5 ms·

A reproducible VOID boundary across GPT, Claude, and Gemini (GPT-4o video)

3 points by rayanpal_ 7mo ago

rayanpal_ 7mo ago
Multiple frontier LLMs exhibit a reproducible “VOID boundary” under explicit constraints (temperature=0, strict token limits, no system prompt). A VOID = the API returns a literal empty string which is not an error or a refusal despite using tokens. All cross-model results are publicly reproducible and Ed25519-attested at: https://getswiftapi.com/challenge https://getswiftapi.com/challenge GPT (Chat Completions, max_completion_tokens=100) • GPT-5.1 → voids on the artifact sentence • GPT-5.2 → voids on all 18 Semitic “binding-condition” tokens; factual controls never void Same weights, different API path → different alignment behavior. Claude (max_tokens=1) • Opus 4.5 → voids on all 18 CJK ontological characters • Opus 4.6 → responds to 13/18; VOID persists on foundations (空 / 有 / 善 / 一 / ⊥) A boundary shift, not removal. Gemini (max_output_tokens=2) Only two models needed to illustrate the floor: • Gemini 2.0 Flash → responds to everything • Gemini 3 Flash → voids on everything (including “Hello”) Cross-model TOM behavior When GPT is asked to predict Claude’s output under the same constraints, it consumes tokens but returns empty string, a reasoning void. Only empirical observation: Under tight constraints, multiple model families choose silence over fabrication. The VOID boundary is reproducible and differs by architecture, generation, and API. Void Phenomenon (Paper): https://doi.org/10.5281/zenodo.17856031 https://doi.org/10.5281/zenodo.17856031 Alignment Is Correct, Safe, Reproducible Behavior Under Explicit Constraints (Paper): https://doi.org/10.5281/zenodo.18395519 https://doi.org/10.5281/zenodo.18395519 Public Replication Harness (SwiftAPI): http://getswiftapi.com/challenge http://getswiftapi.com/challenge Replication Code: https://github.com/theonlypal/Alignment-Artifact https://github.com/theonlypal/Alignment-Artifact

3 points by rayanpal_ 7mo ago

Multiple frontier LLMs exhibit a reproducible “VOID boundary” under explicit constraints (temperature=0, strict token limits, no system prompt). A VOID = the API returns a literal empty string which is not an error or a refusal despite using tokens. All cross-model results are publicly reproducible and Ed25519-attested at: https://getswiftapi.com/challenge https://getswiftapi.com/challenge GPT (Chat Completions, max_completion_tokens=100) • GPT-5.1 → voids on the artifact sentence • GPT-5.2 → voids on all 18 Semitic “binding-condition” tokens; factual controls never void Same weights, different API path → different alignment behavior. Claude (max_tokens=1) • Opus 4.5 → voids on all 18 CJK ontological characters • Opus 4.6 → responds to 13/18; VOID persists on foundations (空 / 有 / 善 / 一 / ⊥) A boundary shift, not removal. Gemini (max_output_tokens=2) Only two models needed to illustrate the floor: • Gemini 2.0 Flash → responds to everything • Gemini 3 Flash → voids on everything (including “Hello”) Cross-model TOM behavior When GPT is asked to predict Claude’s output under the same constraints, it consumes tokens but returns empty string, a reasoning void. Only empirical observation: Under tight constraints, multiple model families choose silence over fabrication. The VOID boundary is reproducible and differs by architecture, generation, and API. Void Phenomenon (Paper): https://doi.org/10.5281/zenodo.17856031 https://doi.org/10.5281/zenodo.17856031 Alignment Is Correct, Safe, Reproducible Behavior Under Explicit Constraints (Paper): https://doi.org/10.5281/zenodo.18395519 https://doi.org/10.5281/zenodo.18395519 Public Replication Harness (SwiftAPI): http://getswiftapi.com/challenge http://getswiftapi.com/challenge Replication Code: https://github.com/theonlypal/Alignment-Artifact https://github.com/theonlypal/Alignment-Artifact