Changes in version 0.6.2                        

Exact binomial test reported with Cohen's h. New test_type = "binomial"
matched via pat_binom_h, anchored on a "binomial p [op] " clause
followed (within ~80 non-period chars) by "Cohen('s)? h = ". When a "
out of " clause is present in the same verbatim, N is recovered
(N_source = "binom_n_out_of_N") and check.R re-computes the two-sided
binomial p via stats::binom.test() assuming p_null = 0.5 (the most
common null in binomial-vs-chance reporting); the recomputed vs reported
delta appears in uncertainty_reasons. When N isn't recoverable, status
routes to NOTE -- the Cohen's h is accepted as reported.

Surfaced by the 2026-05-25 escicheck-iterate corpus expansion against
the CRSP decoy-effect papers (Xiao/Zeng/Feldman 2021 et al), where 2-5
binomial-with-h rows previously fell through to WEAK_GOLD or
OUT_OF_SCOPE. The NOTE-only template (LESSONS.md "NOTE-only test_type
template") was extended cleanly: parse layer adds the pattern + dispatch
branch, check.R adds a tt == "binomial" branch with conditional
recompute. A v0.6.3 follow-up could detect a stated null proportion ("vs
1/3 chance" etc.) to replace the p_null = 0.5 default.

Regression tests in tests/testthat/test-v062-binomial-h.R (7 cases: full
CRSP verbatim with N recovery, bare binomial+h with N=NA NOTE,
80-char-lookahead far-apart rejection, "h" without "binomial p" anchor
guard, chisq+h still routes to chisq, lowercase "cohen h" form, and
uncertainty-message contents when N is recovered).

                        Changes in version 0.6.1                        

Bare t = X, p [op] Y (no df) extraction. Surfaced by the
Lee-Feldman 2025 RSOS Newman-2014 RR replication during the 2026-05-25
escicheck-iterate corpus expansion (24 occurrences in one paper's
Tables 10-15: compact <label> M = m (sd), t = X, p < .001 form where df
lives only in the table header, not the immediate sentence). Before
v0.6.1 such reports returned 0 rows from check_text().

A new pat_t_p_nodf pattern matches t = X followed within ~80 chars by a
p [<=>] clause; (?<![a-zA-Z]) keeps dt =, pt =, etc. from false-positive
matching, and the 80-char lookahead bound prevents a stray t = X from
being yoked to an unrelated downstream p = in long prose. df1 stays NA —
check.R routes to status NOTE because the exact p-check needs df.
Dispatch position: AFTER pat_t_nodf (t = X, df = Y form keeps priority
and yields status=OK with full verification when df is present).

Regression tests in tests/testthat/test-v061-bare-t-p-nodf.R.

                        Changes in version 0.6.0                        

Clinical-trial RR / rdpct / md_hl independent verification, completing
the v0.5.16/17/18 PROSECCO-trial test-type set. Closes the deferred
v0.6.x follow-through promised in the v0.5.16-18 NEWS entries.

Verification (the v0.5.x NOTE rows now compute a comparison)

  - RR -- when the per-arm slash-count clause (<events1>/<total1> ...
    versus <events2>/<total2>) is in the same sentence as the RR clause,
    check_text() computes RR = (events1/total1) / (events2/total2)
    independently and reports the reported-vs-computed delta + a
    Wald-on-log 95% CI in the row's uncertainty message. Fisher-exact /
    chi-square p-value verification remains future work (v0.6.x+).
  - rdpct -- same per-arm cells produce RD = 100 * (events1/total1 -
    events2/total2) and a Wald 95% CI. Farrington-Manning iterative-MLE
    noninferiority p is honestly not-yet-wired and the message says so;
    the Wald approximation is suitable for sanity-checking the point
    estimate, not for noninferiority decisions.
  - md_hl -- the Hodges-Lehmann point estimate cannot be recomputed from
    sentence-level text (needs per-arm rank data), so the row carries
    two sanity checks instead: (a) CI symmetry around the point estimate
    (asymmetric CIs are flagged: |below - above| / width > 0.15); (b)
    p-CI consistency (p < .05 iff 0 outside the 95% CI).
  - New per-row columns: arm1_events, arm1_total, arm2_events,
    arm2_total -- the captured per-arm cells (NA for any row not parsed
    as RR or rdpct, or where the slash-count clause was absent).
    Additive schema change; does not break MetaESCI-critical columns.

Regression tests in
tests/testthat/test-v060-rr-rdpct-mdhl-verification.R. Closes the
2026-05-25-v06x-clinical-trial-compute-branches handoff.

                       Changes in version 0.5.18                        

Median-difference (Hodges-Lehmann) with IQR + CI (escicheck-iterate
cycle 8). Completes the PLOS Med PROSECCO-trial PARSE-MISS punch-list
opened in cycle 1.

New test type

  - Median-difference (test_type = "md_hl"). Parses Hodges-Lehmann
    median-difference reports of the form median difference <val>; 95%
    CI <lo> to <hi>; p[-value]? = <pval>. The HL estimate cannot be
    independently recomputed from a sentence-level extraction (needs
    per-arm rank data), so the row is captured as a NOTE for surface
    transparency. Regression tests in
    tests/testthat/test-v0518-median-diff.R. Caught by the 2026-05-23
    escicheck-iterate validation against the PROSECCO trial AI stats
    gold (10.1371/journal.pmed.1004323).

                       Changes in version 0.5.17                        

Risk-difference percent with CI (escicheck-iterate cycle 7).

New test type

  - Risk-difference percent (test_type = "rdpct"). Parses clinical-trial
    noninferiority risk-difference reports of the form risk difference
    <val>%; 95% [confidence interval (CI)|CI] <lo> to <hi>; ... P =
    <pval>. Full Farrington-Manning noninferiority verification is
    deferred to v0.6.x; this cycle resolves the PARSE-MISS aspect so
    rows appear with status NOTE. Regression tests in
    tests/testthat/test-v0517-risk-diff-pct.R. Caught by the 2026-05-23
    escicheck-iterate validation against the PROSECCO trial AI stats
    gold (10.1371/journal.pmed.1004323).

                       Changes in version 0.5.16                        

Clinical-trial risk ratio with two-proportion slash counts
(escicheck-iterate cycle 7).

New test type

  - Risk ratio (test_type = "RR"). Parses clinical-trial RR reports of
    the form <n1>/<N1> (<pct>%) versus <n2>/<N2> (<pct>%) ... RR
    <val>; 95% CI <lo> to <hi>; p[-value]? = <pval>. The p-clause
    supports both p = 0.15 and the operator-less p-value 0.44 form
    common in PLOS Medicine / NEJM tables. Full verification of RR
    against per-arm cell counts is deferred to v0.6.x; this cycle
    resolves the PARSE-MISS aspect so the row appears with status NOTE
    (extracted but not-yet-fully-verified). Regression tests in
    tests/testthat/test-v0516-rr-slash-counts.R. Caught by the
    2026-05-23 escicheck-iterate validation against the PROSECCO trial
    AI stats gold (10.1371/journal.pmed.1004323).

                       Changes in version 0.5.15                        

Cochran Q meta-analytic heterogeneity test (escicheck-iterate cycle-5,
after user scope decision 2026-05-24 to bring Q in-scope).

