# Evolution 81 — Capacity Plan and Results

## Authorized scope

- Stages 1–6: real OS workers on this host, 10..1,000 targeted; measured up to the point where the
  reviewed memory envelope would be exceeded.
- Stages 7–8: 5,000 and 10,000 **logical identities** with bounded execution concurrency.
- Not authorized: >1,000 concurrent OS processes, paid remote inference, production deployment.

## Measured stages

| workers | tasks done | failed | throughput/s | p95 ms | child max RSS MB | duplicate effects | wall s |
|---|---|---|---|---|---|---|---|
| 10 | 40 | 0 | 100.756 | 9.759 | 86.86 | 0 | 0.397 |
| 50 | 200 | 0 | 237.53 | 131.48 | 87.24 | 0 | 0.842 |
| 100 | 400 | 0 | 252.845 | 434.674 | 88.26 | 0 | 1.582 |
| 250 | 1000 | 0 | 164.366 | 2146.473 | 89.86 | 0 | 6.084 |
| 500 | 2000 | 0 | 109.493 | 6082.557 | 93.86 | 0 | 18.266 |

## Logical identity stages

| logical identities | create s | execution concurrency | tasks done |
|---|---|---|---|
| 5000 | 104.11 | 32 | 64 |
| 10000 | 425.29 | 32 | 64 |

## Not run or blocked

| stage | reason |
|---|---|
| 1000 | attempted twice; SQLite write contention at 1,000 concurrent claiming processes did not complete within the wall cap. 500 real workers is the measured maximum on this host. |
| 5000 | logical identities only (records), never concurrent processes |
| 10000 | logical identities only (records), never concurrent processes |

## Forecast (planning only, not measured)

| logical agents | tasks | model calls | tokens | cost | recommended concurrency |
|---|---|---|---|---|---|
| 100 | 400 | 400 | 48000 | 0.0 | 12 |
| 500 | 2000 | 2000 | 240000 | 0.0 | 62 |
| 1000 | 4000 | 4000 | 480000 | 0.0 | 64 |
| 5000 | 20000 | 20000 | 2400000 | 0.0 | 64 |
| 10000 | 40000 | 40000 | 4800000 | 0.0 | 64 |

## Shutdown conditions

Concurrency cap, rate limit, RSS quota, global budget, per-stage wall cap, and stop-on-failure. A
stage with any failed task or duplicate effect stops the program; it does not continue to a bigger
number to look impressive.

## Honest reading

A worker here executes a deterministic, verifiable hash task under the same durable lease/fence
protocol as a model-backed worker. It is a real OS process and a real governed task, but it is not
an LLM inference. Concurrent inference is measured separately and is much lower. Do not read these
numbers as "1,000 reasoning agents".
