`verification/decision.md`

40 lines

# DTD Verdict — Exo helper integration

## Decision: A — optional shared MCP helper

Confidence: medium · Vote: 2/2 valid voters · Panel availability: 2/6

| Reference | Provider/runtime | Model | Vote | Reason |
|---|---|---|---|---|
| Claude | Anthropic | not invoked | abstain | Canonical DTD ZERO_COST_REQUIRED disables this voter |
| Codex | Signed-in Codex CLI | gpt-6-astra | A | Preserve primary agents; validate the actual integration |
| Qwen | Exo/MLX on M3 Ultra | mlx-community/Qwen3-VL-4B-Instruct-4bit | A | Bounded helper uses an already verified cached model |
| Grok | xAI | not invoked | abstain | Paid API disabled |
| Kimi | Moonshot | not invoked | abstain | Paid API disabled |
| Muse | not invoked | unavailable | abstain | User excludes Ollama |

The task-local panel adapts the skill's Qwen voter to Exo chat completions and omits Muse. It does not call Ollama. The first Qwen answer lacked the required verdict marker and was not counted; a format-corrected rerun returned VERDICT: A. Codex's signed-in CLI resolved gpt-6-astra, as recorded in its log; no claim is made that gpt-5.3-codex ran.

## Dissent

None among the two valid votes. Four voters abstained, so this is limited agreement.

## Post-decision Codex debate

Result: KEEP. The prosecutor objected that basic inference establishes neither coding reliability nor productivity savings. The defender required enforced bounds, no execution tools, optional use, and primary-agent review. The judge retained A for a reversible helper, without claiming improved speed, savings, or distributed performance.

## Why this verdict

A cached model fits available memory and answers requests. Both coding clients can use the same bounded MCP interface while retaining their primary models and approvals. Larger models and automatic delegation are outside this implementation.

## Adversarial review transcript

**PROSECUTOR:** A mistakes feasibility for usefulness. One successful text response proves neither reliable summaries nor acceptable latency under coding workloads. A 4B model can produce plausible errors that cost the primary agents more to review than generating the answer themselves. MCP adds maintenance, resource contention, and another route for sensitive context to enter logs. “Bounded” is undefined: without enforced timeouts, concurrency limits, and input limits, it provides no protection. The strongest alternative is to retain standalone Exo and establish practical value before integrating it. Only two models voted; their agreement does not supply the missing evidence.

**DEFENDER:** Those objections defeat automatic delegation and replacement of the primary agents, but A proposes neither. Its scope is optional, bounded access to an already cached model for reviewed drafts and summaries. Integration can make the relevant limits enforceable: finite input/output sizes, a request deadline, one inference at a time, and explicit failure responses. Exo needs no authority to execute commands or change files. A shared server implementation avoids duplicating the adapter for both clients; it need not mean a shared persistent process. No useful distributed inference or coding competence is assumed. Standalone testing would still leave the actual MCP interaction untested, whereas this limited adapter allows that interaction to be evaluated directly.

**JUDGE:** The prosecutor establishes that productivity gains are unproven and that nominal bounds are insufficient. The defender nevertheless answers the decisive objection: the evidence required for a reversible, optional integration is lower than that required for relying on its outputs. A remains justified within its stated limited scope; broad or automatic delegation would be a different decision.

FINAL: KEEP

The controlling argument is proportionality: verified basic inference supports a small, explicitly bounded MCP integration while hosted agents retain responsibility and review. It does not establish coding reliability, speed improvements, or distributed performance. Enforced limits and optional use make A defensible despite those unknowns; the two valid votes alone do not.