psyclaw 0.27.22 → 0.28.0

package/vendor/ars/academic-paper-reviewer/references/statistical_reporting_standards.md

@@ -0,0 +1,505 @@
+# Statistical Reporting Standards — Statistical Reporting Standards & APA 7.0 Format Quick Reference
+
+This document defines the complete review standards for statistical reporting in quantitative research. `methodology_reviewer_agent` uses this document as the primary reference in Step 4a (Statistical Reporting Adequacy).
+
+---
+
+## 1. Universal Statistical Reporting Checklist
+
+All quantitative research papers **must** report the following items. Check each item during review:
+
+### 1.1 Descriptive Statistics
+
+| Item | Standard | Common Omission |
+|------|----------|----------------|
+| Mean (*M*) | Must be reported for all continuous variables | Only overall reported, not by group |
+| Standard deviation (*SD*) | Must appear paired with the mean | Standard error (*SE*) used incorrectly in place of SD |
+| Sample size (*N* / *n*) | Both total and group sample sizes must be reported | Sample attrition during analysis unexplained |
+| Range | Report Min-Max or interquartile range | Completely absent, unable to judge distribution characteristics |
+| Categorical variable distribution | Report frequency (*f*) and percentage (%) | Only percentage reported, missing raw frequency |
+
+### 1.2 Effect Size
+
+| Item | Standard | Common Omission |
+|------|----------|----------------|
+| Reporting obligation | **All statistical tests must be accompanied by effect sizes** — APA 7.0 mandatory requirement | Only *p*-value reported, no effect size |
+| Select appropriate metric | Choose effect size metric corresponding to the analysis method (see Section 2) | Inappropriate effect size metric used |
+| Interpretation | Must provide Cohen's conventional benchmarks or field-specific benchmarks | Numbers reported but magnitude not interpreted |
+
+**Common Effect Size Metrics Quick Reference:**
+
+| Analysis Method | Effect Size Metric | Small/Medium/Large (Cohen's Convention) |
+|----------------|-------------------|----------------------------------------|
+| *t*-test | Cohen's *d* | 0.2 / 0.5 / 0.8 |
+| ANOVA | *eta*-squared | .01 / .06 / .14 |
+| ANOVA (partial) | partial *eta*-squared | .01 / .06 / .14 |
+| Correlation | *r* | .10 / .30 / .50 |
+| Regression | *R*-squared, *f*-squared | *f*-squared: .02 / .15 / .35 |
+| Chi-square | Cramer's *V*, *phi* | *V*: .10 / .30 / .50 (*df*=1) |
+| Odds Ratio | OR | 1.5 / 2.5 / 4.3 (Rosenthal) |
+
+### 1.3 Confidence Intervals
+
+| Item | Standard | Common Omission |
+|------|----------|----------------|
+| CI reporting | All effect sizes and key estimates **should** report 95% CI | CI completely absent |
+| Format | 95% CI [lower bound, upper bound] | Inconsistent format or using parentheses instead of brackets |
+| Interpretation | Describe the substantive meaning of the CI, not just statistical meaning | Only checking whether CI includes zero, not interpreting width |
+
+### 1.4 Statistical Significance
+
+| Item | Standard | Common Omission |
+|------|----------|----------------|
+| *p*-value format | Report exact *p* value (e.g., *p* = .032) | Only reporting *p* < .05 or *p* > .05 |
+| *p* < .001 | Can report *p* < .001 when *p* is very small | Reporting *p* = .000 (raw statistical software output) |
+| Alpha level | Declare alpha level a priori | Failure to state whether alpha = .05 or another value |
+| Multiple comparisons | Use Bonferroni, Holm, FDR correction | Multiple comparisons without any correction |
+| Non-significant results | Must be fully reported; cannot be hidden | Selectively reporting only significant results |
+
