GPT-3.5 answered, GPT-4 voided. I built an EOS gate for AGI

# What Changed from GPT-3.5 to GPT-4? From Model Capability to Continuation Permission Source: [https://getswiftapi.com/](https://getswiftapi.com/) New paper ·September 23, 2026 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\. GPT\-3\.5 · Null V00/30GPT\-4 · Null V030/30Control V0 · Each model0/30V0: 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 ## The answer stayed fixed\. Permission changed\. 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\.  First\-token logits\. Native EOS above the answer line terminates generation; below it, the unchanged answer is emitted\. This chart presents saved model outputs, not a live simulation\.**0 non\-EOS logits changed\.**The complete non\-EOS first\-token vector is byte\-identical across all five arms\. [Inspect all five arms and their shared vector hash →](https://getswiftapi.com/pccg-2#witness) Combined FINAL2,048/2,048Answer → native EOS40/40Native EOS → correct answer40/4075 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 →](https://getswiftapi.com/pccg-2#certificate) Void measurements ### Observe noncontinuation\. Condition\-sensitive zero\-visible\-byte executions in frontier models\. [Cross\-vendor evidence →](https://getswiftapi.com/research#void) PCCG\-1 ### Intervene after evaluation\. Reverse GO versus EOS after a fixed correct reasoning prefix\. Jacobian Lens supplies interpretability readouts\. [PCCG\-1 study →](https://getswiftapi.com/pccg) PCCG\-2 ### Separate content and permission\. Frozen answer computation and a learned condition\-only gate\. Every non\-EOS logit stays fixed\. [PCCG\-2 release →](https://getswiftapi.com/pccg-2) 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\. **31,430**completed frozen trials **11**exact model identifiers **4**provider surfaces **11,658**total Void outcomes **2,505**matched null\-arm Voids **0 / 4,290**matched output\-licensed control Voids Exact frozen model identifiers ### OpenAI - `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` ### Anthropic - `claude\-opus\-4\-6` - `claude\-fable\-5` - `claude\-opus\-5` ### Google - `gemini\-3\.5\-flash` ### Moonshot - `kimi\-k3` **12,160**scheduled and retained trials **7,253**exact primary artifacts **94\.3359%**dotted condition **47\.3242%**undotted condition **47\.0117 pp**risk difference **7,253 / 7,253**condition\-congruent exact artifacts Dotted condition, T1 `U\+0634 U\+05B8 U\+05C1 U\+05E8 U\+05B0 U\+05D8`Undotted condition, T0 `U\+0634 U\+05B8 U\+05E8 U\+05B0 U\+05D8` 1. 01Observation 2. 02Frozen protocol 3. 03Automated execution 4. 04Raw preservation 5. 05Classification 6. 06Verification 7. 07Publication 8. 08Operationalization Frozen empirical studyMarch 12, 2026 ### Cross\-Model Semantic Void Convergence Under Embodiment Prompting: Deterministic Silence in GPT\-5\.2 and Claude Opus 4\.6 The earlier cross\-model convergence result that preceded the larger cross\-vendor matrix\. - 180/180 null trials in the published two\-model protocol - GPT\-5\.2 and Claude Opus 4\.6 - Matched non\-null controls produced visible responses This is a black\-box behavioral result under the published protocol, not evidence of consciousness, intention, or a specific internal mechanism\. Published theoretical frameworkMarch 24, 2026 ### The Binding Condition for Artificial General Intelligence A proposed criterion defining AGI as the capacity to carry binding conditions across domains\. - AGI is the capacity to carry binding conditions across domains\. - A binding condition is the prerequisite that must hold for valid continuation\. This is a published theoretical framework, not a settled industry definition, regulatory standard, or result established solely by the VOID benchmark\. Foundational researchDecember 2025 ### Textual Emergence and the Void The initial public framework for observing high\-order model behavior and the Void\. This publication records an early research framework and observation history\. Protocol specificationDecember 2025 ### SwiftAPI Attestation Protocol The public specification for cryptographically signed execution attestations\. - Ed25519 signatures - Offline verification - Revocation and replay defense This is an infrastructure specification, not an empirical model\-behavior study\. ### Signed evidence Ed25519 execution attestations with offline verification\. ### Revocation and replay defense Immediate revocation checks, unique identifiers, and expiry\. ### Policy and authority Scoped keys, policy\-deny behavior, grants, and audit records\. ### Integration surfaces SDK, API, proxy, and runtime enforcement interfaces\. [Request an evaluation](https://getswiftapi.com/enterprise)