Publishing is a sequence of editorial decisions made under time pressure: what is this piece about, who should see it, is it accurate, is it satire, and which thirty seconds are worth clipping. The model can answer a bounded version of each question. The editorial judgment, the standards and the publish button stay with people.
Tagging a library is a closed-set decision
JEV Book Tags is a Calibre plugin that assigns categories to books. It sends metadata only, or metadata plus up to eight EPUB samples within a configurable character cap, and Jev chooses from the categories the user defines. The real 32-book demo ended with 31 ready and 1 flagged for review, with a 95% Programming category threshold. The defaults are an 85% assignment threshold, an 8-point closeness margin and 80% evidence confidence, and a candidate must clear 50% even when the threshold is set to 50%.
The review behavior is the useful part. Existing manual tags are preserved and duplicates ignored, automatic application is disabled by default, suggestions must be confirmed, and uncertain or close candidates stay review-only. A restore action only works while books are unchanged since the last application. The plugin requires Calibre 9.0 or newer, and native verification so far is only on Calibre 9.13 on macOS. The README states GPL-3.0-or-later, but GitHub's license field reports NOASSERTION, so the SPDX identity could not be verified. It also sends book metadata or partial EPUB excerpts to TypeSafe's hosted API and needs your own key.
That is a good template for publishing metadata: a closed category set, explicit thresholds, a review queue and a reversible application.
Screening sources is not endorsement
Jev Account Filter is a thinner project with a revealing gap. Its banner describes decision models for video-platform account screening: find accounts relevant to a domain, then select videos for training data. The README markets it as fast, cheap, accurate, flexible and promising and links installation, commands and configuration docs.
The captured README excerpt contains no commands, no configuration and no results. GitHub shows the repository as HTML with no license, created 2026-10-15 and pushed 2026-10-17, with 4 stars. Jev is named, but the specific calls and questions are not documented. That is a reminder that a repository name and a banner are not evidence of an integration. If you need account screening, treat this as a pointer to a workflow, not a component you can audit.
Rank the feed, then summarize
Social Monitor is a full-stack collector for X/Twitter, Reddit, Hacker News, RSS/Atom and GitHub. It ranks the few posts that matter and turns them into summaries with source links, daily updates and a weekly digest. The stack is Flutter/Dart, NestJS, PostgreSQL, RabbitMQ and Redis, licensed Apache-2.0, with 48 stars and a last push on 2026-10-16.
The integration caveat is significant: Jev is named in the GitHub description and topics, but the captured README never names the model, endpoint or decision it makes. Most non-X providers are beta, X uses a private collector, GitHub Issues is manual-only and Telegram is deferred. The ranking is personalized through interest profiles, post ratings and relevance feedback, but how Jev participates is not documented in the excerpt.
For a newsroom, the architecture still suggests the right separation: collect broadly, rank narrowly, summarize with citations, and keep the ranking policy in code. A model can score relevance; it should not decide what is true.
Moderation is a confidence-banded pipeline
Slop Scope asks whether an image is AI-generated. The browser derives a fixed fact dossier from 16 measured windows and a lens-fringe field, plus a whole-image CLIP ViT-B/16 detector. Only that fixed enum dossier leaves the browser; the image, pixels, filename, hash and metadata are never sent. Two typed Jev requests, an evidence panel and a verdict, cost about $0.00013 at September 2026 prices.
The verdict is likely_ai, likely_real or uncertain, and uncertain is a valid outcome. The README states plainly that Jev's weights are not calibrated probabilities and that the detector-dossier-judge combination has not passed its own acceptance steps; accuracy is unknown. The reference-frequency gate fails for the detector's AI-side facts, no paired keep run was made, and no pipeline meets the project's bar of at most 1% misclassified with uncertain counted as a miss. The code is MIT, but the f1 model starts from CLIP ViT-B/16 plus non-commercial training data such as OpenFake.
That candor is more useful than a confident demo. A moderation queue built on this would send uncertain cases to a person, log the evidence, and never present the verdict as a fact. The human review entry is the relevant control.
Satire and fact-checking need order controls
Onion or Not asks whether a headline is Onion-style satire or real news. Three models answer side by side: Jev through TypeSafe's API with your own key, AnyJev L0 and raw Qwen3-8B, both running locally through llama.cpp. The option-swap toggle asks both satire-or-real-news and real-news-or-satire, because a plain LLM sometimes changes its answer when the options arrive in a different order. AnyJev L0 reads both orders, so its answer cannot flip that way, per the README.
The README excerpt shows no accuracy results, only the comparison method, and the setup needs both a TypeSafe key and about 5 GB of local model weights. It is MIT licensed, created and pushed on 2026-10-16. The option order bias problem is the real contribution: if a satire classifier changes its mind when the labels swap, it is not measuring the headline.
The broader question is whether cheap evaluation improves editorial judgment. Judgement in the Age of Jev is a perspective piece, not an empirical study. It formulates a conditional Jevons hypothesis: cheaper machine evaluation may increase organizational consumption of it where latent demand is substantial and complementary costs do not dominate. It distinguishes prediction, machine evaluation, organizational judgement and authorisation, whose costs need not fall together, and warns that evaluations can share evidence, criteria and errors, scale mis-specified rubrics, and change decision rights through thresholds.
That is the right frame for a fact-checking desk. A cheaper check does not automatically produce more truth; it produces more checks. Whether that helps depends on who owns the rubric and who can overrule the threshold.
Clips and highlights are evidence selection
Jev Highlight Cutter turns long interviews, podcasts and livestreams into evidence-based highlight reels. A Codex-style agent skill plus a Python CLI reads the transcript and reviews cut boundaries, Jev scores the text evidence, and local FFmpeg cuts. It reads SRT, VTT and Whisper-style JSON while keeping original absolute timestamps, then picks non-overlapping segments inside a target-duration budget. It keeps scores, selection rationale and source-to-cut time mapping, and renders a clickable local review page, MP4 clips and an export verification report.
It needs Python 3.10+ and FFmpeg with ffprobe, and the core CLI uses only the standard library. Real scoring uses a user-global TYPESAFE_API_KEY; the doctor command does not call Jev or verify quota. The README scopes it to interview podcasts and calls the output a rough cut: no bilingual subtitles, dubbing, BGM or JianYing/CutChat project export, and no benchmarks. That is honest about the product boundary.
The prompt injection threat model and moderation and hate speech pieces cover the adversarial side of media pipelines, where source text is untrusted input. The use-case map lists media examples but calls them example decisions rather than shipped products.
The common thread is that every one of these systems works best when the model returns a label with a confidence and the editorial system decides what to do with it. Tag with review, rank with citations, moderate with uncertain as a valid verdict, and clip with a human watching the rough cut. That is slower than a fully automatic feed, and it is the version that can survive a correction.