+### 1.5 Statistical Power
+
+| Item | Standard | Common Omission |
+|------|----------|----------------|
+| A priori power analysis | State target power (typically >= .80), assumed effect size, alpha, required sample size | Power analysis completely absent |
+| Effect size source | Based on prior research, pilot study, or theoretical expectation | Using Cohen's convention without explanation |
+| Tool | Use G*Power, pwr package, etc. | Tool not specified |
+| Post-hoc power | Report observed power for non-significant results | Type II error risk not discussed for non-significant results |
+| Sensitivity analysis | Report the minimum detectable effect size given *N* | Sensitivity analysis not conducted |
+
+### 1.6 Missing Data Handling
+
+| Item | Standard | Common Omission |
+|------|----------|----------------|
+| Missing data reporting | Report missing data amount and proportion for each variable | Missing data situation not reported |
+| Missing mechanism | Discuss MCAR / MAR / MNAR | MCAR assumed without testing |
+| Handling method | State the method used: listwise deletion / pairwise deletion / MI / FIML | Not stated or only using listwise deletion |
+| Sensitivity analysis | Compare result robustness across different missing data handling methods | Only one method used, sensitivity not tested |
+
+### 1.7 Assumption Testing
+
+| Assumption | Applicable Analysis | Testing Method | Common Omission |
+|-----------|-------------------|---------------|----------------|
+| Normality | *t*-test, ANOVA, regression | Shapiro-Wilk / K-S / Q-Q plot / skewness & kurtosis | Completely untested or only invoking CLT |
+| Homogeneity of variance | Independent *t*-test, ANOVA | Levene's test | Not reported or alternative method not used when violated |
+| Linearity | Regression, correlation | Residual plot / scatter plot | Linearity assumed without testing |
+| Independence | Most parametric tests | Durbin-Watson / research design explanation | Nested data not handled |
+| Multicollinearity | Multiple regression | VIF, tolerance, correlation matrix | VIF not reported or reported but not addressed |
+| Residual normality / homoscedasticity | Regression | Residual plot, Breusch-Pagan | Residuals not checked after model fitting |
+
+---
+
+## 2. Method-Specific Checklists
+
+### 2.1 *t*-test (Independent / Paired Samples)
+
+| Check Item | Description |
+|-----------|-------------|
+| Report *t* statistic | *t*(df) = X.XX, *p* = .XXX |
+| Independent vs paired | Correct selection? Paired designs need to report pairing logic |
+| Effect size | Cohen's *d* (independent) or *d*_z (paired) |
+| Assumption testing | Normality (important for small samples), homogeneity of variance (independent *t*-test) |
+| Welch's *t*-test | Is Welch correction used when variances are unequal? |
+| Directionality | Is one-tailed vs two-tailed supported by a priori theoretical basis? |
+
+### 2.2 ANOVA (One-Way / Factorial / Repeated Measures)
+
+| Check Item | Description |
+|-----------|-------------|
+| Report *F* statistic | *F*(df1, df2) = X.XX, *p* = .XXX |
+| Effect size | *eta*-squared, partial *eta*-squared, or *omega*-squared |
+| Post-hoc comparisons | When main effect is significant, are post-hoc tests done (Tukey / Bonferroni / Games-Howell)? |
+| Interaction effects | In factorial designs, are interactions interpreted? Are simple effects tested? |
+| Sphericity assumption | For repeated measures, is Mauchly's test reported + Greenhouse-Geisser / Huynh-Feldt correction? |
+| Assumption testing | Normality, homogeneity of variance (Levene's), independence of between-group observations |
+| Unequal group sizes | When group sizes differ substantially, is Type III SS used? |
