Loupe EcoCoder Reviewer Portal
Download ecocoder CLI
Provisioned Reviewer Access · Hard $25.00 Spend Cap

Measured Coding Routing & Energy Readout

Run Claude Code, pi, or delegated coding subagents on a 3-tier open-weight ladder (qwen3.6-35b-fast → deepseek-v4.1-flash → kimi-k3.1-pro) paired with static code-graph slicing and deterministic architecture checks. Compare live NVML GPU Joules against Claude 5.5 and 3T Frontier MoE baselines.

Sign in with Invited Email

Authenticate via WorkOS using the email address provisioned for your reviewer account.

Or Sign in with Reviewer Key

Already have an eco_live_... key? Paste it to open your live budget and NVML energy telemetry.

Why Code Quality Governs Multi-Turn Energy

Sustainable Code That Subsequent AI Turns Can Actually Read

Turn-1 Benchmarks Hide Turn-40 Context Rot

Passing a test suite once is cheap if the model dumps untyped dict bags across modules, clones dispatch branches, or swallows failures behind broad except Exception: blocks. On the next feature request, every model has to pull 3×–5× more files into context to trace hidden side effects, retries broken callers, and escalates to frontier weights.

Static Slicing + Deterministic Evolution Checks

EcoCoder pairs its 3-tier ladder with codegraph-slicer v0.3.10 (AST call/import blast-radius slicing) and sub-40ms static checks across four brownfield defect classes: Data-Model (DM001–004), State-Machine (SM001–004), Evolution-Regression (ER001–004), and Error-Contract (EC001–004).

Sample Benchmark 1: Efficiency Triangle v4 (train5 Brownfield Checkpoint Evolution, 522 Tests)

Scenario Served Stack Multi-CP Pass Rate EC Violations Retail Token USD GPU Energy Joules / Pass Provenance
leanlm-eco (EcoCoder v2) qwen3.6-35b-fast → deepseek-v4.1-flash → kimi-k3.1-pro 495 / 522 (94.83%) 0 $3.22 486.3 kJ (135.1 Wh) 982 J NVML metered (5/5)
commercial-base claude-opus-5-5 (unrouted) 510 / 522 (97.70%) 0 $7.95 1,328.7 kJ (369.1 Wh) 2,605 J Modeled token physics
commercial-eco Claude Haiku / Sonnet / Opus 5.5 487 / 522 (93.30%) 0 $13.35 2,132.8 kJ (592.4 Wh) 4,379 J Modeled token physics
neuralwatt-base glm-5.3 (single open-weight) 464 / 522 (88.89%) 0 $6.93 2,755.2 kJ (765.3 Wh) 5,938 J NVML metered (5/5)

Evaluated across 5 brownfield checkpoint transitions (xjq CP4→CP5, file_backup CP2→CP3, code_search CP2→CP3, metric_transform_lang CP2→CP3, dynamic_config_service_api CP2→CP3). In separate 2×2 holdout factorials, adding ER + EC governance cut worst-function complexity (crap_max) by 95.8% (702.0 → 29.2 on file_merger CP3→CP4) and reduced downstream agent input tokens by 46.9% (3.13M → 1.66M on lab_intake).

Methodology Critique Welcome

Open Calibration Questions for Energy Reviewers

  1. NVML empirical anchor: Method A is least-squares fitted (±3.5%) to 1,526 NVML-metered turns across 3B/35B through 45B/1T non-MLA MoEs; how closely does hyperscaler serving track this curve at 100B–300B active?
  2. Boundary & idle attribution: NeuralWatt meters per-request NVML GPU Joules; how should shared-batch idle GPU and host CPU/DRAM draw be harmonized against facility-level estimates?
  3. Cloud PUE defaults: Should frontier comparisons default to hyperscaler PUE 1.09–1.14 (AWS/GCP/Anthropic range) or conservative industry PUE 1.20?
  4. Multi-node batch scaling: EcoLogits 0.11.x fixes batch=64 while multiplying per-GPU energy across all weight-holding GPUs; how should effective cluster batch size scale above 16 GPUs?
  5. Weight precision & HBM topology: Production 2026 frontier MoEs serve FP8/FP4 on 8–16× B200/H200 nodes rather than FP16 on 80 GB H100s; what resident-weight coefficient best captures multi-node HBM static draw?
  6. KV-cache scaling exponent: Active parameters scale as O(L · d²) while KV-cache bytes under GQA/MLA scale sublinearly (~P_act^(2/3) or ~P_act^(1/2)); what exponent best matches long-context HBM traffic?
  7. 2026 Claude 5.5 parameter priors: Are our central priors for Haiku 5.5 (30B / 200B), Sonnet 5.5 (110B / 1.2T), and Opus 5.5 (250B / 2.4T) well-bracketed?