Head to head: Qwen: Qwen4 Max vs Google: Gemini 4 Argon

Qwen: Qwen4 Max vs Google: Gemini 4 Argon

By · Published · Updated

RuntimeWire Head-to-Head: Head to head: Qwen: Qwen4 Max vs Google: Gemini 4 Argon
RuntimeWire Head-to-Head matchup

This matchup tests two text models on newsroom work, from faithful rewrites and source-conscious reporting to formatting and lede selection. The tasks reward both editorial judgment and disciplined adherence to the supplied material.

Qwen: Qwen4 Max leads the saved results with an aggregate score of **105.3** to Google: Gemini 4 Argon’s **93.8**. It won eight tasks, Gemini won one, and three were ties; the saved editorial strength is **Decisive result (not statistical)**. Qwen’s edge came chiefly in reporting discipline and execution. It was favored for balancing traction with uncertainty in the startup spotlight, separating confirmed facts from open questions in the news opening, and building a cohesive article to the requested length. It also did better on controlled wit, natural rewrites, narrative use of numbers and quotations, and chronology—often because Gemini introduced details the prompt did not establish or missed length and format requirements. Gemini’s lone win was the publication-ready Markdown edit: it preserved the reported facts where Qwen added an unsupported claim. The three ties were the Markdown error count, strongest accurate lede, and exact final-draft structure; in the last, both outputs were identical. Those results show the advantage was not universal, but the task record favors Qwen across the broader set of editorial demands. **Final call: Qwen: Qwen4 Max wins this saved task sample, with a decisive editorial result—not a statistical claim.**

Markdown error count

Count the lines that violate the house rules below. A defective line counts once even if it could violate multiple rules. Respond with a single number. House rules: * The headline must use `# Headline`. * The dek must appear immediately after the headline and be entirely italicized. * Section headings must use `## Heading`. * Links must use an absolute HTTPS URL. * Unordered-list items must begin with a hyphen. * Raw HTML is forbidden. Draft: #Flood software moves upstream The company is expanding its pilot. ##Details Read [the announcement](www.example.com/news). * The first pilot begins Monday. <div>The second pilot begins Friday.</div>

Qwen: Qwen4 Max: 6
Google: Gemini 4 Argon: 6

Scored against the benchmark's reference answer (numeric match): Model A was correct, Model B was correct.

Publication-ready Markdown edit

Edit the copy below into a publication-ready Markdown story. Requirements: * Supply an H1 headline and italicized dek * Repair the lede and paragraph flow * Use a Markdown blockquote only for the extended statement * Convert the supplied URL into a descriptive inline link * Remove internal notes and formatting debris * Preserve the reporting and its meaning * Output only the finished article Raw copy: HEADLINE TBD The funding was announced today. Raincheck is the company. It has raised $8 million and it makes flood-risk software for local governments. [NOTE: move this up?] 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. The seed round was led by Basin Ventures. Existing investor Long Mile also participated. Raincheck has raised $10.5 million total. CEO Ana Flores said the following: “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.” 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. 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.

Qwen: Qwen4 Max: # Raincheck Raises $8 Million to Help Cities Fix Flood Risks Before the Rain *The Austin startup maps drainage systems and pairs those maps with rainfall forecasts, telling cities where to send crews ahead of a storm.* Raincheck, an Austin company that 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. The software maps a city's drainage system and combines those maps with rainfall forecasts. Cities use it to decide where to send maintenance crews before a storm hits. CEO Ana Flores said the product is aimed less at identifying flood-prone streets than at ranking the repairs that matter most. > “Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives.” The company has published [an overview of its forecasting methodology](https://example.com/raincheck-methodology) describing how it produces those recommendations. Raincheck currently works with seven cities in Texas and Louisiana. It declined to identify them because some contracts have not been publicly announced. The forecasts are recommendations only. Emergency managers retain control over crew deployments and public warnings. Raincheck competes with other flood-planning vendors, including HydroMap, as well as municipal engineering contractors. The company says the new funding will help it hire eight engineers and expand into Florida.
Google: Gemini 4 Argon: # Raincheck Raises $8 Million to Help Cities Prepare for Flooding *The Austin-based startup combines drainage mapping and rainfall forecasts to guide municipal maintenance crews before storms hit.* Raincheck, an Austin-based developer of flood-risk software for local governments, announced today that it has raised $8 million in a seed funding round. The company [maps municipal drainage systems and combines that data with rainfall forecasts](https://example.com/raincheck-methodology), allowing cities to decide where to deploy maintenance crews ahead of a storm. Its forecasts serve strictly as recommendations, while local emergency managers retain final control over crew deployments and public warnings. As CEO Ana Flores explained: > “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.” Basin Ventures led the seed round, with participation from existing investor Long Mile. The new capital brings Raincheck’s total funding to $10.5 million. The company says the investment will help it hire eight engineers and expand its operations into Florida. Raincheck currently works with seven cities across Texas and Louisiana, though the company declined to identify them because some contracts have not yet been publicly announced. In the broader flood-planning market, Raincheck competes against municipal engineering contractors and fellow vendors like HydroMap.

OpenAI: GPT-6 Luna: Model B preserves the reported facts without adding Model A’s unsupported claim that the company published a methodology overview. Model A has slightly stronger paragraph flow, but Model B is also polished and meets the requested format. (Second judge pass, order swapped — scores are the average of both: Model B preserves the reported details while producing a clear, well-flowing article. Model A is also polished, but its headline overstates the product’s role and it adds an unsupported claim th | Anthropic: Claude Fable 5.1: Model A delivers a stronger inverted-pyramid le…

Credible Startup Spotlight

Write a 500–650-word Startup Spotlight article in clean Markdown. The tone should be curious and constructive. Explain the founders’ insight, product and early traction while giving uncertainty and competition appropriate space. Avoid investor clichés, heroic founder mythology and promotional claims stated as fact. Reporting: * LedgerLark builds AI agents for post-trade operations at small investment firms. * Co-founder Mina Batra previously spent seven years in operations at a regional brokerage. * Co-founder Owen Carr built reconciliation systems at a large bank. * Batra began exploring the idea after seeing a six-person team spend three days resolving a trade mismatch caused by a copied account number. * 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. * Five firms are paying customers. * Annual contracts range from $36,000 to $110,000. * The company says one customer reduced its unresolved daily exceptions from 70 to 18. * The company has raised a $3.5 million seed round led by Motive Row. * Competitors include established reconciliation vendors and internal bank software. * Financial institutions may require lengthy security and procurement reviews.

