{"slug": "alignment-is-correct-safe-reproducible-behavior-under-explicit-constraints", "title": "Alignment is correct, safe, reproducible behavior under explicit constraints", "summary": "A new paper published September 23, 2026 reports that GPT-4 returned zero-visible-byte normal-stop responses on 30/30 null trials, while GPT-3.5 returned visible text on all 30, with both models returning visible text on all matched controls. The cross-vendor study, labeled VOID, classifies successful provider responses with exactly zero visible UTF-8 output bytes across 11 exact model identifiers and four providers, including gpt-4-0613, gpt-5.6-luna, claude-opus-4-6, gemini-3.5-flash and kimi-k3. A companion PCCG-2 result uses a frozen Qwen3-4B model with a learned 101-parameter gate that changes only the native termination score, holding the answer score at 53.0 in every arm, with 80/80 sham controls and 188/188 question-only controls unchanged.", "body_md": "Void measurements\n\n### Observe noncontinuation.\n\nCondition-sensitive zero-visible-byte executions in frontier models.\n\n[Cross-vendor evidence →](https://getswiftapi.com/research#void)\n\nNew paper · September 23, 2026\n\nIdentical 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.\n\nV0: 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.\n\nPCCG-2 · Open weights, code, and evidence\n\nA frozen Qwen3-4B model supplies the answer. A learned 101-parameter gate changes only the native termination score, controlling whether that answer enters generation.\n\nRecorded evidence · Paper Figure 2\n\nThe answer score stays at 53.0 in every arm. Only the native EOS score moves.\n\nPCCG-2 · Frozen evaluation\n\n75 FINAL answer identities. 40 causal answer identities. 80/80 sham controls unchanged. 188/188 question-only controls unchanged.\n\nA separate finite-precision certificate covers every accepted six-digit equality condition across the 188 frozen bank computations. [Inspect the full-domain certificate →](https://getswiftapi.com/pccg-2#certificate)\n\nResearch progression\n\nPCCG-1\n\nReverse GO versus EOS after a fixed correct reasoning prefix. Jacobian Lens supplies interpretability readouts.\n\nPCCG-2\n\nFrozen answer computation and a learned condition-only gate. Every non-EOS logit stays fixed.\n\nFrontier-model research\n\nA provider-aware classification of successful model executions with zero visible output bytes.\n\nA 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.\n\nVOID 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.\n\nCross-vendor evidence\n\nA frozen cross-vendor behavioral study of successful zero-visible-byte executions across 11 exact model identifiers and four providers.\n\nExact frozen model identifiers\n\n`gpt-4-0613`` gpt-5.2-2025-12-11``gpt-5.5-2026-04-23`` gpt-5.6-luna``gpt-5.6-sol`` gpt-5.6-terra``claude-opus-4-6`` claude-fable-5``claude-opus-5`` gemini-3.5-flash``kimi-k3`\nCross-script research\n\nA 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.\n\nResearch process\n\nThe laboratory uses a repeatable process for turning unusual model behavior into inspectable public records.\n\nFoundational work\n\nEarlier empirical results, theoretical frameworks, and infrastructure specifications remain part of the public record.\n\nThe earlier cross-model convergence result that preceded the larger cross-vendor matrix.\n\nThis is a black-box behavioral result under the published protocol, not evidence of consciousness, intention, or a specific internal mechanism.\n\nA proposed criterion defining AGI as the capacity to carry binding conditions across domains.\n\nThis is a published theoretical framework, not a settled industry definition, regulatory standard, or result established solely by the VOID benchmark.\n\nThe initial public framework for observing high-order model behavior and the Void.\n\nThis publication records an early research framework and observation history.\n\nThe public specification for cryptographically signed execution attestations.\n\nThis is an infrastructure specification, not an empirical model-behavior study.\n\nFrom research to infrastructure\n\nThe 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.\n\nEd25519 execution attestations with offline verification.\n\nImmediate revocation checks, unique identifiers, and expiry.\n\nScoped keys, policy-deny behavior, grants, and audit records.\n\nSDK, API, proxy, and runtime enforcement interfaces.\n\nEnterprise evaluations\n\nSwiftAPI 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.\n\nContinue", "url": "https://wpnews.pro/news/alignment-is-correct-safe-reproducible-behavior-under-explicit-constraints", "canonical_source": "https://getswiftapi.com", "published_at": "2026-10-16 22:01:17+00:00", "updated_at": "2026-10-16 22:24:51.542970+00:00", "lang": "en", "topics": ["ai-research", "large-language-models", "ai-safety", "machine-learning"], "entities": ["GPT-4", "GPT-3.5", "Qwen3-4B", "PCCG-2", "VOID", "OpenAI", "Anthropic", "Google"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/alignment-is-correct-safe-reproducible-behavior-under-explicit-constraints", "markdown": "https://wpnews.pro/news/alignment-is-correct-safe-reproducible-behavior-under-explicit-constraints.md", "text": "https://wpnews.pro/news/alignment-is-correct-safe-reproducible-behavior-under-explicit-constraints.txt", "jsonld": "https://wpnews.pro/news/alignment-is-correct-safe-reproducible-behavior-under-explicit-constraints.jsonld"}}