ZICQ
中 Log in / Sign up
ZICQ Info Open Source AI #Deslop #Agent Skills #AI Text Detection #Psycholinguistics #Factual Consistency

Deslop Turns AI-Writing Checks into a Measurable Revision Loop

Avatar of Mr.Xu

By Mr.Xu Community Post

Published:

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:Deslop is a GitHub-hosted skill for reducing AI-generated writing patterns without forcing prose into a fixed Simple English or ISO 24495 template. Drawing on Biber-feature research, a 44-study survey, psycholinguistic readability work and Graphite’s released “AI Tells” data, it scans text before and after revision, reports fixed and newly introduced problems, checks for dropped facts, and flags tells associated with the reviser’s own model family. The author says the skill was evaluated agains


Core mechanism: turning de-slopping into differential measurement

Deslop replaces a one-shot prompt-based rewrite with a revision loop that scans the draft, compares before-and-after results, and checks whether any facts were lost. Rather than forcing every passage into a rigid Simple English or ISO 24495 template, the skill draws on Reinhart et al.’s comparison of Biber grammatical features, Terčon and Dobrovoljc’s survey of 44 studies, Clark, Martin and Mahowald’s analysis of why AI text is difficult to read, and Graphite’s released “AI Tells” data.

Confirmed workflow

The public description supports the following sequence:

  1. Scan the original text to establish a baseline for style, readability and model-specific tells.
  2. Perform the revision while preserving facts and claims that must remain intact.
  3. Scan the revised text again and report which problems were fixed and which were newly introduced.
  4. Check independently for dropped facts and for tells associated with the reviser’s own model family.

The fourth step matters because a scanner and reviser from the same model family may share blind spots. Family-specific rules provide a way to detect the common failure mode in which one form of AI writing is merely replaced with another.

Technical breakdown: from a scalar score to multiple constraints

Based on the exposed behavior, a scan can be abstracted as:

S(x) = [B(x), P(x), T_m(x), F(x)]

Here, B(x) represents corpus-relative Biber-style grammatical and lexical co-occurrence features, P(x) represents psycholinguistic signals associated with cognitive load and readability, T_m(x) contains style tells associated with reviser family m, and F(x) represents fact-retention signals. This vector is an engineering abstraction derived from the feature description, not a disclosure of the project’s internal algorithm.

The scanner can then compute:

ΔS = S(x_after) - S(x_before)

A differential report is more useful operationally than one absolute “AI slop” score. A lower style score does not guarantee better prose if fact retention falls at the same time: the rewrite may merely be hiding information loss in more polished language. Deslop’s strongest engineering idea is therefore to expose style changes and factual risk within one revision audit.

Engineering trade-offs and the limits of the reported evaluation

| Design choice | Benefit | Cost or risk | |---|---|---| | Corpus-derived signals instead of a fixed prose template | Preserves more of the author’s voice | Legitimate human writing can trigger some signals, requiring domain-specific thresholds | | Before-and-after comparison | Locates problems introduced by the rewrite | Misses pre-existing problems that the revision leaves untouched | | Model-family tell lists | Detects characteristic failure modes of the current reviser | Rules become stale after model updates and can create uneven treatment across providers | | Fact-loss checking | Prevents revision from deleting qualifiers, numbers or key claims | Embedding, entailment or additional model calls add latency, Token cost and false positives | | Separation of scanner and reviser | Reduces correlated blind spots between generation and evaluation | Adds orchestration complexity and requires auditable intermediate records |

The project uses three matched-prompt corpora: 9,984 Graphite web-article topics, 24,322 HC3 questions and 3,027 ELEPHANT advice prompts. Graphite’s released data includes tell lists for nine current models. Comparing human and model answers to the same prompts reduces difficulty mismatch, but the public description provides no precision, recall, F1, AUROC, confidence intervals, human blind-test results, or comparisons against unrestricted rewriting, Simple English templates and commercial detectors. The disclosed evidence therefore establishes corpus coverage, not real-world detector accuracy.

The three corpora also leave important genres uncovered, including code comments, legal clauses, technical design documents and long-form reports. Feature distributions learned or measured on web articles, questions and advice prompts should not be assumed to transfer directly to domains where precise wording and factual constraints matter.

Deployment guidance for developers

  1. Use the scanner as a quality gate, not an automatic rewriter. Return issues and deltas for human or agent review by default; block publication only for high-confidence fact loss, rather than allowing a style score to trigger unrestricted rewriting.
  2. Maintain an immutable fact ledger. Extract numbers, dates, entities, causal relations, negations and qualifiers before revision, assign each fact an identifier, and verify it after revision instead of relying only on surface Token overlap.
  3. Calibrate separately by genre and model family. Technical documentation, marketing copy and narrative writing require different thresholds. Update tell rules when models change and version the rule set so benchmark behavior cannot drift silently.
  4. Optimize under constraints rather than a scalar alone. A practical objective is maximize Q_style(x') subject to F(x') >= F(x) - ε, where ε bounds acceptable fact loss.
  5. Run in shadow mode before A/B deployment. Track human preference, critical-fact omission, AI-tell precision and recall, false-positive rate, p95 latency and per-revision cost. Adopt the workflow only if it consistently beats unconstrained rewriting on real tasks.
  6. Audit the supply chain and license. Before production use, verify the repository license, dependency locks, whether the scanner sends text to external services, and whether rule files can be audited and rolled back.

Deslop’s core contribution is to turn “does this sound like AI?” from a vague aesthetic judgment into a measurable revision process. It does not yet demonstrate universal AI-text detection, but its before-and-after diff, fact-preservation check and model-family self-inspection provide a more reliable engineering interface for agentic writing than a single instruction to rewrite the text.


Source: GitHub Projects via Hacker News (2026-11-16)

— END —

Tags: #Deslop #Agent Skills #AI Text Detection #Psycholinguistics #Factual Consistency

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…