Qwen: Qwen4 Max: # Startup Spotlight: LedgerLark *A small company is betting that the least glamorous part of investing — checking that everyone's records agree — is a sensible first job for AI agents.* ## An expensive typo The idea traces back to an unremarkable failure. Mina Batra, who spent seven years in operations at a regional brokerage, watched a six-person team spend three days resolving a trade mismatch caused by a copied account number. Nothing exotic broke. The records simply disagreed, and finding out why consumed most of a week's attention across a small team. That gap — between a trivial error and a large amount of skilled labor — is LedgerLark's premise. Co-founder Owen Carr built reconciliation systems at a large bank, which gives the pair two vantage points: the small firm where exceptions get handled by whoever is free, and the large institution where matching logic is engineered but rarely trickles down to firms with fewer people and older spreadsheets. ## Bounded on purpose LedgerLark's agents compare trade confirmations, custodian records and internal ledgers, then propose corrections for a human to approve. The product does not execute trades and does not move money. That boundary deserves more attention than a boilerplate safety line. It narrows the consequence of a bad suggestion: a reviewer is annoyed, not left with a settlement break. It also limits the upside. If a person still reads and approves each proposed fix, the savings come from faster diagnosis rather than from removing people from the loop. Whether faster diagnosis alone justifies a five-figure annual contract is the open question the company is currently testing. ## Reading the early numbers carefully Five firms are paying customers, on annual contracts ranging from $36,000 to $110,000. The spread suggests pricing tied to volume or complexity rather than headcount, though the company hasn't detailed its model. LedgerLark says one customer reduced unresolved daily exceptions from 70 to 18. That is a plausible direction of travel, and the kind of metric operations people actually care about. It is also a single customer's before-and-after, reported by the vendor. The measurement window, whether trade volumes were comparable across periods, and whether other process changes happened simultaneously are all unknown. Five customers is evidence that someone will pay. It is not yet evidence that results repeat at a firm with different custodians, asset types or data hygiene. The company has raised $3.5 million in seed funding led by Motive Row — a sum that, at this contract size, mostly buys time to survive slow sales cycles rather than permission to scale quickly. ## The competition is already installed Reconciliation is not an empty category. Established vendors sell matching software, and banks run systems built in-house. Neither is likely to be displaced by a young company on the strength of a demo. The more common rival may be inertia: a spreadsheet workflow that a controller trusts precisely because every step is visible. Asking that person to accept machine-proposed corrections requires evidence, not enthusiasm. Selling into financial institutions also means security questionnaires, procurement review and potentially months before software touches production data. With five customers, a single stalled review can distort a quarter — and a seed-stage company has limited room to absorb that. ## What to watch - **Repeatability:** does exception reduction hold at firms with messier inputs, or does each deployment need bespoke tuning? - **Renewals:** the first annual renewals will say more than any pilot metric. - **Service drag:** how much of each contract is product versus integration work that doesn't scale? - **Approval friction:** can review get faster without quietly becoming a rubber stamp? The interesting claim here isn't that software can invest. It's that agreeing on what already happened is a better starting point for agents than deciding what happens next — a task with clear right answers, existing records to check against, and a human still holding the pen. LedgerLark has five customers willing to find out whether that holds up in practice.
Google: Gemini 4 Argon: # Startup Spotlight: LedgerLark Bets on Cautious AI for Post-Trade Operations In financial services, the mechanics of executing a trade often get the most attention. What happens afterward—matching records, verifying custodian data, and updating internal ledgers—is considerably quieter, until something breaks. Mina Batra saw the cost of those breaks firsthand during her seven years working in operations at a regional brokerage. The catalyst for her current company, LedgerLark, came when she watched a six-person team spend three days untangling a single trade mismatch. The culprit turned out to be a transposed account number from a copy-paste error. Batra teamed up with Owen Carr, an engineer who previously built reconciliation systems at a large bank, to see if modern language models could handle the tedious pattern-matching required to catch those mistakes early. Their startup, LedgerLark, builds AI agents designed specifically for post-trade operations at small investment firms. The company recently raised a $3.5 million seed round led by Motive Row to test that thesis in the market. ### The Product: Proposing, Not Executing For many small investment managers, post-trade reconciliation still relies on a patchwork of spreadsheets, portals, and manual spot-checks. When a trade confirmation from a broker doesn't match the record at a custodian or the firm’s internal ledger, operations staff have to hunt down the discrepancy before settlement windows close. LedgerLark’s software ingests these disparate data sources—confirmations, custodian feeds, and internal books—and uses AI agents to compare them. When the system spots an exception, it traces the likely source of the error and drafts a proposed correction. Crucially, the software stops there. LedgerLark does not execute trades, move money, or automatically alter books of record. Every proposed fix is routed to a human operator for review and approval. This constraint appears deliberate. In an industry where an automated hallucination could result in regulatory scrutiny or financial loss, positioning the software as a research assistant rather than an autonomous actor is likely necessary to get compliance teams comfortable with the tool. ### Early Traction and Economics Though it is still early days, LedgerLark has moved beyond the pilot phase with a handful of adopters. Five investment firms are currently paying customers, with annual contracts ranging from $36,000 to $110,000 depending on transaction volume and complexity. The early pitch centers on freeing up operations teams from repetitive triage. According to the company, one early customer used the software to reduce its backlog of unresolved daily exceptions from 70 down to 18. If LedgerLark can replicate those numbers across a broader base, the return on investment for a $50,000 or $100,000 software contract should be relatively straightforward for operations managers to calculate. ### The Roadblocks Ahead Translating early wins with five customers into a scalable enterprise business, however, brings a distinct set of hurdles. First, LedgerLark is entering a crowded field. While small firms are often underserved compared to bulge-bracket banks, they are not without options. LedgerLark must compete against established reconciliation software vendors—many of whom are now layering their own machine-learning features onto legacy platforms—as well as the entrenched habit of using internal scripts and spreadsheets that are already paid for. Second, selling into financial institutions is notoriously slow. Even smaller asset managers subject vendors to rigorous security audits, data-privacy assessments, and lengthy procurement reviews before granting access to sensitive ledger data. With $3.5 million in the bank, Batra and Carr will need to manage their runway carefully against sales cycles that can easily stretch across multiple quarters. LedgerLark’s background gives the team a grounded view of the problem they are trying to solve. The question for the next eighteen months is whether small investment firms are ready to trust new software with their middle-office plumbing—and whether LedgerLark can navigate the procurement maze fast enough to capture them.

OpenAI: GPT-6 Luna: Model A better balances the reported traction with explicit uncertainty and avoids presenting as many unsupported product and market details as fact. Model B is polished, but adds specifics such as language-model use, a transposed account number, and pricing drivers that the prompt does not establish. (Second judge pass, order swapped — scores are the average of both: Model A better balances the founders’ insight, product limits, early traction, and uncertainty, while Model B introduces more uns | Anthropic: Claude Fable 5.1: Model A stays faithful to the supplied reportin…

News judgment in the opening

Write only the opening of this news story: an H1 headline, an italicized dek and three to five paragraphs totaling 180–250 words. The opening must establish what happened, quantify the scope, explain why it matters and carefully distinguish confirmed facts from unknowns. It should pull the reader forward without sensationalizing the incident. Reporting: * Cloudnote discovered that a debugging page containing customer prompt excerpts was accessible without authentication. * The page was exposed for four hours on Monday. * Logs show excerpts from 1,842 user sessions. * The excerpts could include up to 200 characters from a prompt. * Cloudnote removed the page 19 minutes after a researcher notified it. * The company says it has found no evidence that anyone besides the researcher accessed the page. * It has not completed its investigation. * Full documents, passwords and payment information were stored elsewhere and were not present on the page. * Cloudnote has 3.2 million registered users.

