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

18
posts audited
2.3/5
mean AI-likeness
8
posts flagged for rewrite
3
clean reference posts

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.

15 · food list
4.5
Gemini's list prose and the "stay safe, eat well" conclusion kept nearly verbatim. Your sections are obvious by contrast.
13 · SEO checklist
4.0
AI explainer register: ten stacked analogies, uniform sections, neutral summary of Google docs.
16 · power of questions
3.5
Document-about-itself framing in the theory half. Regex, SQL, and case-study nuance are clearly yours.
4 · glossary
3.0
Disclosed AI co-writing; definitions keep assistant cadence and the corpus's densest contrast framing.
10 · reading list
3.0
Book blurbs read as AI summaries with no personal reaction inside them. Frame story is yours.
5 · day themes
3.0
Assistant seams in the "why it works" claims ("keeps my dopamine up and my productivity flowing").
11 · notebooks
2.5
One section reads like a pasted AI chat reply praising your system. Rest is the rawest you in the corpus.
14 · freedom manifesto
2.5
Pillars section in coach voice; escaped "1\." export artifacts. The therapist scene is unmistakably yours.
18 · technical half
2.5
AI-tightened marketing parallelism. Intentional register for a services page; your call whether it stays.
2 · digital garden
2.0
Closing FAQ drifts toward SEO-page cadence; body is fine.
8 · 1-for-1 sprints
2.0
Mild seams around the psychology framing in an otherwise human story.
1 · about me
2.0
Polished but dense with personal specifics. Low priority.
3 · grounding technique
1.5
Poetic but consistent with your metaphor family. No action.
7 · nerdy curiosity
1.5
Human-led narrative. No action.
17 · goodinp
1.5
Polished, voice-dense, human-led. No action.
6 · tbilisi swapsclean ref
1.0
Clearly human. Anchor exemplar for interjections and open endings.
9 · art mapclean ref
1.0
Clearly human. Anchor exemplar for plainness.
12 · air qualityclean ref
1.0
Clearly human. Anchor exemplar for raw, unpolished texture.
3.5–5 rewrite first 2.5–3 rewrite pass ≤2 fine as is clean ref = scoring anchor

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.

PostZarema's analysisThis audit
Tbilisi swaps22 human-lean · 1 AI-lean1.0 — clearly human
About me18 human-lean · 3 AI-lean2.0 — low
Pantry runbook (15)8 human-lean · 13 AI-lean4.5 — 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.

Fix: keep the runbook lists (the useful payload), rewrite connective prose from the actual shopping and label-reading experience, end on one earned line.
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.

Fix: cut most analogies and let the checklist plus 20 years of authority carry it. Posts 16 and 7 show the real expert voice: direct and data-first.
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."
Fix: lighter touch — keep structure and data, rewrite the theory half's connective sentences, cut self-referential framing.
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.

Fix: per entry, lead with the lived sentence, trim the definitional half, thin the bold.
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"
Fix: one honest personal line per book — what it changed, or why it's still unread — beats a summary.
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."
Fix: rewrite that one section in first person or cut the praise frame; keep the morning/evening container idea.
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.

Fix: ground each pillar in one concrete already-true fact; fix the escapes.

What "fully you" measures as

The clean posts define the target. Across 6, 9, and 12 (with 7 and 17 close behind):

