Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
35 commits
Select commit Hold shift + click to select a range
8747e46
Measure only the spoken words when a band file asks for it
ilien-dev Nov 12, 2026
a06ce8b
Run the checklist on the spoken words too
ilien-dev Nov 12, 2026
4fae963
Strip bold-wrapped directions and keep sentences that open with a tim…
ilien-dev Nov 12, 2026
4f73ab0
Add the corpus script and manifest for video scripts
ilien-dev Nov 12, 2026
121188e
Reject unpunctuated captions and leftover caption markup from the scr…
ilien-dev Nov 12, 2026
5e7560c
Add assistant video scripts on the human titles
ilien-dev Nov 12, 2026
1c2a765
Drop the model's notes to the user and word-count trailers from measu…
ilien-dev Nov 12, 2026
cafd8ed
Calibrate bands for video scripts
ilien-dev Nov 12, 2026
b1d7776
Add a narration judge prompt and the judges' reasons on video scripts
ilien-dev Nov 12, 2026
48db7a8
Add the script patterns that hold on held-out scripts
ilien-dev Nov 12, 2026
06c99fd
Show how the combined script checks lean on the comment asks
ilien-dev Nov 12, 2026
f01f06f
Cut every human quote in the judges' reasons to a few words
ilien-dev Nov 12, 2026
3570456
Narrow the comment-ask and stay-to-the-end checks and read wrapped sc…
ilien-dev Nov 13, 2026
c2e6f98
Keep comment asks as a tell and rebuild the script bands on the curre…
ilien-dev Nov 13, 2026
37a1cff
Add end-to-end sets for video scripts and the rules as tested
ilien-dev Nov 13, 2026
80205ee
Add end-to-end sets for video scripts and the rules that held
ilien-dev Nov 13, 2026
4b6530e
Judge the script rewrites again on their whole narration
ilien-dev Nov 13, 2026
85b6401
Keep every narration paragraph when directions are stacked
ilien-dev Nov 13, 2026
3095346
Rebuild the script bands, checks and results on whole narrations
ilien-dev Nov 13, 2026
455a6cc
Measure the shipped script register end to end
ilien-dev Nov 13, 2026
e85335a
Keep check.sh state files out of snapshots and keep line endings in t…
ilien-dev Nov 13, 2026
1382823
Document the video script register and what the end-to-end runs showed
ilien-dev Nov 13, 2026
e50330c
Drop the script rewrite instruction that no run supports
ilien-dev Nov 13, 2026
794b9d1
Name the snapshot behind each end-to-end figure and take the untested…
ilien-dev Nov 13, 2026
e5842ef
Make the script corpus builder safe to run and to fail
ilien-dev Nov 13, 2026
4c654bd
Find the spoken column of a script table by its header
ilien-dev Nov 13, 2026
ecb3b39
Name snapshots and commits in the script results and correct two corp…
ilien-dev Nov 13, 2026
ce0d109
Add notes, samples and an edit-only rule to the script end-to-end sets
ilien-dev Nov 13, 2026
fb87a53
Record the script rewrite round and ship what was confirmed
ilien-dev Nov 13, 2026
c26590c
Confirm the notes result for scripts on held-out titles and say what …
ilien-dev Nov 13, 2026
c55096c
Record the notes and samples combination for Spanish scripts
ilien-dev Nov 13, 2026
15b52c5
Say where the test notes came from and what the second judge saw, and…
ilien-dev Nov 13, 2026
ac28250
Point the script entry at the notes rules below it and give the measu…
ilien-dev Nov 13, 2026
e98a043
Mention the video script register on the site
ilien-dev Nov 13, 2026
1033e90
Version 1.5.0
ilien-dev Nov 13, 2026
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
The table of contents is too big for display.
Diff view
Diff view
  •  
  •  
  •  
2 changes: 1 addition & 1 deletion .claude-plugin/marketplace.json
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,7 @@
{
"name": "quiron",
"source": "./",
"version": "1.4.0",
"version": "1.5.0",
