Skip to content

StealthHumanizer v2.4.2 β€” Sub-1B Specialized NLP & Academic Release

Latest

Choose a tag to compare

@rudra496 rudra496 released this 26 Aug 22:29
· 186 commits to master since this release

Changelog

All notable changes to this project are documented in this file.

The format is based on Keep a Changelog, and
this project follows Semantic Versioning.

[2.4.2] - 2026-07-04

Changed β€” sub-1B specialized humanizer for short text

After a deeper HuggingFace audit + on-VPS testing of every candidate, swapped
the short-text backend from Qwen2.5-0.5B (general instruct) to
cive202/humanize-ai-text-bart-large (BART, 406M params).

  • Peer-reviewed paper (Paneru 2026, "Make it Sound like Human")
  • BERTScore 0.924 on test set β€” measurable fidelity, not marketing
  • 25,140 training pairs with document-disjoint splits
  • MIT license, updated April 2026
  • The ONLY sub-1B specialized humanizer that ACTUALLY preserves length + facts
    on real inputs (verified on 84-word + 319-word test cases)

Auto-routing by input word count:

  • ≀150 words β†’ cive202 BART (sub-1B, 5–12s, faithful)
  • 150 words β†’ Ollama gemma3:4b (4B, 25–90s, only faithful option for long form)

Why the swap: Qwen2.5-0.5B worked on short text but hallucinated fake awards
and publications on long inputs ("Euratom Medal of the Russian Federation",
invented grants, etc.). Ateeqq/Text-Rewriter-Paraphraser produced garbage
tokens. amicus summarized. No sub-1B model is faithful on long text β€” cive202
BART-large is the only one that works on short, gemma3:4b is the only one
that works on long. Auto-routing was the honest compromise.

[2.4.1] - 2026-07-04

Changed β€” swapped humanizer backend to gemma3:4b

The first version of Rudra's Free Usage Model used WhiteRoomProdigy/amicus-humanizer-v1-onnx
(T5-base, 60M params). On real inputs it produced lossy, repetitive summaries
(84 words in β†’ 19 words out, "scalable and maintainable" repeated in the same
sentence). The new backend is Ollama gemma3:4b with a 7-rule humanize
system prompt β€” faithful rewrites that preserve length, facts, names, and
structure. Same external API (no client change needed), same Oracle VPS, same
API keys.

  • Latency: 1.2s β†’ ~25s for a 100-word input (acceptable on free-tier ARM CPU).
  • Quality: input 84 words β†’ output 140 words, all facts preserved, detector
    score drops to ~6% AI probability on the humanized text.
  • Free-tier footprint: stealth-models container dropped from 1.1 GB β†’ 249 MB
    RAM because we no longer load the ONNX model in-process. The Python service
    now proxies /humanize/ to Ollama over the docker bridge.

[2.4.0] - 2026-07-04

Added β€” Rudra's Free Usage Model (zero-config default)

  • New default provider: "Rudra's Free Usage Model." The public Vercel
    deployment now ships with a pre-configured humanizer + detector pair hosted on
    the maintainer's free Oracle Cloud ARM VPS. No API key needed from the user β€”
    just paste text and click Humanize.
    • Humanizer: WhiteRoomProdigy/amicus-humanizer-v1-onnx (T5-base, 60M, ONNX INT8)
    • Detector: fakespot-ai/roberta-base-ai-text-detection-v1 (RoBERTa-base, 125M)
  • Independent AI-detector verdict now shown below every humanize result.
    After each successful humanize, the app automatically fires the hosted
    RoBERTa detector on the output and renders the verdict (Likely Human / Likely
    AI + AI probability) in a separate panel, distinct from the heuristic
    finalScore.
  • Server-side proxy module (lib/server/rudra-free.ts) keeps the API keys
    in Vercel env vars (RUDRA_HUMANIZER_API_KEY, RUDRA_DETECTOR_API_KEY,
    RUDRA_API_BASE_URL). Keys never reach the browser bundle.
  • Settings UI hides the API key field for Rudra's Free Usage Model and
    shows a "Pre-configured" notice instead.
  • Self-hosters can opt out by setting NEXT_PUBLIC_DEFAULT_PROVIDER=gemini
    (or any other provider id). See .env.example for the full set of new vars.

