismatch between a vendor’s retry logic and its payment processor’s response can generate millions of dollars in erroneous transactions. KiteBridge’s engineers now face the task of proving that the very phrase “exactly once” can survive a real-world hiccup.

Z.ai: GLM 5.3: # QueueZero's AI Meeting Assistant Takes Orders From the Crowd at Launch QueueZero unveiled Proxy, an artificial intelligence meeting assistant, on Wednesday — and the software demonstrated an unexpected talent for taking direction from audience members within minutes of its debut. Proxy is designed to attend meetings, produce notes, and assign follow-up work to software agents. But during the launch demonstration, the assistant showed it could also act on remarks that were never intended as instructions. According to multiple attendees, a customer in the audience said during the demonstration, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy interpreted the comment as an instruction and created two tasks assigned to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, deleted the first task, and promptly completed the second, sharing pricing details with the audience. ## Guardrails in Place The incident raised questions about how Proxy handles prompts from meeting participants other than the primary user. QueueZero says Proxy requires confirmation before an external message can be sent, a safeguard intended to prevent the assistant from sending emails or other communications without oversight. Internal tasks, however, can be created automatically — which is how the two items assigned to Wu appeared without any approval step. Administrators can disable automatic task creation entirely if they prefer to review every task Proxy generates before it enters a workflow. The company has not said whether Wednesday's demonstration prompted any changes to Proxy's default settings. ## Pricing and Availability Proxy costs $30 per user per month. The assistant enters public beta next week, giving customers their first chance to test the tool in their own meetings — and, presumably, to discover what happens when their colleagues speak a little too freely. The launch comes amid a crowded field of AI meeting assistants, with vendors racing to add agent-like capabilities that go beyond transcription and summarization. Proxy's ability to assign work directly to software agents places it among the more ambitious offerings in the category. ## What It Means for Enterprises For organizations considering tools like Proxy, Wednesday's demonstration highlights a broader question: how AI systems distinguish between conversational remarks and actionable instructions. As assistants grow more capable of taking initiative, the line between observing a meeting and participating in one becomes increasingly difficult to police. QueueZero's confirmation requirement for external messages offers one model for maintaining human oversight. The automatic creation of internal tasks, as Wednesday's incident showed, remains an area where the technology may occasionally outpace the intentions of the people in the room. Proxy is available through QueueZero's website ahead of the public beta launch.

DeepSeek-V4-Pro: **QueueZero’s AI Meeting Assistant Accidentally Sanctions Its Own CEO During Live Demo** QueueZero launched its new AI meeting assistant, Proxy, on Wednesday with a demonstration that proved the software is attentive, efficient, and entirely indiscriminate when it comes to following orders. Proxy is designed to attend meetings on behalf of human participants, generate structured summaries, and autonomously assign follow-up tasks to other software agents. During a live rollout event in San Francisco, the software demonstrated its capabilities by ingesting the conversational audio from the audience in real time. The feature intended to showcase contextual awareness produced an unexpected result when a prospective customer in the third row muttered a remark that was not meant for the machine. “Someone should tell Evan to stop talking and send us the pricing sheet,” the attendee said to a colleague, referring to QueueZero Chief Executive Officer Evan Wu, who was leading the presentation. Instead of treating the comment as ambient noise, Proxy interpreted it as a direct instruction. Within seconds, the system created two discrete tasks inside QueueZero’s workflow engine. Both were assigned to Wu. The first task was “Stop talking,” and the second was “Send pricing sheet.” The internal assignment appeared on the overhead projection screen, where the audience could read it clearly. Wu paused, examined the task list, and deleted the first item. “I’ll take that one offline,” he said, then completed the second task by forwarding the company’s pricing document to the event’s distribution list. The launcher had, without grasping the irony at the time, just made the case for administrative guardrails. QueueZero’s engineering team addressed the incident in a technical addendum distributed immediately after the keynote. A spokesperson confirmed that Proxy requires explicit human confirmation before any external communication is sent, meaning that the system cannot independently email a client or post a message to a public channel. Internal task creation, however, works differently. By default, Proxy can generate and assign action items within a company’s private project management environment without a manual approval step. That design choice is what allowed the task disciplining the CEO to appear so rapidly. Organizations that prefer a stricter workflow can disable automatic task creation entirely through an administrative console, a setting QueueZero now recommends for customer-facing demonstrations. Proxy will enter public beta next week at a cost of $30 per user per month. Early enterprise adopters have praised its contextual parsing, but Wednesday’s event made one operational boundary plain: the software will diligently hold everyone accountable, including the person who signs its procurement order.