"description": "AI humanizer skill: removes AI slop and the signs of AI writing from prose, then measures the rewrite against human baselines.",
"author": {
"name": "ilien",
Expand Down
6 changes: 3 additions & 3 deletions AGENTS.md
Original file line number Diff line number Diff line change
Expand Up @@ -33,8 +33,8 @@ python3 scripts/aimeter.py --calibrate DIR # rebuild bands.json from human tex
scripts/build-corpus.sh [OUTDIR] [author ...] # fetch pre-2022 dev.to corpus, then calibrate
python3 scripts/evaluate.py --human DIR [--ai-train DIR] NAME=DIR ... # score the skill
python3 scripts/factdiff.py SOURCE REWRITE # facts in the rewrite, not in the source
QUIRON_BANDS=scripts/bands-fiction.json ... # other registers: fiction, es
python3 eval/e2e/harness.py rewrite SNAP REP [--set dev|test|es|fic|fic2] # needs the claude CLI
QUIRON_BANDS=scripts/bands-fiction.json ... # other registers: fiction, es, script[-es]
python3 eval/e2e/harness.py rewrite SNAP REP [--set dev|test|es|fic|fic2|scr|scr2|scr-es|scr2-es] # needs the claude CLI
python3 eval/e2e/report.py OLD NEW [--set ...] # blind judges, paired per title
```

Expand All @@ -44,4 +44,4 @@ python3 eval/e2e/report.py OLD NEW [--set ...] # blind judges, paired per title
- `audit.py` `CHECKS`: one entry per pattern, PASS / FAIL / TELL / READ. FAIL only where ≤~5% of held-out human texts trip it (`evaluate.py` shows the rate).
- `check.sh`: convergence is two consecutive clean passes with READ items ruled on in both.

`SKILL.md`, `references/patterns.md`, `CHECKS` and `AI_REF` describe the same patterns and numbers; change one, update the others. `references/numbers.md` mirrors `bands.json`. `SKILL.md` stays under 500 lines; detail goes in `references/`, one level deep. Band files (`bands.json`, `-fiction`, `-es`) come from `build-corpus*.sh`. Human corpora are not committed; `eval/ai/` samples are.
`SKILL.md`, `references/patterns.md`, `CHECKS` and `AI_REF` describe the same patterns and numbers; change one, update the others. `references/numbers.md` mirrors the band files. `SKILL.md` stays under 500 lines; detail goes in `references/`, one level deep. Band files come from `build-corpus*.sh` (the script one needs yt-dlp). Human corpora are not committed; `eval/ai/` samples are.
4 changes: 2 additions & 2 deletions CITATION.cff
Original file line number Diff line number Diff line change
Expand Up @@ -9,8 +9,8 @@ type: software
authors:
- alias: ilien
website: "https://github.com/ilien-dev"
version: 1.4.0
date-released: "2026-11-12"
version: 1.5.0
date-released: "2026-11-13"
license: AGPL-3.0-only
repository-code: "https://github.com/ilien-dev/quiron"
url: "https://quiron.ilien.dev/"
Expand Down
6 changes: 3 additions & 3 deletions CLAUDE.md
Original file line number Diff line number Diff line change
Expand Up @@ -33,8 +33,8 @@ python3 scripts/aimeter.py --calibrate DIR # rebuild bands.json from human tex
scripts/build-corpus.sh [OUTDIR] [author ...] # fetch pre-2022 dev.to corpus, then calibrate
python3 scripts/evaluate.py --human DIR [--ai-train DIR] NAME=DIR ... # score the skill
python3 scripts/factdiff.py SOURCE REWRITE # facts in the rewrite, not in the source
QUIRON_BANDS=scripts/bands-fiction.json ... # other registers: fiction, es
python3 eval/e2e/harness.py rewrite SNAP REP [--set dev|test|es|fic|fic2] # needs the claude CLI
QUIRON_BANDS=scripts/bands-fiction.json ... # other registers: fiction, es, script[-es]
python3 eval/e2e/harness.py rewrite SNAP REP [--set dev|test|es|fic|fic2|scr|scr2|scr-es|scr2-es] # needs the claude CLI
python3 eval/e2e/report.py OLD NEW [--set ...] # blind judges, paired per title
```

