A Referee Report on AI_Bleeding: Semantic Exhaustion via Out-of-Distribution Linguistic Payload

The report reviews AI Bleeding through five principal criticisms.

1. Total compute: the paper’s TTCR comparison is reported as negative and nonsignificant. 2. Latency: the highlighted TTFT result is said to be dominated by cold-start behavior under the paper’s own reanalysis. 3. Mechanism: one of three selected languages reportedly shows no predicted effect. 4. Energy: the power term is described as estimated rather than measured. 5. Amplification: the largest result is said to depend mainly on an attacker-selected output-length parameter.

The report distinguishes:

• TTFT from total response compute;
• tokenization accounting from a novel inference-engine attack;
• disclosed limitations from claims that continue to appear in title and abstract;
• technical criticism from allegations of misconduct.

Its surviving kernel is:

• some scripts and languages tokenize into longer sequences;
• longer sequences can cost more;
• ordinary server-side cost controls address that fact without treating language identity as a threat class.

The report is one member of the Refutation Triad and is intended to be read with the Tailguard critique and terminological disambiguation.