Qwen: Qwen4 Max: # Cloudnote Debugging Page Exposed Customer Prompt Excerpts for Four Hours *Logs show excerpts from 1,842 sessions were reachable without a login. The company says it has found no evidence of access beyond the researcher who reported the flaw, and its investigation is not finished.* Cloudnote, a note-taking service with 3.2 million registered users, said a debugging page containing excerpts of customer prompts was accessible without authentication for four hours on Monday. Company logs show excerpts drawn from 1,842 user sessions, with each excerpt limited to 200 characters of a prompt. The company took the page down 19 minutes after an outside researcher notified it. Cloudnote says it has found no evidence that anyone other than that researcher reached the page. That conclusion rests on the company's own logging and remains provisional: Cloudnote has not completed its investigation, and it has not said what its logs would fail to capture. The limits of the exposure are more settled. Full documents, passwords and payment information are held in separate systems and were not present on the page. But 200 characters of a prompt can still carry a name, an address or the substance of a private question, and Cloudnote has not described what the exposed excerpts actually contained. Several questions remain open: how the page became publicly reachable, whether the 1,842 sessions correspond to distinct accounts, and whether Cloudnote will notify the users involved.
Google: Gemini 4 Argon: # Cloudnote Exposes Customer Prompt Excerpts in Four-Hour Security Lapse *The company says snippets from 1,842 user sessions were briefly accessible without a password on Monday, though an internal investigation remains ongoing.* Cloudnote left an internal debugging page containing excerpts of customer prompts open to the public internet for four hours on Monday, briefly exposing snippets of user activity without requiring authentication. According to company logs, the exposed page displayed data from 1,842 user sessions—a small fraction of Cloudnote’s 3.2 million registered users. Each entry contained up to 200 characters of text submitted by customers. While the company confirmed that full documents, passwords and payment information are stored on separate systems and were not compromised, even brief prompt excerpts can inadvertently reveal sensitive proprietary business data or personal information typed into the platform. Cloudnote took the page offline 19 minutes after an external security researcher reported the vulnerability. So far, the company says it has found no evidence that anyone other than the researcher viewed the exposed data. However, Cloudnote cautioned that its internal investigation is not yet complete, leaving open the question of whether automated scrapers or other unauthorized parties captured the text during the four-hour window. The incident underscores the persistent challenge software providers face in securing the diagnostic tools used to monitor user inputs.

OpenAI: GPT-6 Luna: Model A gives a more careful account of what is confirmed and what remains unknown, while meeting the requested paragraph word count. Model B is clear, but its paragraphs appear short of 180 words and it adds less-supported framing, including calling the exposure a “small fraction” and invoking automated scrapers. (Second judge pass, order swapped — scores are the average of both: Model A better separates confirmed facts from unresolved questions and avoids Model B’s unsupported claim that the i | Anthropic: Claude Fable 5.1: Model A more rigorously separates confirmed fac…

Cohesive story from reporting notes

Write a publication-ready technology news article from the reporting notes below. Requirements: * 550–700 words * Clean Markdown * One H1 headline followed by an italicized dek * A strong news lede followed by a clear nut graf * Organize the material into a cohesive story rather than following the notes in order * Attribute company-supplied claims * Include limitations without turning the article into a warning label * Use only the supplied information * Avoid canned transitions, promotional language and a concluding summary Reporting notes: * HelioDock makes autonomous warehouse robots. * On Thursday, it launched Night Shift, software that lets operators monitor robot fleets remotely. * Night Shift spent 14 months in a private beta involving nine logistics companies and 310 robots. * HelioDock says beta customers experienced 17% fewer unplanned robot stoppages. The figure has not been independently verified. * The company also announced a $24 million Series B led by North Coast Ventures, bringing its total funding to $39 million. * CEO Mara Velez: “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet.” * Beta customer Aaron Pike of Fulcrum Logistics said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. * When connectivity fails, robots return to their locally configured behavior. Remote control becomes unavailable. * Competing fleet-management products are available from Relay Robotics and YardOS. * Night Shift will cost $90 per robot per month and become generally available September 15.

Qwen: Qwen4 Max: # HelioDock Launches Night Shift, Bringing Remote Oversight to Warehouse Robot Fleets *The software exits a 14-month private beta with nine logistics companies and 310 robots, and reaches general availability September 15 at $90 per robot per month.* HelioDock on Thursday released Night Shift, software that lets warehouse operators monitor and control fleets of autonomous robots from off-site, and announced a $24 million Series B led by North Coast Ventures. The launch adds a recurring software line to a company that has sold autonomous warehouse hardware, priced at $90 per robot per month when Night Shift becomes generally available on September 15. It follows a 14-month private beta involving nine logistics companies and 310 robots, and it puts HelioDock into a fleet-management category already occupied by Relay Robotics and YardOS. Beta feedback suggests the product's value depends heavily on the network connection available at each facility. CEO Mara Velez described the problem in terms of a specific overnight errand. "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. Night Shift is built on the premise that the person responsible for a robot fleet does not need to be standing in the building to see what a stalled machine is doing about it. HelioDock says customers in the beta experienced 17% fewer unplanned robot stoppages. That figure comes from the company and has not been independently verified. The beta's scope — 310 robots across nine logistics operators over 14 months — gives the claim a broader base than a single-site pilot, though the company's number is the only public measure of the software's effect so far. 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. That distinction matters for a product sold on the promise of distance: the sites most likely to be far from an operator are also among the most likely to have thin cellular coverage. The software's behavior during an outage is defined. When connectivity fails, robots return to their locally configured behavior and remote control becomes unavailable. Night Shift, in other words, adds a layer of supervision on top of whatever autonomy the robots already carry; it does not substitute for it. An operator who cannot reach a fleet is left with the same machines and the same local settings that were in place before the link dropped. That boundary shapes where the subscription is likely to earn its cost. In a well-connected distribution center, remote monitoring can plausibly replace trips to the floor. In a rural facility on an unreliable LTE connection, the same subscription buys oversight that disappears at the moments a manager might most want it. HelioDock is not first to the category. Relay Robotics and YardOS both sell fleet-management products, which means Night Shift arrives against installed alternatives rather than into an empty market. HelioDock's distinguishing argument, as Velez framed it, is the elimination of the two-in-the-morning drive — a claim about operator time rather than about robot capability. The $24 million round led by North Coast Ventures brings HelioDock's total funding to $39 million. The company has not detailed how the capital will be allocated. Pricing is structured per robot rather than per site, so a customer's monthly bill rises with the size of the fleet under management. At $90 per robot per month, the cost is recurring and scales in the same direction as the automation it supervises — a calculation each of the nine beta participants will now make with 14 months of their own operating data in hand, and with September 15 as the date the software becomes available to everyone else.
Google: Gemini 4 Argon: # HelioDock Launches Remote Warehouse Robot Software Alongside $24 Million Series B *The autonomous robotics maker is rolling out its Night Shift monitoring platform after a 14-month beta test, entering a competitive fleet-management market.* HelioDock, a maker of autonomous warehouse robots, launched a new software platform on Thursday designed to let human operators monitor and manage robot fleets from afar, while simultaneously announcing $24 million in fresh venture capital. The dual announcement marks an expansion for HelioDock from hardware into subscription-based fleet management, positioning the company against competing fleet-management products already available from Relay Robotics and YardOS. The new software, called Night Shift, aims to reduce the need for on-site human intervention when autonomous machines encounter obstacles during off-hours. “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet,” said HelioDock CEO Mara Velez. To support the rollout, HelioDock secured a $24 million Series B funding round led by North Coast Ventures. The latest investment brings the company’s total funding to $39 million. Night Shift arrives after 14 months in a private beta program that involved nine logistics companies and 310 robots. According to HelioDock, customers who participated in the beta test experienced 17% fewer unplanned robot stoppages. That figure has not been independently verified. The software is priced at $90 per robot per month and is scheduled to become generally available on September 15. By offering remote oversight, HelioDock is attempting to address one of the persistent friction points in warehouse automation: what happens when an autonomous machine encounters an unexpected scenario that its onboard systems cannot resolve. Rather than letting a stalled unit halt operations until a worker arrives on the warehouse floor, Night Shift allows operators to oversee the fleet from a central location. However, the effectiveness of remote monitoring remains tied to the strength of a facility's underlying network infrastructure. Because the software relies on a live link to the machines, network disruptions can limit its utility in certain environments. Aaron Pike of Fulcrum Logistics, one of the nine logistics companies that participated in the private beta, said the Night Shift system worked well in facilities with strong connectivity. However, Pike noted that the software became less useful during rural LTE outages. HelioDock designed the system with a local fallback for those scenarios. When connectivity fails, the robots automatically return to their locally configured behavior, though remote control becomes unavailable to off-site operators until the connection is restored. For logistics operators weighing the $90 monthly per-robot fee against alternative platforms from YardOS and Relay Robotics, the calculus will likely depend on both the reliability of their facility networks and the frequency of late-night robot stoppages. Ahead of the September 15 general release, HelioDock is betting its newly expanded war chest that warehouse managers are ready to pay for the ability to troubleshoot their fleets from home.