New test type

  - Cochran Q (test_type = "cochran_q"). Parses meta-analytic
    heterogeneity tests of the form Q_T [40] = 104.65, p < .001
    (optional subscript, brackets or parens for df). The Q statistic is
    chi-square distributed under the homogeneity null with the reported
    df, so the reported p-value is verified against pchisq(Q, df,
    lower.tail = FALSE) in the same dispatch path as Kruskal-Wallis H.
    No standard effect size is recoverable from Q alone; an uncertainty
    note records that I-squared (when reported) is not independently
    verified. Regression tests in tests/testthat/test-v0515-cochran-q.R.
    Caught by the 2026-05-23 escicheck-iterate validation against the
    Identifiable-Victim AI stats gold (10.1525/collabra.90203, R03).

                       Changes in version 0.5.14                        

Two narrow parse fixes from the 2026-05-24 escicheck-iterate cycle-4
validation against the Collabra canary.

Parse fixes

  - Bayesian model-averaged estimates no longer inherit a global-text N.
    A r = 0.002 (95% CI [0; 0.004]) reported as the output of a RoBMA /
    Bayesian model-averaging / posterior-model-average analysis
    previously fell through the local -> extended -> global N cascade
    and picked up an unrelated paper's N from somewhere later in the
    text (producing a misleading df1 = N-2, N = 1004 attribution on a
    model-averaged estimate with no recoverable per-study sample size).
    The cascade now recognizes "RoBMA", "Bayesian model-averaging",
    "model-averaged", "posterior model average", and "PMA" markers in
    the local + extended context and stops before the global fallback,
    leaving N_source = "bayesian_model_no_n". Regression tests in
    tests/testthat/test-v0513-bayesian-no-n.R.

  - Table-fragment duplicates of body-text statistics now collapse.
    Replication / extension papers commonly print a summary table that
    lists the same correlations / effect sizes already reported in the
    Results body. Each numeric appeared twice in the extracted output:
    once with the full parenthesized form (r(741) = -.43, 95% CI [-.49,
    -.37]) and once as a table cell (r = -.43 [-.49, -.37]). They are
    now collapsed to a single row by (test_type, stat_value, df1, df2,
    N) exact match, keeping the parenthesized body-text version. For
    r-rows, the missing df1 in the table-fragment row is normalized to
    N-2 before matching. Regression tests in
    tests/testthat/test-v0514-dedup-table-fragments.R.

                       Changes in version 0.5.12                        

Recall fix for the Collabra / APA partial-eta-squared convention.

Parse fixes

  - pat_etap2 now recognizes the eta^2p / eta^2_p form (subscript-p
    AFTER the squared symbol) in addition to the previously-supported
    etap^2 form (subscript-p BEFORE). Caught by the 2026-05-23
    escicheck-iterate validation against the AI stats gold: 13+ F-rows
    across two Collabra replications (Identifiable Victim,
    Experiential-vs-Material) dropped their reported partial-eta-squared
    point estimate (CI was captured, name + value null) because every
    Collabra paper writes η^2p = .008 with the p trailing the caret-2.
    The point estimate now binds correctly; status upgrades OK → PASS
    once the reported effect matches the computed. Regression tests in
    tests/testthat/test-v0512-etap2-caret-p-form.R.

                       Changes in version 0.5.11                        

Documentation-only release. The design_ambiguous output flag has always
combined two semantically distinct cases under one name; this release
makes the distinction explicit and parseable without changing behaviour.

Documentation / output-string clarifications

  - ambiguity_reason now carries a stable bracket-tagged category suffix
    when applicable: "[category: structural-design]" for the
    Phase 8A-bis paired-vs-independent case (a t / F(1,df) / z test
    reports d or g and BOTH the independent variant family and the
    paired variant family were computed), or "[category: cross-family]"
    when the reported ES type has no same-type variants in the
    computed-variants set (e.g. a Cohen's d reported on an F(2,df)
    omnibus, or ES type not specified at all). Existing reason
    substrings are preserved untouched (so downstream substring matches
    like the internal "No same-type" check continue to work); the tag is
    appended idempotently just before the output tibble is built.
    Consumers that want to programmatically distinguish the two
    semantics should grep the reason for the bracketed category: tag.
  - design_ambiguous flag semantics are now documented end-to-end. The
    flag is intentionally broad (ambiguity_level != "clear") and covers
    BOTH categories above; downstream consumers that only want the
    narrow paired-vs-independent meaning can filter on the new category
    tag. Documented in the check_text() @return block, in API.md, in the
    frontend /api-docs page, and in LESSONS.md. No code behaviour
    changed.
  - @return for check_text() now enumerates the notable output columns
    inline (previously a single sentence "tibble with comparison
    results"), starting with design_ambiguous, ambiguity_level,
    ambiguity_reason, and matched_variant.

                       Changes in version 0.5.10                        

Bare r = with a confidence interval — a parse fix found by
escicheck-iterate.

Bug fixes

  - A bare correlation r = value reported with a confidence interval but
    no p-value is now extracted. The r = (no-df) pattern previously
    required a nearby p-value before it would emit a result — a guard
    against casual r = .3 mentions. A correlation reported with a CI
    (e.g. r = -.74 [-0.92, -0.30]) is a genuine result even without a p,
    so the guard now accepts a p-value OR a confidence interval,
    mirroring the chi-square (p or df) and Mann-Whitney (p or z) no-df
    guards. An explicitly labelled CI (95% CI [...]) always counts; a
    bare bracketed pair counts only when its bounds bracket the r value,
    so an unrelated bracketed pair (a page range, a citation index) is
    not mistaken for a CI.

                        Changes in version 0.5.9                        

Chi-square chi^2 caret token — a parse fix found by escicheck-iterate.

Bug fixes

  - A chi-square written as chi^2(df) (the word "chi" with a caret
    superscript) is now parsed. The chi-square token alternation was
    duplicated across four call sites — the sub-chunk splitter and
    pat_chi / pat_chi_nodf / pat_chi_two_dfs — and the copies had
    drifted: the symbol forms allowed an optional caret (X^2, and the
    Greek-letter form) but the word form only matched chi2 with no
    caret, and the splitter copy lacked the precomposed superscript
    forms entirely. So chi^2(1) = 3.74 returned zero statistics. The
    alternation is now hoisted to one shared chi_tok definition used by
    every chi path, so the accepted notations can no longer drift apart.
    No behaviour change for the previously-recognised forms (chi2,
    chi-square, X2, the Greek-letter and precomposed-superscript forms).

                        Changes in version 0.5.8                        

Chi-square bare-n sample size — a parse fix found by escicheck-iterate.

Bug fixes

  - A bare lowercase n = is now read as the total N for a chi-square
    when no other sample-size token is present. pat_N deliberately
    matches only N / nobs because a bare n = is commonly a per-group
    size — but a chi-square reporting chi2gof(1) = 31.01, p = ..., n
    = 329 (the JASP goodness-of-fit form) then had N come back NA and
    could not compute its effect size. A chi-square-scoped fallback now
    accepts a single bare n = as the total N, but only when the chunk
    carries no n1 / n2 per-group token and exactly one n = appears (two
    or more are per-group counts, not a total).

                        Changes in version 0.5.7                        

DSCF (Dwass-Steel-Critchlow-Fligner) post-hoc W — a parse +
categorisation fix found by escicheck-iterate.

Bug fixes

  - DSCF (Dwass-Steel-Critchlow-Fligner) post-hoc pairwise comparisons —
    reported as W = ..., the post-hoc test following a significant
    Kruskal-Wallis — are now recognised. A negative DSCF W (W = -3.84, p
    = .018) returned 0 stats because pat_W and the sub-chunk splitter
    both rejected the leading minus; a positive DSCF W (W = 5.99) parsed
    but was mislabelled Wilcoxon's W. pat_W and the splitter now accept
    a leading sign, and a new dscf test type is assigned to a negative
    W, or to a W in an explicit DSCF / Dwass / Kruskal-pairwise context.
    No standard effect size is recoverable from the W statistic alone,
    so a DSCF result is an honest "cannot verify" NOTE — the same
    conservative route as Kendall's W, not the Wilcoxon-W mis-route it
    used to fall into.

