# Extraction report — `buckwalter2014themysterofstakes`

## What the pipeline produced
- `papers/buckwalter2014themysterofstakes/out/bundle.json` shows **no extracted tables** (`out/tables/` empty); numeric condition summaries were therefore not available from `out/tables/*`.
- `papers/buckwalter2014themysterofstakes/out/tei.xml`, `out/fulltext.md`, and `out/text.txt` contained most prose, but the TEI figure caption for the bank-condition descriptives was **truncated** (missing the final assertion cell).

## Manual interventions (required / performed)
### Used the PDF directly (documented per instructions)
Because tables were missing and one TEI caption was truncated, I consulted `papers/buckwalter2014themysterofstakes/pdf/paper.pdf` via text extraction to recover analysis-critical values:
- PDF p.10: sample size line (N=215), scale anchors (1=false … 5=true), and vignette text used for moderator coding.
- PDF p.15–16: typo vignette + prompt wording for moderator coding.
- PDF p.25: full set of bank-condition means/SDs (including the missing High Error/High Stakes assertion cell).
- PDF p.26: typo-condition means/SDs + ANOVA F/df.

These values are cited in `papers/buckwalter2014themysterofstakes/buckwalter2014themysterofstakes.yaml` with `provenance.page` pointing to the PDF page number.

## Key problems / uncertainties

### 1) Bibliographic details (paper-level)
The PDF lacks clear front-matter bibliographic metadata (editors/publisher/DOI). `paper.year` was set to 2014 based on filename/repo context.

**Human fill-in**
- `paper.citation` (full): ___
- `paper.year` (confirm): ___
- `paper.published` (confirm Yes/No): ___
- `paper.doi` (if any): ___

### 2) Split Ns not reported, but analyzed N can be recovered
For both experiments, the paper does not report per-condition Ns. However, the analyzed total `N` can be recovered from the paper’s exclusion notes and confirmed from the omnibus ANOVA denominator df:

- Study 1: page 10 reports `N = 215`; page 25 footnote 10 says `30` participants were removed; omnibus tests are reported as `F(1,177)`, which in an 8-cell between-subject design implies analyzed `N = 185 = 177 + 8`.
- Study 2: page 16 reports `100` participants; page 26 footnote 18 says `10` were removed; omnibus tests are reported as `F(1,86)`, which in a 4-cell between-subject design implies analyzed `N = 90 = 86 + 4`.

Under the analyst-authorized equal-cell assumption already used elsewhere in the repo, this is sufficient to compute `d` and `v` from the reported means/SDs. I therefore added numeric effect sizes and documented the assumption in `papers/buckwalter2014themysterofstakes/analysis/effect_sizes.qmd`.

Equal-cell allocations used:

- Study 1 bank effects: `185 / 8 = 23.125` per cell, so each low-vs-high stakes contrast within fixed error/speech-act uses `n_low = n_high = 23.125`.
- Study 2 typo knowledge effect: `90 / 4 = 22.5` per cell, so the low-vs-high stakes knowledge contrast uses `n_low = n_high = 22.5`.

### 3) Study 2 recruitment/compensation not reported
The bank study explicitly mentions MTurk + Qualtrics; the typo study only reports "100 participants" and does not specify recruitment/compensation in the extracted text.

**Human fill-in**
- Study 2 recruitment source (mTurk? students? other?): ___
- Study 2 compensation details (if any): ___

### 4) Denial vs assertion direction (analysis decision)
Study 1 includes both **knowledge assertions** and **knowledge denials**. The extraction keeps them as separate effects and notes that denial truth judgments may need reverse-scoring depending on the meta-analytic construct.

**Human fill-in**
- Should denial effects be (a) excluded, (b) included as-is, or (c) reverse-scored to align with knowledge-ascription direction? ___