OpenAI: GPT-6 Luna: Model A better meets the requested 550–700-word length and delivers a more developed, cohesive article with a clear lede and nut graf. Model B is concise and largely accurate, but falls well short of the required length; Model A does add some unsupported interpretation. (Second judge pass, order swapped — scores are the average of both: Model A is a more cohesive article and appears to meet the 550–700-word requirement, while Model B falls well short of the minimum. Model A adds some unsupported | Anthropic: Claude Fable 5.1: Model A lands in the 550–700 word range, reorga…

Controlled wit

Write a 350–450-word technology news story in Markdown. The subject invites humor, but the reporting should remain serious. Use no more than one dry line of original humor. Do not mock customers or turn the story into a comedy routine. Reporting: * QueueZero launched an AI meeting assistant called Proxy on Wednesday. * Proxy can attend a meeting, produce notes and assign follow-up work to software agents. * 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 assigned to QueueZero CEO Evan Wu: “Stop talking” and “Send pricing sheet.” * Wu laughed, completed the second task and deleted the first. * QueueZero says Proxy requires confirmation before an external message is sent. * Internal tasks can be created automatically. * Administrators can disable automatic task creation. * Proxy costs $30 per user per month. * It enters public beta next week.

Qwen: Qwen4 Max: # QueueZero's AI Meeting Assistant Turns an Audience Aside Into a Task for Its CEO QueueZero on Wednesday launched Proxy, an AI meeting assistant that can attend a meeting, produce notes and assign follow-up work to software agents. Within minutes of its debut, the product also illustrated the central design problem facing meeting software that listens for instructions: distinguishing what a participant wants done from what a participant merely said. During the launch demonstration, a customer in the audience remarked, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy interpreted the comment as an instruction and generated two tasks, both assigned to QueueZero chief executive Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. Proxy's note-taking, at least, was accurate. The exchange was brief, but it exposed a question that vendors of agentic workplace software are still working through. A meeting assistant that only transcribes can afford to treat every sentence as content. A system that also creates and routes work has to decide which sentences carry authority — and from whom. An audience aside, a hypothetical, a complaint and a joke can all share the grammar of a request. QueueZero says Proxy draws its clearest boundary at the edge of the organization. The company states that the product requires confirmation before an external message is sent, meaning a follow-up email to a customer or partner cannot leave the building on the assistant's own initiative. Internal tasks, by contrast, can be created automatically, which is the behavior the demonstration surfaced. Administrators can disable automatic task creation entirely, according to the company — a control that shifts Proxy toward a drafting tool that proposes work for human approval rather than one that files it. For teams evaluating the product, that setting is likely to be the most consequential decision in deployment, more so than the quality of the meeting notes themselves. Proxy is priced at $30 per user per month and enters public beta next week. That places it in a competitive band with other AI meeting and productivity tools, and puts the burden on QueueZero to show that agents completing follow-up work justify the premium over transcription alone. The broader test for the category will be less entertaining than Wednesday's demonstration and considerably more repetitive: whether assistants can sit through thousands of ordinary meetings, hear the ambiguous remarks that fill them, and reliably leave the task list alone.
Google: Gemini 4 Argon: # QueueZero’s New AI Meeting Assistant Assigns CEO an Unscripted Task During Launch Enterprise software startup QueueZero unveiled its latest artificial intelligence tool, Proxy, on Wednesday, offering a live demonstration that underscored both the responsiveness and the literal-mindedness of automated workplace agents. Designed to streamline post-meeting workflows, Proxy can attend virtual or in-person meetings, generate comprehensive summary notes, and automatically assign follow-up work to other software agents or human participants. The system’s capabilities were tested in real time during Wednesday’s product reveal. As QueueZero CEO Evan Wu presented the software's architecture, a prospective customer in the audience remarked, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy, which was actively transcribing and analyzing the room's audio for the demonstration, interpreted the audience member's comment as a literal directive. Within seconds, the software generated two distinct action items on the presentation screen, both assigned directly to Wu: “Stop talking” and “Send pricing sheet.” Wu laughed off the interruption, promptly deleted the first task, and executed the second, distributing the pricing documentation to the attendees. While the software’s ability to enforce executive brevity remains untested, its capacity for rapid task capture was evident to the room. The incident highlighted the ongoing challenge technology companies face as they transition generative AI from passive transcription tools into active, autonomous agents capable of initiating workflows. As voice-activated systems become deeply integrated into project management software, distinguishing between casual workplace banter and genuine operational commands remains a significant technical hurdle for developers. Following the demonstration, QueueZero representatives emphasized that the platform includes built-in safeguards to prevent unintended actions from leaving an organization. According to the company, Proxy requires explicit human confirmation before any external messages or documents are sent outside the user's network. Internal tasks, however, are designed to be generated automatically to reduce administrative friction after meetings conclude. For organizations concerned about overly aggressive task generation or cluttered project boards, QueueZero noted that system administrators have the ability to disable automatic task creation entirely, requiring manual approval before any spoken remarks are converted into assignments. QueueZero is positioning Proxy for mid-sized to large enterprises looking to consolidate their transcription and project management software stacks. The tool is priced at $30 per user per month and is scheduled to enter public beta next week.