                        Changes in version 0.5.6                        

Bare regression-coefficient lines — a parse fix found by
escicheck-iterate.

Bug fixes

  - A bare regression-coefficient line — b = 0.45, SE = 0.12, p = .001,
    the standard APA form for a coefficient with its standard error and
    p but no t-statistic written out — is now detected. effectcheck's
    regression-type promotion fired only when a t-test had already been
    parsed, so a bare b + SE had no test type to promote and the line
    returned 0 stats. When b, SE and a reported p all co-occur and no
    test statistic was parsed, effectcheck now creates a regression
    result and synthesises the coefficient t = b / SE; all three are
    required so an incidental b/SE co-occurrence cannot spuriously
    create a result. df is unknown (no test statistic was reported), so
    the row is reported as an honest NOTE.

                        Changes in version 0.5.5                        

JASP "nobs" sample-size token — a parse fix found by escicheck-iterate
running effectcheck against the real-article AI gold corpus.

Bug fixes

  - The JASP "nobs" sample-size token (number of observations) is now
    recognised as the total N. A chi-square reporting nobs = 659 had N
    come back NA, so the reported Cohen's w / Cramér's V could not be
    verified (status NOTE). pat_N now accepts nobs alongside capital N.
    A bare lowercase n = is deliberately still not matched — it is
    commonly a per-group size and would be mis-read as the total N.

                        Changes in version 0.5.4                        

Regression-coefficient handling — a categorisation fix found by
escicheck-iterate running effectcheck against the real-article AI gold
corpus.

Bug fixes

  - A standardized regression coefficient (beta) reported on a plain
    t-test is no longer cross-matched to a Cohen's d. (β = 0.83, t(261)
    = 5.82, p < .001) — a mediation / regression path coefficient — had
    its reported beta matched against the t-test's computed Cohen's d
    variants (matched_variant = d_ind_equalN / gav / drm), a meaningless
    cross-family comparison whose PASS/NOTE verdict depended on whether
    the beta value coincidentally resembled the computed d. A beta from
    a multi-predictor / mediation model is not recoverable from the
    t-statistic alone, so it is now left unmatched and reported as an
    honest NOTE — mirroring the Stage 1 Gap 3 treatment of Cohen's h on
    a chi-square.

                        Changes in version 0.5.3                        

Scientific-notation p-values — a parse fix found by escicheck-iterate
running effectcheck against the real-article AI gold corpus.

Bug fixes

  - Scientific (E-notation) reported p-values are now captured. A
    reported p written in E-notation — p = 2.572e-08, p = 1.2e-3, the
    form R / JASP / Python emit — was not parsed: pat_p requires a
    [01].x mantissa (so it rejects 2.572) and pat_p_sci only handles the
    p < 10^-N form. 5 of the 12 chi-square results in the gold
    for 10.1098/rsos.250367 carry an E-notation p, so effectcheck
    silently skipped a checkable p-value (status SKIP). A new
    pat_p_enote pattern captures the mantissa+exponent and converts it
    to a plain decimal; the reported p is now checked against the
    computed p.

                        Changes in version 0.5.2                        

Subscripted chi-square notation — a parse fix found by escicheck-iterate
running effectcheck against the real-article AI gold corpus.

Bug fixes

  - JASP-style subscripted chi-square is now parsed. A chi-square
    written with a subscript label glued to the symbol — chi2gof(2),
    chi2Pearson(1), the form JASP emits — returned 0 stats: parse.R's
    chi-square patterns required the open parenthesis to follow the chi
    token immediately, so a gof / Pearson word between them blocked the
    match. 7 of the 12 chi-square results in the gold
    for 10.1098/rsos.250367 were invisible. An optional subscript group
    (an allowlist of gof / Pearson / Yates / LR / MH / Wald) is now
    accepted in pat_chi, pat_chi_nodf, pat_chi_two_dfs and in the
    sub-chunk splitter — the last so a paragraph of subscripted
    chi-squares splits into one result per statistic rather than
    collapsing into one row.

                        Changes in version 0.5.1                        

Stage 1 validation fixes — four gaps found by validating the v0.5.0
Stage 1 coverage against six real articles (AI gold generated via the
article-finder skill).

Bug fixes

  - One-sample t-test detection (Gap 1) now covers the "mean vs a
    constant" family phrased with than / from / compared to — e.g.
    "higher than chance", "differed from the scale midpoint" — not only
    the "against chance / against the midpoint" forms. A one-sample
    t-test phrased "...were higher than chance" was still mislabelled
    design_inferred = "independent", matched_variant = "dz".
  - Kendall's W (Gap 2) is no longer misparsed as Wilcoxon's W. The bare
    W = token is shared by Wilcoxon's W (a large rank-sum) and Kendall's
    W (the coefficient of concordance, bounded 0-1); a W in [0, 1]
    reported in a "Kendall" / "concordance" context is now classified as
    the new kendall_w test type and recognised as a kendalls_W effect
    size.
  - Cohen's h on a chi-square (Gap 3) is no longer mis-matched to the
    contingency phi/V. A one-proportion / goodness-of-fit chi-square
    that reports Cohen's h as its effect size now yields an honest
    "cannot verify" NOTE — h is a function of two specific proportions
    and is not recoverable from the chi-square statistic alone.
  - Spearman confidence intervals (Gap 4): a bare r(df) correlation now
    carries the Spearman (Bonett & Wright 2000) interval as an
    alternative method in the CI candidate pool alongside the Pearson
    Fisher-z interval. A Spearman correlation whose method was declared
    only in a distant Methods section no longer draws a spurious CI
    mismatch. No reclassification occurs, so papers mixing Pearson and
    Spearman are unaffected; the row stays labelled Pearson r and
    ci_method_match records which method matched.

                        Changes in version 0.5.0                        

Coverage Stage 1 — closes effect-size / test-type gaps from
the 2026-05-16 coverage roadmap (P1, P2, P3, P6, P7).

New features

  - One-sample t-tests are now labelled design_inferred = "one-sample"
    with a d_onesample matched variant. Previously a one-sample t-test
    was mislabelled independent/dz (the recomputed value was correct,
    since the one-sample d formula t/sqrt(N) coincides with dz, so only
    the labels were wrong).
  - Spearman's rho and Kendall's tau are first-class test types. They
    are parsed from their symbol forms (rho(df)=, tau(df)=, Greek
    symbols) and an r(df)= reported in a Spearman/Kendall context is
    reclassified. Each gets a rank-appropriate p-value (Spearman:
    t-approximation; Kendall: normal approximation) and confidence
    interval (Spearman: Bonett & Wright 2000; Kendall: Fisher-z, Fieller
    et al. 1957) — never the Pearson path.
  - Chi-square sub-types are detected (chisq_subtype column) and routed
    correctly: Friedman to Kendall's W, goodness-of-fit to Cohen's w,
    McNemar to an honest "cannot verify". None are silently given a
    contingency-table phi/V.
  - Cohen's h is computed and verifiable from a two-proportion z-test.
  - Confidence intervals now populate for omega-squared, partial
    omega-squared, epsilon-squared, Cohen's f-squared, adjusted
    R-squared, and Cohen's w.

Internal

  - New chisq_subtype output column.
  - Strict design_inferred test assertions: a categorization regression
    to "unclear" now fails the test suite.

                        Changes in version 0.4.2                        

Bug fixes

  - Correlation (r) parsing: a Cohen's-d-family token (d/g/dz/dav/drm)
    is now adopted as an r-test's reported effect size only when it
    appears after the r statistic (APA order: statistic, then effect
    size). A d-family token positioned before the r belongs to a
    preceding clause and is no longer conflated into the r result.
    Previously a two-analysis sentence such as an abstract's
    "...(d=0.39[0.25, 0.54]) ... (r=-.34[-.43, -.24])" produced a single
    row pairing the second clause's r with the first clause's d. A d
    co-reported with the r (r(50)=.40, p=.003, d=0.87) is still matched.
    Found by the escicheck-iterate corpus loop on Chen et al. (2023,
    Collabra).

                        Changes in version 0.4.1                        

Bug fixes

  - t-test sample-size inference: a document-wide ("global text") N that
    is structurally incompatible with a t-test's degrees of freedom --
    e.g. N = 608 applied to a t-test with df = 364, where the design
    requires N = df + 2 -- is now overridden with the df-based N (df + 1
    paired / df + 2 otherwise). Previously such a t-test kept the wrong
    N, producing a wrong recomputed d and a spurious WARN even when the
    reported effect size was consistent. The Welch branch already did
    this; the non-Welch branch only flagged it. Found by the
    escicheck-iterate corpus loop on Chen et al. (2021, JESP) Study 3
    t-tests.

                        Changes in version 0.4.0                        

Breaking changes — extraction layer removed

All file-input functions are now .Defunct() and emit an error directing
callers to extract via docpluck and pass the resulting text to
check_text():

  - read_any_text()
  - check_file(), check_dir(), check_files()
  - checkPDF(), checkPDFdir()
  - checkHTML(), checkHTMLdir(), checkDOCXdir()
  - compare_file_with_statcheck() — replaced by compare_with_statcheck()
    (text input)

The pure-text-analysis API (check_text(), compute_and_compare_one(), the
parsing layer, all effect-size and CI computations, and every output
column) is unchanged.

The package no longer requires poppler-utils, tesseract, magick, or qpdf
system dependencies. SystemRequirements field removed from DESCRIPTION;
corresponding entries removed from Suggests.

Migration: see https://docpluck.app/api-docs for the API contract.
Working R reference implementation in the ESCImate web-app repo at
tests/scripts/docpluck_shootout.R.

New features (carried over from 0.3.6 deception-detection work)

  - New per-row column df_arity_mismatch (logical, default FALSE) flags
    structurally malformed test statistics where the declared test label
    disagrees with the number of df arguments supplied — F(48) (F always
    takes two df), t(36, 10) (t always takes one df), chi2(48, 14)
    (chi-square takes one df), r(50, 30) (r takes one df). Such rows
    previously were silently dropped because the strict regex patterns
    rejected them; v0.3.6 emits the row with df_arity_mismatch = TRUE,
    status = "NOTE", and an explanatory uncertainty message, while
    skipping all recomputation paths (p_computed, effect sizes,
    decision_error are all NA).

  - New tier-5 verification fixture
    (tests/testthat/test-deception-arena.R) documents the ScienceArena
    stats-extraction-v1 adapter contract: every row corresponding to a
    deceptive stat is flagged by at least one of decision_error,
    extraction_suspect, insufficient_data, df_arity_mismatch,
    ambiguity_level == "highly_ambiguous", or status %in% c("WARN",
    "ERROR").

API documentation

  - API.md documents df_arity_mismatch and adds a "Suspicion signals for
    downstream consumers" section listing the six row-level fields a
    benchmark adapter should OR together to derive flagged_suspicious.

                        Changes in version 0.3.5                        

Addresses MetaESCI v0.3.5 request: CI-audit feature pack. Adds CI
computation coverage for previously-uncomputable effect-size families
(OR, R², standardized β, partial r, semi-partial r) and new per-row
metadata for characterizing CI reporting quality at scale (precision
tracking, completeness flags, level mismatch, bounded-parameter
clipping, symmetry classification).

Purely additive — no v0.3.4 behavior changes.

Compute: CI computation coverage gaps closed

  - ci_OR_all() — odds-ratio CI via Wald-on-log(OR). Three sources for
    SE: (1) supplied SE_logOR, (2) Fisher exact CI from a 2×2 cell
    vector, (3) Wald inversion back-derived from a reported p-value when
    neither is available. Resolves MetaESCI 1A.
  - ci_R2_all() — R² CI routed through ci_etap2_all() (R² ≡ partial η²
    in one-predictor / single-omnibus regression). Methods retagged with
    _via_etap2 suffix so the matcher distinguishes R²-routed from native
    η²-routed CIs. Resolves MetaESCI 1B.
  - ci_standardized_beta_all() — normal-approximation CI on standardized
    β. Uses supplied SE_beta when available, else back-derives from
    t-stat.
  - ci_partial_r_all() and ci_semi_partial_r_all() — Fisher-z transform
    CIs for partial and semi-partial correlations. Resolves MetaESCI 1C.

Parse: decimal-place precision tracking

  - New helper count_decimal_places() extracts trailing-digit count from
    raw regex match strings before numify() (which loses trailing
    zeros).
  - Four new output columns capture APA-7 precision:
    effect_reported_decimals, ciL_reported_decimals,
    ciU_reported_decimals, stat_value_decimals. Resolves MetaESCI 2A.

Check: CI audit metadata (Phase 6)

  - ci_expected (logical) — TRUE when row carries an effect size from a
    family for which CIs are normative reporting
    (d/g/r/η²/η_p²/R²/OR/V/φ).
  - ci_reported (logical) — TRUE when both bounds parsed (F-test df
    artifact already excluded at parse time). Resolves MetaESCI 2B.
  - ci_level_mismatch (character) — categorical {match, 90_vs_95_anova,
    implausible, unstated_assumed_95, NA}. Compares parsed level against
    the APA-95% canonical default. Resolves MetaESCI 2C.
  - ci_clipped_to_bound (character) — {none, lower_0, upper_1, both, NA}
    for bounded ES families (η², η_p², R², ω², ε², generalized η², V,
    φ). Resolves MetaESCI 2D.
  - ci_symmetry_class (character) — categorical refinement of the
    existing ci_symmetry ratio: {symmetric_expected,
    asymmetric_expected, symmetric_unexpected, asymmetric_unexpected,
    NA}. Resolves MetaESCI 2E.

Frontend (escimate.app)

  - CI block now renders ci_width_ratio, ci_level_source, the new
    ci_level_mismatch / ci_clipped_to_bound / ci_symmetry_class chips, a
    "CI expected, missing" badge, and an APA-7 precision row with
    precision-mismatch warning. Decision-error reason now appears as a
    tooltip on the badge. Downgrade-reason chips
    (decision_error_downgraded, unknown_groups_downgraded,
    r2_cross_pairing_detected) surfaced as inline status indicators
    instead of being hidden in raw metadata.
  - Notes block surfaces software_notes, best_practice_notes, and
    alternative_formulas (previously visible only in the metadata
    panel).

                        Changes in version 0.3.4                        

Addresses MetaESCI v0.3.4 request: 42 Category A ERROR false positives
where reported eta2/etap2 was cross-matched to cohens_f/cohens_f2
without detection.

Check: Phase 8D Signal 14 — eta/f cross-family detection (E11)

  - Phase 8D entry condition now includes eta2, etap2, generalized_eta2
    alongside the existing R2, adjusted_R2, f2, cohens_f.
  - New Signal 14: when reported eta2/etap2 is matched to
    cohens_f/cohens_f2/f2 with delta > 0.05, downgrades ERROR -> WARN
    with r2_cross_pairing_detected. Standalone (no contextual signals
    needed) — same rationale as Signal 13: both eta2 and cohens_f are
    deterministic from F, so any mismatch means the reported value came
    from a different analysis.
  - Symmetric: also handles cohens_f reported + eta variant matched.
  - Expected impact: MetaESCI Category A ERRORs 42 -> ~6, raw ERROR rate
    0.128% -> ~0.06%.

                        Changes in version 0.3.3                        

Follow-up to 0.3.2 addressing MetaESCI v0.3.3 request: the E8 pre-strip
was a no-op on real docpluck output.

Parse: thousand-sep comma strip now handles spaces after comma (E8 follow-up)

  - The v0.3.2 regexes in parse.R required t(2,758) with no space — but
    docpluck v1.4.4's A4 paren-spacing normalizer always emits t(2, 758)
    with a space. The fix matched the pre-A4 raw text we'd been shown in
    the MetaESCI report, not the actual post-normalizer input. Net
    effect in v0.3.2: zero rows recovered on the PSPB article
    10.1177/0146167220905712.
  - Fix: add \s* after the comma in all three pre-strip regexes
    (t/H/r/z, F, chi-square-N). Single-character change per regex.
  - New tests: parallel with-space cases for all three statistic types,
    plus an end-to-end 5-line fixture copied verbatim from docpluck's
    output on 0146167220905712 (t(2, 758) = -2.96, ...).

                        Changes in version 0.3.2                        

Follow-up to 0.3.1 addressing MetaESCI requests E8 and E10.

Parse: thousand-separator commas in test-statistic parens (E8, HIGH)

  - parse.R / normalize_text() previously let the decimal-comma
    converter mis-normalize t(2,758) as t(2.758), after which parse.R
    silently read it as Welch df=2.758 and back-computed N≈5. In the
    MetaESCI 339-PDF pre-test, PSPB article 10.1177/0146167220905712
    dropped 47 rows due to this, since every subsequent check treated
    the garbage df as genuine and the results were rejected downstream.
  - Fix: normalize_text() now strips thousand-separator commas from
    inside t(...), F(...), F[...], H(...), r(...), z(...) and
    chi-square(df, N = ...) parens before the decimal-comma converter
    runs, the same way N = 1,234 is already pre-stripped.
  - Handles t(2,758), F(2, 1,234), F[1, 2,500], and chi-square(3, N
    = 1,542) — with an iterative pass so N = 12,345,678-style
    multi-comma numbers survive.
  - New tests: test-parse.R now covers the t/F/chi-square cases and an
    end-to-end check_text() assertion that df=2758 round-trips.

Compute: Cohen's dz CI uses noncentral-t inversion (E10, MEDIUM)

  - ci_dz() / ci_dz_all() previously claimed a "noncentral_t" method but
    actually computed qt(alpha/2, df, ncp = dz*sqrt(n)) / sqrt(n) —
    i.e., quantiles of a single noncentral-t distribution, not the
    Algina & Keselman (2003) inversion. For small n this returned bounds
    that could be dramatically wrong: e.g., dz=0.55, n=9 returned a
    one-sided-looking [-1.66, 0.05]-style interval instead of the
    correct [-0.17, 1.24].
  - Fix: new internal ci_dz_noncentral_t() uses MBESS::ci.sm() (the
    reference implementation of standardized-mean CI inversion) when
    available, falling back to a stats::uniroot()-based inversion that
    solves for the noncentrality parameters whose α/2 and 1−α/2
    quantiles equal the observed t = dz·√n. The normal-approximation
    fallback is unchanged and still available when inversion fails.
  - MetaESCI reported 20 divergent rows between the legacy
    run_escicheck.R pipeline and 0.3.1. Under 0.3.2 the new
    implementation agrees with MBESS on the fixture dz = 0.55, n
    = 9, 95% CI, which is what legacy ci.sm returned — so the 20
    CI-width-ratio discrepancies should resolve.
  - Downstream impact: any downstream consumer tracking ci_match_rate
    for Cohen's dz (and ci_dz_all) will see bounds shift. This is a
    correctness fix, not a silent behavior change — flag it in your
    analysis plan.
  - New tests: test-golden-exact.R pins the dz = 0.55, n = 9 fixture and
    adds sanity checks for dz = 0, n = 20 (symmetric) and dz = 0.5, n
    = 100 (narrow).

Parse: decimal-comma no longer corrupts author affiliation markers

  - normalize_text() previously fired the decimal-comma → decimal-dot
    conversion on author affiliation footnotes like Braunstein1,3
    (multi-affiliation) and Wagner1,3,4, rewriting them to Braunstein1.3
    / Wagner1,3.4. The corruption shifted context windows enough to flip
    at least one eLife t-test result from WARN to OK on a real paper.
  - Fix: add a negative lookbehind (?<![a-zA-Z,]) on both decimal-comma
    gsubs so a letter (or a preceding comma, for the middle of
    a 3-affiliation run) blocks the match. The trailing lookahead was
    also tightened from [^0-9] to [^0-9a-zA-Z] to block the 1,3Boryana
    converse case. The second rule's leading quantifier was changed from
    \d* to \d+ so the match is always anchored at a real digit, letting
    the lookbehind check the character before that digit rather than the
    character before the comma.
  - 5 new tests in test-extraction-quality.R cover Braunstein/Wagner
    affiliation blocks, the 1,3Boryana converse, the stat-expression
    case that must still convert (d = 0,45), and the
    thousands-separator-in-N regression guard.

E9 — Smaller parse.R gaps (deferred, needs repro bundle)

  - The 13-row residual across 5 PSPB/JESP/RSOS/MP sources needs the
    staged .txt files from MetaESCI's
    data/results/subset_metaesci_regression_textstaging/ directory,
    which was not available at the time of triage. Will investigate when
    repro bundle is attached.

                        Changes in version 0.3.1                        

This is a housekeeping release packaging the v0.3.0f → v0.3.0n bug-fix
wave with a stable CRAN-style version number, batch-stdout hygiene, a
schema stability test, and a new decision_error_reason diagnostic
column. Addresses MetaESCI requests E1–E4 and E7.

DESCRIPTION version sync (E2)

  - DESCRIPTION Version: bumped from 0.3.0 (which covered every build
    v0.3.0 → v0.3.0n) to 0.3.1. Downstream pipelines can now
    discriminate the v0.3.0n bug-fix wave from earlier v0.3.0 builds via
    packageVersion("effectcheck") alone instead of requiring a git SHA.

Batch stdout: noncentral-t overflow spam silenced (E1)

  - MBESS::ci.smd (via ci_d_ind_noncentral_t) printed a multi-line
    warning to stdout every time the noncentrality parameter exceeded
    R's ~37.62 accuracy limit. At corpus scale this could print hundreds
    of lines per batch and drown out per-PDF progress output.
  - Fix: preempt the overflow by routing |ncp| > 37.62 directly to the
    large-sample normal approximation (which is no less accurate than
    MBESS's iterative fallback at that regime). Remaining MBESS calls
    are additionally wrapped in utils::capture.output() as a
    belt-and-suspenders silencer.
  - Behaviour change: a small number of d-CI computations will now be
    tagged method = "normal_approx" instead of "noncentral_t". The
    numerical difference is below the effect-size tolerance and does not
    affect PASS/WARN/ERROR status assignment.

Schema stability test (E3)

  - Added tests/testthat/test-schema-stability.R. The test asserts that
    check_text() returns a tibble containing every MetaESCI- critical
    column (source, check_scope, check_type, status, uncertainty_level,
    uncertainty_reasons, unknown_groups_downgraded,
    r2_cross_pairing_detected, decision_error_downgraded,
    design_ambiguous, ci_match, ci_check_status, ci_method_match,
    ci_width_ratio, ci_symmetry, decision_error, plus new
    decision_error_reason). An optional second check runs against a
    fixture PDF via the EFFECTCHECK_TEST_PDF env var and asserts the
    column set and element types are identical between check_text() and
    checkPDF(). By construction both paths funnel through
    process_files_internal() → check_text(), so this is an invariant
    guard against future regressions.

New column: decision_error_reason (E7)

  - Every row now carries a decision_error_reason character column. For
    rows where decision_error == FALSE the value is NA. For rows where
    decision_error == TRUE the value is one of:
      - reported_sig_computed_ns — reported p < alpha but recomputed p
        >= alpha (claimed significance does not reproduce).
      - reported_ns_computed_sig — reported p >= alpha but recomputed p
        < alpha (claimed non-significance does not reproduce).
      - ns_label_vs_computed_sig — paper reports "ns"/"not significant"
        but recomputed p < alpha.
      - other — catch-all for future decision-error variants.
  - Downstream analysis (e.g. MetaESCI analysis.Rmd) can now break
    decision errors down by mechanism without reparsing raw_text.

Expected row-count delta vs v0.3.0f (E4 — MetaESCI batch guidance)

On the MetaESCI metaesci_regression 200-PDF frozen benchmark (seed 42),
comparing v0.3.0f (last full batch) to v0.3.0n / 0.3.1:

| subset | v0.3.0f rows | v0.3.0n rows | delta | v0.3.0f ERRORs |
v0.3.0n ERRORs |
|----------------------|-------------:|-------------:|---------------:|---------------:|---------------:|
| meta_psychology (139)| 464 | 464 | 0 | 0 | 0 | | metaesci_regression
| 2,209 | 3,385 | +1,176 (+53%) | 13 | 0 |

The +53% row-count delta on metaesci_regression is driven by parser
gains, not a config-default change (plausibility_filter and try_tables
defaults are unchanged). The new rows come from:

  - Phase 7 multi-predictor regression rows that v0.3.0f short-circuited
    before v0.3.0m's multi-predictor-beta fix landed.
  - F-test rows previously lost to the F ≈ 0 crash (fixed in v0.3.0n).
  - z-test CI checks (1,517 previously UNVERIFIABLE rows now produce
    ci-check output, fixed in v0.3.0m).

Downstream consumers must re-derive all aggregate numbers from a fresh
v0.3.1 batch — old v0.3.0f aggregates are not directly comparable.
The 13 → 0 ERROR reduction on metaesci_regression is real (v0.3.0n's F
≈ 0 crash fix + multi-predictor-beta fix), not artefactual.

No columns were added or removed vs v0.3.0n other than the new
decision_error_reason column described above.

                        Changes in version 0.3.0                        

Bug fixes (MetaESCI v0.3.0m batch deep-dive)

  - Fixed ~100 F-test crashes ('list' object cannot be coerced to type
    'double') for F near zero. The v0.3.0m defensive guard covered
    Phase 5 matching but missed the Phase 6 CI-fallback path at
    check.R:2809, which extracted computed_variants[[eff]]$ci without a
    tryCatch. Now mirrors the guarded pattern already in use above.
  - Fixed ~61 false positive ERRORs from multi-predictor regression rows
    where both unstandardized b and standardized beta are reported with
    different values (e.g., b = 4.12, beta = 0.29). v0.3.0m only
    detected the b == beta masquerade. effectcheck computes
    single-predictor standardized_beta_from_t which cannot match a
    multi-predictor reported beta — the comparison is now skipped with a
    "multi-predictor regression" uncertainty note instead of flagging
    ERROR.

Bug fixes (MetaESCI batch validation)

  - Fixed 153 false positive ERRORs from unstandardized regression
    coefficients (b = 0.29) being compared to computed standardized
    beta. Parser's pat_eta regex matched "eta" inside "beta",
    mislabelling effect sizes. Added negative lookbehind and
    b-masquerade detection.
  - Fixed 27 F-test computation crashes (NaN delta) from list-type
    values in Phase 5 matching and infinite CI bounds from ci_cohens_f
    when eta-squared is near 1.0. Added defensive guards throughout.
  - Added z-test CI computation for d, dz, and r variants. Previously
    1,517 z-test results with reported CIs showed UNVERIFIABLE status.

Enhancements

  - Recognize r as a self-verifying effect size. When r is reported as
    the test statistic (e.g., r(48) = .42), it now routes through
    effect-size checking with PASS status, not p_value_only with OK.

Bug fixes

  - Fixed bind_rows() crash during batch processing when MBESS
    noncentral F-inversion returns non-numeric types under extreme
    noncentrality parameters (>37.62). ciL_computed became a list
    instead of double, crashing dplyr::bind_rows(). Now coerced to
    numeric with NA fallback.

Bug fixes

  - Fixed decimal CI level parsing: papers reporting "99.9% CI", "99.5%
    CI" produced ci_level values of 0.09, 0.05 due to (\d+)% regex
    failing on decimal percentages. Changed to (\d+\.?\d*)% with
    plausibility guard (ci_level < 0.50 falls back to 0.95).

Bug fixes and cleanup

  - Fixed eta-squared (η²) extraction from PDFs using Unicode
    mathematical italic characters.
  - Removed WebR/Private mode (archived to archive/webr/). Frontend is
    Cloud-mode only.
  - Optimized pdftotext extraction pipeline.

Bug fixes and cleanup

  - Fixed bare Cohen's f not recognized for t-tests. t(287.58) = -0.21,
    p = 0.837, f = -0.01 now correctly extracts f = -0.01 for any test
    type (was gated to F-tests only).
  - Removed Shiny app (shiny/ directory, start_shiny.bat). Next.js
    frontend is the sole UI.

Bug fixes and new features

  - Fixed F-test df falsely parsed as CI bounds: F(2, 76) = 3.45 no
    longer produces ciL=2, ciU=76. pat_CI4 now checks if matched values
    equal df1/df2 and skips them.
  - Multi-method CI verification: Phase 6 now tries all available CI
    computation methods (noncentral t, normal approximation,
    Steiger 2004 for eta-squared) and picks the best match. New output
    columns: ci_delta_upper, ci_check_status, ci_method_match,
    ci_width_ratio, ci_symmetry.
  - Removed dead CI pattern matching code.
  - 1286 total tests passing.

Bug fixes

  - Page number artifact guard: strip standalone digits before
    line-break joining to prevent page numbers from being concatenated
    with statistics.

Parser fixes and artifact detection

Addresses 13 false positive ERRORs from MetaESCI v0.3.0c validation
(132,537 results, 24 ERRORs). Expected: 24 -> ~10 ERRORs.

Bug fixes

  - Capital D/G effect sizes now parsed: D = 0.44, Hedges' G = 0.85, Dz
    = 0.40 all correctly matched (was: returned NA). 5 confirmed cases.
  - Generalized eta-squared (geta-squared, Geta-squared, generalized
    eta-squared) now correctly labeled as generalized_eta2 and routed to
    NOTE (was: parsed as plain eta2, producing false ERRORs). 8 cases.
    Generalized eta-squared cannot be computed from F/df (Bakeman 2005).
  - d-vs-t cross-check: when |d| > 3 and far exceeds the maximum
    plausible d from t and df, flags as extraction artifact (NOTE).
    Catches two-column PDF interleaving garbled values.
  - d > 10 rejection extended to dz/dav/drm (was: only d/g). Catches 43
    line-number artifacts (dz=219, dz=388, etc.).
  - d > 5 integer + spurious context guard extended to all d-family
    types.