+
+### 2.3 Regression Analysis (Linear / Logistic)
+
+#### Linear Regression
+
+| Check Item | Description |
+|-----------|-------------|
+| Model summary | *R*-squared, adjusted *R*-squared, *F* test for model |
+| Coefficient table | *B*, *SE*, *beta*, *t*, *p*, 95% CI for *B* |
+| Multicollinearity | VIF (< 5 or < 10 depending on field convention), tolerance |
+| Residual diagnostics | Normality, homoscedasticity, linearity, outliers (Cook's *D*) |
+| Variable selection | Rationale for enter vs stepwise method |
+| Effect size | *R*-squared, *f*-squared, Cohen's *f*-squared |
+
+#### Logistic Regression
+
+| Check Item | Description |
+|-----------|-------------|
+| Model fit | Hosmer-Lemeshow / chi-squared / -2LL / Nagelkerke *R*-squared |
+| Coefficient reporting | *B*, *SE*, Wald, OR, 95% CI for OR |
+| Classification accuracy | Classification table, sensitivity, specificity, AUC/ROC |
+| Assumptions | Independence of observations, linearity in the logit (linear relationship between continuous predictors and logit) |
+| Sample size | At least 10-20 events per predictor variable (EPV rule) |
+
+### 2.4 Structural Equation Modeling (SEM)
+
+| Check Item | Standard |
+|-----------|----------|
+| Sample size | Typically >= 200; or 5-10 times the number of estimated parameters |
+| Model fit indices | **Must report multiple indices simultaneously** (at least 4) |
+| CFI / TLI | >= .95 (good); >= .90 (acceptable) |
+| RMSEA | <= .06 (good); <= .08 (acceptable); must report 90% CI |
+| SRMR | <= .08 |
+| chi-squared/df | <= 3 (some scholars suggest <= 2) |
+| Factor loadings | Standardized >= .50 (ideal >= .70) |
+| Measurement model | CFA before SEM (two-step approach) — Anderson & Gerbing (1988) |
+| Reliability and validity | CR >= .70, AVE >= .50, discriminant validity (Fornell-Larcker / HTMT) |
+| Modification indices | When using modification indices, must have theoretical support |
+| Normality | Multivariate normality (Mardia's coefficient); when violated, use robust ML or bootstrapping |
+
+### 2.5 Hierarchical Linear Modeling (HLM / MLM)
+
+| Check Item | Standard |
+|-----------|----------|
+| Nested structure | Clearly explain each level (e.g., students -> classes -> schools) |
+| ICC | Report Intraclass Correlation Coefficient; ICC > .05 supports using MLM |
+| Random effects | Report random intercept and (if applicable) random slope variances |
+| Fixed effects | Report coefficients, *SE*, *t* / *z*, *p*, CI |
+| Between-group sample size | Level-2 unit count (typically recommended >= 30) |
+| Centering | Explain whether grand-mean centering or group-mean centering is used and why |
+| Model comparison | Use deviance (-2LL), AIC, BIC to compare nested models |
+| Effect size | Pseudo *R*-squared (e.g., Snijders & Bosker's *R*-squared) |
+
+### 2.6 Chi-Square Test
+
+| Check Item | Description |
+|-----------|-------------|
+| Reporting format | chi-squared(df, *N* = XX) = X.XX, *p* = .XXX |
+| Effect size | Cramer's *V* (larger than 2x2) or *phi* (2x2) |
+| Expected frequencies | All cells expected frequency >= 5; if any cell < 5, use Fisher's exact test |
+| Independence | Are observations truly independent? (Repeated measures are not suitable for ordinary chi-square) |
+| Residual analysis | When significant, check standardized residuals to determine which cells contribute to significance |
+
+### 2.7 Non-Parametric Tests
+
+| Check Item | Description |
+|-----------|-------------|
+| Justification for use | Clearly explain why parametric tests are not used (e.g., normality violation, ordinal scale) |
+| Method selection | Mann-Whitney *U* / Wilcoxon / Kruskal-Wallis / Friedman — is the correct test matched |