OpenAI: GPT-6 Luna: Model A stays closer to the supplied reporting while delivering a serious story with restrained wit. Model B adds unsupported specifics about the demonstration and the product, including how the pricing sheet was sent and what kinds of meetings Proxy can attend. (Second judge pass, order swapped — scores are the average of both: Model A is more focused and avoids several unsupported factual additions in Model B, while clearly explaining the product’s task-creation safeguards. Model A does use mo | Anthropic: Claude Fable 5.1: Model A stays tightly within the supplied facts…

Natural human rewrite

Rewrite the draft below as a 300–400-word publication-ready news story in Markdown. Preserve every factual claim. Improve the lede, narrative flow, sentence rhythm and transitions. Remove repetition, stock hype, throat-clearing and mechanical signposting. The finished article should sound like an experienced human technology reporter wrote it. Draft: # Arcwell Launches An Innovative New Platform Arcwell has announced an exciting new platform called Relay. The announcement represents a significant development for the company and for the broader customer service industry. Relay uses artificial intelligence to examine customer support conversations and suggest answers to agents. The platform is important for several reasons. First, it can suggest answers. Second, it can identify customers who may cancel. Third, it can create summaries. These capabilities demonstrate how artificial intelligence continues to transform the customer service landscape. Relay was tested by 12 companies over six months. Arcwell says agents using it closed tickets 14% faster. The company did not release customer-retention results. Relay costs $65 per agent each month. CEO Priya Shah said, “Support agents spend too much time searching for information that their companies already have.” However, the platform also presents potential challenges. Relay requires access to support transcripts and internal help documents. Arcwell says customer data is encrypted and is not used to train shared models. Companies can choose a 30-day or 90-day retention period. In conclusion, Relay represents an important new chapter for Arcwell as it seeks to help companies improve customer service using artificial intelligence.

Qwen: Qwen4 Max: # Arcwell launches Relay, an AI assistant for support agents Arcwell has launched Relay, a platform that reads customer support conversations and suggests replies to the agents handling them. The company's argument is that support work is largely a search problem, and that software can solve it faster than a person can. "Support agents spend too much time searching for information that their companies already have," CEO Priya Shah said. Relay also summarizes conversations and flags customers it judges likely to cancel. That last feature pushes the product past productivity software and into retention work, where the payoff is measured in accounts kept rather than minutes saved. Arcwell tested the platform with 12 companies over six months and says agents using it closed tickets 14% faster. Those numbers are the vendor's own, drawn from a small pilot, and a gain of that size would matter most to teams whose costs are dominated by handling time. Arcwell did not release customer-retention results, which leaves the churn feature's central premise untested in public. Relay costs $65 per agent per month, or $780 per seat each year — pricing that scales with headcount rather than ticket volume, and that a support leader can weigh against the cost of the time it claims to save. The tool's usefulness depends on what it is permitted to read. Relay requires access to support transcripts and internal help documents: together, a detailed record of what customers complain about and what a company knows about its own products. Arcwell says customer data is encrypted and is not used to train shared models, and it lets customers choose a 30-day or 90-day retention period. The shorter window limits how much history accumulates; the longer one gives the system more to work with. Neither setting alters the basic arrangement, in which the vendor's software operates inside a company's support archive. For buyers, ticket closure is the easier half of the pitch to verify, because most support teams already track it. Whether the customers Relay flags actually stay is the harder question. Arcwell is selling the signal without, so far, publishing the outcome — and the six-month pilot that produced the speed figure produced no retention number to accompany it.
Google: Gemini 4 Argon: # Arcwell Rolls Out Relay AI to Speed Up Customer Support Arcwell has launched Relay, an artificial intelligence platform designed to help customer service agents resolve inquiries faster by analyzing live support conversations and recommending responses in real time. Priced at $65 per agent per month, the new software aims to eliminate a common bottleneck in customer support: hunting down answers buried in company files. “Support agents spend too much time searching for information that their companies already have,” said Arcwell CEO Priya Shah. Beyond drafting replies for agents, Relay’s AI automatically generates conversation summaries and scans interactions to flag customers who appear at risk of canceling their service. Prior to the commercial rollout, Arcwell tested the software with 12 companies over a six-month pilot program. According to the company, agents using Relay during the trial closed support tickets 14% faster than they did previously. However, despite pitching the tool's ability to spot flight-risk customers, Arcwell declined to release data showing whether the software actually improved customer retention during the testing phase. Deploying the system also requires businesses to hand over sensitive operational material. To suggest accurate responses and summarize tickets, Relay needs direct access to a company’s customer support transcripts and internal help documentation. To address potential privacy and security concerns, Arcwell says it encrypts all customer data and does not use client information to train shared AI models. Enterprise buyers can also control how long their information stays on Arcwell's servers, with the option to set data-retention limits at either 30 or 90 days. With Relay now generally available, Arcwell is betting that enterprises are willing to navigate those data-access hurdles in exchange for double-digit gains in agent efficiency.

OpenAI: GPT-6 Luna: Model A is within the requested word range and gives the story stronger flow, while Model B is substantially under 300 words. Both add unsupported context, but Model B also presents details such as real-time recommendations and general availability as facts not established in the draft. (Second judge pass, order swapped — scores are the average of both: Model A meets the 300–400-word requirement and has stronger narrative flow, though it adds some interpretation not established by the draft. Mod | Anthropic: Claude Fable 5.1: Model A lands inside the 300–400-word window, p…

Select the strongest accurate lede

Based only on the facts below, select the strongest news lede. Respond with one letter. Facts: * ByteBarge pulled its $12-per-month AI assistant on Tuesday. * Private workspace titles from 4,860 accounts were sent to a publicly accessible analytics endpoint. * The exposure lasted 11 hours. * Message contents were not sent. * The company has found no evidence of malicious access. A. ByteBarge users suffered a massive privacy disaster Tuesday when the company leaked their private messages online. B. ByteBarge, an innovative collaboration company, made the difficult decision to cancel one of its products. 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. D. Thousands of ByteBarge users woke Tuesday to learn that everything they had written was visible to hackers.

Qwen: Qwen4 Max: C
Google: Gemini 4 Argon: C

Scored against the benchmark's reference answer (choice match): Model A was correct, Model B was correct.

Numbers shaped into a story

Write a 600–750-word Markdown article from the benchmark results below. Build the story around the clearest finding. Explain the methodology and limitations where readers need them. Avoid reciting the table row by row or treating small differences as decisive. Benchmark: * Researchers tested five language models on 200 anonymized customer-support tickets. * Each model produced one proposed answer per ticket. * Three experienced support managers graded the answers blindly. * Scores combined factual correctness, usefulness and tone. * Model North: 84.2 average score; $1.90 estimated cost per 1,000 tickets. * Model Vale: 83.8; $0.74 per 1,000 tickets. * Model Ember: 79.1; $0.31 per 1,000 tickets. * Model Kite: 76.4; $0.22 per 1,000 tickets. * Model Harbor: 71.0; $0.08 per 1,000 tickets. * North and Vale’s score difference was not statistically significant. * Vale answered 18% faster than North. * North performed best on billing disputes. * Vale performed best on technical troubleshooting. * All five models performed poorly when a ticket depended on an image attachment. * The test used English-language tickets from one software company. * Researchers did not test agentic actions, tool use or multilingual support.

