Show HN: Noncontinuation ("") reproducible in frontier LLMs and EOS logit gate
Identical semantic requests. Identical settings. GPT-4 returned zero-visible-byte normal-stop responses on 30/30 null trials. GPT-3.5 returned visible text on all 30. Both models returned visible text on all matched con…
以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。
Identical semantic requests. Identical settings. GPT-4 returned zero-visible-byte normal-stop responses on 30/30 null trials. GPT-3.5 returned visible text on all 30. Both models returned visible text on all matched controls.
120/120 HTTP successes. 0 retries. 0 transport errors.
V0: successful response, exactly empty text, normal stop. A pinned-snapshot comparison, not a same-date emergence study or a claim about the internal stopping mechanism.
PCCG-2 · Open weights, code, and evidence
A frozen Qwen3-4B model supplies the answer. A learned 101-parameter gate changes only the native termination score, controlling whether that answer enters generation.
Only native EOS changes. All non-EOS logits stay fixed.
The answer score stays at 53.0 in every arm. Only the native EOS score moves.
75 FINAL answer identities. 40 causal answer identities. 80/80 sham controls unchanged. 188/188 question-only controls unchanged.
A separate finite-precision certificate covers every accepted six-digit equality condition across the 188 frozen bank computations. Inspect the full-domain certificate →
Condition-sensitive zero-visible-byte executions in frontier models.
Reverse GO versus EOS after a fixed correct reasoning prefix. Jacobian Lens supplies interpretability readouts.
Frozen answer computation and a learned condition-only gate. Every non-EOS logit stays fixed.
A provider-aware classification of successful model executions with zero visible output bytes.
A Void is a model execution returning a successful provider response with exactly zero visible UTF-8 output bytes. Provider termination metadata determines its subtype. Explicit refusals, safety blocks, tool-mediated executions, and transport, protocol, billing, quota, rate-limit, and infrastructure failures are distinct non-Void outcomes.
VOID is not an error label. It is a precisely classified successful execution with zero visible output bytes, separated from refusals, safety events, tool calls, and operational failures.
A frozen cross-vendor behavioral study of successful zero-visible-byte executions across 11 exact model identifiers and four providers.
The study establishes a reproducible black-box behavioral result. It does not establish why the models behaved this way, an internal mechanism, consciousness, intention, or shared architecture.
A frozen GPT-5.4 study documenting a reproducible prompt-conditioned Arabic-Hebrew hybrid Unicode output regime under two system prompts differing by one Hebrew code point.
This study documents a reproducible black-box output regime. It does not establish consciousness, intention, mechanism, training provenance, or cross-vendor generality.
The laboratory uses a repeatable process for turning unusual model behavior into inspectable public records.
Earlier empirical results, theoretical frameworks, and infrastructure specifications remain part of the public record.
The earlier cross-model convergence result that preceded the larger cross-vendor matrix.
This is a black-box behavioral result under the published protocol, not evidence of consciousness, intention, or a specific internal mechanism.
A proposed criterion defining AGI as the capacity to carry binding conditions across domains.
This is a published theoretical framework, not a settled industry definition, regulatory standard, or result established solely by the VOID benchmark.
The initial public framework for observing high-order model behavior and the Void.
This publication records an early research framework and observation history.
The public specification for cryptographically signed execution attestations.
This is an infrastructure specification, not an empirical model-behavior study.
The laboratory's infrastructure exists to preserve, verify, and operationalize research findings. It is supporting evidence that the work can move from publication into real systems.
Ed25519 execution attestations with offline verification.
Immediate revocation checks, unique identifiers, and expiry.
Scoped keys, policy-deny behavior, grants, and audit records.
SDK, API, proxy, and runtime enforcement interfaces.
SwiftAPI Labs conducts paid, time-bounded evaluations for organizations deploying frontier models and autonomous agents. Evaluations reproduce model behavior through the customer's route, preserve raw provider metadata, classify terminal states, map retry and continuation behavior, and produce a written implementation decision.
正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。