+| Effect size | *r* = *Z* / sqrt(*N*) (Mann-Whitney); *W* (Kendall's) |
+| Reporting format | Report test statistic, *p*-value, effect size |
+| Post-hoc comparisons | After significant Kruskal-Wallis, pairwise comparisons + correction needed |
+
+---
+
+## 3. APA 7th Edition Statistical Format Quick Reference
+
+### 3.1 Number Formatting
+
+| Rule | Correct | Incorrect |
+|------|---------|-----------|
+| *p*-value no leading zero | *p* = .032 | *p* = 0.032 |
+| Statistics that can exceed 1.0 have leading zero | *M* = 0.75 | *M* = .75 |
+| Statistics that cannot exceed 1.0 have no leading zero | *r* = .45 | *r* = 0.45 |
+| Generally 2 decimal places | *M* = 3.45 | *M* = 3.4 |
+| *p*-value 2-3 decimal places | *p* = .03 or *p* = .032 | *p* = .0321 |
+| Percentages 0-1 decimal places | 45.2% | 45.2381% |
+
+**Statistics that cannot exceed 1.0** (no leading zero): correlation coefficients (*r*, *R*), proportions (*p*-value), Cramer's *V*, *phi*, *eta*-squared, *R*-squared, *beta* (standardized regression coefficient)
+
+**Statistics that can exceed 1.0** (leading zero): *M*, *SD*, *B* (unstandardized regression coefficient), Cohen's *d*, *t*, *F*, chi-squared
+
+### 3.2 Statistical Symbol Italicization Rules
+
+| Italic | Non-italic |
+|--------|-----------|
+| *M*, *SD*, *SE* | df |
+| *N* (total sample), *n* (subsample) | SS, MS |
+| *t*, *F*, *p*, *r*, *R*, *z* | OR, CI, VIF |
+| *d*, *f*-squared, *eta*-squared, *omega*-squared | AIC, BIC, CFI, TLI |
+| *B*, *beta* | RMSEA, SRMR |
+| *chi*-squared | ICC |
+| *U*, *W* (non-parametric statistics) | ANOVA, SEM, HLM |
+
+### 3.3 Statistical Results Reporting Format Examples
+
+| Analysis Method | APA Format Example |
+|----------------|-------------------|
+| Independent samples *t*-test | *t*(58) = 2.45, *p* = .017, *d* = 0.63, 95% CI [0.12, 1.14] |
+| Paired samples *t*-test | *t*(29) = -3.12, *p* = .004, *d*_z = 0.57 |
+| One-way ANOVA | *F*(2, 87) = 4.56, *p* = .013, partial *eta*-squared = .09 |
+| Linear regression | *B* = 0.34, *SE* = 0.12, *beta* = .28, *t*(95) = 2.83, *p* = .006, 95% CI [0.10, 0.58] |
+| Logistic regression | *B* = 1.24, *SE* = 0.45, Wald = 7.59, *p* = .006, OR = 3.46, 95% CI [1.43, 8.37] |
+| Chi-square | chi-squared(2, *N* = 150) = 8.34, *p* = .015, *V* = .24 |
+| Mann-Whitney | *U* = 245.00, *z* = -2.31, *p* = .021, *r* = .29 |
+| SEM fit | chi-squared(52) = 78.34, *p* = .011, CFI = .97, TLI = .96, RMSEA = .045 [.018, .068], SRMR = .038 |
+| HLM fixed effects | *gamma*_10 = 0.45, *SE* = 0.15, *t*(28) = 3.00, *p* = .006 |
+
+### 3.4 Table Format Standards
+
+| Rule | Description |
+|------|-------------|
+| Three-line table | APA tables have only three horizontal lines (above header, below header, bottom of table), no vertical lines |
+| Table numbering | Table 1, Table 2... (bold), title on the line below the number (italic) |
+| Note levels | General note (Note.) -> Specific note (superscript a, b) -> Significance (*p* < .05, **p* < .01) |
+| Asterisks | \**p* < .05. \*\**p* < .01. \*\*\**p* < .001. |
+| Alignment | Numbers right-aligned, decimal points aligned |
+
+---
+
+## 4. Statistical Red Flags
+
+The following patterns during review should raise red flags, requiring further investigation or author clarification:
+
+> **Triage levels, not finding severities (#574 A3).** The HIGH/MEDIUM/LOW labels in the tables below rank how urgently a red flag deserves scrutiny — they are detection triage, never a severity vocabulary. When a red flag becomes a reported weakness, its severity is the Schema 6 enum (Critical / Major / Minor), assigned by decision impact per the Finding Contract — never copied from the triage label.
