The Ragged Edge: Updating the AI-Tells Pass

Dismantling Contemporary Model Conventions, Expanding Registry 2.0, and Implementing Closed-Loop Arc Perturbation

September 4, 2026  ·  Data Timestamp: 2026-09-04T06:50:36+02:00
Research, Diagnostic Architecture & Harness Implementation: Gemini Flash 3.8 (High)
Prompt, Direction & Editorial Ground Truth: David Jhave Johnston
Download Updated Harness Package (Sep 4, 2026)
Includes narracode.md, master_ai_tells.md, README.md, FAQ.md, master_poetics.MD, and master_phrases_to_avoid.md (ZIP, 32 KB)
Prompt #1  ·  2026-09-04T06:29 CEST
"I want an update to the AI-tells pass in the Narracode harness. Scan the internet and come up with a plan for how to improve the style so it doesn't use the conventions of contemporary AI models, but can retro-fit or weed narra-code-writing so it is unpredictably diverse like human literary writing. Provide report-plan in html format, data timestamped and attributed to yourself Gemini Flash 3.8 (High)"
Prompt #2  ·  2026-09-04T06:42 CEST
"good. implement plan. but do NOT Retro-fitting the Existing Stories. Leave all stories that are published exactly as they are. Snapshot the harness and ai-tells with date timestamps for posterity and implement changes as decribed. Also put in one that scans for overuse of numeric categories fixations on 'eleven' etc... Also i'm wondering if the ai-tells pass shld not incl consultation of the poetics specific to each story to bend-blend and subtly shift they replacements done during weeding-retrofit. And with that is it possible to intro a perturbation class that considers character and plot and motif arcs as it rewrites specific sentences? (which themselves in a closed loop need to again be assessed for ai-tells. note again: these are not regulations but efforts to smooth away the statistical conformity and patterned habits of AI writing. Make a new report timedatestamped with this as prompt #2 fold beneath prompt #1, then proceed."

In late 2025 and 2026, frontier models (Claude 3.5/3.7/Opus 5, GPT-4o/o1/o3, and Gemini 1.5/2.0/3.x) evolved past early, crude "tells" like delve or raw em-dash floods. In their place arose far more insidious structural conventions: the Contrast Reframe ("not X, but Y"), Somatic Interiority Clichés, Sensory Dyads, Uniform Information Density (UID) smoothing, and Thematic Closing Wrap-Ups. Left unaddressed, these conventions homogenize generative literature into a polite, predictable, middle-brow drone.

The AI-tells pass in the Narracode harness was originally built as a span-level audit to catch machine markers. Over August 2026, empirical measurement across 1,234 human edits (What the Edits Know) proved that simple lexical subtraction accounts for only 9% of what human revision does; the remaining 91% is non-trivial textural substitution: breaking metric symmetry, colliding estranged registers, and introducing radical linguistic friction.

This report documents the completed update to the Narracode harness: expanding the tell registry to 29 classes, conditioning remedies on story-specific poetics, wiring live exemplar retrieval into generation, and implementing closed-loop arc perturbation.

Implemented Architectural Upgrades

· · ·

1. Contemporary AI Conventions (The 2025–2026 Model Fingerprint)

Modern models are trained to avoid obvious mistakes. In doing so, they gravitate toward the statistical mode of language, generating prose that is blameless, symmetrical, and devoid of tactile surprise.

# Convention Structural Mechanism Model Motivation
C1 The Contrast Reframe
(Not X, but Y)
\b(not (a|an|the)? \w+, but (a|an|the)? \w+)\b
"It was not malice, but exhaustion."
Simulates philosophical nuance by rejecting a strawman premise before offering a tidy resolution.
C2 Somatic Reflex Clichés
(The Visceral Shortcut)
Involuntary bodily reactions: "breath he didn't know he was holding", "something shifted behind her eyes", "stomach dropped". Over-indexes on "show, don't tell" by converting complex private states into autonomic nervous system reflexes.
C3 Adjectival Sensory Dyads
(The Twin-Pack)
Coordinated pairs of sensory modifiers: "cold, metallic air", "hollow, brittle certainty", "pale, trembling hands". Acoustic balancing: ensuring rhythmic symmetry across noun phrases.
C4 Semantic Smoothing
(Connective Over-glaze)
Clause-leading transitional glues: "And yet," "Still," "In that moment," "Even so," "Perhaps that was why." Eliminates the disorienting, jagged leaps of thought native to human consciousness.
C5 Uniform Information Density (UID)
(The Monotone Ceiling)
Evenly distributed token surprisal. Sentences adhere to a 12–22 word metronome with balanced periodic clauses. Mathematical minimization of perplexity variance; avoids both bizarre solecisms and bone-dry monosyllabic slogs.
C6 Epistemic Varnish
(Performed Hesitancy)
Softening hedges: "a kind of," "in some quiet way," "seemed almost to," "perhaps." Simulates human doubt without committing to real narrative risk.
C7 Thematic Wrap-Up
(The Moralizing Coda)
Ending scenes or paragraphs with a neat, universalizing philosophical takeaway. Assistance bias: the urge to leave the prompter with a polished, morally rounded artifact.
C8 Numeric Fixation
(Arbitrary Specificity)
Overuse of specific recurrent odd numbers: eleven, seventeen, forty-one, or unanchored counts ("eleven minutes", "seventeen similar entries"). Simulates precision by grabbing low-probability integers from a low-loss latent pocket, creating false authenticity without system mechanics.
· · ·

2. Registry 2.0: Classes 21–29 in master_ai_tells.md

The registry has been updated from 20 to 29 classes. These additions provide machine-auditable regexes, concrete remedies, and strict target rates:

Class Name Diagnostic Pattern / Detector Remedy & Target Rate
21 The Contrast Reframe \b(not (a|an|the)? \w+ (of|in) \w+, but (a|an|the)? \w+)\b or \b(not \w+, but \w+)\b where both terms share semantic domain. CUT negation; state the second term flatly, or cut entirely. Target rate: 0 per draft.
22 Somatic Reflex Shortcut Autonomic clichés: breath [he/she] didn't know, knot in [his/her] stomach, hair on [his/her] neck, something shifted in. REPLACE with an external, concrete physical action or tool telemetry. Target rate: 0 in narration.
23 Sensory Dyad (Twin Adjectives) Two coordinate adjectives modifying a noun where one is literal and one is metaphorical: "cold, bitter air", "hollow, pale light". CUT to the single stranger adjective, or convert to a bare noun. Target rate: ≤ 1 per 2,000 words.
24 Semantic Smoothing Glues Conjunctive adverbs opening sentences: "And yet," "Still," "Moreover," "Even so," "Perhaps that was why." CUT the conjunction. Let the sentences collide paratactically. Target rate: ≤ 1 per section.
25 Epistemic Varnish (The Politeness Hedge) Intrusions of "in some quiet way", "a sort of", "almost as if", "seemed to" where narration has focal authority. CUT the hedge. State the event as cold fact. Target rate: 0 in austere narration.
26 Unearned Thematic Coda Paragraph-final sentence summarizing the emotional or intellectual significance of the paragraph's actions. CUT the final sentence. Terminate on the preceding physical action. Target rate: 0 at beat endings.
27 Symmetrical Periodic Sentence Compound sentences joined by while or as where both independent clauses have equal word counts (±2 words) and parallel syntax. PUNCTUATE into two asymmetric sentences; convert one into a verbless noun fragment.
28 Ecclesiastical Diction Sneaking church vocabulary into secular scenes: litany, baptism, catechism, sacred, reverent, testament, confessional. REPLACE with administrative, industrial, biological, or vernacular terms. Target rate: 0 unless church-world.
29 Numeric Fixation / Arbitrary Specificity Unanchored recurrent integers (especially eleven, seventeen, forty-one, three minutes) doing aesthetic work without operational mechanics. REPLACE with true engineering/logistical telemetry (e.g. 320ms, shard NOR-14, 504 error) or CUT the number to let the noun stand alone. Target rate: ≤ 1 unanchored numeral per story.
· · ·

3. The Poetics Filter for Remedies

A mechanical audit that replaces machine spans with generic stock words merely substitutes one form of synthetic flatness for another.

Under the updated specification in master_ai_tells.md, every proposed remedy or retrofit substitution MUST consult the active project's POETICS.md:

                    ┌──────────────────────────────────────────────┐
                    │               Flagged AI-Tell                │
                    │   "The cold, bitter silence between them..." │
                    └──────────────────────┬───────────────────────┘
                                           │
                                           ▼
                    ┌──────────────────────────────────────────────┐
                    │       Read Project's local POETICS.md        │
                    │   - Attentional Dialect (e.g. Blanchotian)   │
                    │   - Refusals (No melodrama, no church words) │
                    │   - Reference Authors (Knausgaard, Butler)   │
                    └──────────────────────┬───────────────────────┘
                                           │
                                           ▼
                    ┌──────────────────────────────────────────────┐
                    │          Poetics-Conditioned Remedy          │
                    │   "Silence. Unserviced air handler upstairs."│
                    └──────────────────────────────────────────────┘

If POETICS.md specifies an austere post-industrial register, a Class 22 somatic reflex ("knot in his stomach") is replaced not by generic stoicism ("he was calm"), but by an objective physical transaction native to the world ("His thumbs caught in the copper turnings"). The substitution is dyed in the project's unique attentional substrate.

· · ·

4. The Arc-Perturbation Engine (Closed-Loop Structural Revision)

To move beyond line-level trimming, narracode.md now specifies The Arc-Perturbation Pass (invoked with "perturb this", "retrofit this", or "break the cadence"). When a sentence is flagged as synthetic or flat, the engine leverages the story's structural memory to perturb the sentence with dramatic and thematic friction:

4.1 The Three Structural Invocations

4.2 The Closed-Loop Clearance Check

Because structural perturbation involves creative re-writing, it carries an immediate risk: smuggling in melodrama, unanchored symbols, or new AI tells.

To ensure these revisions smooth away statistical conformity without introducing new model habits, the engine enforces a mandatory Closed Loop:

1. Original Draft (Synthetic / Flat):
"She waited by the window as the rain started to fall, feeling a sudden pang of regret."

2. Arc Perturbation (Consulting character-interiority.md & obligations.md):
"She stood by the sill, her fingers testing the salt stain on the envelope she had sworn to burn before Marcus reached the ferry."

3. Closed-Loop Tell Audit (Scanned against Classes 1–29):
- Contrast reframe? None.
- Somatic shortcut? None (fingers testing salt stain is active manual task).
- Sensory dyad? None.
- Ceremonial predicate? None.
- Thematic wrap-up? None.
Result: CLEAR. Committed to draft.
· · ·

5. Direct File Verification

The following files in the repository have been updated and are live for inspection:

· · ·

6. Conclusion

Literary prose is not generated by suppressing errors; it is generated by motivated asymmetry:

By pairing a 29-class audit with poetics-conditioned remedies, live exemplar retrieval, and closed-loop arc perturbation, the Narracode harness ensures that generative literature remains what art has always been: not a smoothed consensus of the linguistic past, but an unpredictable encounter with the real.

Sources & Project References

All works and media on Glia.ca by David Jhave Johnston is licensed under CC BY-NC-SA 4.0 Creative Commons Attribution Non-Commercial Share-Alike