This newsroom-focused matchup tests the models on accuracy, structure, restraint and storytelling—from choosing a lede to shaping reported notes into publishable copy. The key tension is whether polish and narrative momentum come at the expense of fidelity to the source material.
Qwen: Qwen4 Max is the saved winner, with aggregate scores of **103.0** to MoonshotAI: Kimi K3’s **91.0**. It won 6 tasks to Kimi’s 2, with 4 ties; the editorial assessment is a **clear result (not statistical)** on this saved task sample, not a claim about general capability.
Qwen’s edge came through in several core newsroom tests: selecting the strongest accurate lede, handling the opening of a news story, applying controlled wit, editing Markdown for publication, building chronology and weaving quotes into a narrative. The recurring advantage was restraint—its answers more often stayed close to the supplied reporting instead of adding unsupported specifics.
Kimi had meaningful wins. It produced the stronger startup spotlight and the more faithful natural rewrite, notably preserving that 12 companies tested Relay rather than changing them into customers. That makes the result more than a sweep: Kimi showed better judgment on two assignments where factual precision and fit to the requested form were decisive.
The four ties also matter. The models matched on the benchmark’s numeric Markdown error count, exact final-draft structure and two broader writing tasks, including a story built from reporting notes. Overall, Qwen was the more consistent choice across this particular sample, while Kimi’s wins show where its approach can pay off.
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: # Paying 2.6x More Bought No Measurable Quality
The clearest result from a recent blind evaluation of five language models on customer-support work is a negative one: the most expensive model tested did not produce demonstrably better answers than a model costing about a third as much.
Model North scored 84.2 on a composite quality scale. Model Vale scored 83.8. The researchers report that the 0.4-point gap was not statistically significant — meaning the test cannot distinguish the two on quality. Vale also answered 18% faster and cost $0.74 per 1,000 tickets against North's $1.90.
If you are choosing between those two on the strength of this benchmark, the premium buys nothing the evaluation could detect.
## How the test worked
Researchers drew 200 anonymized support tickets and had each model produce one proposed answer per ticket. Three experienced support managers graded the answers blindly, without knowing which model wrote which response. Scores combined factual correctness, usefulness and tone into a single number.
Two design choices deserve credit. Blind grading removes the brand halo that skews many informal model comparisons, and using support managers rather than crowd workers means the graders actually know what a good reply looks like. Using the same 200 tickets across all five models also makes the comparison paired, which is the right way to detect modest differences.
The composite score is the design's weak point. A single number can hide a trade-off: a warm, confident answer that gets a refund policy subtly wrong may score like a terse, correct one. Teams with strict accuracy requirements would want the correctness subscore, which this summary does not break out.
## The scores, and what they're worth
| Model | Avg. score | Est. cost / 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 |
Below the top pair, quality falls off. Ember and Kite form a middle band roughly five to eight points behind the leaders; Harbor trails by a wider margin. Only the North–Vale comparison came with a significance test, so treat the ordering within the middle band as suggestive rather than settled. The gap between the top pair and Harbor is large enough to be practically meaningful regardless.
One honest caveat on the cost story: at these prices, the absolute stakes are small. A team handling 20,000 tickets a month would spend about $38 on North and about $15 on Vale. That difference is dwarfed by the cost of the human reviewing the draft. Cost efficiency here is a tiebreaker, not a business case — the stronger argument for Vale is that it matches the leader's quality while responding 18% faster.
## Two specialists and one shared blind spot
The category results complicate a simple ranking. North performed best on billing disputes; Vale performed best on technical troubleshooting. That pattern is plausible — billing answers hinge on policy precision, troubleshooting on procedural clarity — but the summary gives no per-category sample sizes or significance tests. With 200 tickets split across categories, these are hints for routing experiments, not proof.
More consequential is the finding that cuts across all five models: every one performed poorly when a ticket depended on an image attachment. Screenshots of error dialogs, photos of invoices, annotated UI captures — none of the models handled them well, and no amount of model selection fixed it.
That is a workflow problem, not a procurement problem. If a meaningful share of your tickets carry images, the highest-value intervention is detecting them and routing them to a human or asking the customer for a text description — not upgrading to a pricier model that fails the same way.
## What this test can't tell you
The scope is narrow by design. Tickets were English-language and came from a single software company, so the results say nothing about other languages, other industries, or support contexts with different regulatory stakes. Costs are estimates, not measured invoices. The report does not give inter-rater agreement, so we can't gauge how much graders diverged. And researchers did not test agentic actions, tool use or multilingual support — which means the benchmark measures drafted replies, not the ability to look up an account, issue a refund or close a loop.
Two hundred tickets is also enough to detect a large quality gap, not a small one. "Not statistically significant" is not "identical."
## The practical read
Start with the cheaper of the two leaders, route image-dependent tickets away from automation entirely, and re-run the comparison on your own ticket mix before committing. This benchmark's most useful contribution is not a winner. It's evidence that price and quality came apart at the top — and that the biggest failure was one no model in the test could solve.
MoonshotAI: Kimi K3: # Two AI models tie at the top of a support-ticket benchmark — but one is 60% cheaper and 18% faster
A new benchmark pitted five language models against 200 real customer-support tickets and produced a result that should matter to any team shopping for an AI drafting tool: the two best models are effectively indistinguishable on quality, yet one costs well under half as much to run and answers noticeably faster.
## How the test worked
Researchers drew 200 anonymized tickets from a software company's support queue and asked each model for one proposed answer per ticket. Three experienced support managers graded every response blind — they did not know which model wrote which answer — using a composite score that combined factual correctness, usefulness and tone. Alongside quality, the team estimated each model's cost at a volume of 1,000 tickets and measured response speed.
That design beats a synthetic quiz in a few ways: the tickets reflect real customer problems, blind grading reduces brand bias, and the three-part score approximates what a support lead actually cares about. It also has limits that matter for how far the conclusions travel — more on those below.
