`verification/deepseek-decision.md`

35 lines
# DTD Verdict — Local DeepSeek option and unused Qwen cleanup

## Decision: A
Confidence: medium · Vote: 2/2 valid · Panel availability: 2/6.

Add the compact DeepSeek-R1-0528-Qwen3-8B-4bit model as explicit selection in the existing Exo MCP bridge. Preserve the existing default helper and configured jobs. Delete only user-authorized unreferenced Qwen caches after fresh state and open-file checks.

| Reference | Provider/runtime | Actual model | Vote | Reason |
| --- | --- | --- | --- | --- |
| Claude | Claude Code | Not invoked | Abstain | Disabled by canonical ZERO_COST_REQUIRED mode. |
| Codex | Signed-in Codex CLI | gpt-6-astra | A | Compact selectable integration, with runtime checks. |
| Qwen | Local Exo | Qwen3-VL-4B-Instruct-4bit | A | Preserve defaults and avoid constrained peers. |
| Grok | xAI | Not invoked | Abstain | Unavailable under zero-cost mode. |
| Kimi | Moonshot | Not invoked | Abstain | Unavailable under zero-cost mode. |
| Muse | Ollama | Not invoked | Abstain | User excludes Ollama. |

## Dissent
None among valid voters. Four abstentions limit confidence.

## Post-decision Codex debate
Result: KEEP. Artifact size alone does not establish runtime feasibility, and a model having no open files does not establish lack of scheduled use. The decision remains conditional on live inference and reference checks; those conditions must be preserved.

**PROSECUTOR:** A mistakes download size for operational feasibility. A 4.61 GB artifact does not establish that inference fits within 11.8 GiB available RAM alongside existing workloads; context, KV cache, and runtime allocations matter. Qwen’s claim that quantization “ensures minimal overhead” is unsupported. Likewise, no running instances and no open files do not prove a cache is dispensable: configured jobs may need it later. Cleanup could disrupt precisely the workloads A promises to preserve.

**DEFENDER:** Those objections defeat unconditional deployment, but A explicitly requires runtime verification and preservation of existing jobs. Adding an explicit selection leaves the default helper intact and permits bounded testing before adoption. Cleanup authorization covers only unreferenced caches; anything required by a preserved configuration fails that condition, even if currently idle. Insufficient memory, missing compatibility evidence, or an unresolved cache dependency must block the corresponding operation—not justify weakening the constraints.

**JUDGE:** The prosecutor exposes unsupported confidence in Qwen’s rationale, but does not defeat the conditional proposal in A.

FINAL: KEEP

The controlling argument is that A authorizes a selectable integration subject to verification, not an assumption that artifact size proves runtime fitness. Retain it with actual memory and response checks, regression checks for the existing helper, and configuration-aware confirmation that every cleanup target is unreferenced. A failed check blocks execution of that step; it does not make the proposed direction wrong.

## Why this verdict
A compact local model provides the requested option within the existing integration. Both active and configured Qwen dependencies were preserved; only the six unreferenced Exo directories were removed through its API. The full DeepSeek model was not installed. Small-model quality and bounded reasoning time remain practical limits.