Hangul Integrity Investigation Pipeline

From field impression to root cause: observation, controlled datasets, layered detection, adversarial rebuttal, verdict

Hangul Integrity Investigation Pipeline From field impression to root cause: observation, controlled datasets, layered detection, adversarial rebuttal, verdict 01 / Observation 02 / Hypotheses + Experiments 03 / LLM Judge 04 / Heuristic Detection 05 / Adversarial Adjudication 06 / Verdict + Actions EX / Rebutted Flags Field Report · Hangul syllable swaps · Observation · Claude Code Field Report Hangul syllable swaps Claude Code Hypotheses · H-fill / H-path / H2 · Hypotheses + Experiments Hypotheses H-fill / H-path / H2 Runs D1-D6 · 1P vs Bedrock · Hypotheses + Experiments · 1,071 runs Runs D1-D6 1P vs Bedrock 1,071 runs Heuristic Probe · high precision · Heuristic Detection Heuristic Probe high precision LLM Judge · high recall, noisy · LLM Judge LLM Judge high recall, noisy Byte Forensics · backslash trichotomy · Adversarial Adjudication Byte Forensics backslash trichotomy 3-Lens Rebuttal Panel · unanimous to survive · Adversarial Adjudication 3-Lens Rebuttal Panel unanimous to survive Class A · escape-miscoding · Verdict + Actions · 45/45 vs 1/60 Class A escape-miscoding 45/45 vs 1/60 Mitigate · literal UTF-8 rule · Verdict + Actions Mitigate literal UTF-8 rule Rejected · 96% of judge flags · Rebutted Flags Rejected 96% of judge flags Legend User UI Agent logic Policy Context / trace Cloud service External system

Flags Are Not Facts

  • • Raw detector and judge flags only nominate candidates
  • • 96% of judge flags died under adversarial review
  • • Only unanimous 3-lens survivors become verdicts

Byte-Level Ground Truth

  • • Backslash trichotomy separates model vs logger escapes
  • • Doubled backslash is the decisive guilty fingerprint
  • • Parsed toolUse.input destroys the fingerprint

Verdict Drives Action

  • • Class A blocked by a literal UTF-8 output instruction
  • • Fill and path hypotheses rejected by flat dose curves
  • • Class B residuals escalated to AWS with RequestIds