# Errata and interpretation corrections

Version: `0.1.0-stage1`  
Status: applies to the preserved upstream files in this public bundle

The report and contract files in this bundle are preserved evidence artifacts. They have not been silently rewritten. The corrections below govern their public interpretation.

## 1. Legacy endpoint-count boilerplate

`common-shock-contract.md` contains an interpretation-boundary sentence saying that the primary result applies to ten named endpoints. The analyzed legacy confirmation used **nine endpoints**. Qwen3-30B/Alibaba was excluded in whole before outcome access after the original ten-endpoint execution failed its infrastructure-support gate. The common-shock report, machine-readable legacy summary, methodology page, and PDF use the nine-endpoint population.

## 2. Mistral provider route

The original endpoint list in `common-shock-contract.md` names `mistral-small32-24b-deepinfra-fp8`. A provider-route-restart amendment changed the executed route to `mistral-small32-24b-mistral-eu` before pilot outcomes. The pilot gate, legacy summary, methodology page, and PDF report the executed Mistral EU route.

## 3. Capital-multiplier label

`base-study-report.md` retains a column labeled “Capital multiplier.” The later common-shock contract explicitly says that no `n_eff` ratio is to be called a capital or absolute-variance multiplier. Treat those values only as the report’s stated equal-exposure, square-root-of-variance, benchmark-relative illustrations. They are **not capital requirements, regulatory multipliers, financial advice, or deployment rules**. The public PDF omits that column.

## 4. Goel et al. citation title

The source manuscript package described the Goel et al. paper as “Chance-adjusted probabilistic agreement for language-model errors.” That is not the paper’s title. The public briefing and PDF use the verified title **Great Models Think Alike and this Undermines AI Oversight**, identified as an ICML 2025 poster and arXiv:2502.04313.

## 5. Retrospective pair-selection count

The execution-side selection audit reported **3,253 support-qualified model pairs from 77,815 possible unordered pairs**. The Stage 1 bundle does not contain the event-level pair-support matrix needed to regenerate that count independently. The count is therefore reported with an explicit reproduction boundary, not presented as independently reproduced from this bundle.