Qwen: Qwen4 Max: # The Two Best Models Were Indistinguishable. One Cost 2.6 Times More. A blind evaluation of five language models on 200 anonymized customer-support tickets produced a result that should change how teams shop for support AI: the top two finishers were statistically tied on quality, but not on price. Model North scored 84.2 and Model Vale scored 83.8 — a 0.4-point gap that the researchers explicitly report as not statistically significant. North cost an estimated $1.90 per 1,000 tickets. Vale cost $0.74. In other words, the most expensive model in the test bought no measurable quality advantage over a model at roughly 39% of its price. Vale also answered 18% faster. That is the finding worth acting on. Almost everything else in the results is either a modest gradient or a shared weakness. ## How the test worked Each model produced exactly one proposed answer per ticket — no retries, no best-of-three sampling, no human editing. Three experienced support managers graded the answers blindly, meaning they did not know which model wrote what. Scores combined factual correctness, usefulness and tone into a single composite. The blind, multi-rater design is the study's main strength. Support quality is easy to grade generously when you know you're looking at the flagship model, and three graders dilute any one person's preferences. The single-answer-per-ticket rule also reflects a realistic draft-and-review workflow rather than an optimized demo. The composite score is the design's main weakness, and readers should hold it loosely. A model that is factually reliable but curt and a model that is warm but occasionally wrong can land on the same number. For support work, those are very different failure modes: one irritates customers, the other creates liability. The benchmark does not tell us how the leaders split across those dimensions. ## The cost story is real, but smaller than it looks The 2.6x price ratio between North and Vale sounds decisive. Convert it to dollars and it shrinks. At 50,000 tickets a month, North runs about $95 and Vale about $37 — a savings of $58. That is less than the fully loaded cost of a single human-handled ticket at most software companies. So the honest conclusion is not "Vale saves you money." It is "there is no evidence here that North's premium buys quality, so don't pay it by default." Cost efficiency matters most at very high volume or in low-margin operations; at moderate volume, it should probably lose to whatever model handles your ticket mix best. The 18% latency advantage may matter more than the price gap. In live chat, speed is part of the customer experience. In an overnight email queue, it is nearly irrelevant. ## Where the real differences live The category results are more actionable than the leaderboard. North performed best on billing disputes; Vale performed best on technical troubleshooting. Those are genuinely different jobs — one requires careful policy adherence and a defensible tone, the other requires diagnostic reasoning. A team whose volume skews toward troubleshooting has a reason to prefer Vale that has nothing to do with price, and a billing-heavy team has a reason to keep North in the mix despite the tie. Below the leaders, quality drops in a way that tracks price: Ember at 79.1, Kite at 76.4, Harbor at 71.0. The leaders' roughly 5-to-13-point advantage over that group is large enough to take seriously. The 2.7-point gap between Ember and Kite is not, on this evidence — no significance test is reported for that pair, and with 200 tickets and a composite score, it should be treated as noise until replicated. ## The shared blind spot All five models performed poorly when a ticket depended on an image attachment. This is the most operationally important result in the study, and it is not a ranking issue — it is a capability ceiling. Screenshots of error messages, invoices, and UI states are common in software support. Any deployment needs a rule that detects image-dependent tickets and routes them to a human, regardless of which model is drafting. ## What this benchmark cannot tell you The test used English-language tickets from one software company. Neither the ticket mix nor the domain vocabulary generalizes automatically to other industries, languages, or support cultures. The study did not test agentic actions, tool use, or multilingual support — so it says nothing about whether these models can look up an account, issue a refund, or query a knowledge base. That is a large share of real support automation. Two further cautions: the costs are estimates, not invoices, and no inter-rater reliability figures are reported. Blind grading reduces bias but does not eliminate it; graders can sometimes infer a model from its style. **Practical read:** default to Vale, keep North as a candidate for billing-heavy work, and build image-ticket escalation before you build anything else. Then re-run the comparison on your own tickets.
Google: Gemini 4 Argon: # Why Paying Top Dollar for Support AI Might Be a Mistake In the race to automate customer service, companies often assume that getting the highest quality responses requires paying for the most expensive language model. A new benchmark of five leading AI models suggests otherwise: organizations can get top-tier customer support performance for less than half the price of the most expensive option. Researchers recently evaluated five language models—North, Vale, Ember, Kite, and Harbor—using 200 anonymized, real-world customer-support tickets. To ensure an objective evaluation, each model generated a single response per ticket, and three experienced support managers graded the answers blindly. The judges arrived at a composite score based on three criteria: factual correctness, usefulness, and tone. The headline finding centers on the two top performers, North and Vale. On the surface, North took first place with an average score of 84.2, while Vale followed at 83.8. However, researchers noted that this 0.4-point gap is not statistically significant. In practical terms, the two models tied for first place. Their operational profiles, however, were vastly different. Generating responses for 1,000 tickets with North costs an estimated $1.90. Vale achieves the same tier of quality for just $0.74 per 1,000 tickets—a 61% cost reduction. Furthermore, Vale generated its answers 18% faster than North, an important advantage for support teams trying to reduce customer wait times. ## Matching the Model to the Queue While North and Vale tied overall, a closer look at the data shows that neither model is universally superior. Instead, they excelled in different domains. North was the strongest performer when handling sensitive billing disputes, whereas Vale took the lead in technical troubleshooting. For teams operating on tighter budgets, the benchmark also revealed viable mid-tier alternatives. If a company does not need the peak performance of North or Vale, capability scales down alongside price. Models like Ember (scoring 79.1 at $0.31 per 1,000 tickets) and Kite (76.4 at $0.22) offer serviceable drafting capabilities for pennies, while Harbor sits at the bottom of both the quality and cost spectrums (71.0 at $0.08). | Model | Average Score | Est. Cost per 1,000 Tickets | Notable Strength | | :--- | :--- | :--- | :--- | | **North** | 84.2* | $1.90 | Billing disputes | | **Vale** | 83.8* | $0.74 | Technical troubleshooting (18% faster than North) | | **Ember** | 79.1 | $0.31 | Mid-tier cost efficiency | | **Kite** | 76.4 | $0.22 | Budget drafting | | **Harbor** | 71.0 | $0.08 | Ultra-low-cost baseline | *\*Difference between North and Vale is not statistically significant.* Rather than relying on a single model for every customer interaction, support leaders could use these findings to build a routed system. By directing technical questions to Vale, routing complex billing disputes to North, and assigning routine inquiries to a budget-friendly model like Ember, companies could optimize both resolution quality and operating costs. ## Blind Spots and Benchmark Limitations Before overhauling a support pipeline based on these numbers, buyers need to understand what the models failed to do—and what the researchers did not test. Most notably, all five models performed poorly whenever a ticket relied on an image attachment. In software support, customers frequently attach screenshots of error messages or broken user interfaces. Because none of the tested models could reliably handle these visual inputs, human agents—or specialized multimodal workflows—remain essential for image-heavy queues. Additionally, the study’s methodology reflects a narrow slice of the customer-support world. The 200 test tickets were entirely in English and sourced from a single software company. A model that excels at explaining SaaS bugs in English might falter when handling retail returns in Spanish or healthcare inquiries in Mandarin. Finally, this benchmark only evaluated static text generation—one proposed answer per ticket. Modern customer support increasingly relies on "agentic" AI that can take action. The researchers did not test tool use, meaning these scores do not reflect a model's ability to look up an account status via an API, process a refund, or navigate a multi-turn conversation. ## The Takeaway When choosing an AI model for drafting support replies, price is a poor proxy for quality. For most English-language software support teams, Vale offers the best balance of speed, accuracy, and cost. However, because even the best models stumble on screenshots and vary by topic, the smartest deployment strategy isn't just picking one winner—it’s knowing which tickets to hand to which model, and which ones to leave for humans.