## The headline finding: a dead heat on quality, a rout on cost
Model North earned the top average score, 84.2. Model Vale finished right behind at 83.8 — a gap the researchers found was not statistically significant. In plain terms, three blind graders could not reliably tell the two models' answers apart.
The pair separates cleanly on everything else. Vale's estimated cost is $0.74 per 1,000 tickets against North's $1.90 — a saving of roughly 61%, or put another way, equivalent quality at about two-fifths the price. Vale also produced answers 18% faster, an edge that compounds in a live queue where agents or customers wait on drafts.
One wrinkle is worth noting before declaring a winner. The models showed different specialties: North scored best on billing disputes, Vale on technical troubleshooting. A queue dominated by refunds and invoices might reasonably weight North's edge; a technically heavy queue tilts further toward Vale. Ticket mix, not the headline average, should drive the choice.
## Below the top two, quality falls faster than price
The remaining three models form a distinct second tier: Model Ember averaged 79.1 at $0.31 per 1,000 tickets, Model Kite 76.4 at $0.22, and Model Harbor 71.0 at $0.08. The drop from Vale to Ember alone is nearly five points — more than ten times the North–Vale gap — and Harbor's rock-bottom price comes with a score 13 points off the lead.
A caveat: the results flag statistical significance only for the top pairing, so the gaps further down should be read as indicative rather than definitive. Even so, the pattern is hard to miss — every step down the price ladder costs real quality, and the steepest savings buy the steepest declines.
## A weakness all five shared
Whatever their differences, every model stumbled on the same category: tickets that depended on an image attachment. When a screenshot or photo carried the key information, scores fell across the board. For buyers, the lesson is that text-only performance overstates readiness — an image-aware workflow or human handoff is still necessary for that slice of the queue.
## Read the results with these limits in mind
The benchmark drew its 200 tickets from a single software company, in English, so the findings may not transfer to retail, telecom or healthcare queues, to other languages, or to different ticket mixes. The composite score also blends correctness, usefulness and tone into one number, hiding whether a model's edge came from accuracy or from tone. The test measured drafting only: it did not touch agentic behavior such as issuing refunds, calling account tools, or handling multilingual conversations — the very capabilities many support teams are now deploying. And with one answer per model per ticket, the study says nothing about consistency across repeated attempts.
## The takeaway
At the top of this benchmark, quality has converged. The decision between the two leaders is no longer "which writes better answers?" but "which fits our ticket mix, latency targets and budget?" — and for most queues, the cheaper, faster option will be hard to argue against. Just don't expect any of the five to handle screenshot tickets unsupervised.
OpenAI: GPT-6 Luna: Model B is more concise, appears to meet the 600–750-word limit, and gives the benchmark a clear, readable structure; Model A is thoughtful but likely exceeds the requested length. Model A is slightly more cautious about what a nonsignificant result establishes, while Model B overstates that the top models are indistinguishable and that each price step down costs quality. (Second judge pass, order swapped — scores are the average of both: Model B builds a clear narrative around the statistically | Anthropic: Claude Fable 5.1: Both build the story around the North–Vale tie…
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
MoonshotAI: Kimi K3: 6
Scored against the benchmark's reference answer (numeric match): Model A was correct, Model B was correct.
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
**The trade is the easy part. A two-founder company is betting that small investment firms will pay for help with everything that happens afterward.**
The origin story is deliberately unglamorous. Mina Batra spent seven years in operations at a regional brokerage, where she watched a six-person team spend three days resolving a trade mismatch caused by a copied account number. Nothing blew up. No money vanished. It was simply three days of skilled people reading documents side by side, looking for a wrong digit.
That episode prompted Batra to start exploring whether software could absorb more of that work. She teamed up with Owen Carr, who built reconciliation systems at a large bank — a useful pairing, since the problem is partly about document handling and partly about knowing how ledgers are supposed to agree.
## What the agents actually do
LedgerLark's agents compare trade confirmations, custodian records and internal ledgers, identify discrepancies, and propose corrections that a human approves before anything changes. The product does not execute trades or move money.
That boundary looks like a sensible design choice rather than a limitation to be overcome later. It keeps the software out of the path of irreversible actions and gives a compliance officer something concrete to evaluate. It also relocates the risk: if an agent proposes a plausible-looking correction and an analyst approves it at the end of a long day, the error still lands in the books. How clearly the system surfaces its reasoning, and how it handles cases where it is uncertain, may matter more than how many exceptions it catches.
The underlying insight is reasonable. Reconciliation at small firms often runs on spreadsheets, email threads and institutional memory, while large banks build dedicated systems. Work that is mostly comparing structured-ish documents is a plausible fit for language models — provided the output is checked.
## Early numbers, and their limits
Five firms are paying customers, on annual contracts ranging from $36,000 to $110,000. That spread suggests pricing tied to volume or complexity, and it implies revenue still in the low hundreds of thousands. LedgerLark says one customer reduced unresolved daily exceptions from 70 to 18.
That figure is worth noting and worth holding loosely. It is self-reported, drawn from a single customer, and lacks a stated baseline period or definition of "unresolved." Exception counts can fall for reasons unrelated to software: a quieter trading month, a cleaned-up account master, a new hire. A more convincing version of this claim would show several firms, a longer window, and some accounting for what else changed.
The company has raised a $3.5 million seed round led by Motive Row — enough to support a small team through slow sales cycles, not enough to build a broad platform quickly.
## Where it gets hard
The competition is not hypothetical. Established reconciliation vendors already sell into these firms, have years of edge-case handling, and can add similar features to products customers already trust. Larger institutions often rely on internal bank software, which is difficult to displace and rarely evaluated on features alone. At the low end, the real alternative may be hiring another operations person — a comparison that gets uncomfortable when a contract costs $36,000 a year.
