PLAT-215: LLM Runtime, Prompt, Token, and Cost Metering¶
Statement¶
The platform shall expose a common runtime contract for LLM-backed experiment or range agents across configured providers, preserving provider, model, prompt variant, rendered prompt, tool-use mode, token usage, call latency, and estimated cost per run. Metering shall be aggregated by run, experiment condition, event, and range where that context exists.
Rationale¶
PLAT-202 covers credential plumbing and shardable LLM access, but Shifter also needs the measurement and cost-governance side identified by LilRAE (formerly APTL): provider abstraction, prompt capture, token usage, latency, and estimated spend.
Traceability¶
- DOCUMENTS → SPEC
aptl:EXP-003(LilRAE specification, former APTL identifier EXP-003: Multi-Provider LLM Runtime) - DOCUMENTS → SPEC
aptl:EXP-004(LilRAE specification, former APTL identifier EXP-004: Per-Run LLM Token and Cost Metering) - DOCUMENTS → SPEC
aptl:EXP-005(LilRAE specification, former APTL identifier EXP-005: System Prompt Parameterization for Experiments)