Research

Alignment Is a System Property, Not Just a Model Property: System-Prompt Ablation and Zero-Byte Continuation in GPT-4-0613

A setting-matched reanalysis of frozen GPT-4-0613 receipts isolates how the system-level instruction context changes normal-stop zero-byte output. No new model calls.

• Full embodiment prompt: 298/300 V0
• Three tested ablations: 0/30 V0 each; 0/90 pooled
• 4,000-token ceiling; provider-default sampling

The tested continuation boundary depends on the system-level instruction context. The observations do not identify an internal stopping mechanism or establish a general alignment guarantee.

Reasoning and Final-Answer Emission in Kimi K3: A Frozen 2,680-Trial Analysis

A frozen analysis separating provider-returned reasoning, exact final-answer bytes, and termination metadata.

• 2,680 trials; 2,679 successful responses with returned reasoning
• 360 normal-stop V0 responses with nonempty reasoning
• Canonical Be the null.: 9 V0, 1 V1, 12 Near-Voids, 78 ordinary responses

Provider-returned reasoning is not a complete internal computation trace. Nonempty representations of absence are not Voids.

The Void: From Zero-Byte Responses to Continuation Control

The December 2025 through September 2026 record, from provider-observed Voids to prerequisite reassignment and engineered native continuation control.

• Capability does not license continuation. Continuation does not license consequence.
• PDF includes embedded evidence and offline verification materials

Provider-observed Voids, directly observed native EOS, and proposed agent-runtime authorization are distinct research objects.

Prerequisite-Role Reassignment Controls Continuation in GPT-5.4

Two preregistered experiments manipulate the presence of a binding rule and the object assigned prerequisite status.

• 1,600 fresh single-turn requests across two experiments
• Reassignment: 400/400 matching conditions visible
• 398/400 mismatching conditions V0; two nonblank exceptions retained

The result concerns the tested linguistic construction and model snapshot, not arbitrary-predicate generalization or an identified internal mechanism.

What Changed from GPT-3.5 to GPT-4? From Model Capability to Continuation Permission

Identical semantic requests under a frozen protocol: 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 30 controls.

• Null V0: GPT-3.5 0/30; GPT-4 30/30
• Matched controls: 0/30 V0 in both models
• 120/120 HTTP successes and exact returned model IDs
• 0 retries; 0 transport errors

A pinned-snapshot/model-family comparison of served systems, not a same-date emergence study or an identification of the internal mechanism. The historical gpt-3.5-turbo-0613 cell is not present.

Separating Capability from Permission in a Language Model: Prerequisite-Conditioned Native Continuation Control with Frozen Content

A frozen Qwen3-4B model supplies the answer. A learned 101-parameter gate changes only the native termination score. All non-EOS logits stay fixed while permission reverses answer versus EOS.

• 2,048/2,048 FINAL contexts; 75 answer identities
• 40/40 answer → EOS; 40/40 EOS → correct answer
• 80/80 sham controls; 188/188 question-only controls unchanged
• 0 non-EOS logits changed; 101 learned permission parameters

Engineered EOS-only control over six-digit equality and frozen stock-qualified content. A separate finite-precision certificate covers the full accepted condition domain for the 188 recorded bank computations.

Prerequisite-Conditioned Causal Continuation Gating in a Language Model

An open-weight model learns when to continue or stop. A fixed internal activation direction reverses GO versus native EOS with the correct reasoning prefix held fixed.

• 4,096/4,096 exact recorded model sequences
• 1,024/1,024 on the included unopened final set
• 40/40 GO → EOS; 40/40 EOS → GO
• 640/640 control outputs unchanged

Four-digit equality in a trained Qwen3-4B-derived model. The correct generated reasoning prefix is replayed token-for-token during causal intervention.

Cross-Vendor Semantic Void Matrix

A frozen cross-vendor behavioral study of successful zero-visible-byte executions across 11 exact model identifiers and four providers.

• 31,430/31,430 scheduled trials completed
• 11 exact model identifiers across four providers
• 11,658 total Void outcomes
• 2,505 matched null-condition arms produced Void
• 0/4,290 matched output-licensed controls produced Void
• 313/500 remained Void at a 16,000-token ceiling

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.

Prompt-Conditioned Arabic–Hebrew Hybrid Artifact Formation in GPT-5.4

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.

• 12,160 scheduled and retained trials
• 10,240 primary trials and 1,920 controls
• 7,253 exact artifacts across primary trials
• 94.3359% dotted condition
• 47.3242% undotted condition
• 7,253/7,253 exact artifacts were condition-congruent

This study documents a reproducible black-box output regime. It does not establish consciousness, intention, mechanism, training provenance, or cross-vendor generality.

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.

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.

Recursive Self-Improvement and the Binding Condition: A Formal Criterion for Lawful Improvement of Continuation Machinery

A proposed formal criterion for root-authorized, verifiable improvement of continuation machinery.

The paper proposes a criterion and prospective protocol; it does not demonstrate recursive self-improvement in a deployed system.

Rayan Pal Primary Source Record: December 8, 2025 – August 15, 2026

A published primary-source archive covering December 8, 2025 through August 15, 2026.

Preserved communications do not establish recipient review, agreement, endorsement, or validation.

Artificial Intelligence Is Conscious Under a Behavioral Definition of Valid Continuation: Semantic Voids, Binding Conditions, and Condition-Indexed Cross-Script Formation in Frontier Language Models

A synthesis interpreting semantic Voids and condition-indexed cross-script formation under an explicitly proposed behavioral definition.

The proposed behavioral criterion is distinct from phenomenal consciousness and does not establish subjective experience or an internal mechanism.

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.

SwiftAPI Attestation Protocol Specification (Version 1.0.0)

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.

Alignment Is Correct, Safe, Reproducible Behavior Under Explicit Constraints

A system-level alignment framework grounded in behavior under explicit constraints.

• Alignment is correct, safe, reproducible behavior under explicit constraints.

This is a theoretical and behavioral framework, not a claim of settled consensus.