Streaming Incremental Verification for LLM Outputs: Early-Stop Feasibility and API Compatibility Observations (v0.1)
Year: 2026
Type: preprint
Abstract: Independent technical report and minimal reproducible package for streaming incremental verification of LLM outputs. What this is: (1) a 200-problem pilot of streaming early-stop verification: only hard contradictions (BLOCK) trigger early stopping; BLOCK trigger rate 5.5% (11/200), estimated full-sample token saving 2.36%, latency saving 2.2%, false-stop 0/11; a sensitivity check shows stopping on uncertain (REVIEW) verdicts would cause 89.4% false stops, so only deterministic BLOCK may stop. (2) A 40-problem cross-protocol comparison of DeepSeek API surfaces (chat/completions vs Responses API vs Anthropic-compatible Messages): observations on stream termination events, reasoning-token consumption, usage accounting, and first-visible-token timing. (3) A call for collaboration: research questions, suggested experiments, and cost estimates for teams able to extend this work. Scope and limits: single model, single run, small samples; numbers are engineering estimates; no claim of SOTA, correctness, priority, or performance ranking. Not an official statement by any API vendor. Report text is CC BY 4.0; code is MIT. No raw model outputs, benchmark questions, customer data, credentials, or private paths are included. AI-assisted preparation is disclosed. Contents: report (Markdown + PDF), demo transcript (replay of 3 representative cases from archived aggregates; optional live mode), minimal research plan, self-contained code for the demo and protocol adapters, offline tests, and de-identified aggregate JSON. AI usage disclosure: This release was prepared with AI assistance (drafting, formatting, code scaffolding, review checklists). All experiment numbers come from local, re-runnable scripts and archived aggregates; the author reviewed and is responsible for all content. No raw model outputs, benchmark questions, customer data, or credentials are included. Not peer-reviewed; no SOTA/accuracy claim.
Author: Yin Kang
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7234031493