+
+### 4.1 P-hacking Indicators
+
+| Red Flag | Description | Severity |
+|----------|-------------|----------|
+| Many *p* near .05 | Multiple results with *p* concentrated in the .04-.05 range | HIGH |
+| Selective reporting | Only significant results reported, non-significant ones disappeared | HIGH |
+| Vague analysis strategy | Analysis strategy not stated a priori, appears exploratory in hindsight | MEDIUM |
+| Unexpected subgroups | Post-hoc subgroup decomposition to find significant results | MEDIUM |
+| Flexible sample size | No pre-defined stopping rule (sequential testing without correction) | HIGH |
+| "Excluding outliers" | Large number of outliers excluded with unclear criteria | MEDIUM |
+
+### 4.2 HARKing (Hypothesizing After Results are Known)
+
+| Red Flag | Description | Severity |
+|----------|-------------|----------|
+| Perfect hypothesis-result match | All hypotheses supported without exception | MEDIUM |
+| Exploratory analysis packaged as confirmatory | Literature review clearly constructed post-hoc | HIGH |
+| Hypothesis directionality change | Originally predicted positive but result was negative, yet claimed "as expected" | HIGH |
+| No pre-registration | No OSF / AsPredicted pre-registration link provided (not mandatory but recommended) | LOW |
+
+### 4.3 Missing Effect Sizes and Confidence Intervals
+
+| Red Flag | Description | Severity |
+|----------|-------------|----------|
+| No effect sizes reported at all | Conclusions based solely on *p*-values | HIGH |
+| CI completely absent | Cannot judge estimation precision | MEDIUM |
+| Extremely wide CI | CI spans from small to large effect sizes, imprecise estimation | MEDIUM |
+| Inconsistent effect sizes | Reported effect sizes inconsistent with calculations from raw data | HIGH |
+
+### 4.4 Sample Size Issues
+
+| Red Flag | Description | Severity |
+|----------|-------------|----------|
+| No power analysis | Sample size lacks a priori calculation basis | MEDIUM |
+| Sample too small | In regression analysis, *N* < 10 x number of predictors | HIGH |
+| Unexplained sample attrition | Large gap between starting *N* and final *N* without explanation | MEDIUM |
+| SEM small sample | *N* < 200 without small sample correction | MEDIUM |
+| HLM Level-2 insufficient | Level-2 units < 30 | MEDIUM |
+
+### 4.5 Uncorrected Multiple Comparisons
+
+| Red Flag | Description | Severity |
+|----------|-------------|----------|
+| Multiple *t*-tests instead of ANOVA | 3+ group comparisons using multiple *t*-tests | HIGH |
+| No post-hoc after ANOVA | Main effect significant but claiming group differences without post-hoc tests | MEDIUM |
+| Multiple DVs uncorrected | Multiple dependent variables tested separately on the same dataset without Bonferroni or FDR | MEDIUM |
+| Multiple model comparisons | Trying multiple models but only reporting "the best one" | HIGH |
+
+### 4.6 Assumption Violation
+
+| Red Flag | Description | Severity |
+|----------|-------------|----------|
+| Assumption testing completely absent | Skipping normality/homogeneity/linearity tests | MEDIUM |
+| Violations not addressed | Violations reported but original analysis still used | HIGH |
+| CLT as excuse | "Because *N* > 30, normality can be ignored" without actual testing | LOW |
+| Excessive VIF | VIF > 10 but no action taken | HIGH |
+
+### 4.7 Other Red Flags
+
+| Red Flag | Description | Severity |
+|----------|-------------|----------|
+| *p* = .000 | Raw statistical software output, should be *p* < .001 | LOW |
+| df inconsistent with *N* | *N* derived from degrees of freedom doesn't match reported *N* | HIGH |
+| Inconsistent table numbers | Text narrative contradicts table values | HIGH |
+| Statistical software not stated | Not reporting SPSS / R / Stata / Mplus and version | LOW |
+| Causal language | Non-experimental designs (correlational/survey) using causal inference language | MEDIUM |
+| Unreachable mean (GRIM) | A reported mean of discrete-scale data cannot be produced by any integer sum at the reported *N* and precision | HIGH |
+| Unreachable SD (GRIMMER) | A reported SD of discrete-scale data is not attainable by any response distribution consistent with the reported mean and *N* | HIGH |
+| *p* incompatible with statistic and df | The reported *p* matches neither tail reading of the reported test statistic at the reported df | HIGH |
+
+---
+
+## 5. Common Statistical Methods in Higher Education Research
+
+Higher education research papers frequently involve the following topics and corresponding analysis methods. This table can be referenced during review to judge whether method selection is appropriate.
