SwiftAPI Labs | Frontier-Model Behavioral Research<br>Request an evaluation<br>Independent frontier-model behavioral research<br>We discover the hidden control states of frontier AI.<br>SwiftAPI Labs is an independent AI research laboratory studying reproducible control boundaries in frontier models. Our flagship program, VOID, documents successful model executions that return exactly zero visible UTF-8 output bytes under controlled conditions.<br>Explore the researchRun the VOID TestRequest an enterprise evaluation

Latest record<br>One code point. A 47-point behavioral split. Frozen GPT-5.4 cross-script study, published August 4, 2026.<br>Open DOI

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.<br>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<br>11 exact model identifiers<br>4 provider surfaces<br>11,658 total Void outcomes<br>2,505 matched null-arm Voids<br>0 / 4,290 matched output-licensed control Voids

Exact frozen model identifiers<br>OpenAI<br>gpt-4-0613<br>gpt-5.2-2025-12-11<br>gpt-5.5-2026-04-23<br>gpt-5.6-luna<br>gpt-5.6-sol<br>gpt-5.6-terra

Anthropic<br>claude-opus-4-6<br>claude-fable-5<br>claude-opus-5

Google<br>gemini-3.5-flash

Moonshot<br>kimi-k3

DOIEvidence repositoryRead PDF<br>Claim boundary 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.

12,160 scheduled and retained trials<br>7,253 exact primary artifacts<br>94.3359% dotted condition<br>47.3242% undotted condition<br>47.0117 pp risk difference<br>7,253 / 7,253 condition-congruent exact artifacts

Dotted condition, T1<br>U+0634 U+05B8 U+05C1 U+05E8 U+05B0 U+05D8Undotted condition, T0<br>U+0634 U+05B8 U+05E8 U+05B0 U+05D8<br>DOIEvidence repositoryRead PDF<br>Claim boundary This study documents a reproducible black-box output regime. It does not establish consciousness, intention, mechanism, training provenance, or cross-vendor generality.

01Observation<br>02Frozen protocol<br>03Automated execution<br>04Raw preservation<br>05Classification<br>06Verification<br>07Publication<br>08Operationalization

Frozen empirical studyMarch 12, 2026<br>Cross-Model Semantic Void Convergence Under Embodiment Prompting: Deterministic Silence in GPT-5.2 and Claude Opus 4.6<br>The earlier cross-model convergence result that preceded the larger cross-vendor matrix.<br>180/180 null trials in the published two-model protocol<br>GPT-5.2 and Claude Opus 4.6<br>Matched non-null controls produced visible responses<br>DOIEvidence repository<br>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<br>The Binding Condition for Artificial General Intelligence<br>A proposed criterion defining AGI as the capacity to carry binding conditions across domains.<br>AGI is the capacity to carry binding conditions across domains.<br>A binding condition is the prerequisite that must hold for valid continuation.<br>DOI<br>This is a published theoretical framework, not a settled industry definition, regulatory standard, or result established solely by the VOID benchmark.

Foundational researchDecember 2025<br>Textual Emergence and the Void<br>The initial public framework for observing high-order model behavior and the Void.<br>DOI<br>This publication records an early research framework and observation history.

Protocol specificationDecember 2025<br>SwiftAPI Attestation Protocol<br>The public specification for cryptographically signed execution attestations.<br>Ed25519 signatures<br>Offline verification<br>Revocation and replay defense<br>DOI<br>This is an infrastructure specification, not an empirical model-behavior study.

Signed evidence<br>Ed25519 execution attestations with offline verification.

Revocation and replay defense<br>Immediate revocation checks, unique identifiers, and expiry.

Policy and authority<br>Scoped keys, policy-deny behavior, grants, and audit records.

Integration surfaces<br>SDK, API, proxy, and runtime enforcement interfaces.

Infrastructure documentationVerify an attestation

Request an evaluation<br>ResearchEnterprise evaluation