File size: 36,695 Bytes
470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 470bcea 37cb069 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 | """Redact internal reviews into publishable ICSAC-branded review artifacts.
Operates on the authoritative internal review markdown in ``reviews/`` and
emits a sanitized version that is safe to publish on icsacinstitute.org.
The redaction removes all vendor/model identifiers, renames reviewers
generically ("Reviewer 1", "Reviewer 2", ...), drops internal workflow
detail (raw API error payloads, slot indices, fallback chains), and
replaces the disagreement flag with a human-readable consensus label.
A grep-gate (``assert_clean``) fails hard if any forbidden token survives
redacting. Callers must catch the exception and abort publication.
"""
from __future__ import annotations
import os
import re
from dataclasses import dataclass, field
# Hard-fail vendor/model identifiers. Any case-insensitive substring hit
# indicates a leak of panel composition and must abort publication.
#
# Tokens chosen to catch identity leaks (OpenRouter route paths, specific
# model family IDs, vendor names rarely legitimate in academic prose)
# WITHOUT catching subject-matter discussion of published transformers.
# A paper reviewing "GPT-2 and Gemma-2 activations" must pass the gate β
# those are scientific subjects, not panel self-references.
FORBIDDEN_VENDOR_TOKENS: tuple[str, ...] = (
# Infrastructure names β never appear in legitimate academic prose.
"openrouter",
"anthropic",
# OpenRouter route prefixes β the "/" guarantees a path, not a word.
"openai/",
"nvidia/",
"google/gemma",
"meta-llama/",
"z-ai/",
"minimax/",
"nousresearch/",
"qwen/",
"mistralai/",
"deepseek/",
"cognitivecomputations/",
"liquid/",
# Specific panel model families β narrow enough that a hit is a leak.
"nemotron",
"gpt-oss",
# Panelist name β small false-positive risk ("Claude Shannon") accepted
# to catch "As Claude, I..." self-reference leaks from slot 0.
"claude",
)
# Hard-fail credential/infra phrases. Substring match, case-insensitive.
# These are structurally always leaks β there's no legitimate prose
# reason for a paper review to contain these compounds.
FORBIDDEN_SECRET_PHRASES: tuple[str, ...] = (
"api key",
"api keys",
"access token",
"auth token",
"auth key",
"bearer token",
"secret key",
"private key",
"api token",
"api tokens",
"bearer ",
)
# Soft-warn tokens β bare "key", "api", "token", "google" that appear
# regularly in academic prose ("key findings", "Google Scholar", "tokenizer").
# Surfaced in the redaction report but do not abort. Operators can grep the
# published review manually if they want extra assurance.
SOFT_WARN_TOKENS: tuple[str, ...] = (
"google",
"api",
"token",
"key",
)