Notes

  • The detector's verdict is intentionally shown as a separate panel rather than
    blended into the heuristic score β€” RoBERTa detectors have a known high
    false-positive rate on formal prose, so the two numbers measure different
    things.

[2.3.0] - 2026-06-28

Fixed β€” humanization quality overhaul

This release fixes the root causes of broken output: incomplete sentences,
grammar/punctuation errors, and over-casual rewrites that "changed everything."

  • Re-humanize prompt rewritten. The old prompt instructed the model to inject
    casual openers ("Honestly,"), em-dashes, rhetorical questions, and to split every
    long sentence β€” which produced orphan fragments ("Honestly?", "But get thisβ€”?").
    The new prompt forbids fragments and disconnected interjections, demands complete,
    grammatical sentences, and preserves meaning, facts, and register.
  • Post-processing no longer creates orphan sentences. disruptFlow previously
    spliced in standalone interjections ("Right.", "Funny enough.") and rhetorical
    questions as their own sentences. It now only softens stiff AI openers in place.
  • Em-dash stripping is now punctuation-aware. "wordβ€”?" no longer becomes the
    broken "word,?"; trailing/leading and parenthetical dashes are all handled cleanly.
    Numeric ranges (2020–2025) are still preserved.
  • Sentence-start capitalization safety net. A final capitalizeSentenceStarts
    pass restores correct casing after every transformation, with abbreviation guards.
  • Removed casing-mangling transforms. The random semicolon-merge and the
    short-sentence-merge both lowercased the following word, which broke proper nouns
    and safety-critical phrasing ("Never share keys" β†’ "and never share keys").
  • Collocations now use word boundaries. "facilitate" no longer matches inside
    "facilitated" leaving "make easierd"; inflected words are left intact. Literal
    "..." placeholders that injected stray ellipses were removed.
  • thingβ†’stuff and similar context-blind safe-synonym swaps removed.
  • injectHumanVoice no longer breaks proper nouns (removed the conjunction
    prepend that forced-lowercased the next word).
  • Empty-string AI-phrase replacements (in conclusion β†’ '') that left dangling
    commas now produce clean transitions.

Changed

  • Removed the GPTZero API scoring panel from the UI. With no GPTZERO_API_KEY
    configured it only ever showed a local fallback, which was confusing. Detection
    still runs locally (Human Score, heatmap, readability); external checking is a
    one-click link to GPTZero, ZeroGPT, and Copyleaks for anyone who wants it.
  • Humanize is now a single pass by default. The automatic multi-round
    "refine to 90%+" loop that ran after every humanize (up to 8 extra API calls,
    blocking, and silently copying to the clipboard) is gone. Refinement is opt-in
    via the "Refine further" button (bounded to 3 rounds, no silent clipboard write).
  • Simpler, more professional UI. Collapsed Advanced options by default, replaced
    the redundant 7-card stats grid with a slim footer bar, cleaned emoji-heavy action
    labels, and grouped external detector links. The sentence diff folds stray
    fragments into the previous row instead of showing misleading pairings.

Fixed β€” tooling / CI

  • npm run lint works again. Next.js 16 removed next lint; switched to a
    flat ESLint 9 config (eslint.config.mjs) using the TypeScript-ESLint tooling
    bundled with eslint-config-next (FlatCompat hit a circular-structure bug with
    this version combo).
  • CLI runners work on Windows. cli-providers.ts now spawns .mjs/.js
    runners through node, avoiding Windows spawn EFTYPE on shebang scripts.
  • Browser extension least-privilege. Removed the redundant always-on
    <all_urls> content script (the web app already reads ?text= itself); kept
    activeTab as the minimal permission for tab messaging.