New features

  - Phase 8G: heuristic generalized eta-squared detection. When reported
    eta2 < computed partial eta2 with ratio 0.10-0.95, downgrades ERROR
    to WARN with explanatory note.

  - Phase 14: cross-result effect size sweep. When a result has ERROR,
    tries matching the reported effect size against ALL other test
    statistics in the same article. Reports all attempts to the user. If
    a match is found with a different statistic, downgrades to WARN with
    cross-pairing note. Covers eta2/omega2/f from F, d/g/dz from
    t/F(1,df), V/phi from chi-square, r from t.

  - Cross-type effect size conversions: t-test now computes eta2,
    omega2, Cohen's f, and R2 as alternatives (t-test = F(1,df)
    equivalence). r-test computes d = 2r/sqrt(1-r^2). z-test computes r
    = z/sqrt(z^2+N). Chi-square computes Cohen's w, contingency
    coefficient C, and d from phi (for 2x2 tables). All cross-type
    matches are alternatives — they activate when the author reports an
    unconventional effect size for the test type.

Tests

  - 54 new tests (1236 total, 0 failures, 0 regressions)

                        Changes in version 0.2.8                        

Design ambiguity improvements

Addresses 399 remaining ERRORs from MetaESCI v0.2.7 audit (132,499
results). Philosophy: compute ALL plausible alternatives under different
design assumptions; if ANY alternative matches, downgrade severity.

New features

  - z-test paired variants: added dz = z/sqrt(N) (paired/Wilcoxon
    assumption) alongside existing d = 2z/sqrt(N)
    (independent/Mann-Whitney). Also computes dav, drm via r-grid sweep
    and gz, gav, grm Hedges-corrected variants.
  - Phase 8A-bis: structural design ambiguity detection for t-tests,
    F(1,df), and z-tests. When both independent and paired variant
    families are computed, promotes ambiguity_level to "ambiguous"
    regardless of which variant matches best.
  - Phase 8D Signal 9: large R-squared delta (>0.5) with contextual
    signals now triggers cross-pairing detection. R-squared is bounded
    [0,1], so delta>0.5 means F and R-squared are almost certainly from
    different models.
  - Cramer's V multi-m: when df allows multiple table dimensions, tries
    all m candidates and picks the one producing V closest to reported
    value.

Bug fixes

  - Relaxed extraction_suspect guard in Phase 8B for design-ambiguous
    cases. Large deltas are expected (d-from-t vs d-from-raw differs ~2x
    for paired designs) and should not block the design-ambiguous
    downgrade. Range guard ensures genuinely wrong values still produce
    ERROR.
  - Extended Phase 8C (unknown group sizes downgrade) to cover z-tests.
  - Relaxed Phase 8D ANOVA context guard for extreme R-squared deltas
    (>0.5) when regression signals are also present.

Internal

  - New compute function: dz_from_z(z, N) — paired d from z-statistic
  - 27 new tests in test-v028-design-ambiguity.R (1013 total, 0
    failures)

                        Changes in version 0.2.7                        

Bug fixes and API improvements

Bug fixes

  - Fixed devtools::load_all() calls in test files that broke R CMD
    check in CI (devtools is not available in CI environment)
  - Fixed codoc mismatch: unknown_groups_action parameter was missing
    from Rd documentation for check_text() and compute_and_compare_one()
  - Added min_confidence parameter forwarding in plumber.R API

Documentation

  - Removed working documents from public repo (FEEDBACK_RESPONSE,
    METAESCI_REPORT, DEVELOPMENT, testingai) — moved to archive/
  - Updated API.md to version 0.2.7 with unknown_groups_action and
    min_confidence parameter documentation
  - Version bump from 0.2.6 to 0.2.7 (aligns DESCRIPTION with frontend)

                        Changes in version 0.2.6                        

Design ambiguity + decision error fixes

Based on MetaESCI analysis of 132,499 results from 8,415 articles. These
changes reduce the ERROR false positive rate from ~3.9% to ~0.8%.

Design-ambiguous t-test downgrade (check.R)

  - New design_ambiguous_action parameter (default "WARN"). When a
    t-test or F(1,df) effect size ERROR occurs with ambiguous variant
    matching, the status is downgraded to WARN with confidence capped
    at 4. This reflects the known limitation that d computed from
    t-statistics systematically differs from d computed from raw data
    (means/SDs).

Decision error requires reported p-value (check.R)

  - Decision errors now require an explicitly reported p-value. Without
    one, there is no author's significance decision to check. Fixes
    false decision errors for regression z-statistics from coefficient
    tables and other extraction-only results.

r-test global N guard (check.R)

  - Decision errors suppressed for r-tests when sample size was inferred
    from global text (N_source == "global_text"). The globally-inferred
    N may not apply to this specific correlation (e.g., subgroup
    analysis).

API changes

  - New parameter: design_ambiguous_action (forwarded via plumber.R)
  - Fixed: method_context_action was missing from plumber.R option map

                        Changes in version 0.2.5                        

PDF extraction quality improvements

Based on MetaESCI extraction analysis of 121,040 results from 8,415 PDFs
across 7 journals. These changes reduce PDF extraction artifacts
affecting statistical parsing from ~6.5% to ~0.6%.

Header/footer stripping (utils-pdf.R)

  - New strip_headers_footers() function removes repeated lines (5+
    occurrences, 15-120 chars) from pdftotext output. Fixes
    page-number-appended-to-p-value artifacts.

Dropped decimal recovery (parse.R)

  - p < 001 now corrected to p < .001 during normalization.
  - p = NNN where NNN has 3+ digits (e.g., p = 484) corrected to p =
    .NNN (e.g., p = .484). Flagged as extraction_suspect with assumption
    note in uncertainty_reasons.
  - New p_decimal_corrected column in parsed output tracks which
    p-values were corrected.

General line-break joining (parse.R)

  - Lines ending with =, <, or > followed by a digit on the next line
    are now joined. Catches edge cases like F(1, 30) =\n4.425 that
    existing stat-specific patterns missed.
  - Lines where ( is followed by a line break then a digit are joined
    (broken df).

Standalone page number removal (parse.R)

  - Lines containing only 1-3 digits are removed during normalization
    (page numbers).

Computation-guided decimal recovery (check.R — Phase 5B)

  - When extraction_suspect is triggered by an extreme delta, the
    pipeline now tries all possible decimal placements of the reported
    effect size (e.g., 615 → 61.5, 6.15, 0.615) and checks if any
    matches the computed value within tolerance.
  - If a match is found, the effect size is recovered with a
    decimal_recovered flag and detailed assumption note. Status is
    re-evaluated (may become PASS/WARN).
  - Uses computed values as oracle — self-verifying, zero risk to
    correct data.
  - Also flags p-values that were decimal-corrected during
    normalization.

New columns

  - decimal_recovered: TRUE when Phase 5B successfully recovered a
    dropped decimal
  - p_decimal_corrected (parse output): TRUE when normalization
    corrected a dropped decimal in p-value

Tests

  - 39 new extraction quality tests in test-extraction-quality.R
  - 878 total tests passing (0 failures)

                        Changes in version 0.2.4                        

Validation-driven improvements

Based on comprehensive validation of 19,690 results across 7 journals
(MetaESCI).

