The Salad Story That Wasn’t in the File

The assigned theme promises cyclospora, U.S. lettuce, and the sort of consumer-facing food-safety guidance that should be boring enough to print on a refrigerator. The supplied evidence, however, contains no produce recall, regulatory notice, epidemiological investigation, case count, farm, distributor, or confirmed Cyclospora finding. It contains software benchmarks, a terminal application, a glacier map, and a security contact file. A salad outbreak has apparently been replaced by a packet of nerds arguing with their tools.

That is not a trivial clerical inconvenience. In food safety, the difference between a confirmed contamination finding and a plausible-sounding narrative is the difference between useful advice and an internet panic wearing a lab coat. There is no evidence here for a consumer to avoid any lettuce, discard any product, seek medical care, or infer a U.S. outbreak. The responsible practical guidance is therefore unusually short: do not make food decisions on the basis of a topic label that the underlying material fails to support.

The broader lesson is less glamorous but more durable. Risk communication must preserve the chain from claim to evidence, especially when the claim is likely to alter behavior. A system that can confidently produce an outbreak bulletin from irrelevant material is not being helpful; it is merely automating the ancient human talent of seeing meaning in the wrong pile of documents. The rest of this issue is about that same problem in a different industry: how measures, interfaces, and impressive counters can conceal what is actually happening.

A Quesma benchmark report tested RTK, a tool that filters terminal output before an AI coding agent reads it. Its marketing-adjacent appeal is obvious: fewer tokens! Lower bills! A future in which command-line detritus is compressed into a tasteful artisanal reduction. But the authors found that displayed output savings did not reliably translate into lower spending in their tested configurations.

Their central distinction is worth carrying into every metrics-heavy product demo. RTK’s reported gain is derived from removed command-output bytes divided by four, rather than from billed tokens or money. In the report’s task-level comparison, Fable was 1% more expensive with RTK, a difference the authors say was not clearly distinguishable from zero; DeepSeek was 17% more expensive on average. For DeepSeek, compressed terminal output coincided with more prompt tokens and 18% more turns overall. Saving text in one corner of a workflow can alter the agent’s behavior everywhere else.

This is not a verdict on RTK in all settings. The tests covered particular tool versions, models, routes, and Terminal-Bench 2.1 tasks; one costly failure involved a bug fixed after the runs. It is, however, a healthy warning against treating a dashboard’s preferred numeral as economic reality. The project’s own README puts it with admirable restraint: “RTK cuts up to 90% of the bash output your agent reads. […] it is not the same as cutting your bill by 90%. ” In both food tracing and software procurement, the denominator is where the bodies are buried—metaphorically, one hopes.

The same measurement problem appears after the agent has shipped something that passes its tests. In a post on measuring code sloppiness, Earendil’s Sebastian argues that functional correctness does not settle whether code is comprehensible, cohesive, or fit to survive its next maintainer. This is the software equivalent of a package clearing one inspection while the supply chain remains a black box: a narrow test may be real, but it is not the whole system.

Using SlopCodeBench-derived measures, the post reports average verbosity of 0.33 for agent-generated code versus 0.15 for established repositories, and average erosion of 0.68 versus 0.31. Those numbers are not final scientific tablets: the underlying comparisons and methodology are not independently reproduced in the supplied material, and the verbosity measure partly uses handcrafted heuristics. Still, they target a real operational concern. An agent can generate a locally valid patch at machine speed, while the accumulated abstractions, duplication, and complicated functions make future change slower, stranger, and more expensive.

The author’s caution is also the important part: optimizing any one proxy invites Goodhart’s law to arrive with a clipboard. Reducing lines of code can be useful until teams begin worshipping brevity; asking an LLM to grade its own work merely appoints the defendant as the building inspector. “Asking LLMs to judge the code they write is not a substitute for a proper evaluation.” The mature response is not to reject automation, but to keep independent checks, meaningful samples, and humans who can still explain why the numbers moved.

The gPTY repository offers a compact case study in why an interface is never merely an interface. The Godot-and-Rust terminal workspace advertises panes controlled through JSON-RPC, a CLI, and an MCP server, alongside persisted history, search, and agent-facing controls. It also describes concept capture that routes recognized terminal output into adjacent panes. The era’s preferred machine is no longer just a terminal; it is a terminal that has acquired a control plane and a biography.

The project says its concept features display captured material but do not inject shell input, and that its reasoning view observes agent lifecycle events rather than orchestrating agent state. Those boundaries matter. They are useful design claims, not proof of security. The repository’s own documentation says most of its codebase was generated with LLMs and may include unidiomatic patterns or bugs; it also notes deferred audit follow-ups in areas including an unbounded PTY channel, peer credentials, PATH-resolved programs, and supply-chain issues. Candor is better than vaporware, but candor is not an audit.

This is where the absent lettuce story and the very present agent-tool story meet. Safety depends on knowing what happened, where it happened, and what a signal actually establishes. The Global Glacier Extinction Explorer models projected glacier disappearance with medians and Q25–Q75 ranges rather than laundering uncertainty into a single crisp prophecy. That habit should travel. Whether one is tracking contamination, coding costs, code quality, or terminal automation, the first public-service feature is not confidence. It is a visible boundary between what is known, what is measured, and what somebody would very much like you to assume.