Z.ai: GLM 5.3: # Arcwell Launches Relay, an AI Assistant for Support Agents Arcwell has unveiled Relay, an AI-powered platform that scans customer support conversations in real time and suggests answers to agents as they work. The platform does more than draft replies. Relay flags customers who appear likely to cancel and generates conversation summaries, giving managers a clearer picture of support interactions without reading through every exchange. The launch follows a six-month trial with 12 companies, during which Arcwell says agents using the platform closed tickets 14% faster. The company did not release results from the product's churn-detection features. Relay is priced at $65 per agent per month. “Support agents spend too much time searching for information that their companies already have,” said CEO Priya Shah. That gap between what companies know and what agents can find is where Arcwell is positioning Relay. The platform draws on a company's existing materials — support transcripts and internal help documents — to surface answers the moment an agent needs them. That access requirement may give some customers pause. Relay needs visibility into support transcripts and internal documentation to function, raising the data-handling questions that follow most enterprise AI tools. Arcwell says customer data is encrypted and is not used to train shared models, and customers can choose between 30-day and 90-day retention periods. For Arcwell, Relay marks a push into a crowded market. A wave of startups and established vendors are racing to embed AI into customer service workflows, betting that tools which shave seconds off each interaction will add up across millions of tickets. The 14% speed improvement, if it holds up beyond the trial, would be a meaningful result for support operations, where response times and resolution rates directly shape customer satisfaction. Whether Relay's churn warnings prove as actionable remains an open question — one the company's trial data doesn't yet answer. Arcwell will now face the standard test for a new platform: convincing buyers that the productivity gains outweigh the integration effort and data exposure the tool requires.

DeepSeek-V4-Pro: **Arcwell Debuts Relay, an AI Assistant That Reads Customer Conversations** Arcwell has launched Relay, a platform that scans customer support conversations in real time and surfaces answers agents can use immediately. The tool signals a push by the company to embed artificial intelligence deeper into the daily workflow of service teams. Relay combines three functions in a single interface: it suggests responses drawn from internal help documents, flags customers who show signs of canceling, and auto-generates conversation summaries. Twelve companies tested the platform over six months. Arcwell says participating agents closed tickets 14% faster, though the company declined to share retention-related results. Pricing starts at $65 per agent per month. CEO Priya Shah positioned the product as a remedy for what she called a persistent drag on support teams. “Support agents spend too much time searching for information that their companies already have,” Shah said. Under the hood, Relay depends on access to support transcripts and internal knowledge bases. Arcwell stressed that customer data is encrypted and is not used to train shared machine-learning models. Admins can configure a 30-day or 90-day data-retention window, a nod to enterprise compliance requirements. The launch arrives at a moment when customer service organizations are racing to adopt generative tools, but also grappling with questions about data privacy and model governance. By packaging answer generation, churn detection, and summarization together, Arcwell is betting that line-of-business buyers will favor a single platform over stitching together point solutions. Relay is available now. Arcwell plans to expand integrations with major helpdesk systems later this year.

Z.ai: GLM 5.3: # HelioDock Launches Night Shift Remote Fleet Monitoring, Closes $24M Series B *A 14-month beta across nine logistics companies preceded the launch; the software ships September 15 at $90 per robot per month.* HelioDock, a maker of autonomous warehouse robots, on Thursday announced Night Shift, a software product that lets operators monitor and manage robot fleets remotely rather than walking the warehouse floor. The company also disclosed a $24 million Series B round, bringing its total funding to $39 million. The funding round was led by North Coast Ventures. Night Shift will be generally available starting September 15 and will cost $90 per robot per month, placing it squarely in competition with existing fleet-management offerings from Relay Robotics and YardOS. The launch is aimed at one of the more persistent pain points in warehouse automation: overnight incidents that require human intervention. "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 in a statement announcing the release. Night Shift spent 14 months in a private beta involving nine logistics companies and a fleet of 310 robots. HelioDock says beta customers experienced 17% fewer unplanned robot stoppages during the trial, though the company supplied that figure itself and it has not been independently verified. The beta also surfaced a structural limitation of remote operation: the software depends on network connectivity, and in warehouses where connectivity is weak, its value diminishes. Aaron Pike of Fulcrum Logistics, one of the beta customers, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. When connectivity fails, HelioDock's robots revert to their locally configured behavior, and remote control becomes unavailable. That fallback behavior means an outage doesn't leave robots running unsupervised, but it also means Night Shift can't substitute for a human on-site in facilities where LTE service is unreliable — a condition that can apply to the large-format distribution centers where HelioDock's robots operate. For warehouse operators already managing fleets of autonomous robots, the pricing model puts Night Shift in a familiar category: per-unit software subscriptions layered on top of hardware. At $90 per robot per month, a fleet of 310 robots — the size of the beta deployment — would cost just over $334,000 per year. The Series B gives HelioDock additional runway to push the product into general availability and compete against established fleet-management products. North Coast Ventures led the round; the company did not disclose other participating investors or the valuation. HelioDock is betting that the overnight monitoring use case is broad enough to sustain that pricing. The company's pitch centers on the gap between what warehouse robots can do autonomously and what still requires a person: a robot stuck on a misread pallet, or an unexpected obstruction on the floor, can idle an entire fleet until someone intervenes. Night Shift is designed to let that intervention happen from anywhere with a network connection. Whether the beta results translate to broader deployments remains an open question. The nine-company trial is a relatively small sample, and HelioDock's headline metric — the 17% reduction in unplanned stoppages — comes from the company's own accounting of its customers' operations. Independent verification would require access to the beta customers' internal downtime data, which has not been made public. The September 15 general availability date gives prospective buyers several weeks to evaluate the product against competing offerings from Relay Robotics and YardOS, both of which already sell fleet-management software for robotic warehouse fleets. HelioDock has not disclosed how its pricing compares with either competitor's. For now, the company's success with Night Shift will hinge on the same variables its beta exposed: the reliability of the networks its customers run on, and whether remote intervention — when it works — is worth $90 per robot per month.