Bug fixes (Category A — 673 results)

  - warn_tiny_delta (48 results): Decision error no longer upgrades
    PASS→WARN when effect size match is excellent (delta < 0.5x
    tolerance) and ambiguity is clear.
  - method_context_in_result (527 results, 58 ERRORs): New
    method_context_in_chunk flag distinguishes method keywords IN the
    stat's sentence vs nearby context. ERROR status capped at NOTE for
    in-chunk method contexts (power analysis, meta-analysis, etc.).
  - cross_type_error (68 results, all ERROR): Phase 5 same-type matching
    now includes alternatives (e.g., g_ind for t-tests). Previously only
    computed_variants were searched, missing valid same-type matches.
  - effect_not_in_text (18 results, 15 ERRORs): Parse-time rejection of
    impossible effect sizes: R2/V/phi/eta2 > 1.0, round-integer d/g > 5.
  - suspicious_decision_error (25 results, 4 ERRORs):
    effect_test_mismatch flag now caps ERROR→NOTE for type-incompatible
    effect sizes (e.g., chi2 with R2=52).
  - garbled_p_threshold (4 results): Extended garbled p-value detection
    to non-inequality p > 0.5 with large computed discrepancy.

Extraction guards (Category B — 41 PDF extraction artifacts)

  - Computed-side plausibility: If computed effect size exceeds
    plausibility bounds (e.g., computed d=13.49 from garbled t-stat),
    flags extraction_suspect and caps ERROR→NOTE.
  - Stat value plausibility: Flags |t| > 100 and F > 10000 as possible
    artifacts.
  - DF plausibility: Flags df ≤ 0 or df > 50000 as possible artifacts.
  - Tightened bounds: d/g/dz/dav/drm plausibility from 10→5.

New features

  - Confidence score (confidence column, 0-10 integer): Deterministic
    quality score aggregating ambiguity level, match type, delta
    distance from threshold, design inference, and extraction quality.
  - Result context (result_context column): "study" or "method"
    classification.
  - method_context_action parameter: Controls behavior for
    method-context stats ("NOTE", "WARN", or "SKIP"). Default: "NOTE".
  - min_confidence parameter: Minimum confidence score for output
    filtering. Results below threshold are dropped. Default: 0 (no
    filtering).

                 Changes in version 0.2.3 (2026-03-25)                  

New features

  - User configuration: New cross_type_action, ci_affects_status, and
    plausibility_filter parameters for check_text().
  - Plausibility bounds: Implausibly large effect sizes (e.g., d > 10, r
    > 1) flagged via extraction_suspect column. Configurable via
    EFFECT_PLAUSIBILITY in constants.
  - Effect size families: OR, RR, IRR, h added to EFFECT_SIZE_FAMILIES.
  - Completeness: Cohen's f² = r²/(1-r²) for r-tests.
  - Beginner-friendly UI: Legend, tooltips, plain English verdicts,
    narrative report mode in frontend.
  - Paste-text input: Direct text input mode in frontend.

Bug fixes

  - Fixed F→t conversion missing Hedges' g variant (g_ind).
  - Added d_ind_min/d_ind_max bounds for F-test conversions.
  - Fixed chi-square inline N parsing and multi-stat sentence handling.
  - Fixed section-number false positives in p-value extraction.
  - Fixed ns (non-significant) notation parsing.

API changes

  - Options forwarding in plumber.R via do.call().
  - Summary and version fields included in all API responses.

                        Changes in version 0.2.2                        

Bug fixes

  - Cross-type matching: Best match now selected across effect size
    types, reducing false ERROR rate from 47.6% to ~10%.
  - Paired N fix: Paired designs no longer double the sample size
    incorrectly.
  - NOTE→PASS: Results matching within tolerance now correctly report
    PASS instead of NOTE.
  - check_type column: Added to output for transparency in variant
    matching.
  - extraction_suspect flag: Extreme deltas flagged for manual review.

New features

  - SKIP status: Extraction-only results (no p-value or effect size to
    verify) now get status "SKIP" instead of misleading WARN.
  - Universal tail fallback: Phase 9 tries both one-tailed and
    two-tailed when decision error occurs, resolving to NOTE with
    explanation.
  - Two-tailed detection: New two_tailed_detected flag overrides
    one-tailed when both present in text.
  - Method context detection: method_context_detected flag suppresses
    decision_error for p-curve, equivalence test, TOST contexts.
  - N candidates for r-tests: Multiple N values extracted from context;
    best match selected with assumption note.
  - One-tailed scope fix: one_tailed_detected now searches chunk only,
    preventing cross-chunk bleeding.
  - Garbled p-value detection: p-values like p < 0.645 flagged as
    extraction_suspect.

                        Changes in version 0.2.1                        

Bug fixes

  - Fixed thousands-separator N parsing (e.g., N = 1,182 no longer
    misread as N = 1).
  - Fixed Welch t-test N estimation.
  - Removed officer dependency; DOCX extraction via pandoc only.
  - Fixed DOCX segfault crash.

Improvements

  - Cold start progress indicator for Render free tier.
  - Page-load health check and server status banner for classroom use.
  - Paste-text input mode.
  - Parser robustness: 7 fixes from AI verification testing across 12
    journal styles.

CRAN-related

  - Fixed invalid URI, APA spelling, broadened package scope
    description.
  - Fixed effectcheck namespace refs for shinyapps.io deployment.
  - Removed pdftools dependency.
  - Added development disclaimers throughout package and app.

                        Changes in version 0.2.0                        

New features

  - Nonparametric tests: Mann-Whitney U, Wilcoxon W, and Kruskal-Wallis
    H test parsing and effect size computation (rank-biserial r, Cliff's
    delta, epsilon-squared, Kendall's W).
  - Regression support: b/SE coefficient parsing, t = b/SE verification,
    regression type inference, standardized beta computation.
  - HTML report generation: generate_report() produces self-contained
    HTML reports with executive summary, color-coded rows, and
    interactive tables. render_report() provides a convenience wrapper.
    PDF fallback available.
  - CSV and JSON export: export_csv() and export_json() for
    machine-readable output.
  - statcheck comparison: compare_with_statcheck() and
    compare_file_with_statcheck() for side-by-side comparison with
    statcheck results.
  - Variant helper functions: get_variants(), get_same_type_variants(),
    get_alternatives(), format_variants(), compare_to_variants(),
    get_variant_metadata(), get_effect_family().

Bug fixes

  - Fixed missing r p-value computation.
  - Fixed missing CI for paired designs (ci_dz).
  - Corrected eta-squared CI formula (partial eta-squared).
  - Fixed MBESS ci.sm -> ci.smd call.
  - Fixed drm_from_dz formula.
  - Fixed Cohen's f regex false positives.
  - z-test pattern now checked last to avoid capturing auxiliary
    z-values from U/W tests.

Parser improvements

  - Ultra-robust Unicode normalization (186+ character mappings).
  - PDF column interleaving handled via pdftotext.
  - Sub-chunking for multi-stat sentences.
  - Greenhouse-Geisser corrected df support.
  - Comprehensive handling of PDF extraction artifacts.

                        Changes in version 0.1.0                        

  - Initial release.
  - Support for t-tests, F-tests/ANOVA, correlations (r), chi-square,
    and z-tests.
  - APA-style parsing from text, PDF, HTML, and DOCX documents.
  - Conservative variant-matching approach computing all plausible
    effect sizes.
  - statcheck-compatible API (checkPDF, checkHTML, checkPDFdir, etc.).
  - Decision error detection (significance reversals).
  - S3 class with print, summary, and plot methods.