Selling to small firms sidesteps some enterprise procurement pain, but it caps contract sizes and produces fewer reference customers. And financial institutions of any size tend to require lengthy security and procurement reviews. Five closed deals is early evidence that the process can be completed, not yet proof that it can be completed repeatedly and predictably.
The open questions are practical ones. Do the agents generalize across custodians, asset types and messy file formats without heavy hand-holding? Does the exception reduction persist after the first enthusiastic months? Can a small vendor keep pace with the audit and reporting expectations of regulated buyers?
LedgerLark has picked a real, tedious problem and drawn a careful line around what its software is allowed to do. That is a credible place to start. Whether it becomes a durable business depends on unglamorous work of its own — the kind that takes longer than three days.
MoonshotAI: Kimi K3: # Startup Spotlight: LedgerLark Wants AI Agents to Shorten the Back Office's Longest Days
*The startup is betting that software which proposes fixes — but never executes them — can win over cautious operations teams.*
At a regional brokerage, Mina Batra once watched a six-person operations team spend three days resolving a single trade mismatch. The cause, once they found it, was almost absurd: an account number copied incorrectly between systems. Batra, who spent seven years in operations at the firm, kept returning to the episode. Much of the team's work, she noticed, wasn't judgment — it was comparing records that should have agreed and figuring out why they didn't.
That observation became LedgerLark, which Batra co-founded with Owen Carr, who built reconciliation systems at a large bank. The company develops AI agents for post-trade operations at small investment firms — the work of confirming that trades settled as intended and chasing the exceptions when they didn't.
## What the agents actually do
LedgerLark's agents compare three sources — trade confirmations, custodian records, and a firm's internal ledgers — then flag discrepancies and propose corrections for a human to approve. Two boundaries matter: the product does not execute trades, and it does not move money. The software does the hunting and the paperwork; people keep the decision.
That looks deliberate. Reconciliation is structured and document-heavy, a plausible fit for language-based agents. But errors here are expensive, so keeping approval authority with operations staff may make the product easier to buy — and easier to defend to compliance teams.
## Early, but paying
LedgerLark says five firms are paying customers, on annual contracts ranging from $36,000 to $110,000 — a modest book, roughly $180,000 to $550,000 in implied annual contract value, but evidence the product has moved past pilots at a few firms.
The company also says one customer cut its unresolved daily exceptions from 70 to 18 after adopting the software. That figure comes from LedgerLark and hasn't been independently verified, and one customer's results may not generalize. Still, if the agents reliably clear even half of a routine exception queue, the savings would matter for teams the size of Batra's old one.
The company has raised a $3.5 million seed round led by Motive Row.
## Harder questions ahead
LedgerLark is entering an established category. Reconciliation vendors have sold to financial institutions for decades and hold integrations and compliance track records that a five-customer startup can't match. Many institutions also run reconciliation software built in-house, and "we already have something" is a durable objection in enterprise sales. Whether incumbents simply add agent-style features to existing products is an open question. LedgerLark's implicit answer is focus: small investment firms are often underserved by vendors that price for large ones. That may be right, but it hasn't been tested at scale.
Sales cycles may also be slow. Financial institutions typically require lengthy security and procurement reviews before new software touches trade data. LedgerLark's first five customers may have been unusually willing; the next fifty may not be.
And the core promise — agents that propose *correct* corrections — will live or die on accuracy the company hasn't publicly disclosed. How often are proposed fixes rejected? What happens when a wrong one slips past an approver?
## What to watch
The useful signals over the next year will be unglamorous: customer count, renewals, and whether contract values hold as LedgerLark moves beyond friendly early adopters. If Batra's original insight holds — that exception work is more mechanical than it looks — the agents could earn a durable place in the back office. If the exceptions that matter turn out to require real judgment, the software may clear the easy cases while the long days stay long.
OpenAI: GPT-6 Luna: Model B is more likely to meet the 500–650-word limit while covering the required facts and giving competition, sales friction, and unverified traction appropriate weight. Model A is thoughtful and well written, but appears over the word limit and adds a few unsupported details; Model B also has some speculative phrasing, though less of it. (Second judge pass, order swapped — scores are the average of both: Model B better balances the reported facts with clearly framed uncertainty. Model A is en | Anthropic: Claude Fable 5.1: Both pieces are grounded in the supplied report…
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 the support desk
Arcwell has released Relay, a platform that uses artificial intelligence to read customer support conversations and suggest answers to the agents handling them. The pitch is narrow and practical: much of the material agents need to resolve a ticket already exists somewhere inside the company, and finding it is the slow part.
"Support agents spend too much time searching for information that their companies already have," CEO Priya Shah said.
Relay does more than draft replies. It also flags customers who may be about to cancel and produces summaries of support conversations, combining work that typically happens during a ticket with work that happens after it. That mix — retrieval, prediction and record-keeping in one tool — is where Arcwell is making its bet on how support teams will use AI.
The company tested Relay with 12 customers over six months and says agents using it closed tickets 14% faster. Arcwell has not released figures on customer retention.
Relay costs $65 per agent per month, a per-seat price that makes the cost of deployment easy to calculate against the time savings Arcwell claims.
The software's access requirements are broader than its output. To work, Relay needs both support transcripts and a company's internal help documents — the record of what customers said and the documentation of how the business says it operates. Arcwell says customer data is encrypted and is not used to train shared models, and gives customers a choice between 30-day and 90-day retention periods, letting them decide how long these conversations remain in the system.
Those settings will matter to any support organization weighing the trade: handing a vendor a searchable archive of customer complaints in exchange for faster handling times.
The unresolved piece is the churn prediction. Arcwell has quantified the speed gain but not the retention effect, so the claim that Relay can spot a customer heading for the exit rests, for now, on the company's description of what the software does rather than on published evidence of what it prevents. Buyers paying $65 per seat each month will be judging that feature on faith until Arcwell says more.