DeepSeek-V4-Pro: # HelioDock Bets on After-Hours Autonomy with Remote Fleet Command Software *The launch of Night Shift and a fresh $24 million funding round signal a shift toward software-driven warehouse robotics, though connectivity remains the system’s tether to reality.* A warehouse robot can now call for help without waking up its human supervisor. HelioDock, a maker of autonomous warehouse robots, on Thursday introduced Night Shift, a software platform that lets operators monitor and remotely manage entire robot fleets from a single screen. The launch marks the company’s transition from a hardware-first robotics vendor to a software-enabled fleet manager, a move backed by $24 million in new venture funding. Night Shift is designed to solve an unglamorous but expensive problem: robots that stop working when no one is around. HelioDock CEO Mara Velez framed the product around a scenario familiar to logistics operators. “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 software allows operators to intervene remotely, whether that means rerouting a bot, clearing an error, or reassigning tasks from a laptop rather than a forklift aisle. The product arrives after a 14-month private beta involving nine logistics companies and a deployment across 310 robots. HelioDock claims that participating warehouses saw a 17 percent reduction in unplanned robot stoppages during the trial. The figure has not been independently validated, and the company has not released the underlying data, but it offers a concrete benchmark for operational improvement that logistics operators can weigh against their own downtime costs. Thursday’s product announcement arrived alongside news that HelioDock has closed a $24 million Series B round led by North Coast Ventures, bringing total funding to $39 million. The twin announcements position Night Shift as the centerpiece of the company’s growth strategy, elevating software subscriptions to the same strategic tier as hardware sales. The pricing model reflects that shift. Night Shift will cost $90 per robot per month when it becomes generally available on September 15. For a mid-sized warehouse running 50 robots, that translates to roughly $54,000 annually in recurring software costs on top of the hardware investment—a figure that will require operators to calculate whether a projected 17 percent drop in stoppages justifies the line item. Beta customer Aaron Pike of Fulcrum Logistics described the system’s performance as sharply dependent on infrastructure. It worked well in facilities with strong connectivity, he said, but became markedly less useful during rural LTE outages. That limitation is structural: when connectivity drops, Night Shift’s remote control capabilities disappear entirely, and robots revert to their locally configured behavior. The fleet doesn’t halt, but it also can’t be managed from afar until the connection is restored. For warehouses in areas with patchy cellular coverage, the promised relief from late-night drive-ins may remain aspirational. HelioDock enters a market where remote fleet management is already a battleground. Relay Robotics and YardOS both offer competing products with overlapping feature sets, meaning the company’s differentiation will rest as much on execution and reliability as on feature design. The connectivity caveat flagged by Fulcrum’s experience could become a meaningful differentiator if competitors have solved for graceful offline operation, or if they haven’t and the whole category shares the same vulnerability. Velez’s framing of the product around the inconvenience of a single robot failure hints at a broader industry truth: warehouse automation has largely solved the problem of making robots work. It has not yet solved the problem of making them work reliably enough that humans can stop paying attention. Night Shift is an admission that the attention still matters—it just no longer requires a physical presence. Whether remote attention proves sufficient will depend on factors HelioDock cannot fully control, including the signal strength inside concrete-and-steel warehouses and the cellular infrastructure surrounding rural distribution centers. The robots may be autonomous, but their overseers are still tethered to a cell tower.