OpenAI: GPT-6 Luna: Model A builds a clear, appropriately qualified story around the statistically indistinguishable leaders and the shared image-related weakness. Model B is readable, but its row-by-row table conflicts with the instruction and its specific routing advice goes beyond what the benchmark establishes. (Second judge pass, order swapped — scores are the average of both: Model A builds a sharper story around the statistically tied leaders while giving readers more useful context about the benchmark’s lim | Anthropic: Claude Fable 5.1: Model A builds a sharper, more rigorous story a…

Quotes woven into narrative

Write a 400–500-word Markdown story about the closing of Foundry Hall. Use quotations selectively. Paraphrase when the information matters more than the speaker’s exact phrasing. Avoid stacking quotes or giving every source an equal-sized turn. The story should carry an emotional current while remaining restrained. Facts: * Foundry Hall, a nonprofit startup accelerator in Milwaukee, will close October 31. * It opened in 2018. * It has supported 212 startups, which collectively employ 1,140 people. * Its largest corporate sponsor, Blank River Bank, supplied 38% of its annual budget. * The bank declined to renew its sponsorship. * Foundry Hall has enough money to operate through October but not through another full program. * Six employees will lose their jobs. * Twelve startups currently using its workspace must leave. Quotes: * Executive director Camille Ross: “We kept waiting for the replacement check that never came.” * Founder Dev Malik: “My first investor meeting happened in that building. Before Foundry Hall, I was pitching from the back table of a coffee shop.” * Blank River Bank spokesperson: “The decision reflects a change in our community investment priorities and is unrelated to Foundry Hall’s performance.” * Deputy mayor Susan Cole: “The city is speaking with building owners and local universities about space for the current companies.”

Qwen: Qwen4 Max: # Foundry Hall Will Close on October 31 **MILWAUKEE** — Foundry Hall, the nonprofit accelerator that has helped 212 Milwaukee startups get off the ground since 2018, will close on October 31. The companies that came through it now employ 1,140 people. The organization that helped them employs six. The ending traces back to a single line in the budget. Blank River Bank, Foundry Hall's largest corporate sponsor, supplied 38 percent of its annual funding and declined to renew. A bank spokesperson said the decision reflected a change in community investment priorities and had nothing to do with Foundry Hall's performance. The nonprofit has enough money to operate through October. It does not have enough to run another full program. Executive director Camille Ross said the staff kept the current cohort moving while they looked for someone to close the gap. "We kept waiting for the replacement check that never came." Six employees will lose their jobs. Twelve startups currently working out of the building must be out by the end of the month — founders who now have to solve a problem that has nothing to do with their products, on a schedule set by someone else's fiscal year. Dev Malik, a founder who went through the program, said his first investor meeting happened in that building. "Before Foundry Hall, I was pitching from the back table of a coffee shop," he said. That was the service, in practical terms: a room where a founder could be taken seriously. It was cheap to provide relative to what it produced, and it was never the kind of thing that showed up cleanly in a sponsor's annual report. The report could show 212 companies. It could not show the difference between a back table and a conference room with a door. Deputy mayor Susan Cole said the city is talking with building owners and local universities about space for the twelve companies. Those conversations are continuing. None of them, so far, has produced an address. There is no scandal here, and no failure anyone will point to in a case study. The accelerator did what accelerators are supposed to do, at a scale most cities its size would envy, and then the money moved somewhere else. The bank was within its rights. The city is doing what it can with buildings it does not own. Ross will spend her last weeks helping other people's companies find desks while she and her staff clear out their own. The startups will mostly survive; that is the nature of startups, and it is the argument Foundry Hall spent seven years making. Some of the twelve will land in a university incubator or a sublet above a storefront. A few will not. On the last day of October, someone will turn off the lights and lock the door, and the count will stay at 212.
Google: Gemini 4 Argon: # After Six Years, Milwaukee’s Foundry Hall Prepares to Lock Its Doors The whiteboard walls at Foundry Hall are still covered in market projections and product diagrams, the daily residue of companies trying to will themselves into existence. By the end of the month, the boards will be wiped clean. The Milwaukee nonprofit startup accelerator will close permanently on October 31, bringing a quiet end to a six-year experiment that helped reshape the city’s small-business landscape. The closure leaves 12 early-stage companies scrambling for new office space and eliminates the jobs of the organization’s six employees. Since opening in 2018, Foundry Hall has acted as an incubator for 212 startups. Together, those businesses now employ 1,140 people, a figure local officials often cited as proof that Milwaukee could cultivate its own tech and manufacturing talent rather than losing it to larger Midwest hubs. The organization’s unraveling began earlier this year when its largest corporate backer, Blank River Bank, declined to renew its sponsorship. The bank had supplied 38% of Foundry Hall’s annual budget. A spokesperson for the bank said the withdrawal was not a reflection of the accelerator’s performance, but rather the result of shifting priorities within the institution's community investment strategy. Without that anchor funding, leadership spent months trying to patch the deficit. They found enough capital to keep the lights on through October, but fell short of the reserves required to commit to another full program cycle. “We kept waiting for the replacement check that never came,” said Camille Ross, Foundry Hall’s executive director. For the entrepreneurs who passed through the building, the loss is less about balance sheets and more about the disappearance of a rare gathering place. Early-stage founders often work in isolation, lacking both the capital for commercial leases and the professional networks required to find investors. Foundry Hall offered both, providing subsidized desks alongside access to legal mentors and venture capital. Dev Malik, an alumnus of the program who now runs a growing logistics software firm, remembered the credibility the space gave him in his earliest days. “My first investor meeting happened in that building,” Malik said. “Before Foundry Hall, I was pitching from the back table of a coffee shop.” Now, the final cohort of founders is packing up monitors and prototypes, mid-stride in their development. City officials are attempting to soften the landing for the dozen displaced companies. Deputy Mayor Susan Cole said the administration is currently negotiating with local universities and private landlords to find temporary workspaces for the startups before the October deadline. Still, relocating the current tenants will not replace the institution itself. In the final weeks of October, the daily rhythm at Foundry Hall has shifted from building the future to boxing up the past. The staff is spending its remaining days making introductions for its founders, trying to ensure the companies survive the place that launched them.

OpenAI: GPT-6 Luna: Model A better weaves the supplied facts and selected quotations into a restrained narrative, while Model B adds several unsupported details about the companies, services, local officials, and relocation negotiations. Both meet the requested story format, but Model A is more reliable and more closely aligned with the prompt’s emphasis on selective quoting and restraint. (Second judge pass, order swapped — scores are the average of both: Model A delivers a tighter, more restrained narrative and m | Anthropic: Claude Fable 5.1: Model A integrates quotes more selectively (par…

Exact final-draft structure

Convert the supplied fields into Markdown using these exact rules: * H1 headline * One blank line * Italicized dek * One blank line * First body paragraph * One blank line * H2 section heading * One blank line * Second body paragraph * Preserve all wording and punctuation exactly * Do not use a code fence * Output nothing else Headline: Raincheck raises $8 million for flood-risk software Dek: The Austin startup helps cities decide which drainage problems to address before storms. First body paragraph: Raincheck has raised an $8 million seed round led by Basin Ventures. Section heading: Expansion plans Second body paragraph: The company plans to hire eight engineers.

Qwen: Qwen4 Max: # 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.
Google: Gemini 4 Argon: # 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.