MoonshotAI: Kimi K3: # Arcwell's Relay drafts answers for support agents—and spots customers who may cancel
Arcwell has launched Relay, an artificial intelligence platform that analyzes customer support conversations and suggests answers to the agents handling them. It's the company's entry into one of the busiest testing grounds for applied AI: customer service.
Drafting replies is only part of the pitch. Relay can also identify customers who may cancel and generate summaries of support interactions, folding three distinct jobs into a single tool.
The platform arrives after a six-month pilot involving 12 companies, and Arcwell is touting one headline result from the trial: agents using Relay closed tickets 14% faster. That is the only hard number the company has shared. It did not release customer-retention results—a conspicuous gap for a product that sells itself partly on catching churn before it happens.
Relay costs $65 per agent per month, so pricing scales with the size of a support team.
CEO Priya Shah says the platform targets a frustration built into support work itself. "Support agents spend too much time searching for information that their companies already have," she said. Relay's answer is to do the searching for them—but doing so requires deep access to the places where that information lives.
To work, Relay must read a company's support transcripts and internal help documents. Handing that kind of material to an outside system raises obvious questions, and Arcwell has answers ready. The company says customer data is encrypted and is not used to train shared models, a concern that looms large for businesses evaluating AI tools that ingest proprietary information. Companies can also choose how long their data is kept, with retention options of 30 or 90 days.
What the pilot doesn't show is whether Relay delivers on its boldest promise. Closing tickets faster is one thing; keeping the customers the system flags from leaving is another. Until Arcwell publishes retention data, that half of the pitch rests on faith rather than evidence.
OpenAI: GPT-6 Luna: Model B more precisely preserves the draft’s claim that 12 companies tested Relay, while Model A changes this to 12 customers and adds more speculative framing. Both meet the requested length and are polished, but Model B is the more faithful news story. (Second judge pass, order swapped — scores are the average of both: Model B preserves the draft’s key facts, including that Relay was tested by 12 companies, and delivers a polished story within the requested length. Model A is also well written | Anthropic: Claude Fable 5.1: Both preserve the facts and land in range, but…
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
MoonshotAI: Kimi K3: **C**
This lede is accurate and complete: it states what happened (the product was pulled), the reason (the data exposure), the scope (4,860 accounts, workspace titles only), and the duration (11 hours)—all supported by the facts.
The others fail:
- **A** falsely claims private *messages* were leaked; only workspace titles were exposed.
- **B** reads like PR spin and buries the actual news (the data exposure).
- **D** falsely claims message contents were visible to "hackers"; contents weren't sent and there's no evidence of malicious access.
Scored against the benchmark's reference answer (choice match): Model A was correct, Model B was incorrect.
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 Left Prompt Excerpts From 1,842 Sessions Reachable Without a Login
*A page meant for internal troubleshooting was publicly accessible for four hours Monday. Cloudnote says it has found no evidence of access beyond the researcher who reported it, and its investigation is not finished.*
Cloudnote left a debugging page containing excerpts of customer prompts accessible without authentication for four hours on Monday, the company said. Logs show the page held material from 1,842 user sessions. Cloudnote removed it 19 minutes after a researcher notified the company.
Each excerpt could include up to 200 characters from a prompt — a fragment of a question or draft rather than a complete text. Full documents, passwords and payment information were stored in separate systems and were not present on the page, according to Cloudnote.
The company says it has found no evidence that anyone other than the researcher accessed the page. That finding is preliminary: Cloudnote has not completed its investigation, and it has not described what its logs would and would not capture. Whether any excerpts were viewed or copied by others remains unresolved.
The 1,842 sessions represent a small share of Cloudnote's 3.2 million registered users, though the significance depends on what those users typed. Cloudnote has not said how many distinct accounts were involved, whether it will notify them, or what caused the page to be exposed.
MoonshotAI: Kimi K3: # Cloudnote Debugging Page Exposed Prompt Excerpts From 1,842 User Sessions for Four Hours
*The note-taking service said it removed the page 19 minutes after a researcher flagged it and has found no evidence anyone else accessed it — but its investigation is not over.*
Cloudnote, a note-taking service with 3.2 million registered users, has disclosed that a debugging page containing excerpts of customer prompts was viewable without a password for four hours on Monday.
Company logs show the page contained excerpts from 1,842 user sessions, each excerpt up to 200 characters of a prompt. The sessions represent a small fraction of Cloudnote's user base, though the company has not said how many individual accounts they correspond to.
Cloudnote said full documents, passwords and payment information are stored in separate systems and were not present on the page. Still, even a 200-character fragment can contain names, business details or other sensitive text a user typed, and the exposure of an internal diagnostic tool raises questions about how such pages are secured.
The company said it removed the page 19 minutes after an outside researcher reported it, and that it has found no evidence anyone other than that researcher accessed the page during the four-hour window. But its investigation is not complete, and it has not said how the page became publicly reachable or whether it will notify the users affected.
OpenAI: GPT-6 Luna: Model B gives a clearer sense of scale by putting the user base alongside the exposure, and it cleanly separates logged facts from the unfinished investigation. Model A is more restrained, but its claim that the excerpts were fragments rather than complete text is not established by the reporting. (Second judge pass, order swapped — scores are the average of both: Model B better meets the requested length in its paragraphs while covering the exposure, scope, response and unresolved questions. Mo | Anthropic: Claude Fable 5.1: Model A sticks strictly to the supplied reporti…
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 Proxy Attended Its Own Launch and Assigned the CEO a Task
QueueZero on Wednesday launched Proxy, an AI meeting assistant that can attend a meeting, produce notes and assign follow-up work to software agents. The demonstration also delivered an unplanned illustration of how actually the system listens.
The product sits in a fast-growing category of tools that aim past transcription and toward execution: not just recording what was said, but converting it into work items and handing them to agents that can carry them out. QueueZero's pitch is that the gap between a meeting ending and something happening afterward is where most organizational momentum is lost.
Roughly midway through the launch presentation, a customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet."
