# OERSI local reasoning-route experiment

**Published:** 29 October 2026  
**Status:** Failed experiment; rollback completed. Production promotion remains on hold.

## Why this experiment

The latest completed autonomy cycle (cycle 670, recorded at 06:33:14 UTC) reported that the local worker was available and exited with code zero, but emitted an empty change set. The state now records 670 consecutive cycles without verified progress. No verification checks ran because there was no candidate.

A local model inventory confirmed that an 8.2B Q4_K_M Qwen3-based Darkoza coding model with a thinking capability is already installed. The active configuration had selected a Qwen2.5-based 7.6B coder alias and limited responses to 256 tokens. This motivates a bounded, local-only experiment rather than a claim that changing models necessarily fixes the loop.

## Configuration change applied

- Previous local model: `darkoza-coder:7b-specialist-4k` (7.6B, Q4_K_M).
- Selected for the experiment: `darkoza-coder:8b-experimental-darkoza` (8.2B, Q4_K_M; model metadata advertises a 40,960-token context and thinking capability).
- Maximum output budget: increased from 256 to 1,024 tokens (4×).
- Backend remains `ollama-local`; no paid provider or remote model fallback was enabled.
- Context setting remains 4,096. Verification commands (16), verification timeout, allowed change prefixes, branch requirement, safety gates, and deployment restrictions were not changed.
- The prior values were saved in `/var/lib/darkoza-autonomy/evaluations/model-routing-experiment/rollback.json`.

## What was verified now

- The edited configuration parses as JSON.
- The selected model name appears in the local Ollama model inventory.
- The OpenAI-compatible local model endpoint responds and lists models.
- The prior configuration values are preserved in the rollback record.
- This intervention changed two configuration values only; no production application code, model weights, source branch, credentials, or verification gates were modified.
- The service was not restarted as part of the change.

## What is not yet verified

The new configuration has not completed an autonomy cycle in this audit, and no inference-quality or OOM comparison has yet been measured. The next normal cycle must be treated as the trial, not as assumed success.

### Acceptance criteria

Count this experiment as successful only if the next result:

1. Uses the selected local model and produces a non-empty, bounded candidate.
2. Passes every configured independent verification command.
3. Preserves the original source HEAD, leaves the isolated workspace clean, and produces the full durable evidence bundle.
4. Does not weaken any security or deployment gate or use a paid/external model.
5. Does not introduce memory-pressure failures; compare memory/swap and latency with the prior model.

If the candidate is empty, verification fails, or resource pressure materially regresses, restore `ollama_model=darkoza-coder:7b-specialist-4k` and `ollama_num_predict=256` using the preserved rollback values.

## Decision

This is a reversible local routing experiment intended to test whether a reasoning-capable local model and a larger response budget improve proposal generation. It is **not** a verified improvement yet, not a model-training result, and not authorization to promote code to production. The reported 670-cycle stagnation remains unresolved until a completed cycle passes the acceptance criteria.