[2.2.0] - 2026-06-02

Added

  • Full CLI tool (stealthhumanizer / stealth-humanize) with humanize, detect,
    and providers commands. Supports stdin, file I/O, JSON output, all rewrite levels,
    styles, tones, languages, and domains.
  • Two local CLI-runner providers (claude-code, codex) that use existing CLI logins
    instead of API keys β€” confined with disabled tools, read-only sandbox, and isolated
    working directories.
  • Batch humanization API (POST /api/humanize-batch) for up to 20 texts concurrently
    with 3 parallel API calls.
  • GPTZero API integration with local heuristic fallback for AI detection scoring.
  • Grammar checking API (POST /api/grammar) with LLM-backed analysis and local fallback.
  • Protected region extraction/restoration β€” code blocks, links, URLs, and mentions are
    preserved verbatim during humanization.
  • Rehumanize API (POST /api/rehumanize) with regression guard that falls back to
    lighter rewrites when semantic drift is detected (Jaccard + n-gram overlap).
  • Logistic regression human-likeness scorer (POST /api/model-score) with 11 text
    features.
  • File upload with magic-byte validation for PDF (pdfjs-dist), DOCX (mammoth), and TXT
    extraction (max 10 MB).
  • Standardized API error handling module (lib/api-response.ts) with consistent error
    response format across all routes.
  • Shared text utilities module (lib/text-utils.ts) deduplicating sentence splitting
    and word counting.
  • Unit test suites for detector engine (25+ tests), rate limiter, and input validation.
  • CLI regression test suite (18 tests) with mocked provider fetch and fake CLI runners.
  • SEO: JSON-LD structured data, FAQ schema, programmatic robots.txt and sitemap.xml.
  • .env.example documenting all 15 provider API key variables plus CLI overrides.
  • SECURITY.md supported versions table and responsible disclosure policy.
  • Comprehensive troubleshooting guide (docs/TROUBLESHOOTING.md).
  • GitHub issue templates for bug reports, feature requests, and API key issues.
  • Release workflow for tag-based GitHub releases with changelog validation.

Changed

  • Security hardening: Content-Security-Policy headers, X-Frame-Options DENY, HSTS,
    X-Content-Type-Options nosniff, per-IP rate limiting (30 req/min default), SSRF
    prevention via Cloudflare account ID validation, XSS-safe error responses.
  • API routes enforce CLI-only provider rejection β€” browser and serverless contexts
    cannot spawn claude-code or codex subprocesses.
  • Detector calibrated with corpus-derived thresholds instead of static heuristics.
  • Architecture doc updated to reflect current provider count and module structure.
  • CLI binary path configurable via STEALTHHUMANIZER_CLAUDE_CODE_BIN and
    STEALTHHUMANIZER_CODEX_BIN environment variables.

Fixed

  • Performance: detectAI() no longer re-runs on every React render (useMemo fix).
  • EPIPE crash on CLI stdin when child process exits early.
  • PDF upload parsing stability (pdfjs-dist replacing pdf-parse).
  • Score regressions in rehumanize loop β€” regression guard prevents semantic drift.
  • pnpm-lock.yaml synced with package.json (dotenv dependency added).
  • CI/CD workflow branch triggers updated for master branch.
  • Client bundle crash from node:fs import in browser-targeted style-model.ts.
  • LinkedIn URL corrected in Footer.tsx.

Security

  • Per-IP rate limiting on all API routes with configurable window.
  • Input validation on all endpoints (length, type, format checks).
  • File upload magic-byte validation prevents content-type spoofing.
  • CLI runners confined: claude-code runs with --tools "" (all tools disabled);
    codex runs with --sandbox read-only, isolated directory, no env inheritance.
  • SSRF prevention: Cloudflare account IDs validated as 32-char hex strings.
  • API error responses sanitized to prevent information leakage.

[2.1.0] - 2026-04-11

Added

  • Style-aware humanization engine calibrated from academic writing statistics
    (OpenAlex, 11 domains, 2018–2025).
  • Style model capturing real human writing patterns: sentence length distribution
    (mean 20.5, median 20, stddev 9.4), burstiness (0.426), vocabulary diversity
    (69.4%), passive voice ratio (18.1%), transition word frequencies, and sentence
    starter distributions.
  • Style-aware prompt builder that injects corpus-calibrated statistical targets
    directly into LLM rewrite prompts.
  • Corpus-calibrated detection thresholds β€” dynamic baselines derived from real
    academic writing rather than hardcoded values.
  • Corpus-aware post-processing engine for output refinement.
  • Expanded collocation database with 50+ new AI phrase replacements (150+ total).
  • Bulk paper download and ingestion pipeline (scripts/papers/).