Expand All @@ -44,4 +44,4 @@ python3 eval/e2e/report.py OLD NEW [--set ...] # blind judges, paired per title
- `audit.py` `CHECKS`: one entry per pattern, PASS / FAIL / TELL / READ. FAIL only where ≤~5% of held-out human texts trip it (`evaluate.py` shows the rate).
- `check.sh`: convergence is two consecutive clean passes with READ items ruled on in both.

`SKILL.md`, `references/patterns.md`, `CHECKS` and `AI_REF` describe the same patterns and numbers; change one, update the others. `references/numbers.md` mirrors `bands.json`. `SKILL.md` stays under 500 lines; detail goes in `references/`, one level deep. Band files (`bands.json`, `-fiction`, `-es`) come from `build-corpus*.sh`. Human corpora are not committed; `eval/ai/` samples are.
`SKILL.md`, `references/patterns.md`, `CHECKS` and `AI_REF` describe the same patterns and numbers; change one, update the others. `references/numbers.md` mirrors the band files. `SKILL.md` stays under 500 lines; detail goes in `references/`, one level deep. Band files come from `build-corpus*.sh` (the script one needs yt-dlp). Human corpora are not committed; `eval/ai/` samples are.
16 changes: 15 additions & 1 deletion README.es.md
Original file line number Diff line number Diff line change
Expand Up @@ -135,7 +135,9 @@ este repo. Nada entra porque "se lee mejor".
- **Bases humanas.** Las bandas salen de 167 posts de dev.to y 157 cuentos de
WritingPrompts escritos antes de que existiera ChatGPT, y de 241 posts de dev.to en
español de 118 autores. Desde noviembre de 2026 hay también bandas para descripciones
de pull request, mensajes de commit, README y correo de trabajo, en inglés. Una cuarta parte de cada conjunto se reserva y nunca se usa para
de pull request, mensajes de commit, README y correo de trabajo, en inglés, y para la
narración de guiones de video de YouTube, en inglés y en español, sacadas de subtítulos
manuales de antes de que existiera ChatGPT. Una cuarta parte de cada conjunto se reserva y nunca se usa para
elegir nada.
- **Texto de IA de 2026 con los mismos títulos.** Claude Opus, Sonnet, Haiku y GPT
escribieron los textos de comparación que hay en `eval/ai/`, a partir de la petición
Expand Down Expand Up @@ -196,6 +198,18 @@ scripts/check.sh ARCHIVO # una pasada del ciclo completo
Para ficción en inglés se usa `scripts/bands-fiction.json`, y sin la variable se mide
contra posts de blog en inglés. Los textos de menos de unas 120 palabras no se miden.

Para la narración de un guion de video de YouTube se pone `QUIRON_BANDS=scripts/bands-script-es.json`, o
`QUIRON_BANDS=scripts/bands-script.json` si el guion está en inglés. El medidor lee solo las palabras
que se dicen y deja quietas las acotaciones. Mide el guion y busca los patrones que
comparten los guiones de los asistentes. No se ha visto que un guion reescrito suene más a
persona. Reescrito a partir de notas que eran esquemas del guion humano ya terminado del
mismo video, el P(IA) del juez Claude bajó con títulos reservados de 71,0 a 24,7 en inglés y
a 20,3 en español, mientras que un juez GPT vio unos 15 puntos y siguió puntuándolos en torno
a 65. Sin notas, una reescritura ganó entre 4,5 y 10,5 puntos y cerca de nueve de cada diez
se siguieron puntuando como IA, y los jueces puntuaron como más IA un guion escrito por una
persona después de reescribirlo. Si el guion es tuyo, mídelo y cambia lo que tú decidas
cambiar.

Otros idiomas reciben la misma lista como mejor esfuerzo, porque todavía no tienen una base
humana. El medidor lo dice y solo muestra los números que no dependen de las palabras, como
el largo de los párrafos o los encabezados.
Expand Down
14 changes: 13 additions & 1 deletion README.md
Original file line number Diff line number Diff line change
Expand Up @@ -131,7 +131,8 @@ repo. Nothing ships on "this reads better".
- **Human baselines.** The bands come from 167 dev.to posts and 157 WritingPrompts
stories, all written before ChatGPT existed, plus a Spanish set. Since November 2026
there are also bands for pull request descriptions, commit messages, READMEs and work
email. A quarter of each is
email, and for the narration of YouTube video scripts in English and Spanish, built from
manual captions uploaded before ChatGPT existed. A quarter of each is
held out and never used to pick anything.
- **2026 AI text on the same titles.** Claude Opus, Sonnet, Haiku and GPT wrote the
comparison texts in `eval/ai/`, from the plain prompt a user would type. They were
Expand Down Expand Up @@ -185,6 +186,17 @@ scripts/check.sh FILE # one pass of the full loop
For fiction or Spanish, set `QUIRON_BANDS=scripts/bands-fiction.json` or
`scripts/bands-es.json`. Texts under about 120 words get no measurement.