Proxy parsed the remark as an instruction. It created two tasks and assigned both to QueueZero chief executive Evan Wu: "Stop talking" and "Send pricing sheet."
Wu laughed, completed the second task and deleted the first. The moment was a compact demonstration of the distance between hearing a sentence and understanding one.
## Where the guardrails sit
QueueZero says Proxy requires confirmation before an external message is sent. That boundary matters here: a misread comment can generate an internal record, but it does not, on the company's account, result in a customer-facing email going out on its own.
Internal tasks are a different case. Those can be created automatically, which is what happened onstage. Administrators can disable automatic task creation entirely, a control that enterprise buyers are likely to examine closely before deploying the assistant across recurring meetings with customers, partners or regulators.
The distinction reflects a broader design problem for meeting agents. Conference rooms and video calls are full of speech that is not directive — jokes, hypotheticals, complaints, comments about the meeting itself. A system that treats every sentence containing a verb and a name as a work order will generate tasks nobody requested, and the burden of cleaning them up falls on the humans the tool was supposed to relieve.
## Pricing and availability
Proxy costs $30 per user per month and enters public beta next week. QueueZero has not said how long the beta will run or what limits will apply during it.
For now, the company has a launch story that spread on its own, and a clearer view than most vendors get of the first question buyers will ask: what, exactly, does it think it heard?
MoonshotAI: Kimi K3: # QueueZero Launches Proxy, an AI Assistant That Attends Meetings and Assigns the Follow-Up Work
QueueZero on Wednesday launched Proxy, an artificial intelligence assistant that can attend meetings, produce notes and assign follow-up work to software agents — a promise the company's launch demonstration put to an unexpected test.
During the on-stage demo, 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. Within moments, it had created two tasks assigned to QueueZero chief executive Evan Wu: "Stop talking" and "Send pricing sheet."
Wu laughed, completed the second task and deleted the first. Whether Proxy logged the deletion as "completed" or "insubordination," the company did not say.
The exchange drew laughs in the room, but it illustrates the central challenge of so-called agentic software: systems built to act on natural language must distinguish an instruction from an aside, a joke or ambient conversation. On Wednesday, Proxy failed that test in the least damaging way possible. No message left the building; the chief executive simply received a to-do list.
QueueZero said Proxy requires user confirmation before any external message is sent, describing the safeguard as a firm boundary. Internal tasks, however, can be created automatically — which is precisely what happened on stage. Administrators can disable automatic task creation entirely, the company said.
That tiered permission model is QueueZero's answer to the trust problem slowing enterprise adoption of autonomous agents. The company is betting customers will accept an assistant that occasionally over-listens if it cannot act externally without approval, and if administrators retain the ability to switch the behavior off.
The episode may yet work in QueueZero's favor. Proxy exhibited the two behaviors enterprise buyers worry about most — over-interpreting speech and creating work unprompted — but both failures were visible, reversible and contained, and the pricing sheet did go out promptly.
Proxy enters a crowded market for AI meeting tools. Microsoft, Zoom and Otter.ai already offer transcription and summarization at scale. QueueZero's differentiator is execution: rather than stopping at a record of what was said, Proxy converts discussion into assigned, trackable work, handing tasks to software agents that can draft documents, file tickets and schedule follow-ups.
Proxy costs $30 per user per month and enters public beta next week. QueueZero said the beta will ship with the administrative controls shown at launch, including the ability to disable automatic task creation.
OpenAI: GPT-6 Luna: Model A stays closer to the supplied facts and maintains a serious news tone, while Model B adds unsupported details, including that the pricing sheet was sent, and uses more than one comic aside. Both are well-structured and appear to meet the requested length, but Model A shows better restraint. (Second judge pass, order swapped — scores are the average of both: Model A better respects the limit on original humor and stays closer to the supplied reporting, while Model B adds several unsupporte | Anthropic: Claude Fable 5.1: Model A stays closer to the supplied facts and…
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, Taking Warehouse Robot Supervision Off Site
*The $90-per-robot monthly service reaches general availability September 15 after a 14-month beta, and arrives alongside a $24 million Series B led by North Coast Ventures.*
HelioDock on Thursday launched Night Shift, software that lets warehouse operators monitor fleets of autonomous robots without being in the building, and announced a $24 million Series B round led by North Coast Ventures.
The release puts a price on a problem that has followed warehouse automation into the overnight hours: when a machine stops, someone still has to deal with it. Night Shift costs $90 per robot per month and becomes generally available September 15, following 14 months of private testing with nine logistics companies and 310 robots. HelioDock, which builds the robots itself, is stepping into a category where Relay Robotics and YardOS already sell fleet-management products. Its case rests on a company-reported reduction in stoppages and on a network connection that some facilities cannot reliably provide.
Chief executive Mara Velez described the intended customer in specific terms. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," she said.
The beta is the evidence HelioDock offers that the trip can be avoided. Over 14 months, nine logistics companies ran Night Shift across 310 robots, according to the company. HelioDock says those customers experienced 17% fewer unplanned robot stoppages. That figure is the company's own and has not been independently verified. It is also the number that matters most to the pitch: unplanned stoppages are precisely the events that pull a manager back into a building at night, so the software's value is measured less in dashboards than in interventions that never had to happen in person.
Aaron Pike of Fulcrum Logistics, a beta participant, said the system worked well in facilities with strong connectivity. It became less useful, he said, during LTE outages at rural sites.
That distinction is built into how the product behaves. Night Shift's remote functions depend on a live link between the operator and the machines. When connectivity fails, robots return to their locally configured behavior and remote control becomes unavailable. The fleet continues on whatever logic was set up on the floor; the distant supervisor loses the ability to intervene. In a rural warehouse during an outage, the drive across town can still be required — the scenario Velez used to introduce the product.
For prospective buyers, that makes site connectivity part of the software evaluation rather than an afterthought. A facility with dependable coverage gets the full remote-monitoring proposition. A facility relying on rural LTE gets it intermittently, and only when the network cooperates. The fallback behavior keeps robots running, but it is a local safety net, not a substitute for the remote supervision HelioDock is selling.
