`verification/dtd-verdict.md`

31 lines
# DTD Verdict — Carnegie repair scope

## Decision: A — verify and stage the scoped restoration

Confidence: medium. Vote: 2/2 valid voters; availability: 2/6.

| Reference | Runtime/model | Vote | Reason |
|---|---|---|---|
| Claude | Anthropic/sonnet | Abstain | Disabled by zero-cost guard |
| Codex | OpenAI CLI/configured model (CLI artifact does not establish gpt-5.3-codex) | A | Preserve identities and verify scope before approved execution |
| Qwen | Ollama/qwen3:14b | A | Fix recurrence and prepare exact comparisons |
| Grok | xAI/grok-4.5 | Abstain | Unavailable with paid calls disabled |
| Kimi | Moonshot/kimi-k2.5 | Abstain | Unavailable with paid calls disabled |
| Muse | Ollama/muse-spark-1.2 (environment override) | Abstain | Unavailable |

## Dissent

None among valid panel votes. Post-decision prosecutor objected that aggregate
identity collisions do not prove each individual replacement or product mapping.

## Post-decision Codex debate

Result: KEEP. The recommendation is valid only with exact product mappings,
canonical values, unchanged exclusions, and explicit approval before writes.
Counts alone do not establish correctness. The plan verifier therefore checks
each canonical ID/SKU/value, each exact Shopify variant and product generation,
both metafield namespaces, scalar-versus-grouped identity, and full-table
uniqueness after the proposed restoration. Five negative fixtures fail closed.

The recommendation authorizes no writes. Full responses are in `dtd/`.