# Extraction report: dingeszakkou2019muchstakeknowledge

## Pipeline inputs used (in priority order)

- `papers/dingeszakkou2019muchstakeknowledge/out/bundle.json`
- `papers/dingeszakkou2019muchstakeknowledge/out/fulltext.md`
- `papers/dingeszakkou2019muchstakeknowledge/out/text.txt`
- `papers/dingeszakkou2019muchstakeknowledge/out/tei.xml` (for citation/DOI/volume)
- `papers/dingeszakkou2019muchstakeknowledge/out/tables/tabula_stream_p13_t1.csv` was inspected but not needed for the extracted stakes effects.
- Direct PDF verification via `pdftotext -layout papers/dingeszakkou2019muchstakeknowledge/pdf/paper.pdf -` was used to confirm the printed means/SDs, the reported Tukey HSD `d` values, and the Study 3 means inconsistency.

## Reproducible computations

- Recomputed pairwise between-groups SMDs (`d`) and sampling variances (`v`) for Study 1 and Study 2 from the published Neutral/Stakes group descriptives via `esc::esc_mean_sd` in `papers/dingeszakkou2019muchstakeknowledge/analysis/effect_sizes.qmd`.
- Retained Study 3 `d` from the paper and recomputed only `v` from reported `d` + `t(df)` in `papers/dingeszakkou2019muchstakeknowledge/analysis/effect_sizes.qmd`.
- Added a denominator check in `analysis/effect_sizes.qmd` showing that the paper's reported Tukey HSD `d` values for Studies 1-2 align with the omnibus three-group residual SD rather than the pairwise pooled SD used for standard between-groups SMDs.

## Issues / uncertainties

### FIELD: STUDY_1_2_REPORTED_D_NOT_USED_FOR_META_ANALYSIS
**Problem:** The paper reports Tukey HSD effect sizes for Studies 1-2 (`d = .50` and `d = 1.06`), but these do not equal the ordinary pairwise SMD implied by the published Neutral/Stakes means, SDs, and inferred cell Ns.

**What I did:** I abandoned the paper-reported `d` values for the meta-analytic `effect_size` field in Studies 1-2 and recomputed standard pairwise SMDs from the two-group descriptives:
- Study 1: `d = 0.572183831390`, `v = 0.041228607181`
- Study 2: `d = 0.928919454942`, `v = 0.043879582512`

I kept the paper values under `reported_test.reported_d` for provenance and documented the decision in `effect_size.notes`.

**Human action needed:** None unless the project explicitly prefers Tukey-standardized effect sizes over pairwise pooled-SD SMDs for meta-analysis.

### FIELD: STUDY_1_2_REPORTED_D_DENOMINATOR_CHECK
**Problem:** It was unclear whether the discrepancy between the paper's reported `d` values and the simple two-group SMD reflected a transcription error or a different denominator.

**What I did:** I added a denominator check in `analysis/effect_sizes.qmd`. For both studies, the reported `d` is closely reproduced by dividing the Neutral/Stakes mean difference by the omnibus three-group residual SD from the ANOVA, not by the pairwise pooled SD:
- Study 1: pairwise `d ≈ 0.572`, omnibus-residual `d ≈ 0.497`, reported `d = 0.50`
- Study 2: pairwise `d ≈ 0.929`, omnibus-residual `d ≈ 1.054`, reported `d = 1.06`

**Human action needed:** None. This was treated as sufficient justification to use recomputed pairwise SMDs in the extraction.

### FIELD: STUDY_3_GROUP_MEANS_SWAP
**Problem:** The Study 3 results sentence in the PDF reports `STAKES [M = 4.75, SD = 3.63] vs. NEUTRAL (M = .76, SD = 5.28)` while also stating higher retraction in STAKES (45% vs 12%) and reporting `t(168)=5.74` and `d=.88`.

**What I did:** I treated the direction claim plus `t(168)=5.74` and `d=.88` as the authoritative effect-size inputs; I did **not** record Study 3 per-condition means/SDs in the YAML (`effects[].groups` left empty) and retained `quality_flags: ["group_means_extraction_inconsistent"]`.

**Human action needed:** Optional: if the journal has a correction or supplementary source clarifying the swapped Study 3 labels, the per-condition means/SDs can be added later.

### FIELD: PER_CONDITION_NS_INFERRED_STUDY_1_2
**Problem:** Studies 1-2 do not explicitly list `n` per condition alongside the composite-score means/SDs in the results text.

**What I did:** I inferred per-condition Ns from the reported pairwise χ² comparison totals (Study 1: `n_neutral=51`, `n_stakes=50`, `n_evidence=50`; Study 2: `n_neutral=51`, `n_stakes=50`, `n_evidence=51`). These inferred Ns are now used in the Study 1-2 pairwise SMD computations and are recorded in `effects[].groups[].n`.

**Human action needed:** Optional: confirm condition Ns from the PDF if a table or supplement later provides them explicitly.

### FIELD: EVIDENCE_RELIABILITY_RECODED
**Problem:** The existing extraction left `moderators.evidence_reliability` as `null` for all three effects, but the current instructions require coding the perceived reliability of the epistemic basis whenever it is classifiable from the vignette.

**What I did:** I recoded `evidence_reliability` as follows:
- Study 1: `High` (repeated proofreading + dictionary use)
- Study 2: `Medium` (memory + prediction from prior Saturday bank visit, no extra verification)
- Study 3: `Medium` (same bank-memory basis as Study 2)

I also updated the corresponding `moderators_coding.evidence_reliability.reason` fields to match the current instructions.

**Human action needed:** None unless the project adopts a different reliability convention for verification procedures versus memory-based inference.

### FIELD: PAGE_NUMBERS_MISSING
**Problem:** The pipeline text/markdown outputs used do not provide reliable PDF page numbers for all extracted quotes.

**What I did:** Most `provenance.page` fields remain `null`.

**Human action needed:** Optional: map key quotes to PDF page numbers if page-anchored provenance is required later.