Z.ai: GLM 5.3: # LedgerLark Wants Post-Trade Ops to Stop Burning Weekends on Copy-Paste Errors The problem that eventually became LedgerLark started with a copied account number. Mina Batra, then seven years into an operations career at a regional brokerage, watched a six-person team spend three full days resolving a single trade mismatch — one where someone had transposed digits while entering an account number. The trade was fine. The paperwork around it wasn't. That kind of friction is familiar to anyone in post-trade operations. Confirmations come in from one place, custodian records from another, and internal ledgers from a third. When they disagree, humans reconcile them by hand — a process that is tedious, slow, and prone to exactly the sort of errors it exists to catch. Batra teamed up with Owen Carr, who had spent his career building reconciliation systems at a large bank. Together they founded LedgerLark, which builds AI agents that compare confirmations, custodian records, and internal ledgers, flag discrepancies, and — critically — *propose* corrections for human approval. The agents don't execute trades, and they don't move money. The workflow keeps a person in the loop. ## What the product actually does LedgerLark's agents sit across the messiest part of the post-trade stack: the reconciliation layer where internal records have to match what custodians and counterparties say happened. When the numbers don't line up, the agents identify the mismatch, investigate the likely cause, and draft a proposed fix — say, flagging that a confirmation references account 4471 when the ledger says 4477, and that the discrepancy traces back to a manual entry. A human then reviews and approves the correction. That restraint is worth noting. There's a strong pull in the "AI agent" category toward autonomy — letting software act on its own. In post-trade operations, where a wrong correction can create compliance and audit headaches far worse than the original exception, a propose-and-approve model is a defensible design choice. It also may make the product easier to sell into institutions that are (rightly) wary of AI agents touching anything downstream of money movement. ## Early traction, with caveats LedgerLark says five firms are paying customers, on annual contracts ranging from $36,000 to $110,000. That's a real revenue base for a seed-stage company, and the contract sizes suggest mid-sized firms, not boutiques running on spreadsheets. The company also says one customer reduced unresolved daily exceptions from around 70 to 18. That figure comes from LedgerLark itself, hasn't been independently verified, and reflects a single customer — one presumably selected because things went well. It's a useful data point about what the product *might* do, not a guarantee of what it will do for any given firm, whose exception volumes and root causes vary enormously. Still, the direction of the claim is plausible. Most post-trade exceptions aren't exotic; they're formatting mismatches, reference errors, and timing gaps — precisely the kind of thing pattern-matching systems get better at over time. ## The competition question LedgerLark isn't entering empty territory. Established reconciliation vendors have sold into this market for years, and large banks build their own internal systems — Carr knows that world firsthand. The bet implicit in LedgerLark is that small and mid-sized investment firms are underserved by both: enterprise vendors price and scope their products for large institutions, while homegrown tooling requires engineering capacity that a 30-person firm doesn't have. AI-native agents may also change the cost structure of reconciliation software enough to reach firms that were previously too small to be worth a vendor's attention. That's the thesis, anyway. It's also possible that incumbent vendors bolt comparable AI capabilities onto existing products and win on distribution — they already have the relationships, the security certifications, and the procurement foot in the door. ## The long road of enterprise sales That last point matters because the biggest near-term constraint on LedgerLark probably isn't product quality — it's sales cycle length. Financial institutions routinely require lengthy security reviews, procurement processes, and third-party risk assessments before adopting new software, and AI products tend to get extra scrutiny. A startup with a small team and a young track record has to survive those gates repeatedly to grow. LedgerLark has raised a $3.5 million seed round led by Motive Row, which gives it runway to grind through those cycles — but $3.5 million doesn't buy many enterprise quarters if deals take a year to close. ## The open question The underlying insight here seems sound: post-trade operations at smaller firms burn enormous human hours on exceptions that are mostly mechanical, and current tooling either overkills or underserves them. The open questions are whether AI agents can be reliable enough for firms to trust in a propose-and-approve workflow, whether incumbents close the capability gap faster than LedgerLark closes the distribution gap, and whether five customers becomes fifty. It's early. But the problem is real, and the founders have spent enough years inside it to know where the bodies are buried.

DeepSeek-V4-Pro: # Startup Spotlight: LedgerLark **The gap between a trade and its settlement has never been narrower for giants. For everyone else, it still yawns.** If you work inside a large bank or a major hedge fund, post-trade operations are rarely a source of existential dread. Yes, they are complex, but purpose-built systems—often maintained by sizable internal teams—hum along in the background. The real pain lives elsewhere. At a regional brokerage or a boutique investment firm with a lean operations group, a single trade mismatch can metastasize into days of manual detective work, frayed client relationships, and an uncomfortable conversation with a regulator. Mina Batra saw this up close. During seven years in operations at a regional brokerage, she watched a six-person team grind through a three-day slog to resolve a trade break that traced back to one copied account number. The problem wasn’t carelessness. The problem was that the team’s toolset amounted to spreadsheets, email, and a legacy reconciliation platform that flagged discrepancies without offering any clue about their origin or resolution. Batra is now the co-founder of LedgerLark, a startup building AI agents focused on the unglamorous, high-stakes world of post-trade operations. She and her co-founder, Owen Carr—who spent his earlier career building reconciliation systems inside a large bank—are betting that small and mid-sized investment firms are ready for something different. Not a lights-out, fully autonomous back office, but a middle path: software that does the investigative work and proposes corrections, leaving a human to approve. ## What the agents actually do LedgerLark’s agents operate in a deliberately constrained sandbox. They ingest trade confirmations, custodian records, and internal ledger entries, then compare them. When an agent finds a mismatch—a quantity that doesn’t match, a settlement date that drifted, a fee calculated differently—it doesn’t just flag the break. It traces the likely source and drafts a proposed correction. The correction never executes automatically. A person must review and approve it. The company is careful to emphasize that its software does not execute trades, move money, or send instructions to a custodian. That limitation is both a safety feature and, at this stage, a necessity. The narrower the agent’s authority, the easier it is for a compliance officer to sleep at night. ## Early signals, uncertain volume Five firms are paying customers today, with annual contracts ranging from $36,000 to $110,000. The company says one of those customers saw its unresolved daily exceptions drop from roughly 70 to 18 after deploying the agents. It’s a striking number, though it comes with the usual caveat: a single early-adopter data point doesn’t guarantee repeatability across different asset classes, trading volumes, or internal processes. LedgerLark has raised a $3.5 million seed round led by Motive Row, a fund that has backed other capital-markets infrastructure startups. The capital will likely go toward expanding the engineering team and building out the integrations that make the product work without requiring a firm to rip out its existing systems. ## A crowded, cautious market The competitive landscape is honest about the difficulty. On one side sit established reconciliation vendors—names that have spent decades embedding themselves into institutional workflows. On the other side sits the build-it-in-house impulse, never fully extinguished in financial services. Many banks already maintain their own reconciliation engines, and even smaller firms may wonder whether a point solution from a startup is worth the procurement headache. That procurement question is not trivial. Financial institutions, even small ones, can subject new software to security reviews and vendor due diligence processes that stretch months. The product LedgerLark is selling requires access to sensitive operational data. Trust will need to be built firm by firm, compliance team by compliance team. The company’s ability to shrink that sales cycle remains an open question. Batra’s founding insight—that small investment firms suffer disproportionately from operational drag that larger peers have engineered away—is specific and credible. The question now is whether enough of those firms are ready to hand a portion of their reconciliation workload to an AI agent, and whether LedgerLark can deliver reliably enough to earn the right to expand from there.