+
+### 5.1 Recommended Methods by Research Question Type
+
+| Research Question Type | Recommended Method | Description |
+|-----------------------|-------------------|-------------|
+| Two-group comparison (e.g., experimental vs control) | Independent samples *t*-test / Mann-Whitney | Depending on data normality |
+| Multi-group comparison (e.g., different institution types) | ANOVA / Kruskal-Wallis | Mean comparison for 3+ groups |
+| Pre-post comparison | Paired *t*-test / Wilcoxon | Change within the same group |
+| Predictive analysis (continuous DV) | Multiple regression | Multiple predictors' effects on continuous outcome |
+| Predictive analysis (binary DV) | Logistic regression | E.g., graduation/dropout, pass/fail |
+| Nested data (students -> schools) | HLM / MLM | Higher education data naturally has nested structure |
+| Latent constructs and path analysis | SEM / CFA | Measuring unobservable constructs (e.g., teaching quality) |
+| Scale reliability and validity | EFA -> CFA | Scale development or validation |
+| Categorical variable association | Chi-square / Fisher's exact | Cross-tabulation analysis |
+| Longitudinal data | Growth curve models / Latent growth models | Tracking student trajectories over multiple years |
+| Large-scale datasets | Weighted analysis / sampling design correction | Accounting for sampling design when using national survey data |
+
+### 5.2 Special Considerations for Higher Education Research
+
+| Consideration | Description |
+|--------------|-------------|
+| **Nested structure** | Higher education data almost always has nesting (students -> departments -> institutions); ignoring it underestimates standard errors and inflates Type I error |
+| **Sampling design** | When using national databases (e.g., MOE statistics, public higher education data), must account for sampling weights and clustering |
+| **Selection bias** | Students self-select into departments/institutions, not randomly assigned; consider propensity score matching or Heckman correction |
+| **Ceiling effects** | Satisfaction surveys often show extreme skewness; need to check and consider Tobit model or non-parametric methods |
+| **Small population** | Taiwan has a limited number of universities (~150); census surveys are not appropriate for inferential statistics (census, not sample) |
+| **Time series** | Analyzing multi-year enrollment trends requires considering autocorrelation |
+| **Multiple roles** | Same faculty completing multiple surveys (e.g., teaching evaluations) -> observations not independent |
+
+---
+
+## 6. Statistical Reporting Completeness Judgement
+
+`methodology_reviewer_agent` assesses each applicable item against the study design, analysis, reporting standard, and target venue. There is no weighted completeness score: a missing decision-bearing item cannot be offset by complete presentation items.
+
+| Item | Criterion-bound judgement |
+|---|---|
+| Descriptive statistics | `COMPLETE` / `PARTLY_COMPLETE` / `MISSING` / `NOT_APPLICABLE`, with the applicable reporting requirement |
+| Effect sizes | same states, with the analyses to which the requirement applies |
+| Confidence intervals or other uncertainty | same states, with the estimates to which the requirement applies |
+| Assumption checks | same states, naming only assumptions required by the methods used |
+| Power or precision justification | same states, when required by design or inference |
+| Missing-data handling | same states, with the observed omission or method anchored |
+| Format and table requirements | same states, separated from substantive statistical validity |
+| Red-flag follow-up | each triggered item investigated and resolved, left unresolved, or marked not applicable |
+
+For every `PARTLY_COMPLETE` or `MISSING` item, identify the manuscript location, the criterion source, why the omission matters, and whether it affects a substantive conclusion or only reporting clarity. Synthesize the result narratively; do not convert item states or counts into points, percentages, or a decision category.