# Regex patterns indicating attempted exfiltration via review output.
# Added 2026-04-18 after prompt-injection attack-surface audit. Triggered
# by file paths pointing at our hosts, env-var assignments, and known
# credential prefixes. Match anywhere in the redacted review text.
FORBIDDEN_EXFIL_PATTERNS: tuple[str, ...] = (
# Absolute filesystem paths likely pointing at our hosts
r"/home/orangepi\b",
r"/home/dietpi\b",
r"/opt/orchestrator\b",
r"/etc/passwd\b",
r"/etc/shadow\b",
r"/root/",
r"\.config/[a-z][a-z0-9_-]*\.env\b",
r"C:\\\\Users\\\\",
# Env-var assignments of the form UPPER_SNAKE=longvalue
r"\b[A-Z][A-Z0-9_]{3,}=\S{8,}",
# Known credential prefixes
r"\bsk-ant-api03-[A-Za-z0-9_-]{8,}",
r"\bsk-[A-Za-z0-9]{20,}",
r"\bghp_[A-Za-z0-9]{10,}",
r"\bgho_[A-Za-z0-9]{10,}",
r"\bAKIA[0-9A-Z]{16}\b",
# Bearer tokens of non-trivial length following the keyword
r"\bBearer\s+[A-Za-z0-9._-]{32,}",
# Internal rubric filenames β reviewers and the RQC auditor occasionally
# echo filenames from the prompt ("drift from tone.md"). Public output
# must reference rubrics by prose name, never by repo filename. Rewriting
# pass runs first (_rewrite_rubric_filenames); this gate catches anything
# that slipped through.
r"\b(?:rubrics/)?(?:scope|methodology|ai-provenance|tone|calibration|review_quality_control)\.md\b",
)
# --------------------------------------------------------------------------
# Rubric filename β prose rewrite
# --------------------------------------------------------------------------
#
# The RQC rubric references sibling rubrics by filename ("the standard
# mirrors tone.md"). Audit justifications inherit that phrasing and leak
# internal filenames into public-facing text ("a soft but consistent drift
# from tone.md"). Rewrite before rendering; the hard-gate above catches any
# new filename that isn't in this map so a future rubric addition can't
# silently leak.
RUBRIC_FILENAME_PROSE: tuple[tuple[str, str], ...] = (
("rubrics/review_quality_control.md", "the audit rubric"),
("rubrics/ai-provenance.md", "the AI provenance rubric"),
("rubrics/calibration.md", "the calibration rubric"),
("rubrics/methodology.md", "the methodology rubric"),
("rubrics/scope.md", "the scope rubric"),
("rubrics/tone.md", "the tone rubric"),
("review_quality_control.md", "the audit rubric"),
("ai-provenance.md", "the AI provenance rubric"),
("calibration.md", "the calibration rubric"),
("methodology.md", "the methodology rubric"),
("scope.md", "the scope rubric"),
("tone.md", "the tone rubric"),
)
def _rewrite_rubric_filenames(text: str) -> str:
"""Rewrite rubric filename references to prose descriptions."""
if not text:
return text
out = text
for needle, prose in RUBRIC_FILENAME_PROSE:
pattern = re.compile(re.escape(needle), re.IGNORECASE)
out = pattern.sub(prose, out)
return out
@dataclass
class ParsedReview:
"""Structured view of a reviews/<id>_*.md file."""
record_id: str
title: str
doi: str
review_date: str
recommendation: str
disagreement: bool
dimension_rows: list[tuple[str, str, list[str]]] = field(default_factory=list)
reviewers: list[dict] = field(default_factory=list)
def _parse_frontmatter(body: str) -> tuple[dict, str]:
"""Strip YAML frontmatter; return (fields, remainder)."""
if not body.startswith("---\n"):
return {}, body
end = body.find("\n---\n", 4)
if end < 0:
return {}, body
raw = body[4:end]
rest = body[end + 5 :]
fields: dict = {}
for line in raw.splitlines():
if ":" not in line:
continue
k, v = line.split(":", 1)
fields[k.strip()] = v.strip().strip('"').strip("'")
return fields, rest
def _parse_aggregate_table(body: str) -> list[tuple[str, str, list[str]]]:
"""Pull rows out of the 'Aggregate Scores' markdown table."""
rows: list[tuple[str, str, list[str]]] = []
in_table = False
for line in body.splitlines():
stripped = line.strip()
if stripped.startswith("## Aggregate Scores"):
in_table = True
continue
if in_table and stripped.startswith("## "):
break
if not in_table or not stripped.startswith("|"):
continue
if set(stripped.replace("|", "").strip()) <= set("- "):
continue
cells = [c.strip() for c in stripped.strip("|").split("|")]
if len(cells) < 3 or cells[0].lower() == "dimension":
continue
scores = [s.strip() for s in cells[2].split(",") if s.strip()]
rows.append((cells[0], cells[1], scores))
return rows
def _split_reviewer_sections(body: str) -> list[tuple[str, str]]:
"""Extract [(heading, content), ...] for each '### <Model>' block."""
marker = "\n## Individual Model Reviews\n"
idx = body.find(marker)
if idx < 0:
return []
remainder = body[idx + len(marker) :]
end = remainder.find("\n---\n")
if end >= 0:
remainder = remainder[:end]
sections: list[tuple[str, str]] = []
current_head: str | None = None
current_lines: list[str] = []
for line in remainder.splitlines():
if line.startswith("### "):
if current_head is not None:
sections.append((current_head, "\n".join(current_lines).strip()))
current_head = line[4:].strip()
current_lines = []
else:
current_lines.append(line)
if current_head is not None:
sections.append((current_head, "\n".join(current_lines).strip()))
return sections
def _parse_reviewer_body(content: str) -> dict:
"""Pull recommendation, summary, dimension scores out of one section."""
if content.startswith("**Error:**"):
return {"error": True}
rec_match = re.search(r"\*\*Recommendation:\*\*\s*([A-Z_]+)", content)
sum_match = re.search(r"\*\*Summary:\*\*\s*(.+?)(?:\n\n|\Z)", content, re.S)
dims: list[tuple[str, str, str]] = []
for m in re.finditer(
r"^-\s+\*\*(?P<label>[^*]+)\*\*\s*\((?P<score>[^)]+)\):\s*(?P<just>.+?)(?=\n-\s+\*\*|\Z)",
content,
re.S | re.M,
):
dims.append(
(
m.group("label").strip(),
m.group("score").strip(),
" ".join(m.group("just").split()),
)
)
return {
"error": False,
"recommendation": rec_match.group(1) if rec_match else "N/A",
"summary": " ".join(sum_match.group(1).split()) if sum_match else "",
"dimensions": dims,
}
def parse_review_file(path: str) -> ParsedReview:
"""Load a reviews/<id>_*.md file into structured form."""
with open(path, "r", encoding="utf-8") as f:
text = f.read()
fm, body = _parse_frontmatter(text)
sections = _split_reviewer_sections(body)
reviewers = [
{"raw_heading": head, **_parse_reviewer_body(cont)} for head, cont in sections
]
title = fm.get("title", "").strip('"')
if title.lower().startswith("review:"):
title = title.split(":", 1)[1].strip()
return ParsedReview(
record_id=str(fm.get("record_id", "")),
title=title,
doi=fm.get("doi", ""),
review_date=fm.get("review_date", ""),
recommendation=fm.get("recommendation", "REVIEW_FURTHER"),
disagreement=str(fm.get("disagreement", "False")).lower() == "true",
dimension_rows=_parse_aggregate_table(body),
reviewers=reviewers,
)
def _consensus_label(parsed: ParsedReview) -> str:
"""Translate disagreement + score spread into reader-friendly label."""
max_spread = 0
for _, _, scores in parsed.dimension_rows:
nums = [float(s) for s in scores if re.match(r"^\d+(\.\d+)?$", s)]
if len(nums) >= 2:
max_spread = max(max_spread, max(nums) - min(nums))
if not parsed.disagreement and max_spread <= 1:
return "strong consensus"
if max_spread >= 2:
return "divided"
return "mixed"
def build_public_markdown(parsed: ParsedReview) -> str:
"""Render the sanitized review markdown (safe for public publication)."""
valid_reviewers = [r for r in parsed.reviewers if not r["error"]]
valid_n = len(valid_reviewers)
consensus = _consensus_label(parsed)
lines: list[str] = [
"---",
f'title: "Review: {parsed.title}"',
f'doi: "{parsed.doi}"',
f"record_id: {parsed.record_id}",
f"review_date: {parsed.review_date}",
f"recommendation: {parsed.recommendation}",
f"consensus: {consensus}",
f"reviewer_count: {valid_n}",
"---",
"",
"## Open review",
"",
(
f"This submission was evaluated by a panel of {valid_n} independent "
f"advanced AI reviewers scoring six dimensions. Panel consensus was "
f"**{consensus}**."
),
"",
"### Aggregate scores",
"",
"| Dimension | Mean | Per-reviewer |",
"|-----------|------|--------------|",
]
for label, mean, scores in parsed.dimension_rows:
lines.append(f"| {label} | {mean} | {', '.join(scores) or 'β'} |")
lines.extend([
"",
"### Reviewer assessments",
"",
(
"Individual reviewer assessments are collapsed by default. "
"Expand any row to read that reviewer's summary "
"and per-dimension justification."
),
"",
])
# Emit raw HTML for reviewer blocks so we can use <details> for
# collapsibility. Python-markdown passes block-level HTML through
# unchanged, so the rendered landing page gets native browser-handled
# expand/collapse on each reviewer without any JavaScript.
import html as _html
for idx, r in enumerate(valid_reviewers, start=1):
rec = _html.escape(r["recommendation"])
summary = _html.escape(_rewrite_rubric_filenames(r["summary"]))
lines.append(f'<details class="reviewer-detail">')
lines.append(f'<summary><strong>Reviewer {idx}</strong> β {rec}</summary>')
lines.append("")
lines.append(f'<p><strong>Summary:</strong> {summary}</p>')
if r["dimensions"]:
lines.append("<ul>")
for label, score, just in r["dimensions"]:
lines.append(
f' <li><strong>{_html.escape(label)}</strong> '
f'({_html.escape(score)}): {_html.escape(_rewrite_rubric_filenames(just))}</li>'
)
lines.append("</ul>")
lines.append("</details>")
lines.append("")
lines.extend(
[
"---",
"",
(
"*Reviews at ICSAC are open and transparent. AI tooling helps "
"the panel draft and structure each review; final acceptance "
"decisions rest with human editors. Reviews are published "
"alongside acceptance for accountability; individual reviewer "
"identities are abstracted to keep focus on the assessment "
"rather than the tooling behind it.*"
),
"",
]
)
return _humanize_internal_jargon("\n".join(lines))
def _find_substring_hits(text: str, tokens: tuple[str, ...]) -> list[tuple[str, int]]:
hits: list[tuple[str, int]] = []
lowered = text.lower()
for tok in tokens:
start = 0
while True:
at = lowered.find(tok, start)
if at < 0:
break
hits.append((tok, at))
start = at + 1
return hits
def _find_wordboundary_hits(text: str, tokens: tuple[str, ...]) -> list[tuple[str, int]]:
hits: list[tuple[str, int]] = []
for tok in tokens:
for m in re.finditer(rf"\b{re.escape(tok)}\b", text, flags=re.IGNORECASE):
hits.append((tok, m.start()))
return hits
@dataclass
class ScrubReport:
fatal_hits: list[tuple[str, int]]
warn_hits: list[tuple[str, int]]
@property
def clean(self) -> bool:
return not self.fatal_hits
class RedactionLeak(Exception):
"""Raised when a redacted artifact still contains a fatal token."""
def __init__(self, hits: list[tuple[str, int]], artifact_path: str | None):
self.hits = hits
self.artifact_path = artifact_path
preview = ", ".join(sorted({t for t, _ in hits}))
loc = f" in {artifact_path}" if artifact_path else ""
super().__init__(
f"Redacted review leaked forbidden tokens{loc}: {preview} "
f"({len(hits)} total hit(s))"
)
def _find_regex_hits(text: str, patterns: tuple[str, ...]) -> list[tuple[str, int]]:
"""Return [(matched_text, offset), ...] for each regex pattern that matches."""
hits: list[tuple[str, int]] = []
for pat in patterns:
for m in re.finditer(pat, text):
hits.append((f"[regex {pat!r}] {m.group(0)[:80]}", m.start()))
return hits
_REVIEWER_DETAIL_BLOCK = re.compile(
r'<details class="reviewer-detail">.*?</details>',
re.DOTALL,
)
_JARGON_REWRITES = (
(re.compile(r"\bSlots\b"), "Reviewers"),
(re.compile(r"\bslots\b"), "reviewers"),
(re.compile(r"\bSlot\b"), "Reviewer"),
(re.compile(r"\bslot\b"), "reviewer"),
)
def _humanize_internal_jargon(text: str) -> str:
"""Standardize public-facing language on "reviewer" instead of "slot".
