# Local Model / Coding-Agent Benchmark Worksheet

Original blank worksheet accompanying Fouad Salkini's investigation.
Not results, not an upstream benchmark suite, and not a hardware guarantee.

## Source and model identity
- Repository / publisher:
- Checkpoint revision:
- Quantized file name, byte size and checksum:
- Base model / community fine-tune or distill:
- Model card limitations and license:

## Machine and runtime
- CPU:
- System RAM capacity / speed:
- GPU and VRAM:
- Storage:
- OS / driver / runtime build and backend:
- Exact command and chat template:
- Expert placement / GPU offload:
- Weight quantization / cache types:
- Slots / parallelism:
- Sampling / reasoning / output budget:

## Prompt and performance
- Actual prompt tokens (not only configured context):
- Context capacity setting:
- Uncached / cached / partially cached:
- Time to first token:
- Prompt-processing rate:
- Generation rate:
- Total generated and reasoning tokens:
- Peak RAM / VRAM / swap or page faults:
- Cold versus warm run:
- Repeated-run variation:

## Agent task acceptance
- Task / repository revision:
- Harness and tool-parser versions:
- Allowed tools and permissions:
- Expected behavior and tests:
- Actual edits made:
- Commands and exit status:
- Loops / output-cap / timeout / invalid tool-call events:
- Total time and retries:
- Final pass/fail and unresolved issues:

## Comparison boundary
- Same tasks, tests, budgets and prompt treatment:
- Different quantization/runtime choices, and why:
- Historical correction or regrading applied:
- What was not measured:
- Evidence files and provenance: