Pilot 1 fixes: parenthesized DOIs, docx routing in verify_references,

protocol truncation guard, exact-match counts (0.11.1)

All four surfaced by the first full blinded audit of the demo fixture
(11/12 planted defects caught, zero false positives, four unplanned
real defects found):

- extract_dois (new helper, R/citations.R): Crossref-recommended DOI
  character class keeps parentheses, so legacy Elsevier DOIs like
  10.1016/S1364-6613(03)00028-7 no longer truncate into false 404s;
  trailing punctuation and unbalanced closers from prose are stripped
- verify_references_impl routes file input through read_as_text_lines,
  so Pass 4 line-range mode on a .docx extracts text instead of
  returning raw zip bytes
- cat_protocol: reviewer_zero_prompt and referee_prompt write the
  composed protocol to a temp file and announce the path first, so a
  truncated console printout is one read_file from recovery
- reviewer_zero.md Step 3b: integer counts (Ns, dfs, tallies) must
  match exactly; all.equal tolerance is for continuous statistics only

CI: extract_dois functional test. R CMD check: Status OK.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: IMNMV

.github/scripts/checks.R

@@ -437,5 +437,17 @@ r <- tryCatch({
 }, error = function(e) conditionMessage(e))
 if (isTRUE(r)) pass("cross-restart history: past logs parsed, agent filter works") else fail("past history:", r)
 
+# --- 14. DOI extraction: parenthesized DOIs, trailing junk, prose parens ---
+r <- tryCatch({
+  d <- env$extract_dois(paste(
+    "Monsell, S. (2003). Task switching. TiCS. https://doi.org/10.1016/S1364-6613(03)00028-7",
+    "Also see (doi: 10.1037/a0019842) and https://doi.org/10.1126/science.1201068.",
+    sep = "\n"
+  ))
+  all(c("10.1016/S1364-6613(03)00028-7", "10.1037/a0019842",
+        "10.1126/science.1201068") %in% d) && length(d) == 3
+}, error = function(e) conditionMessage(e))
+if (isTRUE(r)) pass("extract_dois: parens kept, prose parens and trailing dots stripped") else fail("extract_dois:", r)
+
 if (!ok) quit(status = 1)
 cat("\nAll checks passed.\n")

DESCRIPTION

@@ -1,6 +1,6 @@
 Package: ClaudeR
 Title: R Integration for Claude AI
-Version: 0.11.0
+Version: 0.11.1
 Authors@R: person("Nykko", "Vitali", email = "nykvt@icloud.com", role = c("aut", "cre"))
 Description: Connects RStudio with Claude AI to enable interactive coding sessions.
 License: MIT + file LICENSE

R/citations.R

@@ -75,6 +75,33 @@ get_bibtex_impl <- function(doi) {
   bib
 }
 
+# Extract DOIs from text using the Crossref-recommended character class,
+# which allows parentheses: legacy Elsevier DOIs like
+# 10.1016/S1364-6613(03)00028-7 are common in psychology bibliographies and
+# were truncated at the paren by the old pattern, producing false 404s.
+# Trailing punctuation and unbalanced closing brackets (from prose like
+# "(doi: 10.1037/a0019842)") are stripped after matching.
+extract_dois <- function(text) {
+  pattern <- "10\\.\\d{4,9}/[-._;()/:a-zA-Z0-9]+"
+  dois <- regmatches(text, gregexpr(pattern, text, perl = TRUE))[[1]]
+  dois <- unique(trimws(dois))
+  dois <- sub("[.,;:]+$", "", dois)
+  strip_unbalanced <- function(d) {
+    repeat {
+      last <- substr(d, nchar(d), nchar(d))
+      if (!last %in% c(")", "]")) break
+      opener <- if (last == ")") "(" else "["
+      n_open <- lengths(regmatches(d, gregexpr(opener, d, fixed = TRUE)))
+      n_close <- lengths(regmatches(d, gregexpr(last, d, fixed = TRUE)))
+      if (n_close > n_open) {
+        d <- sub("[.,;:]+$", "", substr(d, 1, nchar(d) - 1))
+      } else break
+    }
+    d
+  }
+  unique(vapply(dois, strip_unbalanced, character(1), USE.NAMES = FALSE))
+}
+
 # GET a Crossref API URL with retry/backoff. Crossref rate-limits bursts
 # (HTTP 429); a failed lookup mid-audit silently truncates the reference
 # check, so wait and retry before giving up.

R/ui.R

@@ -2179,13 +2179,15 @@ verify_references_impl <- function(file_path = NULL, text = NULL,
     return("Error: jsonlite package is required. Install with install.packages('jsonlite')")
   }
 
-  # Get text from file or direct input
+  # Get text from file or direct input. Manuscripts (.docx/.pdf) route
+  # through the structured extractor; a raw readLines() on a .docx returns
+  # zip bytes, which is how the Pass 4 line-range mode broke in the field.
   if (!is.null(file_path)) {
     file_path <- path.expand(file_path)
     if (!file.exists(file_path)) {
       return(paste0("Error: File not found: ", file_path))
     }
-    lines <- readLines(file_path, warn = FALSE)
+    lines <- read_as_text_lines(file_path)
     if (!is.null(start_line)) {
       end_l <- if (!is.null(end_line)) min(end_line, length(lines)) else length(lines)
       lines <- lines[max(1, start_line):end_l]
@@ -2195,11 +2197,9 @@ verify_references_impl <- function(file_path = NULL, text = NULL,
     return("Error: Either file_path or text must be provided")
   }
 
-  # Extract DOIs using regex
-  doi_pattern <- "10\\.\\d{4,9}/[^\\s,;\\]\\)>\"']+"
-  dois <- regmatches(text, gregexpr(doi_pattern, text, perl = TRUE))[[1]]
-  dois <- unique(trimws(dois))
-  dois <- sub("[\\.,;]+$", "", dois)
+  # Extract DOIs
+  dois <- extract_dois(text)
+  doi_pattern <- "10\\.\\d{4,9}/[-._;()/:a-zA-Z0-9]+"
 
   # Cap per call: each lookup blocks the R session, and the MCP bridge times
   # out at 120s. Large bibliographies should be paged via start_line/end_line.
@@ -2546,7 +2546,7 @@ reviewer_zero_prompt <- function(prereg_path = NULL, robustness = FALSE,
     txt <- paste0(txt, "\n", build_referee_text())
   }
 
-  cat(txt, "\n")
+  cat_protocol(txt)
   invisible(txt)
 }
 
@@ -2706,10 +2706,25 @@ referee_prompt <- function(lenses = c("logic", "methods", "consistency",
                             reviewers_per_lens = reviewers_per_lens,
                             model = model, cross_vendor = cross_vendor,
                             stance = stance)
-  cat(txt, "\n")
+  cat_protocol(txt)
   invisible(txt)
 }
 
+# Print a protocol, but write the composed text to a file first and announce
+# the path. Long protocols exceed the console-output cap (agents saw "262
+# lines elided" mid-protocol in the field); the file makes the full text one
+# read_file call away.
+cat_protocol <- function(txt) {
+  proto_file <- tempfile(pattern = "clauder_protocol_", fileext = ".md")
+  ok <- tryCatch({ writeLines(txt, proto_file); TRUE }, error = function(e) FALSE)
+  if (ok) {
+    cat("[Full protocol saved to:", proto_file,
+        "-- if this printout is truncated, read that file completely before starting.]\n\n")
+  }
+  cat(txt, "\n")
+  invisible(NULL)
+}
+
 #' Print the R Best Practices prompt template
 #'
 #' Displays the built-in R statistical analysis protocol based on

inst/prompts/reviewer_zero.md

@@ -257,6 +257,11 @@ claim_registry$status[i] <- if (is_match) "match" else if (is_rounding) "roundin
 
 R sets the status. You do not. This prevents eyeballing "close enough" values.
 
+Integer counts are exempt from tolerance: sample sizes, degrees of freedom,
+cell counts, and exclusion tallies must match EXACTLY. A count that differs
+by one is a discrepancy, never rounding. Apply `all.equal` tolerance only to
+continuous statistics.
+
 For claims with multiple values (e.g., "t(38) = 2.12, p = .041, d = 0.34"),
 test each value separately. If any single value is a discrepancy, the whole
 claim is a discrepancy.