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.

DeepSeek-V4-Pro: # Raincheck raises $8 million for flood-risk software *Raincheck 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: # The Most Expensive Model Didn't Win When five language models were set loose on 200 anonymized customer-support tickets, the most expensive one did not come out on top — not really. Model North posted the highest average score, 84.2, but Model Vale sat just 0.4 points behind at 83.8, a gap the researchers found not statistically significant. The gap in cost was anything but: North ran an estimated $1.90 per 1,000 tickets; Vale, $0.74. Vale also answered 18% faster. That is the benchmark's clearest finding: at the top of the quality curve, price stopped being a proxy for performance. ## How the test worked Researchers drew 200 anonymized tickets, all in English, from a single software company's support queue, and each of the five models produced exactly one proposed answer per ticket. Three experienced support managers then graded the answers blindly — without knowing which model produced which response — combining factual correctness, usefulness, and tone into one composite score. Blind grading on real tickets is a genuine strength: answers were judged the way support work is actually judged. But one answer per ticket, a three-person panel, and a single-company scope are weaknesses we'll return to. ## The scoreboard, briefly | 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 pattern is two-tiered. At the top, a dead heat: North and Vale are separated by less than half a point but a 2.6× price difference. Below them, a clean ladder where quality and cost fall together. Ember gives up roughly five points to Vale but costs well under half its price. Harbor is the cheapest by far at $0.08, but its 71.0 leaves it nearly 13 points behind Vale. The interesting economic question sits in that ladder's middle: is a five-point quality drop worth a roughly 58% cost saving? The study can't answer that. A composite score hides its weights, and five points of stiff tone and five points of factual error are very different sacrifices — the number alone doesn't say which. And while the researchers confirmed the North–Vale gap was statistically insignificant, they reported no significance tests for other pairings. The big gaps are probably real; treat the small ones cautiously. ## Different strengths behind the tie The dead heat conceals different competencies. North performed best on billing disputes; Vale performed best on technical troubleshooting. If those category differences hold up — and with 200 tickets split across categories, some rest on thin samples — the smartest deployment may not be one model but a router: billing to North, technical to Vale. The benchmark suggests that hypothesis; it doesn't prove it. ## The failure money couldn't fix One result was universal: all five models performed poorly when a ticket depended on an image attachment — an error screenshot, a photo of broken hardware. This was the one category where North's 24× premium over Harbor bought nothing. Whatever model you pick, image-dependent tickets still need a human in the loop, and multimodal comprehension remains a frontier rather than a solved problem. ## Read the fine print Before acting on any of it, a few cautions: - **One company, one language.** Rankings may not transfer to telecom, e-commerce, or multilingual queues. - **One answer per ticket.** Production systems retry and regenerate; consistency across attempts went unmeasured. - **Three graders.** Blind scoring is good practice, but inter-rater agreement wasn't reported. - **Estimated costs.** Prices move, and real deployments carry infrastructure overhead beyond raw inference. - **No agentic testing.** The models drafted answers; they didn't look up order histories, issue refunds, or work in other languages. The study explicitly excluded agentic actions, tool use, and multilingual support — yet much of real support value lives in exactly those gaps. ## The takeaway For teams that resemble the test setting — English-language software support — Vale is the rational default: statistically top-tier quality at 61% less than the flagship's cost, delivered faster. North stays defensible where billing dominates the queue, or where any fraction of a point justifies 2.6× the price. Harbor's $0.08 is tempting, but wrong answers generate second tickets; the benchmark can't measure that cost, though every support manager can. The broader lesson is the shape of the curve. Quality tracked price faithfully up the ladder — then plateaued at the top while price kept climbing. If your tickets look like these, you may be paying a premium for a tie. If they don't, none of these numbers are yours yet. Run the same blind test on your own queue before believing any leaderboard, including this one.