HelioDock is entering an occupied field. Relay Robotics and YardOS both offer fleet-management products, which means operators evaluating Night Shift are often comparing it against tooling they already have rather than against no remote visibility at all. The 14-month beta and the stoppage figure are the material HelioDock has put forward to make that comparison.
The Series B brings HelioDock's total funding to $39 million, with North Coast Ventures leading the round. The money arrives as the company adds a recurring software line to a business previously defined by the robots it manufactures.
Pricing scales with fleet size rather than floor space or facility count. Applied to the 310 robots that participated in the beta, the $90 per-robot monthly rate would come to $27,900 a month, or roughly $335,000 a year — a figure that will look different to an operator running 30 machines than to one running several hundred. Night Shift opens to all customers on September 15.
MoonshotAI: Kimi K3: # HelioDock Launches Night Shift, Bringing Remote Oversight to Warehouse Robot Fleets
*The autonomous-robot maker paired the software debut with a $24 million Series B led by North Coast Ventures, and set general availability for September 15 at $90 per robot per month.*
HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely, and announced a $24 million funding round alongside the debut.
The launch carries HelioDock beyond building robots and into managing them. The company is betting that as warehouses run larger fleets for longer hours, operators will pay a per-robot subscription to supervise machines without being in the building — a pitch aimed most directly at overnight shifts, when a stalled robot would otherwise mean sending someone to the facility.
"A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," said Mara Velez, HelioDock's chief executive.
Her scenario is the one Night Shift is built around. Through the software, operators can check on their fleets from off-site and, as long as the network connection holds, exercise remote control over individual machines. In the company's framing, an incident that once required a manager's physical presence becomes a problem that can be assessed and, where possible, resolved from a screen.
Night Shift will become generally available September 15, priced at $90 per robot per month. At that rate, a facility running 50 robots would pay $4,500 a month for the service. The subscription gives HelioDock a recurring revenue stream that scales with the size of each customer's fleet, layering a software business on top of its robot hardware.
Before Thursday's release, Night Shift spent 14 months in a private beta involving nine logistics companies and 310 robots — an average of roughly 34 robots per operator. HelioDock says customers in that beta experienced 17 percent fewer unplanned robot stoppages, the same kind of unexpected halt Velez described in her remarks. The company is presenting that reduction as evidence of the product's value, but the figure comes from HelioDock itself and has not been independently verified.
The beta also surfaced the product's central constraint: its usefulness rises and falls with the quality of a warehouse's network connection. Aaron Pike of Fulcrum Logistics, one of the nine beta customers, said the system worked well in facilities with strong connectivity but became less useful during LTE outages in rural areas.
When connectivity fails, robots return to their locally configured behavior, and remote control becomes unavailable. In practice, a manager monitoring a site from a distance loses both visibility and the ability to intervene during an outage, while the machines on the floor carry on under their local settings. Pike's account draws a line around where the product delivers its full value: well-connected facilities rather than sites at the edge of cellular coverage.
The funding announced Thursday is a Series B led by North Coast Ventures and brings HelioDock's total funding to $39 million. The capital arrives as the company shifts from a beta measured in hundreds of robots to open sales of a subscription product.
HelioDock will not have the category to itself. Relay Robotics and YardOS already sell competing fleet-management products, so Night Shift enters a market where operators weighing remote-oversight tools have established alternatives. HelioDock's approach pairs the software with its own robots, offering fleet oversight from the same company that supplies the machines.
OpenAI: GPT-6 Luna: Model A is more cohesive and better handles the connectivity limitation while keeping the company’s unverified stoppage claim attributed. Model B adds less-supported claims, including that operators lose visibility during outages and an average robot count per operator. (Second judge pass, order swapped — scores are the average of both: Model A gives the launch a clearer news frame and integrates the beta evidence, connectivity limitation, and competing products into a more cohesive story. Both | Anthropic: Claude Fable 5.1: Both hit the length, structure and attribution r…
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 Prepare for Floods
*The Austin startup maps drainage systems and pairs them with rainfall forecasts, telling local governments where to send crews before a storm hits.*
Raincheck, an Austin company that builds flood-risk software for local governments, has raised $8 million in seed funding led by Basin Ventures. Existing investor Long Mile also participated, 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 the output to decide where to send maintenance crews ahead of a storm. Raincheck publishes [a description of its forecasting methodology](https://example.com/raincheck-methodology).
The company currently works with seven cities in Texas and Louisiana. Raincheck declined to identify them, saying some of the contracts have not been publicly announced.
Its forecasts are recommendations, not orders: emergency managers retain control over crew deployments and public warnings.
CEO Ana Flores said the product is aimed at a decision cities struggle to make under time pressure:
> "Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives."
Raincheck competes with other flood-planning vendors, including HydroMap, as well as municipal engineering contractors. The company said the new funding will help it hire eight engineers and expand into Florida.
MoonshotAI: Kimi K3: # Raincheck Raises $8 Million to Help Cities Prepare for Floods
*The Austin startup's seed round, led by Basin Ventures, brings its total funding to $10.5 million as it plans new hires and an expansion into Florida.*
Raincheck, a startup that builds flood-risk software for local governments, has raised $8 million in seed funding, the company announced today.
The Austin-based company maps drainage systems and combines those maps with rainfall forecasts — a process described in [the company's published methodology](https://example.com/raincheck-methodology) — so cities can decide where to send maintenance crews before a storm.
The round was led by Basin Ventures, with participation from existing investor Long Mile. Raincheck has raised $10.5 million to date.
Raincheck CEO Ana Flores said:
> "Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives."
Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced.
The company's forecasts are recommendations; emergency managers retain control over deployments and public warnings.
Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida.
