Abstract
This defensive publication discloses a runtime architecture for anonymous, unauthenticated public large-language-model (LLM) surfaces — the "ask a question" box on a marketing website, a public demo endpoint, an embedded help widget — that simultaneously preserves token-by-token streaming user experience and enforces an absolute zero-leak output-safety guarantee. The core mechanism is hold-then-release: the final model answer is streamed from the inference backend into an in-process delta buffer and is never forwarded to the client during generation; output safety executes against the fully assembled answer text; only a clean (or explicitly flagged-but-tolerated) verdict causes the buffered deltas to be replayed to the client in their original chunk granularity. A blocked answer therefore leaks exactly zero tokens, while a clean answer is visually indistinguishable from naive pass-through streaming. Around this core sit three further disclosed elements: (1) a dual-lane classifier — a deterministic regex pre-filter and a neural content-safety classifier of the Llama-Guard class — run concurrently on both the input and the output and fused by a confidence-thresholded three-outcome lattice (block / flag-and-proceed / pass); (2) a provenance-complete abuse-harvest lane that writes every blocked or flagged turn as a labelled training row carrying a four-way verdict, forming a self-replenishing retraining corpus; and (3) an inverted failure posture in which all safety components are strictly soft-fail and availability is governed by a single fail-closed database kill switch, with an unconditional bot-challenge gate on the first message of every conversation. The publication is enabling: it includes data models, control flow, worked examples, and a clean-room reference implementation.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Assuncao, gustavo matthew, "Hold-Then-Release Safety-Gated Streaming with Dual-Lane Classification and Abuse-Sample Harvesting for Anonymous Public LLM Surfaces", Technical Disclosure Commons, (July 13, 2026)
https://www.tdcommons.org/dpubs_series/10880