# Evolution 80 — Resource and Cost Plan

Written 2026-11-15 before any expensive experiment. The build host is also the production host,
so the budget is deliberately small.

## Money

| Item | Budget | Spent |
|---|---|---|
| Cloud model inference | $0 — every remote provider stays BLOCKED | $0 |
| Model credits / automatic billing | prohibited | none |
| External services | none | none |

## Compute (measured before planning, load average 12–23 on 8 vCPU)

| Model | Measured speed (4 threads, batch 16, nice 19) |
|---|---|
| Qwen2.5-0.5B-Instruct | ≈ 3.7 s per call |
| Qwen2.5-1.5B-Instruct | ≈ 10.7 s per call |
| SmolLM2-135M-Instruct | ≈ 1.2 s per call |

## Planned experiments

| Experiment | Tasks | Planned calls | Planned CPU wall time |
|---|---|---|---|
| Execution paths, train split (0.5B, 1.5B, 135M) | 40 | ≈ 175 | ≈ 18 min |
| Execution paths, test split | 48 | ≈ 160 | ≈ 20 min |
| Ten coordination architectures, test split (0.5B) | 48 | ≈ 850 | ≈ 55 min |
| Replication split, separate process | 24 | ≈ 80 | ≈ 10 min |
| Memory evaluation (4 arms × 28 questions, 0.5B) | 28 | 112 | ≈ 7 min |
| LoRA skill (SmolLM2-135M): train + evaluate | 160 + ~190 | ≈ 380 evals | ≈ 10 min |
| Failure-injection drill, security lab | — | 1–2 | ≈ 3 min |
| **Total** | | **≈ 1,750 calls** | **≈ 2 hours** |

## Limits that apply to every run

- `nice -n 19`, at most 4 inference threads, one model in memory at a time (peak ≈ 7 GB for 1.5B).
- No run is started on the second host for inference (2 vCPU, 3.4 GB).
- Runs are resumable: every stage's raw outputs are cached, so an interrupted run never repeats a
  model call.
- If production latency degrades because of these runs, they are stopped.

## What the budget rules out (reported as BLOCKED or NOT EVALUATED, never simulated)

- Models of 3B parameters and above, and any Llama / Gemma / Phi / Mistral / DeepSeek model
  (not on the host; not downloaded onto a production machine).
- llama.cpp, Ollama and vLLM engines (not installed).
- LongMemEval, LoCoMo, MemoryAgentBench, AMA-Bench, MultiAgentBench, AgentBench.
- Endurance campaigns of 6, 24, 72 and 168 hours.
- Cross-host model execution.

Host time is not priced. Reported cost is $0 cloud spend, not $0 total cost.
