colemanok.com

Standalone · AI-systems forensics

Machine Pidgin

An AI-generated manga-recap video — twelve and a half hours of it — narrated itself into nonsense. Armor became blindage; a count became the condi; a woman’s reasoning became the female logic. Instead of laughing it off, I treated the garbage as evidence. From the corrupted output alone — no source code, no access to the tools that made it — I reconstructed the five-stage AI pipeline behind the video, built a working decoder for its private dialect, and pinned down which stage introduces each kind of failure. The short version of the skill: I find why a multi-stage AI system quietly produces garbage, working backward from the garbage.


The full analysis

Document purpose

This is a real-time linguistic analysis conducted over a twelve-hour AI-generated manga recap. What began as amusement at the phrase “not in danger of life” became a systematic decoding of machine-generated pidgin — revealing predictable failure modes, genre-weighted vocabulary selection, and emergent patterns that illuminate both the AI’s limits and the structure of the pipeline that produced it. The discovery process is preserved because how the patterns were identified matters as much as the patterns themselves.

Source material

Primary artifact: an AI-generated recap of Reborn as a Space Mercenary (12.5 hours), from a manhwa-recap channel. The hypothesized pipeline — five AIs in series, no human at any interface:

  1. Manga panels → image-parsing AI (object / scene recognition)
  2. Japanese text → translation AI (OCR + machine translation)
  3. Translated text → an LLM cleanup / formalization pass
  4. Cleaned text → text-to-speech
  5. Audio → caption AI (speech-to-text)

The critical observation: each AI is optimized for a different purpose, and each one’s output becomes the next one’s degraded input. Nobody is minding the seams.

Part 1 — The decoder dictionary

Confirmed mappings, built up by cross-referencing recurrence against scene context:

AI termActual meaningSeenDerivation
BlindageArmor / plating3+French WWI military terminology
CondiCount (noble title)3+Truncated Romance language (conte / conde)
ProthesisProsthesis1One letter dropped; confirmed by context (Duke’s cybernetic arm)
MaculatedStained / tainted1Correct archaic Latin, perfectly deployed
ContrasCons (pros and cons)1Latin root kept instead of the English contraction
Bright weapon riderKnight in shining armor1Idiom decomposition — each word translated separately
The sword’s throwSheath1Unknown derivation, possibly “throw” as covering
Immune handFilthy hand1Multimodal fusion — visual “sword bounced off” + text “filthy hand”
CarinchedChagrined + pinched1Vocabulary fusion — an AI-generated portmanteau
PrestativeAttentive / presenting1+Generated word that merely sounds formal
Rendered[object missing]1Terrifyingly incomplete — implies processing without stating the outcome
The female logicHer reasoning1Editorial insertion + gendered abstraction
TunedStunned1Phonetic truncation (st- dropped)
RepairFix (as in “eyes fixed”)1Polysemy failure — wrong definition selected
The High-roHiro (protagonist)manyName read as a title with an article
The elephantMimi (chibi-art character)1Image parser saw simplified features and guessed pachyderm
MotherboardMothership1Wrong technical domain — computer vs. naval
Data deviceBrooch1“Small important object in sci-fi = technology”
CreaturesMaids / servantsmanyNon-human features + servant role = zoological classification
RiderKnightmanyEtymology kept (ritter = rider), connotation destroyed

Caption AI variants (stateless phonetic rendering)

The final speech-to-text stage has no memory, so a single name shatters into several spellings:

Correct termCaption attempts
SkidbladnirSkid Blitner → Skidblader → Skidleneir
DarenwaldDarren Wald (a TTS micropause split the name)
ElfElfah → Alpha
Adieu“A Jews” (phonetic disaster)
SacrificingSacrifying (a plausible English pattern, faithfully preserved)

Part 2 — Failure-mode taxonomy

Seven tiers, ordered from comprehensible to unrecoverable. The tier tells you which stage to suspect:

TierWhat it isExampleUnderlying pattern
1 · SimpleComprehensible errors“stunned” → “tuned”Phonetic truncation, tense slips, trackable swaps
2 · ArchaicReverse-engineerable substitutionarmor → blindageReaches for formal / archaic register when context weights suggest nobility or military
3 · FusionConcept blending (needs context)“immune hand,” “carinched”Multiple inputs averaged into a chimeric output
4 · DomainWrong categorymothership → motherboardPulls from a higher-weight domain regardless of actual context
5 · EditorialThe AI has opinions“the female logic”Adds evaluative content not present in the source
6 · Semantic voidGrammatical, meaningless“The attackers were rendered.”Sentence-shaped objects with the load-bearing meaning absent
7 · Asspull landNo discernible source“the elephant” (chibi Mimi)Image parser having a film-studies moment, or outright hallucination

Part 3 — Pipeline architecture

Five non-communicating AIs, each competent at its own job:

ComponentOptimized forFailure mode
Image parserObject identificationNo narrative comprehension; chibi = elephant
Translation AILiteral accuracyNo idiom awareness; decomposes set phrases
LLM cleanupFormal proseOver-formalizes; reaches for archaic register
TTSPronunciationMicropause variation seeds downstream errors
Caption AIPhonetic plausibilityStateless; one name becomes three

The key insight: failures emerge at interfaces, not components. Each AI may be competent at its specific task. The catastrophe occurs when one system’s output becomes another’s input without shared context or error correction. The TTS pronounces “Skidbladnir” correctly — it knows Norse mythology. The caption AI has never heard the word and writes “Skid Blitner.” Both are functioning as designed. Together they produce nonsense.

The optimization-from-reality principle: the more an AI “knows” a word, the more likely it is to “help” by transforming it. Unknown words (fantasy names) pass through intact; common words get processed, associated, and mangled. “Darenwald” (made-up) is preserved; “Count” (common) becomes Condi. The AI is most dangerous with the words it recognizes.

Part 4 — The genre-weight hypothesis

“Blindage” recurred only in armor context; “condi” only for nobility; “maculated” and “contras” only in formal social scenes. The hypothesis: the pipeline accumulates context weights from story elements and pulls vocabulary from whatever register matches the perceived genre. Reborn as a Space Mercenary mixes four at once, and each lights up a different cluster:

Story elementWeight activatedVocabulary pulled
Spaceships, techSci-fi modernMotherboard, data device
Counts, dukes, duelsRomance-language nobilityCondi, Candessa, maculated
Fleet battles, armorWWI military technicalBlindage, formations
Love confessionsContemporary casualGirlfriend (from “maiden”)

Bidirectional normalization: the pipeline chases an imagined middle register — elevating anything too plain (“armor” → “blindage”) and modernizing anything too archaic (“maiden in love” → “passionate girlfriend”). That yields a testable prediction: a pure fantasy isekai through the same pipeline should trigger a different archaic cluster (medieval, ecclesiastical), let “knight” survive with no sci-fi pulling it toward “rider,” and produce different visual misreadings — while the gender roulette persists, because that’s pipeline-level, not genre-level.

Part 5 — Worked examples

“Mei recounts the prothesis with alarm.” Combat with a noble wielding a metal-cutting sword. “Recounts” is the wrong verb; “prothesis” sounds real; the Duke has a cybernetic arm. Decoded: Mei notices the prosthetic arm with alarm. Lesson: apparent Tier-7 nonsense is often Tier-2 corruption — context is what separates them.

“The duel has been maculated.” “Maculated” is a real word (the opposite of immaculate) and correctly describes a duel tainted by outside interference — nobody has used it since Shakespeare. The pipeline has access to deep vocabulary; deployment is context-weight driven.

“His eyes repair in pure panic.” “Fix” is polysemous — repair vs. become locked. Decoded: his eyes fixed in pure panic. Polysemy without disambiguation produces plausible-but-wrong output.

“A mouth-to-mouth smile.” Template “[body part] to [body part] smile” — the correct idiom is “ear to ear.” The AI knew the structure and grabbed the wrong slot filler.

“The attackers were quickly subjugated or rendered.” “Subjugated” is a complete thought; “rendered” is not — rendered into what? The missing object implies processing without outcome, and is accidentally the most chilling line in the video. A grammatically confident sentence with a semantic void at the center.

Part 6 — Vocabulary generated in analysis

Terms coined while decoding, offered for circulation:

TermDefinition
Machine pidginThe emergent language produced by AI translation pipelines
Optimized from realityThe drift that occurs when systems “improve” without understanding
Rendered[verb, terminal] to process without a stated outcome; ominous
Genre fingerprintThe error patterns that identify a story’s genre through its mistranslations
Stateless phonetic renderingCaption-AI behavior: no memory, no consistency, pure sound-to-grapheme
Vocabulary fusionPortmanteau creation through concept-blending (carinched)
Light rectangleWindow (AI-style household vocabulary)
The female logic[ironic] the unknowable; that which the AI advises us not to attempt to understand
Not giving blindage[phrase] said of a position that is indefensible

Part 7 — Methodology

How the decoding actually ran, step by step:

  1. Notice the anomaly — “that’s a weird word.”
  2. Form a hypothesis — “could this be [X]?”
  3. Request context — what was happening in the scene?
  4. Cross-reference — does the proposed meaning fit?
  5. Watch for recurrence — one instance is a possible error; three is a vocabulary entry.
  6. Build the decoder — add the confirmed mapping to the dictionary.
  7. Test predictively — does knowing “blindage = armor” help parse the next instance?

The pattern-recognition (across 10+ hours of exposure), the scene context, and the quality control (“that doesn’t match what I saw”) were mine; the archaic-vocabulary identification, cross-language reference, and systematic documentation came from working the problem in dialogue with an LLM. The same take-the-clock-apart impulse behind “why does this institution drift?” applied to “why does the AI say blindage? What’s the mechanism?”

Part 8 — Future research

The analysis is a first pass on a live investigation. Planned comparisons and open questions:

Appendix — the complete sentence

The demonstration that the decoder is complete, not just a catalog of jokes: a sentence that was pure noise at the top of this page is now fully parseable. If you’ve read the dictionary, you can read this.

“The condi’s blindage was maculated when the High-ro, acting as her rider and bright weapon rider, rendered the attackers with his immune hand while Mei recounted the prothesis with alarm, and one of the creatures prestatively offered him the sword’s throw as his eyes repaired in recognition of the female logic.”

Translation: The Count’s armor was stained when Hiro, acting as her protector and knight in shining armor, [processed] the attackers with his [insulted] hand while Mei noticed the prosthetic arm with alarm, and one of the maids attentively offered him the sheath as his eyes fixed in recognition of her reasoning.

You are now fluent in machine pidgin.


← All studies