AI presence in zmoki.xyz posts
Audit of all 18 published feed posts · 2026-08-09 · method adapted from StoryScope (arXiv:2604.03136), scoring discourse-level features rather than word choice
AI-likeness by post
Score 1–5: how much unedited AI-drafted text likely survives in the post. Bar shade follows the score. Hover a row for the verdict.
Two independent analyses agree
Zarema separately ran a 30-feature StoryScope-style analysis on 3 posts. The verdicts converge with this audit's blind scoring, which is decent evidence the feature approach works on this corpus.
| Post | Zarema's analysis | This audit |
|---|---|---|
| Tbilisi swaps | 22 human-lean · 1 AI-lean | 1.0 — clearly human |
| About me | 18 human-lean · 3 AI-lean | 2.0 — low |
| Pantry runbook (15) | 8 human-lean · 13 AI-lean | 4.5 — highest in corpus |
What the AI text looks like, post by post
The top offenders, with the passages that gave them away. Click to expand.
15 · Safety high-protein food list — 4.5
The human core is easy to spot: the constraints list and the shopping story ("For quinoa I chose the red one, cause it's my favorite"). Around it, Gemini's prose survives nearly unedited:
"Canning locks in freshness and allows proteins to last anywhere from 2 to 5 years on the shelf."
"plant proteins are fantastic. They also come loaded with dietary fiber"
"Life has taught me that being prepared is one of the most powerful expressions of self-care… Stay safe, eat well, and keep that emergency stash full."
The conclusion is the classic AI tidy-bow: performed uplift instead of plainly named feeling.
13 · SEO checklist — 4.0
One analogy is voice; ten is a template. This post has an analogy per section:
"SSL is like a lock on your front door" · "Security headers are like security guards" · "It's like a bouncer at a club" · "like creating a business card for your webpage" · "It's like the difference between a house and a home"
Plus rhetorical-question filler ("You wouldn't want medical advice from someone who isn't qualified, right?") and an E-E-A-T section in neutral documentation register.
16 · Power of questions — 3.5
The data half (regex filters, SQL, both case studies) is clearly yours. The theory half carries document-about-itself framing:
"This document outlines the theory, its application to SEO, and the data-driven proof of its importance."
"Theory is good, but data is undeniable."
4 · Neurodivergent glossary — 3.0
Disclosed AI co-creation, and the disclosure itself is part of the voice. But the definitions keep assistant patterns: bolded phrase pairs in every entry and the corpus's densest contrast framing ("not a disease or a deficit", "isn't sadness; it's the absence of feeling"). The lived sentences ("My favorite movies are boring, food is bland") are clearly yours.
10 · Reading list — 3.0
The frame story is yours; the four book blurbs are back-cover summaries with no first-person reaction inside them:
"It provides practical models and strategies to design a life that joyfully incorporates all of your interests"
5 · Day themes — 3.0
"Neurodivergent brains thrive when they can pivot inside a flexible boundary. The freedom to switch tasks within a theme keeps my dopamine up and my productivity flowing."
The cadence of an assistant explaining your own system back to you, plus a closing that restates the lesson three ways.
11 · Notebooks — 2.5
Mostly the rawest writing in the corpus. One section flips register and reads like a pasted AI chat reply — an assistant would praise the system as "beautiful"; you wouldn't praise your own:
"This is a beautiful evolution… Doing it after breakfast is just smart biology… That's having a safety valve."
14 · Freedom manifesto — 2.5
The therapist scene is unmistakably yours. The three-pillar section shifts to coach voice ("an engine of leveraged income", "life is a massive, beautiful experiment"), and escaped 1\. artifacts suggest export from another tool.
What "fully you" measures as
The clean posts define the target. Across 6, 9, and 12 (with 7 and 17 close behind):
- Opens from a lived moment or personal context, never a thesis or generic hook.
- Emotions named plainly ("I cried from the cold", "I was scared as fuck"), never performed through embodied metaphor.
- One earned takeaway at most; the lesson never repeats in intro, middle, and conclusion.
- Endings stay open ("Maybe one day I'll add more swap spots. Or maybe not").
- Conversational transitions ("So," "And," "Naturally,"); "Moreover" appears once in ~19,300 words.
- Small grammar quirks survive ("cause", lowercase "i") — texture, not errors.
- Dense concrete specifics: dates, prices, named venues, exact numbers.
The full feature map — all 20 properties
Every property the audit scored. For 1–5 scales: the blue dot is your measured baseline, the orange band is the AI zone — where AI-assisted drafting typically pushes the score. For the rest: how often the trait appears across your 18 posts.