Changed

  • Style model (public/corpus-style-model.json) now loads via fetch() for
    client-side compatibility (no node:fs in browser bundle).
  • AI detector uses dynamic corpus-calibrated thresholds instead of static
    heuristics.

Fixed

  • LinkedIn URL corrected in Footer.tsx.
  • CI/CD workflow branch triggers updated from main to master.
  • Client bundle crash resolved by isolating node:fs usage from browser-targeted
    style-model.ts.
  • Markdown lint configuration fixed for Quality workflow.

[2.0.0] - 2026-03-31

Added

  • 13 AI provider support with configurable API keys
  • 4 rewrite levels including multi-pass ninja mode
  • 13 preset writing tones and style controls
  • Integrated AI detection heuristic scoring
  • Readability analysis layer (Flesch-Kincaid, Gunning Fog)
  • PDF and DOCX file upload support
  • Grammar check integration
  • Multi-language humanization support
  • Browser-first API key storage (no server-side persistence)
  • Side-by-side humanizer/detector workflow
  • Dark/light theme toggle
  • Research pipeline scripts for benchmark and training
  • Comprehensive docs site with GitHub Pages deployment

Changed

  • Major rewrite from initial prototype to production Next.js application
  • Upgraded to Next.js 16, React 19, TypeScript 6
  • Modernized CI with Node 20/22 matrix testing
  • Overhauled UI with Tailwind CSS and glass-morphism design

Security

  • All user data (API keys, history) stored client-side only
  • No server-side persistence of user prompts or outputs

[1.0.0] - 2026-04-09

Added

  • Comprehensive governance and support documentation (SUPPORT.md, improved
    contribution/security policies).
  • Professionalized GitHub community templates and automation (CODEOWNERS,
    Dependabot, workflow modernization).
  • Structured Markdown docs information architecture under docs/ for onboarding
    and usage.
  • Release workflow for tag-based GitHub release publishing.

Changed

  • Overhauled README.md for professional onboarding, architecture overview, and
    policy links.
  • Modernized CI and Pages workflows for main branch and clearer quality gates.
  • Added repository hygiene baselines (.editorconfig, .gitattributes, refined
    .gitignore).

Removed

  • Removed duplicate root-level FUNDING.yml in favor of .github/FUNDING.yml.