For the narration of a YouTube video script, set `QUIRON_BANDS=scripts/bands-script.json`,
or `scripts/bands-script-es.json` in Spanish. The meter reads only the spoken words and
leaves the stage directions alone. It measures a script and checks it for patterns that
assistant scripts share. It has not been shown to make a rewritten script read as
human-written. Rewritten from notes that were outlines of the finished human script of the
same video, the Claude judge's P(AI) fell on held-out titles from 71.0 to 24.7 in English and
to 20.3 in Spanish, while a GPT judge saw about 15 points and still rated them around 65.
Without notes a rewrite gained 4.5 to 10.5 points and about nine in ten were still rated AI,
and judges rated a script a person wrote as more AI after a rewrite. For your own script,
measure it and change what you choose to change.

Other languages get the same checklist as a best effort. There is no human baseline for
them yet. The meter says so and shows only the numbers that don't depend on the words,
such as paragraph length and headings.
Expand Down
20 changes: 18 additions & 2 deletions SKILL.md
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
---
name: quiron
description: Writes, rewrites or reviews prose so it reads as a person wrote it, and verifies the result with measurements (a meter against human baselines and a pattern checklist) instead of impressions. Use it whenever text a person will read needs to stop sounding like AI, such as blog posts, articles, essays, stories, guides, docs, READMEs, emails, messages to a boss or a team, PR descriptions. That includes removing "AI slop", making text sound less like ChatGPT or a chatbot, humanizing a draft, and reviewing or checking prose for AI tells such as em dashes, "not just X but Y", lists of three, headings everywhere or a summary at the end. Works in any language; English and Spanish have measured baselines, other languages get the same patterns as a best effort.
description: Writes, rewrites or reviews prose so it reads as a person wrote it, and verifies the result with measurements (a meter against human baselines and a pattern checklist) instead of impressions. Use it whenever text a person will read needs to stop sounding like AI, such as blog posts, articles, essays, stories, video scripts, guides, docs, READMEs, emails, messages to a boss or a team, PR descriptions. That includes removing "AI slop", making text sound less like ChatGPT or a chatbot, humanizing a draft, and reviewing or checking prose for AI tells such as em dashes, "not just X but Y", lists of three, headings everywhere or a summary at the end. Works in any language; English and Spanish have measured baselines, other languages get the same patterns as a best effort.
license: AGPL-3.0-only. See LICENSE and NOTICE.
compatibility: Python 3 standard library only. The scripts need a shell and file access.
---
Expand All @@ -22,7 +22,8 @@ prefix them with the full path: `python3 <skill-dir>/scripts/aimeter.py FILE`.
Read these when the step you are on needs them. Each is one level deep and self-contained.

- [references/patterns.md](references/patterns.md): the 33 structural patterns in full,
with the measured rate and fix for each. Read the entry when `audit.py` flags it.
with the measured rate and fix for each, and the checks for video scripts. Read the
entry when `audit.py` flags it.
- [references/word-choice.md](references/word-choice.md): the lexicons, the plain words
models leave out, and the Claude-specific words.
- [references/spanish.md](references/spanish.md): bands, checks and word lists for Spanish.
Expand Down Expand Up @@ -66,6 +67,21 @@ sentences than human ones, while assistant fiction has choppier ones (both run s
English. They measure a draft; the rewrite rules were not tested end to end in these
registers. Most commit messages and many emails are under the meter's 120 words, and
for those the patterns apply by eye.
- `scripts/bands-script.json` (English), `scripts/bands-script-es.json` (Spanish): the
narration of a YouTube video script. Only the spoken words are measured: scene headings,
stage directions, speaker labels, timestamps and source lists are dropped. Rewrite only
what is said and keep those where they are; whoever records the video needs them. A script
a person wrote is measured, not rewritten: rewriting human narration made judges rate it
more AI. For a draft an assistant wrote, first ask the creator for their notes about that
video (what it covers, names, numbers, their opinions and asides), build the script from
them as the notes-first rules below say, and run `check.sh --source` with the draft and the
notes. In testing that was the only thing that moved blind judges: on held-out titles the
Claude judge went from 71.0 for a rewrite without notes to 24.7 in English and to 20.3 in
Spanish, while a GPT judge saw about 15 points and still rated those rewrites around 65.
The test notes were outlines written from the finished human script of the same video, so
a creator's rough notes may do less. Without notes a rewrite gains 4.5 to 10.5 points and still
reads as AI: tell the user so. No script-specific rewrite rule held; the numbers and what
was tried are in `references/numbers.md`.