Core — human-vs-AI separators
Fingerprint — distinctively Zarema
Per-post matrix — 20 features × 18 posts
Cell color is the verdict: green = human-lean (inside your baseline), orange = AI-lean, gray = neutral or not applicable. Hover a column header for the post title. Scroll sideways for all 18 posts.
| 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| C1 opening mode | pc | pc | pc | pc | pc | lm | lm | lm | pc | pc | pc | lm | pc | lm | th | th | pc | pc |
| C2 specificity | 4 | 3 | 3 | 3 | 4 | 5 | 5 | 4 | 4 | 3 | 5 | 5 | 4 | 4 | 5 | 5 | 4 | 4 |
| C3 over-explanation | 2 | 2 | 2 | 2 | 3 | 2 | 3 | 3 | 1 | 2 | 3 | 1 | 3 | 2 | 4 | 3 | 2 | 1 |
| C4 resolution | op | op | pa | pa | ti | op | pa | ti | op | op | pa | op | ti | op | ti | pa | op | pa |
| C5 emotional register | pl | mu | pl | pl | mu | pl | pl | pl | mu | mu | pl | pl | mu | pl | pe | mu | pl | mu |
| C6 hedging | 1 | 2 | 1 | 1 | 2 | 1 | 1 | 2 | 1 | 2 | 1 | 1 | 2 | 1 | 2 | 2 | 1 | 1 |
| C7 transitions | cv | cv | cv | cv | cv | cv | cv | cv | cv | cv | cv | cv | mx | cv | mx | fm | cv | cv |
| C8 contrast framing | 1 | 1 | 1 | 4 | 2 | 0 | 1 | 1 | 0 | 0 | 1 | 0 | 2 | 1 | 1 | 2 | 1 | 1 |
| C9 rhythm variance | 4 | 3 | 4 | 4 | 3 | 5 | 4 | 5 | 3 | 3 | 4 | 4 | 2 | 5 | 3 | 2 | 4 | 5 |
| C10 paragraph uniformity | 2 | 2 | 2 | 3 | 2 | 2 | 2 | 2 | 2 | 3 | 2 | 2 | 4 | 2 | 3 | 4 | 2 | 3 |
| F1 garden cross-links | 3 | 5 | 2 | 2 | 0 | 4 | 2 | 4 | 5 | 1 | 1 | 4 | 1 | 0 | 0 | 1 | 3 | 2 |
| F2 identity threads | 5 | 4 | 2 | 2 | 2 | 3 | 2 | 2 | 4 | 3 | 3 | 4 | 1 | 4 | 3 | 1 | 3 | 2 |
| F3 AI transparency | – | – | – | ✓ | – | – | ✓ | ✓ | – | – | – | – | – | – | ✓ | – | – | – |
| F4 system naming | – | – | – | – | ✓ | – | – | ✓ | ✓ | – | ✓ | – | – | ✓ | – | – | ✓ | – |
| F5 computing metaphors | 0 | 2 | 1 | 1 | 1 | 0 | 0 | 0 | 0 | 0 | 3 | 0 | 0 | 0 | 0 | 0 | 0 | 1 |
| F6 interjections | 1 | 1 | 0 | 1 | 0 | 3 | 1 | 2 | 0 | 0 | 0 | 1 | 0 | 1 | 0 | 0 | 1 | 0 |
| F7 quoted real people | – | – | – | – | – | ✓ | ✓ | – | – | ✓ | – | ✓ | – | ✓ | – | – | – | – |
| F8 reader address | ✓ | ✓ | – | ✓ | – | ✓ | – | – | – | – | – | – | ✓ | – | ✓ | – | ✓ | ✓ |
| F9 unpolished texture | 2 | 2 | 1 | 2 | 2 | 3 | 2 | 2 | 2 | 2 | 4 | 4 | 1 | 1 | 2 | 1 | 2 | 1 |
| F10 earned takeaway | – | – | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | – | ✓ | – | ✓ | ✓ | ✗ | – | ✓ | – |
| human · AI | 16·0 | 13·0 | 13·0 | 15·1 | 10·0 | 17·0 | 16·0 | 15·0 | 13·0 | 10·0 | 15·0 | 15·0 | 6·4 | 15·0 | 7·6 | 3·5 | 16·0 | 10·0 |
Cell codes — C1: lm lived moment, pc personal context, th thesis. C4: op open, pa partial, ti tidy. C5: pl named plainly, mu muted, pe performed. C7: cv conversational, fm formal, mx mixed. ✓ present, ✗ present but AI-flavored, – absent or not applicable. Numbers are 1–5 scores or counts.
What the matrix shows that the ranking can't: the structural verdicts alone catch 13, 15, and 16 (the orange columns), while posts 5, 11, and 14 stay green — their AI seams are cadence inside sentences, not structure, which is why they scored mid-range in the audit but look clean here. The pantry runbook column (7·6) echoes Zarema's independent 8·13 verdict.
Core features are directional: a draft drifting into an orange zone is a flag. Fingerprint features have no AI side; their absence only matters when the topic invited them. The most diagnostic cluster in this audit was C3 + C4 + C5 + C7 drifting together. Baselines were measured on the pre-cleanup corpus and will be re-derived after the rewrites.
Next steps
- Ground truth. Zarema marks which posts actually had AI help and how much survived; corrections to this audit also validate the method.
- Rewrite in priority order — 15, 13, 16, then 4 / 10 / 5, then 11 / 14 — one post at a time, bumping
contentModifiedDateper edit. - Re-derive the baseline in
.claude/skills/voice-fingerprint/baseline.mdfrom the cleaned corpus, then use/voice-fingerprinton every future draft.
Method: 20 discourse-level features (10 human-vs-AI separators, 10 personal fingerprints) induced from the corpus, scored per post in a single blind LLM pass. Scores are cadence judgments, not proof. Files: .claude/skills/voice-fingerprint/ — SKILL.md, baseline.md, audit-2026-08-09.md.