v0.4.0 — correctness hardening + scientific rigor

Additive hardening of the measured case study. Floor intact.

P0 — correctness & honest framing

• Useful work now includes prefill (prompt + generation tokens, goodput-style), recomputed offline by recompute_frontier.py. Granite 4192 → 154 J/1k (27.2× span preserved); Mistral-7B 3552 → 135. Including prefill lowers every J/1k ~10%, ratio unchanged.
• Reframed as throughput-dominated: power is measured flat at the L4 TDP under load. Dual-method shown honestly — the counter-delta runs ~5–20% above the power method (was described as symmetric ±13%); short conc=32 counter point omitted.
• Prefill-aware waste detector (flink/08): measures useful tokens and guards prefill-heavy/long-context windows.

P1 — positioning over the prior art

• research/METHODOLOGY.md + research/metrics_cross_check.ipynb: implements MFU/MBU/goodput/energy-per-token on our own data (single-digit MFU, ~70–85% MBU → memory-bandwidth bound), with an explicit no-cross-comparison caveat and datasheet-cited constants. States plainly the metric and finding are not novel.
• Build-vs-buy table vs Kubecost/OpenCost; DCGM sampling caveat (arXiv:2312.02741).

P2 — roadmap: fleet matrix, real-traffic F, forecast confidence bands, Flink-ML maturity/CFU cost.

CI green (5 jobs, both notebooks reproduced).

v0.3.0 — measured efficiency frontier + comparative + closed-loop governance

Real-time GPU cost governance on Confluent Cloud for Apache Flink — energy-per-useful-token (J/1k) detection, forecasting, and a closed remediation loop.

Highlights

• Measured case study — IBM Granite-3.3-8B on a real NVIDIA L4 (vLLM + NVIDIA DCGM, 100% real telemetry). Dual-method efficiency frontier: joules_per_1k_tokens falls ~27× from 4,639 (conc 1) → 173 (conc 32), NULL at idle. Reproducible notebook + raw data committed.
• Comparative — same L4 sweep on Mistral-7B-Instruct-v0.3 (~12-14% lower J/1k — a 7B-vs-8B size effect, framed honestly).
• Closed-loop governance, all in Flink SQL — built-in ML_DETECT_ANOMALIES + ML_FORECAST, a "utilization-lies" waste detector, and a rule-based remediation recommender (no LLM; reference-aligned with Confluent Streaming Agents).
• Governance & connectors — Schema Registry (BACKWARD), Stream Lineage, three Amazon S3 sinks, least-privilege RBAC.
• Measured unit economics ($/1M tokens) + a component-by-component pipeline walkthrough with live screenshots.

Apache-2.0 · Built by Lutflow