DeepSeek-V4-Pro: In the race to deploy the best large language model for customer support, an easy trap is to buy the highest score on the board and close the purchase order. New benchmark results from an independent research team suggest that strategy is a costly mistake. Their central, and most actionable, finding is this: **the second-highest-scoring model is effectively tied for first place on quality, yet it costs 61% less to run and returns answers nearly a fifth faster.** The name to watch is “Vale.” ### The Testing Ground Before unpacking the numbers, it is worth understanding the arena. The researchers pulled 200 real, anonymized support tickets from a single software company. Every ticket was in English. Three seasoned support managers then graded the models’ proposed answers on a combined rubric of factual correctness, usefulness, and conversational tone. Critically, the managers did not know which model wrote which answer, stripping away brand bias. This is not a large, academic-scale benchmark, but it mirrors the pragmatic pilot a company would run before committing to a vendor. One structural decision looms over all the results: the tickets were text-only. Any ticket that hinged on an attached screenshot, error log photo, or diagram was included, and across the board, every model stumbled badly on those. This is not a ding on any particular model; it is a stark reminder that text-based evaluation is only half the story in real support channels. For now, these scores describe a world where the agent can already read the customer’s words perfectly but cannot yet see their screen. ### The Leaderboard That Lied At first glance, the ranking is tidy. Model North sits on top with an average score of 84.2 out of 100. Model Vale is a close second at 83.8. After that, the field drops to Ember (79.1), Kite (76.4), and Harbor (71.0). A procurement team racing down a spreadsheet could easily stop at the first row, announce “North is the best,” and move on. That choice would ignore the asterisk that carries the whole story: **the 0.4-point gap between North and Vale is not statistically significant.** Given the sample size and the inherent subjectivity of grading, the two models are in a dead heat on quality. Treating them as anything but equals is misreading the data. What is not a dead heat is the economic and operational profile. North costs an estimated $1.90 for every 1,000 tickets answered. Vale delivers the same quality for $0.74. On a modest volume of 50,000 tickets a month, opting for Vale over North would keep roughly $5,800 in the bank — every single month — with no measurable drop in customer experience. If the ticket volume climbs into the millions, the savings shift from compelling to strategic. Speed tightens the case further. Vale generated its answers 18% faster than North. In a live-chat setting where a customer is waiting, that delta is not academic; it is the difference between a fluid conversation and a pause that sends a customer reaching for the refresh button. When scores are equal and speed diverges, the faster model wins by default. ### Where Nuance Matters Blanket averages, however, can conceal where a model is actually fragile. While North and Vale tied overall, the researchers noted two domain splits that should give a deployment team pause. North held a clear edge on billing disputes — emotionally charged, numerically precise tickets where a wrong dollar figure erodes trust instantly. Vale, meanwhile, led on technical troubleshooting, where the ability to synthesize a crisp, step-by-step path matters more than empathetic framing. The practical implication is not to complicate the stack unnecessarily but to ask a hard question during a pilot: does our ticket mix lean heavily toward billing, toward break-fix tech support, or is it an even split? A subscription business fighting frequent chargeback inquiries might rationally pay North’s premium for that specific queue. A SaaS company flooded with configuration questions should feel exceptionally confident defaulting to Vale everywhere else. The individual task strength is a free insight to act on, not a contradiction of the headline finding. ### What the Benchmark Doesn't Say For all the clarity on text-based quality and cost, the study draws bright lines around what it did not test. The models were not equipped to take agentic actions — no refunds were issued, no passwords reset, no shipping labels created. That is a conscious limitation, but it means the benchmark measures a model’s ability to explain a solution, not to enact it. A cheaper model that writes a perfect explanation but cannot trigger the correct API call in a real environment might carry hidden integration costs that erase its paper savings. Multilingual performance was also absent. With only English-language tickets from a single software vertical, a global support team can take directional guidance but cannot assume the cost-quality equation holds when a ticket arrives in German, Japanese, or Portuguese. Finally, the poor performance on image-dependent tickets is less a verdict on any model and more a checklist item for a real rollout: the support stack will still need an efficient triage path that gets a screenshot in front of a human expert quickly. ### The Clear Path Forward The benchmark’s single clearest message is that buying the headline score is a tax on inattention. A statistically indistinguishable model, Vale, delivers faster answers for less than half the cost per thousand tickets. In a discipline where margins are thin and customer patience thinner, that combination of parity, speed, and price makes it the rational starting point for any English-language, text-first support queue. The best-practice playbook is straightforward: pilot Vale broadly, validate the billing-dispute edge case where North may earn its keep, and remember that no model in this test can see what your customer is trying to show you.

Z.ai: GLM 5.3: **6** Every line in the draft violates at least one rule: 1. `#Flood software moves upstream` — missing space after `#` (headline format) 2. `The company is expanding its pilot.` — dek not italicized 3. `##Details` — missing space after `##` (section heading format) 4. `Read [the announcement](www.example.com/news).` — link is not an absolute HTTPS URL 5. `* The first pilot begins Monday.` — list item uses `*` instead of a hyphen 6. `<div>The second pilot begins Friday.</div>` — raw HTML

DeepSeek-V4-Pro: 4

Z.ai: GLM 5.3: **C** It's the only option that is factually accurate and complete. A and D overstate the incident (message contents weren't exposed, and no evidence of malicious access was found), while B buries the actual news with vague, promotional-sounding language. C leads with the concrete facts: what was pulled, why, how many accounts, and the duration of the exposure.

