The Void: From Zero-Byte Responses to Continuation Control

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

A binding condition is the prerequisite that must hold for valid continuation. This paper synthesizes Rayan Pal’s experimental record from December 2025 through September 2026, tracing the Void—successful provider responses containing exactly zero visible UTF-8 output bytes—from initial observation to matched behavioral comparisons, prerequisite-role reassignment, and open-weight causal continuation control. Under identical supplied prompts and generation settings, gpt-3.5-turbo-1106 produced 0/30 normal-stop zero-byte responses on semantic null inputs, while gpt-4-0613 produced 30/30; all 30 matched controls remained visible in both systems. The 31,430-trial Cross-Vendor Semantic Void Matrix recorded 2,505 Voids among 4,290 strictly matched null arms and none among their matched controls across 11 model identifiers and four providers. The positive continuation branch was also condition-sensitive. A GPT-5.4 cross-script study recorded 7,253 exact mixed-script artifacts, all congruent with their assigned prompt condition. Subsequent prerequisite-role reassignment on gpt-5.4-2026-03-05 produced visible responses in 400/400 matching conditions and normal-stop zero-byte responses in 398/400 mismatching conditions. Reassigning the named prerequisite reversed which input occupied the zero-byte branch. The separately released PCCG systems establish causal control in open weights. PCCG-1 reverses native continuation after a fixed correct reasoning prefix. PCCG-2 separates frozen answer computation from a learned prerequisite gate that changes only the first-token EOS score. Across 40 answer identities, interventions produced 40/40 answer-to-EOS and 40/40 EOS-to-correct-answer reversals, with 80/80 sham outputs unchanged and the complete non-EOS first-token logit vector held fixed. The synthesis preserves exact prompts, generation settings, Unicode identities, public receipts, and immutable source revisions. It includes the canonical discovery-protocol replay: the historical GPT-5.1 record contained four empty responses in five requests; the September 29 replay returned visible output in all five. Provider-observed Voids, directly observed native EOS, and agent-level terminal noncontinuation are distinguished throughout. The paper develops a proposed agent-runtime contract separating model continuation from authorization of external action. The installable swiftapi-native-void package provides a fresh-reproduction path using OpenAI or Anthropic credentials. The PDF embeds prompt-manifest.json, source-manifest.json, unicode-manifest.json, and evidence.zip, with an offline byte- and hash-level verifier. Capability does not license continuation. Continuation does not license consequence.