Select a file by putting `QUIRON_BANDS=<skill-dir>/scripts/bands-fiction.json` (or
another of these files) in front of every command, including `check.sh`. The scripts print a
Expand Down
9 changes: 8 additions & 1 deletion docs/es/index.html
Original file line number Diff line number Diff line change
Expand Up @@ -66,6 +66,11 @@
"name": "¿Funciona en otros idiomas?",
"acceptedAnswer": { "@type": "Answer", "text": "El español y el inglés tienen bandas medidas y sus propias listas de palabras. Los demás idiomas reciben la lista de patrones como mejor esfuerzo." }
},
{
"@type": "Question",
"name": "¿Sirve para guiones de video?",
"acceptedAnswer": { "@type": "Answer", "text": "Los mide. Hay bandas para la narración de guiones de YouTube en inglés y en español, sacadas de los subtítulos de 193 y 107 videos subidos antes de 2022, y el medidor lee solo lo que se dice. La nueva versión suena mucho menos a IA solo cuando se arma con tus propias notas sobre el video. Sin ellas mejora poco." }
},
{
"@type": "Question",
"name": "¿Se inventa cosas para sonar humano?",
Expand Down Expand Up @@ -120,7 +125,7 @@ <h2 id="que-detecta">Qué detecta</h2>
<p>Las expresiones famosas, como <em>cabe destacar</em> o <em>es importante señalar</em>, casi no aparecen ya. En la prueba de este proyecto salieron en 2 de 32 posts de IA en español, y en un 4 a 10% de los humanos. Lo que delata a un modelo en 2026 es bastante más discreto.</p>

<h2 id="medidor">Un medidor, no una autoevaluación</h2>
<p>La mayoría de los skills para humanizar texto le dan al modelo una lista de patrones y confían en que siga las instrucciones. Quirón también tiene su lista, pero además incluye un script que cuenta 23 tasas en el texto y compara cada una con el rango de 241 posts de dev.to en español escritos antes de 2022. El inglés y la ficción tienen sus propias bandas.</p>
<p>La mayoría de los skills para humanizar texto le dan al modelo una lista de patrones y confían en que siga las instrucciones. Quirón también tiene su lista, pero además incluye un script que cuenta 23 tasas en el texto y compara cada una con el rango de 241 posts de dev.to en español escritos antes de 2022. El inglés, la ficción y los guiones de video tienen sus propias bandas.</p>
<p>El medidor atrapa algo que una lista no ve. Si le dices a un modelo que escriba como una persona, normalmente se pasa: las frases salen más cortadas y las palabras más simples que las de cualquier persona. El medidor avisa de eso con la misma fuerza que avisa del lado de la IA, así que sabes cuándo parar.</p>
<p>En un tutorial de Stripe escrito por un asistente encuentra 13 de 23 tasas dentro de la banda humana. La nueva versión, hecha a partir de las notas del autor, consigue las 23.</p>
<pre><code>$ python3 scripts/aimeter.py eval/ai/blog/e2e-raw/aspittel-782713.md
Expand All @@ -143,6 +148,8 @@ <h2 id="preguntas">Preguntas</h2>
<dd>No. Esas herramientas leen probabilidades de tokens, y Quirón no las toca. Trabaja únicamente sobre lo que realmente nota alguien al leer.</dd>
<dt>¿Funciona en otros idiomas?</dt>
<dd>El español y el inglés tienen bandas medidas y sus propias listas de palabras. Los demás idiomas reciben la lista de patrones como mejor esfuerzo.</dd>
<dt>¿Sirve para guiones de video?</dt>
<dd>Los mide. Hay bandas para la narración de guiones de YouTube en inglés y en español, sacadas de los subtítulos de 193 y 107 videos subidos antes de 2022, y el medidor lee solo lo que se dice. La nueva versión suena mucho menos a IA solo cuando se arma con tus propias notas sobre el video. Sin ellas mejora poco.</dd>
<dt>¿Se inventa cosas para sonar humano?</dt>
<dd>Sus instrucciones se lo prohíben. Una nueva versión solo puede usar datos de tu texto o de tus notas, y un script lista los números y enlaces de la versión nueva que no están en el original. Cuando al texto le falta un detalle, el skill te lo pregunta.</dd>
</dl>
Expand Down
Loading
Loading