Field notes
We are building a pocket reader and a pocket tavern. These are the potholes, trade-offs, and field tests we hit along the way — written by the people doing the work.
If your JanitorAI DeepSeek proxy died this week, the cause has a timestamp: July 24, 2026, 15:59 UTC, when DeepSeek retired the deepseek-chat and deepseek-reasoner model names. If your configs came up blank instead, that's a separate April incident with an official Recover button. The one-line fix for each, plus the full triage table: 401s, 402s, the save-then-refresh ritual, and region age checks.
Read the post →
SillyTavern is a Node.js server, and Android kills servers. There are now four real ways to run it on a phone in 2026: a Termux install, ST-android (a community APK that packages the actual server — the 'SillyTavern for Android' making the rounds on Reddit), a cloud VPS, or a native tavern app. The exact commands, the failure modes, and who should pick which road.
Read the post →
Word-list AI detection is a lost arms race — flagship models already pass slop scans clean. The StoryScope paper (arXiv:2604.03136, UMD + Google DeepMind) found something sturdier: narrative structure alone separates human from AI fiction at 93.2%, and professional-grade rewriting moves that by 1.6 points. We projected its lexically visible features into a browser tool and validated on 78 passages of our own corpus: five axes separate (direct emotion naming leads at d=1.46), every direction matches the paper — and our own show-don't-tell discipline turned out to carve a brand-new fingerprint. Also: how we killed our strongest number because it was fake.
Read the post →
Impersonation — the bot writing your dialogue, your actions, your feelings — is the complaint JanitorAI's own help center calls one of the most common and frustrating. The root cause usually isn't the model: it's second-person sentences baked into the card's greeting and example dialogue. We scanned the 30 most-starred cards on Chub with a static checker: 14 clean, 11 with minor habits, 5 needing surgery — including the board's most-starred roleplay card at 19,568 stars. Here's the anatomy, with before/after rewrites, and a browser-local scanner to check your own.
Read the post →
A lorebook that hums in SillyTavern can go silent in JanitorAI and fire backwards in RisuAI — without anyone editing an entry. We read the open-source importers and matchers line by line while building a migration checkup: the 1% probability default, disabled entries that revive, NOT_ANY guards that invert into requirements, a 0% entry that becomes always-on, and keyword normalizers that eat CJK keys whole. Every behavior claim in this post is labeled source-verified, docs-backed, or honestly untested.
Read the post →
Three different diseases get filed under “looping”: verbatim re-paste, pet imagery, and fossilized openers. A triage guide built from our repetition detector: 12-token shingle receipts on Grok 4.5 (three straight rounds at 100% old text) and DeepSeek (one 72% slip, self-corrected), why history contamination beats every sampler knob — corrective instructions lost 9 of 10 rounds in our test — which penalty settings backfire, and the fix list in the order that actually works.
Read the post →
Nine models, two genres, 20 consecutive rounds each — every chain archived, every repetition figure recomputed on the raw text. Long-run collapse comes in exactly three shapes: the restart loop (20 rounds, 20 copies of the same opening), the mid-chain freeze (three straight rounds of 100% recycled text), and the ending rewind — 0% to 1.8% verbatim overlap, invisible to detectors, while the plot quietly rewinds to round one. Plus the cold number behind our methodology: AI judges scored 12% on known-answer anchors, agreeing with each other 86% of the time — unanimously wrong.
Read the post →
You edit a preset block and the replies don't move. A world info entry triggers in the test panel but never shows up in the request. Example dialogues have been dead since day one. All three symptoms live in the same place: the assembly step between pressing send and the request leaving. A walkthrough of SillyTavern 1.18's real assembly semantics — defined does not mean ordered, a disabled marker discards while a missing marker appends to the very end, the main prompt gets silently replaced by the card, and example blocks without name prefixes are dropped without a word.
Read the post →
The open-source state of the art fine-tunes an 8B model on 669 novels and still misattributes roughly one line in eight. Our web tool goes the other way: a deterministic rule compiler — five attribution rules plus a standing right to abstain — that found 16 characters with zero false names in a Dream of the Red Chamber excerpt, and turns a 5MB, 258-chapter book into a stage report in 264 milliseconds. No model, no upload, works offline.
Read the post →
Same model, same instruction: a fresh chat obeyed 10 rounds out of 10, a chat with polluted history disobeyed in 9. Long-chat degradation isn't the model getting dumber — it's the model imitating its own old output, with the imitation weight growing every message. The mechanism, what our 300-turn benchmark measured (image-level repetition, +2% to +4.9% reply inflation per turn, rule decay from 253 to 631 words), how the Chat Archaeology lab draws the creep as a curve with a detected inflection floor, and a recovery playbook ordered by surgery size.
Read the post →