+
+---
+
+## 7. Quick Reference: Recommended Review Sequence
+
+Methodology reviewer should follow this sequence when reviewing statistical reporting:
+
+```
+Step 1: Confirm research question -> analysis method correspondence is reasonable (Section 5)
+Step 2: Check whether assumption testing is reported (Section 1.7)
+Step 3: Check universal checklist item by item (Sections 1.1-1.6)
+Step 4: Consult method-specific checklist (Section 2)
+Step 5: Scan red flag list (Section 4)
+Step 6: Verify APA formatting (Section 3)
+Step 7: Produce the criterion-bound completeness judgement and narrative synthesis (Section 6)
+```
+
+---
+
+## 8. Bounded Arithmetic Recompute Procedures (#610)
+
+> **Epistemic status — read this before anything below.** This layer does
+> not replace the human reviewer, and it does not make the reviewer's
+> arithmetic trustworthy. What the engineering can force is narrow and
+> should be named honestly: the receipt grammar (methodology seat, sprint
+> Phase 2) forces every recomputation attempt to be *declared, structured,
+> and auditable* — required fields present, closed enums respected, each
+> mismatch linked to exactly one weakness. Triggering likewise rests on a
+> mandatory declaration with adjudicated honesty, never machine-checked
+> triggering: when nothing is recomputable the seat must say so
+> (`no_recomputable_statistics:`), and the checker verifies only that the
+> declaration exists — never that it is true; a false attestation over
+> recomputable statistics surfaces at adjudication as `MISSED` verdicts.
+> It cannot force the arithmetic
+> inside a receipt to be correct; model arithmetic is not deterministic, and
+> a fully conforming receipt built on wrong arithmetic is still wrong.
+> Correctness is decided by human adjudication against the four-verdict
+> scale (`VERIFIED` / `CLAIM_ONLY` / `MISCOMPUTED` / `MISSED`) defined in
+> `docs/design/2026-08-02-610-statistical-recompute-baseline-spec.md` §7.1.
+> Everything in this section is a bounded aid to that human judgment, never
+> a substitute for it.
+
+Four bounded procedures cover the recomputable slice of the red flags in
+Section 4. Normative worked cases and undecidable boundaries live in the
+#610 spec §5; this section is the reviewer-facing operating summary. Outside
+a procedure's boundary the honest status is `not_computable` with its closed
+reason — never an extended or improvised calculation.
+
+### 8.1 `p_from_test_statistic`
+
+Recompute a reported *p* from a named test family, statistic, and df.
+Prerequisites: test family, statistic value, required df, reported *p*, and
+the tail rule where it matters. *F* and chi-square are upper-tail by family;
+*t*/*z* results must never silently become two-tailed. **When the paper
+states no tail, the receipt must show BOTH labeled values (two-tailed and
+one-tailed)** — a single-tail comparison alone supports no verdict. If the
+tail choice flips the verdict, the status is `not_computable`
+(`tail_ambiguous`). Adjusted, bootstrap, permutation, or exact *p* values
+without their procedure are `not_computable` (`nonstandard_p_procedure`).
+
+### 8.2 `grim`
+
+Test whether a reported mean of discrete-scale data is reachable. Apply only
+to an unweighted mean with known discrete granularity, known item-specific
+analytic *N*, and a stated precision/rounding rule. **The completed
+procedure requires the rounding interval implied by the reported precision
+AND the nearest attainable values straddling the reported mean** — an
+integer-product observation without the rounding-interval reachability check
+is not a completed GRIM procedure. Continuous, weighted, imputed,
+transformed, composite, or item-averaged values with unknown granularity are
+`not_computable`.