What's Changed

  • build(deps): bump next from 16.2.6 to 16.2.7 by @dependabot[bot] in #136
  • build(deps): bump react and @types/react by @dependabot[bot] in #140
  • build(deps): bump pdfjs-dist from 5.7.284 to 6.0.227 by @dependabot[bot] in #137
  • build(deps-dev): bump tsx from 4.22.3 to 4.22.4 by @dependabot[bot] in #138
  • build(deps-dev): bump eslint-config-next from 16.2.6 to 16.2.7 by @dependabot[bot] in #139
  • ci: add CodeQL security scanning and dependency review by @rudra496 in #141
  • Potential fix for code scanning alert no. 14: Server-side request forgery by @rudra496 in #142
  • Potential fix for code scanning alert no. 14: Server-side request forgery by @rudra496 in #143
  • Potential fix for code scanning alert no. 14: Server-side request forgery by @rudra496 in #144
  • build(deps): bump esbuild from 0.28.0 to 0.28.1 by @dependabot[bot] in #153
  • build(deps-dev): bump tailwindcss from 4.3.0 to 4.3.1 by @dependabot[bot] in #154
  • build(deps): bump next from 16.2.7 to 16.2.9 by @dependabot[bot] in #152
  • build(deps-dev): bump eslint-config-next from 16.2.7 to 16.2.9 by @dependabot[bot] in #150
  • build(deps): bump actions/stale from 9 to 10 by @dependabot[bot] in #145
  • build(deps): bump actions/labeler from 5 to 6 by @dependabot[bot] in #146
  • build(deps): bump actions/dependency-review-action from 4 to 5 by @dependabot[bot] in #147
  • build(deps): bump actions/first-interaction from 1 to 3 by @dependabot[bot] in #148
  • build(deps): bump github/codeql-action from 3 to 4 by @dependabot[bot] in #149
  • build(deps-dev): bump @types/node from 25.9.2 to 25.9.3 by @dependabot[bot] in #151
  • security: patch transitive js-yaml, esbuild, postcss vulns by @rudra496 in #155
  • build(deps): bump lucide-react from 1.17.0 to 1.18.0 by @dependabot[bot] in #156
  • build(deps-dev): bump @tailwindcss/postcss from 4.3.0 to 4.3.1 by @dependabot[bot] in #157
  • build(deps): bump dompurify from 3.4.8 to 3.4.10 by @dependabot[bot] in #158
  • build(deps): bump lucide-react from 1.18.0 to 1.20.0 by @dependabot[bot] in #159
  • build(deps): bump actions/checkout from 6 to 7 by @dependabot[bot] in #161
  • build(deps): bump lucide-react from 1.20.0 to 1.21.0 by @dependabot[bot] in #162
  • build(deps): bump dompurify from 3.4.10 to 3.4.11 by @dependabot[bot] in #163
  • build(deps): bump recharts from 3.8.1 to 3.9.0 by @dependabot[bot] in #166
  • build(deps-dev): bump autoprefixer from 10.5.0 to 10.5.1 by @dependabot[bot] in #165
  • build(deps-dev): bump @types/node from 25.9.3 to 26.0.0 by @dependabot[bot] in #164
  • fix: expand AI phrase blacklist by @vku2018 in #173
  • build(deps): bump DavidAnson/markdownlint-cli2-action from 23 to 24 by @dependabot[bot] in #174
  • build(deps): bump actions/first-interaction from 1 to 3 by @dependabot[bot] in #169
  • πŸ”’ Resolve all 39 open Dependabot alerts + add free security automation by @rudra496 in #189
  • build(deps): bump actions/labeler from 6 to 7 by @dependabot[bot] in #182
  • build(deps): bump actions/stale from 10 to 11 by @dependabot[bot] in #186
  • build(deps): bump actions/setup-node from 6 to 7 by @dependabot[bot] in #180
  • build(deps): bump pdfjs-dist from 6.0.227 to 6.2.108 by @dependabot[bot] in #190
  • build(deps): bump lucide-react from 1.21.0 to 1.28.0 by @dependabot[bot] in #192
  • build(deps-dev): bump tsx from 4.22.4 to 4.23.4 by @dependabot[bot] in #196
  • build(deps-dev): bump autoprefixer from 10.5.1 to 10.5.4 by @dependabot[bot] in #199
  • build(deps): bump react and @types/react by @dependabot[bot] in #197
  • build(deps-dev): bump @tailwindcss/postcss from 4.3.1 to 4.3.3 by @dependabot[bot] in #198
  • build(deps-dev): bump tsx from 4.23.4 to 4.23.5 by @dependabot[bot] in #201
  • build(deps-dev): bump tsx from 4.23.5 to 4.23.11 by @dependabot[bot] in #202
  • build(deps): bump lucide-react from 1.28.0 to 1.30.0 by @dependabot[bot] in #204
  • build(deps-dev): bump @types/node from 26.0.0 to 26.2.0 by @dependabot[bot] in #203
  • build(deps): bump mammoth from 1.12.0 to 1.12.1 by @dependabot[bot] in #205
  • build(deps): bump lucide-react from 1.30.0 to 1.31.0 by @dependabot[bot] in #206
  • build(deps-dev): bump tsx from 4.23.11 to 4.23.12 by @dependabot[bot] in #207
  • build(deps): bump actions/upload-pages-artifact from 3 to 5 by @dependabot[bot] in #209
  • build(deps): bump actions/deploy-pages from 4 to 5 by @dependabot[bot] in #211
  • build(deps): bump next from 16.3.0 to 16.3.1 by @dependabot[bot] in #213
  • build(deps-dev): bump eslint-config-next from 16.3.0 to 16.3.1 by @dependabot[bot] in #214
  • build(deps): bump lucide-react from 1.31.0 to 1.32.0 by @dependabot[bot] in #215
  • build(deps): bump next from 16.3.1 to 16.3.2 by @dependabot[bot] in #216
  • build(deps): bump lucide-react from 1.32.0 to 1.33.0 by @dependabot[bot] in #218
  • build(deps-dev): bump eslint-config-next from 16.3.1 to 16.3.2 by @dependabot[bot] in #217

New Contributors

Full Changelog: v2.2.0...v2.4.2