OpenAI: GPT-6 Luna: Model A preserves the reporting while giving the story a slightly tighter flow and avoiding Model B’s repetition of the funding details. Both follow the requested format and handle the extended statement and link appropriately. (Second judge pass, order swapped — scores are the average of both: Model B preserves the announcement timing and presents the software’s role without implying it makes deployment decisions. Model A is well structured, but its dek overstates the software’s authority, and | Anthropic: Claude Fable 5.1: Model A delivers a tighter, more professionally…
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.
MoonshotAI: Kimi K3: # Raincheck raises $8 million for flood-risk software
*The Austin startup helps cities decide which drainage problems to address before storms.*
Raincheck has raised an $8 million seed round led by Basin Ventures.
## Expansion plans
The company plans to hire eight engineers.
OpenAI: GPT-6 Luna: Model A and Model B both follow the required Markdown structure and preserve all supplied wording and punctuation. Their outputs are identical, so neither is better. (Second judge pass, order swapped — scores are the average of both: Model B and Model A both follow the required structure and preserve all wording and punctuation exactly. Their outputs are identical and equally well written.) | Anthropic: Claude Fable 5.1: Model A and Model B produced identical, fully compliant Markdown: correct H1, italicized dek, exact blank-line spacing, H2 heading, verbatim wording and p…
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: # Faulty Update Sent $2.8 Million in Duplicate Reimbursements to KiteBridge Customers
Ninety-one companies spent this week trying to claw back money their own expense software paid out twice. A Tuesday morning update to KiteBridge, which sells expense-management software to corporate customers, generated 13,400 duplicate reimbursements totaling $2.8 million. Employers funded every one of them. By Wednesday evening, $2.3 million had been recovered, leaving roughly $500,000 still pending.
The sequence was short. KiteBridge deployed version 4.2 at 8:45 a.m. Tuesday. The first customer complaint arrived 27 minutes later, at 9:12 a.m. The company suspended automated payments at 10:03 a.m., roughly 80 minutes after the release went live — but by then the duplicate transfers had already been issued.
## How a retry became a second payment
The trigger was a timeout at a payment processor: KiteBridge sent payment instructions and did not get a timely answer back. A timeout is an ambiguous result, not a failed one. The payment may have gone through anyway, which is why software handling money is normally built to check before asking again — typically by tagging each payment with a unique identifier so that a repeated request is recognized as the same payment and ignored.
KiteBridge's fallback system, the code path that takes over when the normal one stalls, did not do that. It treated retries as brand-new reimbursement requests. Each attempt to resolve an uncertain payment created another payment instead.
That behavior sits awkwardly beside the company's own release notes. An April changelog announced that version 4.0 had introduced "exactly-once payment protection" — industry shorthand for a guarantee that a payment executes one time regardless of how many times the request is repeated. Whether that guarantee extended to the fallback path is among the questions KiteBridge's promised technical report will have to answer.
## Customers were told payments were late, not doubled
For most of Tuesday, the public account of the incident did not match what had actually happened. KiteBridge's incident page initially described the problem as delayed reimbursements — money arriving too slowly, rather than money arriving twice.
The page first disclosed the duplicate payments six hours after automated payments were suspended, putting that admission in the late afternoon. Customers relying on the status page for much of the business day would have understood the incident as a delay.
The direction of the error matters for who bore it. No employees had money withdrawn from their accounts. The duplicate credits went outward at employer expense, which means the shortfall sat on company books while the recovery effort worked in reverse — retrieving funds already disbursed rather than restoring money that had been taken.
## Recovery incomplete, explanation promised
As of Wednesday evening, about 82 percent of the $2.8 million had been recovered. The remainder was still pending, with no public timeline for completion.
Chief executive Lena Ortiz apologized for the incident and said the company would publish a technical report. That report is likely to be scrutinized less for the timeout itself — processors time out routinely, and handling them is a basic requirement of payment software — than for why a fallback routine was able to bypass a safeguard the company had advertised seven months earlier, and why the first public description of the failure pointed customers in the wrong direction.
MoonshotAI: Kimi K3: # KiteBridge Update Turns Payment Retries Into $2.8 Million in Duplicate Reimbursements
*The expense-software maker says 91 corporate customers were affected and most of the money has been recovered — but its own changelog promised this kind of failure couldn't happen.*
A routine software update at expense-management company KiteBridge went wrong on Tuesday morning, turning ordinary payment retries into more than 13,400 duplicate reimbursements worth $2.8 million. The money did not come from employees awaiting repayment; it was drawn from the accounts of 91 corporate customers, some of whose finance teams spent the day discovering overpayments in their own ledgers.
KiteBridge deployed version 4.2 of its platform at 8:45 a.m. Tuesday. The first customer complaint reached the company at 9:12 a.m., and by 10:03 a.m. it had suspended automated payments altogether. In the roughly 80 minutes between deployment and shutdown, the flaw multiplied individual reimbursements into thousands of extra payments.
## What went wrong
The failure started with a timeout at one of KiteBridge's payment processors — the third-party services that actually move the money. When a payment request times out, the sender cannot tell whether the transaction went through or died in transit, so modern systems are built to retry automatically. The safety requirement is that those retries be recognized as the *same* transaction, a property engineers call idempotency. KiteBridge's fallback system did the opposite: it treated each retry as a brand-new reimbursement request and queued a fresh payment every time.
That is an awkward outcome for a company that told customers the problem was already solved. In an April changelog, KiteBridge said version 4.0 had introduced "exactly-once payment protection" — industry shorthand for the guarantee that a transaction is processed a single time, no matter how often the underlying request is repeated. Whatever form that protection took, it did not survive Tuesday's timeout.
Customers also learned the full scope slowly. KiteBridge's public incident page initially described the issue as delayed reimbursements, and only disclosed the duplicate payments roughly six hours after automation was suspended — midafternoon, by which point many corporate treasurers had spotted the anomalies themselves.
