Episode 06 · Planned
Speculative decoding and structured outputs
Evaluate compatible proposal paths and constrained generation without assuming speedup.
Compare a target-only baseline with a compatible draft/verification path while sweeping proposal length and load. Separately compare prompt-only structured output with constrained generation under equivalent contracts. Parse or schema validity and independently judged semantic correctness remain separate outcomes.
Evidence: accepted tokens per step and its definition, draft and verification duration, added memory, TTFT/TPOT/goodput, retries, schema validity, semantic correctness, unsupported claims and total cost.