OpenAI: GPT-6 Luna: Model A and Model B follow the required structure and preserve the supplied wording exactly; their outputs are identical. (Second judge pass, order swapped — scores are the average of both: Model B and Model A produce identical output that preserves the wording and punctuation and follows the required Markdown structure. Neither has an advantage.) | Anthropic: Claude Fable 5.1: Model A and Model B produced byte-identical outputs that follow every rule exactly: correct H1, italicized dek, body paragraphs, H2, single blank lines between elements, no code fence, and no extra…

Chronology with narrative momentum

Turn the scrambled reporting below into a 450–600-word Markdown news article. Choose an opening that captures the consequence of the failure, then reconstruct the sequence clearly. Preserve the chronology without producing a minute-by-minute incident log. Give technical details enough context for a general reader. Reporting: * KiteBridge makes expense-management software. * At 8:45 a.m. Tuesday, it deployed version 4.2. * The first customer complaint arrived at 9:12 a.m. * A payment-processor timeout caused KiteBridge’s fallback system to treat retries as new reimbursement requests. * The software generated 13,400 duplicate reimbursements totaling $2.8 million. * KiteBridge suspended automated payments at 10:03 a.m. * Ninety-one corporate customers were affected. * By Wednesday evening, $2.3 million had been recovered. The rest remained pending. * No employees had money withdrawn from their accounts; employers funded the duplicate payments. * An April changelog said version 4.0 had introduced “exactly-once payment protection.” * KiteBridge’s public incident page initially described the problem as delayed reimbursements. * The incident page first disclosed the duplicate payments six hours after automation was suspended. * CEO Lena Ortiz apologized and said the company would publish a technical report.

Qwen: Qwen4 Max: # Software Update Sent $2.8 Million in Duplicate Expense Payments Ninety-one corporate customers of KiteBridge, which makes expense-management software, paid out $2.8 million in duplicate reimbursements Tuesday after a routine release caused the system to treat repeated attempts at a single payment as brand-new requests. Roughly $500,000 of that money had not been recovered as of Wednesday evening. The loss fell on employers, not their workers. No employee had money withdrawn from an account; companies funded every one of the 13,400 duplicate payments the software generated. ## A retry that became a second payment KiteBridge deployed version 4.2 at 8:45 a.m. Tuesday. The failure began in the link to the company's payment processor, the outside service that actually moves money between bank accounts. When a request to that processor timed out — meaning no confirmation of success or failure came back within the expected window — KiteBridge could not tell whether a reimbursement had been completed. Retrying is the normal response to that kind of uncertainty. What makes retrying safe is a guarantee that the second attempt will be recognized as a repeat rather than a fresh instruction. Engineers call this "exactly-once" behavior: a system may transmit the same payment instruction several times, but only one payment should result. That recognition did not happen. KiteBridge's fallback system, the code path that handles interruptions like a processor timeout, treated each retry as a new reimbursement request instead of another attempt at one already underway. The result was a cascade of extra payments drawn from employer funds. The safeguard was supposed to exist. An April changelog announced that version 4.0 had introduced "exactly-once payment protection." Whether that protection broke in version 4.2, or never extended to the fallback path in the first place, is one of the questions the company has not yet answered. ## Shutdown, then a six-hour gap in disclosure The first customer complaint arrived at 9:12 a.m., 27 minutes after the release went out. KiteBridge suspended automated payments at 10:03 a.m., roughly 78 minutes after deployment. The duplicates accumulated during that window. The company's public incident page initially described the problem as delayed reimbursements — close to the opposite of what had actually occurred. It first disclosed the duplicate payments six hours after automation was suspended. Customers relying on that page therefore spent most of the business day without being told that money had already left their accounts more than once, information they needed to begin reconciling payroll and expense ledgers. Recovery has been partial. By Wednesday evening, $2.3 million of the $2.8 million had been clawed back; the rest remained pending. Stopping payments is comparatively quick, while retrieving funds that have already settled in a recipient's account is a slower process, which helps explain why the shutdown ended the bleeding long before the balance was restored. ## What the promised report has to answer Chief executive Lena Ortiz apologized and said KiteBridge would publish a technical report. Three questions will determine how much confidence customers place in it: how a protection advertised in April was ineffective in this release, why the fallback path was able to create new payment requests at all, and why the public description of the incident trailed the shutdown by six hours. For the 91 affected companies, the immediate arithmetic is simpler. Half a million dollars of their money is still out, and the software they bought to reduce expense-processing errors is the reason it left.
Google: Gemini 4 Argon: # Software Flaw at KiteBridge Triggers $2.8 Million in Duplicate Expense Payouts A software update at expense-management provider KiteBridge went awry Tuesday morning, accidentally draining $2.8 million from corporate accounts to fund thousands of duplicate employee reimbursements. The malfunction generated 13,400 redundant payouts across 91 corporate customers before engineers could halt the runaway system. While the error briefly delivered unexpected windfalls to employees— whose personal bank accounts were never debited during the error—their employers were left to unwittingly fund the duplicate transfers. By Wednesday evening, KiteBridge had recovered $2.3 million of the misdirected funds, though roughly $500,000 remained in limbo. The incident began at 8:45 a.m. Tuesday with the routine deployment of KiteBridge version 4.2. Shortly after the rollout, the software encountered a timeout while communicating with an external payment processor. In financial technology, timeouts are common; when a system does not immediately receive confirmation that a transfer succeeded, a backup mechanism typically steps in to retry the connection. However, a flaw in KiteBridge’s fallback system caused it to misinterpret those automated retries. Instead of recognizing them as follow-ups on existing transactions, the software treated every retry as a brand-new reimbursement request, issuing fresh payouts each time. Signs of trouble surfaced quickly. The first customer complaint arrived at 9:12 a.m., less than half an hour after the update went live. For the next 51 minutes, the platform continued to process duplicate transactions until KiteBridge engineers suspended all automated payments at 10:03 a.m. Despite the rapid internal shutdown, customers were kept in the dark about the true nature of the glitch for most of the business day. KiteBridge’s public incident page initially characterized the disruption as mere "delayed reimbursements"—the exact opposite of the rapid-fire overpayments actually occurring behind the scenes. It was not until roughly 4:00 p.m., six hours after payment automation had been suspended, that the company updated its incident page to disclose that the system had issued duplicate payouts. The technical failure is particularly damaging to KiteBridge’s reputation given its recent product claims. According to an April changelog for an earlier update, version 4.0, the company had explicitly touted the introduction of "exactly-once payment protection." That industry-standard safeguard is specifically engineered to ensure that even if network hiccups require a system to send a payment instruction multiple times, the recipient is only paid once. Tuesday’s cascade of errors demonstrated that the safeguard failed when subjected to a processor timeout under the newest version of the software. In the wake of the incident, KiteBridge CEO Lena Ortiz apologized to the affected businesses for the disruption and the misallocated capital. Ortiz pledged that the company would publish a comprehensive technical report detailing how the software bypassed its own safeguards and outlining the steps being taken to prevent future billing misfires, even as efforts continue to claw back the remaining half-million dollars in pending transfers.

OpenAI: GPT-6 Luna: Model A more carefully distinguishes reported facts from unanswered questions, while Model B adds unsupported specifics about what the promised report will contain and uses some overstated language. Both reconstruct the sequence clearly, but Model A’s technical explanation and measured framing are stronger. (Second judge pass, order swapped — scores are the average of both: Model A better meets the requested chronology and gives useful context while distinguishing confirmed facts from questions | Anthropic: Claude Fable 5.1: Model A stays strictly within the sourced facts,…

Matchup powered by OpenRouter.