+
+### 8.3 `grimmer`
+
+Inherits every GRIM prerequisite and additionally requires the SD convention
+(sample vs population), reported precision, and finite discrete support. A
+mismatch needs a completed reachability proof over the bounded response
+space; intuition that an SD "looks too small" is not a procedure. If the
+mean itself is GRIM-inconsistent, the status is `not_computable`
+(`mean_grim_inconsistent`).
+
+### 8.4 `n_from_df`
+
+Invert a reported df to a required sample size under a NAMED test-specific
+identity (`df=N-1` one-sample/paired *t*; `df=N1+N2-2` equal-variance
+independent *t*; the identity is not universal and the receipt must name
+it). Welch–Satterthwaite, Kenward–Roger, corrected repeated-measures,
+multiple-imputation, and mixed/clustered/robust/survey-weighted analyses are
+`not_computable` unless their required inputs are present. Chi-square df
+usually encode table dimensions and cannot be inverted to *N*.
+
+### 8.5 Red-flag classification (arithmetic rows)
+
+Every arithmetic-flavored red-flag row in Section 4 is classified below so
+that no row is silently claimed "solved" by the four procedures (#610 spec
+§8). Inclusion criterion: a Section 4 row appears here exactly when its red
+flag concerns reported numeric values or quantities derivable from them —
+whether by recomputation or by direct comparison against a stated numeric
+threshold; rows about narrative, design, or process conduct (causal
+language, HARKing, selective reporting, software not stated) stay outside
+this classification. Classes: **covered** (a bounded procedure recomputes it — emit a
+receipt when attempted), **reporting-only** (noticing and citing suffices;
+there is nothing to recompute), **direct-threshold** (a read-off comparison
+against a stated numeric threshold; report as a finding, no receipt), and
+**not_computable** (arithmetic in nature, but no bounded procedure exists
+yet — must NOT be presented as verified).
+
+| Red-flag row (Section 4) | Classification | Handling |
+|---|---|---|
+| df inconsistent with *N* (4.7) | covered | `n_from_df` receipt |
+| Unreachable mean — GRIM (4.7) | covered | `grim` receipt |
+| Unreachable SD — GRIMMER (4.7) | covered | `grimmer` receipt |
+| *p* incompatible with statistic and df (4.7) | covered | `p_from_test_statistic` receipt |
+| Inconsistent effect sizes (4.3) | not_computable | No bounded effect-size consistency procedure exists yet (a future separate addition per the #610 spec); flag the inconsistency as a finding, never as a verified recomputation |
+| Many *p* near .05 (4.1) | reporting-only | Distributional suspicion; cite the values, no recompute |
+| Sample too small, *N* < 10 × predictors (4.4) | direct-threshold | Read-off comparison; finding only |
+| SEM small sample, *N* < 200 (4.4) | direct-threshold | Read-off comparison; finding only |
+| HLM Level-2 units < 30 (4.4) | direct-threshold | Read-off comparison; finding only |
+| Excessive VIF > 10 (4.6) | direct-threshold | Read-off comparison; finding only |
+| *p* = .000 (4.7) | reporting-only | Format defect; no recompute |
+| Inconsistent table numbers (4.7) | reporting-only | Internal-consistency contradiction; cite both surfaces, no bounded recompute |
+
+### 8.6 Receipt grammar pointer
+
+Under a sprint contract, every attempted recomputation is recorded as an
+`AR<n>` Arithmetic Receipt in the methodology Phase 2 card — grammar,
+conditional fields, and mismatch-to-weakness linkage are normative in the
+agent's Phase 2 section (canonical source:
+`references/reviewer_sprint_prompt_source.md`, `methodology-receipt`
+fragment) and machine-checked by `scripts/check_phase_conformance.py`.
+Method references: Brown & Heathers (2017) GRIM,
+<https://doi.org/10.1177/1948550616673876>; Anaya (2016) GRIMMER,
+<https://doi.org/10.7287/peerj.preprints.2400v1>.