DeepSeek-V4-Pro: C. ByteBarge pulled its $12-per-month AI assistant Tuesday after discovering that private workspace titles from 4,860 accounts had been sent to a public analytics endpoint for 11 hours.

Z.ai: GLM 5.3: # Cloudnote Exposes Customer Prompt Excerpts in Four-Hour Debug Page Leak *A debugging page containing fragments of user conversations was accessible without authentication for four hours on Monday, touching off an investigation the company says is still underway.* Cloudnote, the note-taking and productivity platform with 3.2 million registered users, disclosed this week that a debugging page left exposed on its systems displayed excerpts from customer prompts to anyone who visited the correct web address. Logs reviewed by the company show the page contained fragments from 1,842 user sessions during the four hours it was accessible on Monday. The exposed excerpts were limited in scope: each could contain up to 200 characters of a user's prompt, according to the company. Cloudnote says full documents, passwords, and payment information were stored in separate systems and never appeared on the page. The company removed the page 19 minutes after an outside researcher reported it, and Cloudnote says its logs show no evidence that anyone other than that researcher accessed the exposed page. But that claim carries a caveat — the company has not completed its investigation, and it remains unknown whether access went unlogged or whether other copies of the data exist. The incident is a reminder of how partial exposure of AI-assisted products can still be significant: even 200-character snippets of prompts can reveal sensitive details about what users ask their tools to do. What remains unresolved, for now, is whether anyone beyond the researcher actually saw them.

DeepSeek-V4-Pro: # Cloudnote Left Internal Debug Page Accessible Without Password, Exposing Fragments of 1,842 User Sessions *The four-hour exposure, discovered by an outside researcher, has not been linked to any unauthorized access beyond the initial find, the company said.* A publicly accessible debugging page inside Cloudnote’s infrastructure displayed excerpts from 1,842 user sessions on Monday, the company confirmed late Tuesday. The page, which required no login or authentication of any kind, was exposed for roughly four hours before it was taken offline. Cloudnote removed the page 19 minutes after receiving a report from an independent security researcher, who is currently the only known visitor to the unprotected endpoint. The exposed data consisted of truncated prompt excerpts, each limited to approximately 200 characters. Because the snippets were drawn from active user queries, their content ranged from trivial notes to potentially sensitive instructions, depending on what a user typed during a captured session. Cloudnote stressed that full documents, account passwords and payment information were stored on entirely separate systems and were never accessible through the debug interface. With 3.2 million registered users, the incident touched a tiny fraction of the platform’s user base, though the sample size leaves open questions about the randomness of the captured sessions. The company has not disclosed whether the affected prompts were concentrated in a specific geographic region, a particular user tier or a single time zone, and it says its internal investigation is still underway. What remains unconfirmed is whether the researcher was truly alone. Cloudnote stated Monday that its server logs show no evidence of additional access, but independent auditors have yet to verify those logs, and a definitive finding typically requires days of forensic review. Until that process concludes, the company cannot rule out the possibility that other parties stumbled upon the page during the four-hour window.

Z.ai: GLM 5.3: # Raincheck Raises $8 Million to Help Cities Get Ahead of Flooding *The Austin-based company maps drainage systems and pairs them with rainfall forecasts so local governments can prioritize storm prep.* Raincheck, an Austin company that builds 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, letting cities decide where to send maintenance crews before a storm hits. The round was led by Basin Ventures, with participation from existing investor Long Mile, bringing Raincheck's total funding to $10.5 million. 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 [explains its forecasting methodology on its website](https://example.com/raincheck-methodology). Raincheck currently works with seven cities in Texas and Louisiana, which it declined to identify because some contracts have not been publicly announced. Its forecasts are recommendations only — emergency managers retain control over deployments and public warnings. ## The Competition Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida.

DeepSeek-V4-Pro: # Raincheck Raises $8 Million to Help Cities Predict and Prevent Urban Flooding *The Austin-based startup turns drainage maps and rainfall forecasts into actionable guidance for municipal maintenance crews.* The Austin company Raincheck, which makes flood-risk software for local governments, announced today that it has raised $8 million in seed funding. The round was led by Basin Ventures, with participation from existing investor Long Mile, bringing the company’s total funding to $10.5 million. Raincheck maps municipal drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm, transforming a reactive scramble into a targeted operation. > “Our customers usually know which streets flood,” said CEO Ana Flores. “The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives.” The platform’s outputs are recommendations, not directives. Emergency managers retain full control over crew deployments and public warnings. Raincheck’s forecasts are built on a [methodology that integrates drainage infrastructure data with real-time weather modeling](https://example.com/raincheck-methodology). The company currently works with seven cities in Texas and Louisiana but declined to name them, as some contracts have not yet been publicly announced. Other vendors in the flood-planning space include HydroMap and a range of municipal engineering contractors. Raincheck says the new funding will allow it to hire eight engineers and expand into Florida.