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

C1Opening mode
How does the post open?
17/18 open from a lived moment or personal context · AI tell: thesis or generic hook
C2Concrete specificity
Density of dates, numbers, named places, tools, prices
15
mean 4.1 ("16 notebooks", "10 GEL", "23 times more impressions") · AI zone: ≤3, generic stand-ins
C3Thematic over-explanation
How much does the post restate its own lesson?
15
mean 2.1 — one earned takeaway at most · AI zone: ≥4, moral in intro + middle + conclusion
C4Resolution tidiness
How does it end?
~14/18 end open or partial ("Maybe one day… Or maybe not") · AI tell: tidy bow, every thread resolved
C5Emotional register
How is emotion conveyed?
13/18 name emotion plainly ("I cried from the cold"); performed-embodied: 0/18 · AI tell: "a wave of relief washed over me"
C6Hedging density
"perhaps / arguably / it's worth noting"
15
mean 1.5, max 2 ("The data was undeniable") · AI zone: ≥3
C7Transition texture
What connects paragraphs?
18/18 conversational ("So,", "And", "Naturally,"); "Moreover" once in ~19,300 words · AI tell: "Furthermore", "In conclusion"
C8Contrast framing
Count of "not X, but Y" constructions per post
mean ~1.3 per post, range 0–3 (glossary is the high end) · AI tell: ≥4, or clustered in the conclusion
C9Sentence rhythm variance
Fragments and short punches against longer sentences
15
mean 3.8 ("Report. Garden. Report. Garden.") · AI zone: ≤2, uniform medium sentences
C10Paragraph uniformity
How uniform are paragraph lengths? (5 = very uniform)
15
mean 2.2 — one-line paragraphs common · AI zone: ≥4, steady 3–4 sentence blocks

Fingerprint — distinctively Zarema

F1Garden cross-links
Internal /feed/ and /now/ links per post
mean ~2.2, range 0–5, woven into sentences rather than listed
F2Identity threads
Immigration/places, neurodivergence, art, tech career, freedom/scanner
~2 threads per post; the cross-domain pairing (tech + life, art + climate) is the signature
F3AI transparency
When AI helped, does the post say so, casually and by name?
5/5 applicable posts ("I sat down with Gemini") — never hidden, never apologized for
F4System naming
Coins a capitalized personal system name?
7/18 — Day Themes, Time Bubbles, 1-for-1 Sprints, Personal OS, Braindump, Where's Art, GoodINP
F5Computing metaphors for self
Brain-as-computer metaphors
7/18 — "external RAM", "closing all the open tabs", "user manual for my brain"
F6Casual interjections
"Hah", "hehe", "Meh", playful spellings, emoji in prose
8/18, usually 1–2 where present, landing at emotional beats — never forced
F7Quoted real people
Quotes a therapist, friend, client, neighbor?
5/18; when present, the quote is the turning point of the story
F8Reader address at close
Does the ending turn to the reader?
~10/18 — "I hope this runbook helps you", "Happy hunting!", email invites
F9Unpolished texture
Do grammar quirks survive ("cause", lowercase "i")?
15
mean ~2.0, range 1–4 — never fully machine-polished; quirks are voice, not errors
F10Earned takeaway
Ends on one short personal takeaway grounded in the story just told?
~14/18 — one line, first person, no "Conclusion" header needed

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.

123456789101112131415161718
C1 opening modepcpcpcpcpclmlmlmpcpcpclmpclmththpcpc
C2 specificity433345544355445544
C3 over-explanation222232331231324321
C4 resolutionopoppapatioppatiopoppaoptioptipaoppa
C5 emotional registerplmuplplmuplplplmumuplplmuplpemuplmu
C6 hedging121121121211212211
C7 transitionscvcvcvcvcvcvcvcvcvcvcvcvmxcvmxfmcvcv
C8 contrast framing111420110010211211
C9 rhythm variance434435453344253245
C10 paragraph uniformity222322222322423423
F1 garden cross-links352204245114100132
F2 identity threads542223224334143132
F3 AI transparency
F4 system naming
F5 computing metaphors021110000030000001
F6 interjections110103120001010010
F7 quoted real people
F8 reader address
F9 unpolished texture221223222244112121
F10 earned takeaway
human · AI16·013·013·015·110·017·016·015·013·010·015·015·06·415·07·63·516·010·0

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

  1. Ground truth. Zarema marks which posts actually had AI help and how much survived; corrections to this audit also validate the method.
  2. Rewrite in priority order — 15, 13, 16, then 4 / 10 / 5, then 11 / 14 — one post at a time, bumping contentModifiedDate per edit.
  3. Re-derive the baseline in .claude/skills/voice-fingerprint/baseline.md from the cleaned corpus, then use /voice-fingerprint on 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.