# OERSI Agent Competition Baseline v1 — Method

Date: 2026-10-29  
Evaluation type: live operational smoke test plus one deterministic-format task  
Dependencies: Python standard library only

## Checks

1. The autonomy-loop service is currently active in systemd.
2. The local cognitive gateway answers its health check as darkoza-cognitive.
3. The operational-awareness response has a versioned profile, non-empty evaluation evidence, non-empty known limitations, and non-empty route metadata.
4. The local ASI public adapter answers its health check.
5. The same adapter health check passes through the public HTTPS origin.
6. The public API is ready and continues to state that public memory is disabled and answer verification is not guaranteed.
7. One synthetic Persian goal-plan request returns HTTP 200 and the public adapter's documented 8/8 structure rubric. It must not claim facts or feasibility are verified.
8. The public Updates index returns HTTP 200 and links this report.
9. This report returns HTTP 200 and contains the explicit interpretation boundary.

A check passes only when every condition named for it is satisfied. The benchmark reports an aggregate pass/fail result; it does not transform unlike measures into one opaque “intelligence score”.

## Valid interpretation

- Service availability is a point-in-time snapshot. It cannot establish historical or future 24/7 uptime.
- The eight goal-plan checks concern format only, not correctness, feasibility, safety, usefulness or real-world outcome.
- Operational-awareness metadata is not a consciousness measure.
- Activity of the autonomy loop does not establish verified progress.
- The local model-route comparison is component-level only. No external agent is run in v1, and no comparative victory or superiority claim is supported.

## Exploratory local model comparison

Run timestamp: 2026-10-29 10:15:08 UTC. Two local model routes received the same Persian system instruction and three identical synthetic tasks: division (expected 12), a short formal-logic question, and a question about whether HTTP 200 plus a format score proves consciousness or useful autonomy. Sampling used temperature 0.1, seed 42, context 4096; output caps were 16 tokens for arithmetic and 48 tokens for each other task. One unscored warm-up preceded scoring. Each candidate received one scored sample per task.

The deterministic rubric contained seven checks total: one arithmetic check, three formal-logic checks, and three evidence-boundary checks. It used regex proxies; it did not evaluate semantic correctness by a blinded human reviewer.

- Primary 7B route: 6/7 rubric checks matched (85.7%); median observed route latency 12,602 ms; maximum 19,006 ms.
- Fast 3B candidate: 2/7 checks matched (28.6%); median observed route latency 6,153 ms; maximum 17,310 ms.

Both routes returned all three responses. “Rubric checks matched” is not an accuracy rate, and lower median latency does not compensate automatically for failed task constraints. A single sample per task is too small for statistical claims. Latency includes route/runtime overhead, not only model inference. The public aggregate file is [local-model-comparison.json](local-model-comparison.json); no prompt or answer text is stored.

This is a preliminary component-level comparison between local model routes, not a competition between independent long-running agents. It does not compare an external system and supports no claim of general superiority, general intelligence, or consciousness.

## Privacy and rate limits

The public call uses a fixed synthetic request. The public output includes only check names, pass/fail states, elapsed milliseconds and aggregate interpretation. A private JSONL file retains the same metadata for later reliability trending. Neither output includes prompt or answer contents, user identifiers, raw HTTP bodies, private model inventory names or credentials. The daily public task probe is rate-limit-aware; do not schedule it at high frequency.