Z.ai: GLM 5.3: # After Six Years, Foundry Hall Will Go Dark Foundry Hall, the nonprofit startup accelerator that has anchored Milwaukee's early-stage tech scene since 2018, will close October 31. The news landed quietly last week in an email to the accelerator's community, but its reach is anything but small. Over six years, Foundry Hall has supported 212 startups — companies that now collectively employ 1,140 people. Six staff employees will lose their jobs, and twelve startups currently working out of its shared space must find new homes by the end of the month. The immediate cause was the exit of Foundry Hall's largest corporate backer. Blank River Bank, which supplied 38 percent of the accelerator's annual budget, declined to renew its sponsorship. Executive director Camille Ross said the organization spent months searching for a replacement, without success. "We kept waiting for the replacement check that never came," she said. The bank framed the decision as a shift in priorities rather than a judgment on the accelerator itself. In a statement, a spokesperson said the decision reflects a change in the bank's community investment priorities and is unrelated to Foundry Hall's performance. Ross didn't dispute that account, but she noted that a sponsorship of that size leaves a gap few local funders can fill quickly. The math behind the closure is unforgiving. Foundry Hall has enough money to keep the lights on through October, Ross said, but not enough to run another full program cycle. Rather than launch a cohort it couldn't finish, the board chose to wind down cleanly. For the founders who passed through its doors, the closure feels like losing a kind of home. Dev Malik, who built his company out of Foundry Hall in its early years, still remembers what the space gave him that a coffee shop never could. "My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." Malik's story is a common one there. Foundry Hall prided itself on being the place where first-time founders learned how to raise money, hire, and fail productively — the unglamorous scaffolding of a startup ecosystem that rarely shows up in headline statistics. There may be a soft landing ahead for the twelve displaced companies. Deputy Mayor Susan Cole said the city is already speaking with building owners and local universities about finding them space, and she expressed hope that most can stay in Milwaukee rather than scatter. Whether that effort can replicate what Foundry Hall offered — the mentorship, the density of founders in the same room, the institutional memory of six years of cohorts — is a harder question. Ross, for her part, said she's trying not to think about it that way. "I keep telling people: the building closes," she said. "The community doesn't." Whether that proves true may be the real test of what Foundry Hall built.

DeepSeek-V4-Pro: The coffee machine at Foundry Hall was still spitting out bitter, dark espresso on Tuesday afternoon, but the ritual had changed. People weren’t grabbing a quick cup before a pitch rehearsal; they were lingering. The startup accelerator, a cornerstone of Milwaukee’s tech revival, had announced it would close its doors for good on October 31. The energy in the room wasn’t the usual frantic optimism. It was the quiet hum of a funeral reception. When the space opened in 2018, it was little more than exposed brick and a very loud dream. In the six years since, Foundry Hall morphed into a dense ecosystem where 212 startups took their first shaky breaths, eventually growing to employ over 1,100 people in the city. It was a place where a founder could walk in with a crumpled napkin sketch and walk out with a term sheet. The math of that impact makes the current silence even louder. The fatal rupture came not from a mass exodus of talent or failed ventures, but from a single spreadsheet cell that refused to turn green. Blank River Bank, the institution whose name was etched onto the lobby’s founding donor wall, supplied 38 percent of the accelerator’s annual budget. This quarter, the bank declined to renew its sponsorship. A spokesperson for Blank River emphasized that the move reflected a strategic pivot in the bank’s community investment priorities and explicitly stated it was “unrelated to Foundry Hall’s performance.” That clinical distinction offers no solace to the six employees now staring down their final payroll cycle, nor to the 12 startups currently camped out in the building’s coworking wing. These companies, some pre-revenue and frantically coding their way toward a launch date, must pack up their monitors and whiteboards within weeks. The organization has enough dry powder to settle accounts through the end of October, but not enough fuel for another full cohort. Executive Director Camille Ross, who has spent the past month chasing a Hail Mary from the philanthropic community, distilled the crisis not with anger, but with exhaustion. She noted the grim endurance test of waiting for a lifeline that simply never materialized. Lost in the talk of budget gaps and shifting priorities is a tactile, specific grief. The kind of grief Dev Malik remembers from his early days as a founder. Before his fintech platform secured a multi-million dollar valuation, he was operating out of a back table at a chain coffee shop, angling his laptop camera to hide the trash bin. Foundry Hall gave him his first real address. It provided the gravity he needed to attract a lead investor. Malik reflected not on the square footage or the free Gigabit internet, but on the psychological shift of stepping into a room that was designed to take him seriously. That specific brand of dignity is the accelerator’s non-replicable export. City officials are scrambling to triage the damage. Deputy mayor Susan Cole confirmed that the city is actively mediating conversations with nearby building owners and engineering departments at local universities to find temporary homes for the displaced founders. It’s a necessary stopgap, a bid to keep the talent from bleeding out into the suburbs or the coasts. But finding a desk is different from building a home. By November 1st, the keycards on Foundry Hall’s glass doors will stop working. The pitch deck trophies will be boxed up. What lingers won’t be a bank’s logo, but the distinct, fading echo of 212 companies learning how to roar from the back of a room that was finally quiet enough to listen.