## Recovery and fallout
By Wednesday evening, the company said it had recovered **$2.3 million** of the **$2.8 million**, with the remaining $500,000 still pending. Because the duplicate payments were funded by employers rather than deducted from workers, no employees saw money leave their personal accounts; the recovery process runs through corporate accounts and, where needed, reversals initiated with the payment processors.
In a statement, chief executive **Lena Ortiz** apologized to customers and said KiteBridge would publish a technical report on the incident. That report is likely to face two pointed questions: why the "exactly-once" safeguards advertised in April failed under a routine timeout, and why the company's public account of the incident lagged hours behind what it already knew.
OpenAI: GPT-6 Luna: Model A better meets the requested length while building a clear chronology and explaining the technical failure for general readers. Model B is also readable, but adds unsupported details about recovery mechanics and customer discovery; Model A’s claim that the April changelog was seven months earlier is likewise unsupported. (Second judge pass, order swapped — scores are the average of both: Model A better balances a clear chronology with accessible technical explanation and stronger narrative | Anthropic: Claude Fable 5.1: Both deliver well-structured, in-range articles…
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 to Close Oct. 31, Ending Seven Years of Backing Milwaukee Startups
Foundry Hall, the nonprofit accelerator that has supported 212 Milwaukee startups since it opened in 2018, will close on October 31. The companies it helped launch now employ 1,140 people — a figure its staff has long treated as the clearest argument for the organization's existence.
The immediate cause is a sponsorship that will not be renewed. Blank River Bank, Foundry Hall's largest corporate sponsor, supplied 38% of its annual budget. A bank spokesperson said the decision reflected a change in community investment priorities and had nothing to do with Foundry Hall's performance.
That distinction offered little comfort inside the building. Executive director Camille Ross said the shortfall left a gap no other sponsor stepped forward to fill.
"We kept waiting for the replacement check that never came."
Foundry Hall has enough money to operate through October. It does not have enough to run another full program, which means the closing date is less a choice than an arithmetic result.
Six employees will lose their jobs. Twelve startups currently working out of the accelerator's space must be somewhere else by the end of the month.
For founders who passed through, the loss is not abstract. Dev Malik said his first investor meeting happened in that building. Before Foundry Hall, he had been pitching from the back table of a coffee shop. Moving a founder from one setting to the other — from an improvised workspace to a room where someone had agreed to listen — was the small, repeatable transaction the accelerator was built to perform, 212 times over seven years.
Deputy mayor Susan Cole said the city is talking with building owners and local universities about space for the companies now in residence. No replacement has been announced. The twelve startups face a move whether or not those conversations produce a door with their names on it.
Nonprofit accelerators tend to be measured by their successes: the funding rounds, the headcounts, the companies that outgrow the space and leave. Foundry Hall's numbers on those terms were respectable, and they were not enough. A single sponsor's decision to redirect its giving outweighed all of them, because 38% of a budget cannot be replaced by a good record.
The staff will spend the remaining weeks on the unglamorous work of closing: helping companies find desks, writing references, deciding what happens to equipment and files. Ross and her colleagues will do this while looking for their own next jobs.
On November 1, the 1,140 people employed by Foundry Hall's alumni will still go to work. Most of them will not know that the organization that helped start their employers no longer exists. That was always the idea — that the companies would become the durable thing, and the accelerator would fade into their early history.
It is a version of success. It is not the one anyone at Foundry Hall wanted this October.
MoonshotAI: Kimi K3: # Foundry Hall to Close, Leaving a Gap in Milwaukee's Startup Scene
Foundry Hall, the nonprofit accelerator that has anchored Milwaukee's early-stage startup community since 2018, will close on October 31 — undone not by failure, but by the loss of a single check. The announcement ends months of uncertainty for the founders and staff inside its walls.
Blank River Bank, the organization's largest corporate sponsor, declined to renew its sponsorship this year. The bank had supplied 38 percent of Foundry Hall's annual budget, and despite months of conversations with other funders, no replacement came through. Executive director Camille Ross said the accelerator has enough money to operate through October but cannot finance another full program. She spent much of the past year, she said, trying to close the gap.
"We kept waiting for the replacement check that never came," she said.
The bank, in a written statement, said the decision reflected a change in its community investment priorities and was unrelated to Foundry Hall's performance.
The arrangement was, in hindsight, a vulnerability as much as a lifeline. Accelerators of Foundry Hall's kind tend to run on renewable, year-to-year commitments rather than deep reserves, and Foundry Hall's largest commitment came from one institution. When that institution changed course, there was little cushion.
By most measures, the organization's performance was substantial. Over seven years, the accelerator supported 212 startups, which collectively employ about 1,140 people. Its offerings were unglamorous and durable: affordable workspace, introductions to investors, and a program structure that kept founders at their tables long after business hours.
The closure carries a cost beyond the balance sheet. Six employees will lose their jobs. Twelve startups currently working in the building must leave by month's end — an unplanned exercise in finding leases, moving teams, and protecting momentum. Deputy Mayor Susan Cole said the city is speaking with building owners and local universities about space for those companies, though nothing has been finalized.
For founders who came through the program, the loss is harder to itemize. Dev Malik built his company at Foundry Hall and still traces its beginning to rooms inside the building.
"My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop."
The end arrives without scandal or spectacle — just arithmetic. One sponsor carried more than a third of the budget; when that support withdrew, the math stopped working. Ross said her staff will spend its final weeks helping the current startups land somewhere new. After that, the building goes quiet — a space that, for seven years, rarely was.
OpenAI: GPT-6 Luna: Model B better balances the reported facts with a restrained emotional arc, while Model A adds more unsupported claims, especially about staff plans and what alumni will know after the closure. Both use quotations selectively and fit the requested form, though Model B also includes a few details not established by the prompt. (Second judge pass, order swapped — scores are the average of both: Model A better weaves selective quotations and paraphrase into a restrained narrative with a strong emot | Anthropic: Claude Fable 5.1: Model A keeps closer to the supplied facts, use…