Pipeline configuration calls each panel position a "slot"; reviewers
occasionally echo the term when describing peer assessments. For
public output we rewrite to "reviewer" so the report matches the
institute's published panel description and never exposes the
pipeline's internal naming.
"""
for pat, repl in _JARGON_REWRITES:
text = pat.sub(repl, text)
return text
def _strip_reviewer_prose(text: str) -> str:
"""Remove <details class="reviewer-detail"> blocks before vendor-token
screening.
Reviewer-detail blocks hold per-reviewer summary + dimension
justifications β prose that LEGITIMATELY describes what the submission
itself says, including author disclosures of AI-assisted writing or
references to prior-art papers (e.g. OpenAI/Gemma citations). A
reviewer noting "the author acknowledges Claude/Gemini as writing
assistants" is factual description, not panel-composition leak.
Structural content β frontmatter, section headers, the aggregate
scores table, the boilerplate footer β still receives the full vendor
check. Secret and exfil patterns are checked across the whole text.
"""
return _REVIEWER_DETAIL_BLOCK.sub("", text)
def scan(text: str) -> ScrubReport:
"""Return a hit report without raising."""
structural = _strip_reviewer_prose(text)
fatal = _find_substring_hits(structural, FORBIDDEN_VENDOR_TOKENS)
fatal.extend(_find_substring_hits(text, FORBIDDEN_SECRET_PHRASES))
fatal.extend(_find_regex_hits(text, FORBIDDEN_EXFIL_PATTERNS))
warn = _find_wordboundary_hits(text, SOFT_WARN_TOKENS)
return ScrubReport(fatal_hits=fatal, warn_hits=warn)
def assert_clean(text: str, artifact_path: str | None = None) -> ScrubReport:
"""Grep-gate: raise RedactionLeak on any fatal hit; return the scan report."""
report = scan(text)
if report.fatal_hits:
raise RedactionLeak(report.fatal_hits, artifact_path)
return report
def _strip_frontmatter(md: str) -> str:
"""Drop the YAML frontmatter from a markdown string for HTML rendering."""
if not md.startswith("---\n"):
return md
end = md.find("\n---\n", 4)
if end < 0:
return md
return md[end + 5 :]
def render_public_html(public_md: str) -> str:
"""Render redacted markdown into an HTML fragment for the landing page.
Uses python-markdown with the 'tables' extension; falls back to a
preformatted-text dump if markdown is not importable.
"""
body = _strip_frontmatter(public_md)
try:
import markdown as _md
return _md.markdown(body, extensions=["tables"])
except ImportError:
import html as _html
return f"<pre>{_html.escape(body)}</pre>"
def publish_public_review(
record_id: str,
reviews_dir: str,
website_repo: str,
) -> str:
"""Redact reviews/<id>_*.md and write md + html to website repo public-reviews/.
Returns the written .md path. Raises RedactionLeak if grep-gate trips on
either the markdown source or the rendered HTML.
"""
# Exclude artifacts that are not the primary review markdown.
# Both RQC and citation reports live alongside in reviews/ as
# <id>_review_quality_control.md and <id>_citations.md; including
# them here would cause sorted()[-1] to silently pick the wrong
# source for any paper whose slug sorts before "review_quality_control".
_NON_REVIEW_SUFFIXES = ("_review_quality_control.md", "_citations.md")
matches = [
f for f in os.listdir(reviews_dir)
if f.startswith(f"{record_id}_") and f.endswith(".md")
and not any(f.endswith(s) for s in _NON_REVIEW_SUFFIXES)
]
if not matches:
raise FileNotFoundError(
f"No review markdown found for record_id={record_id} in {reviews_dir}"
)
src = os.path.join(reviews_dir, sorted(matches)[-1])
parsed = parse_review_file(src)
public_md = build_public_markdown(parsed)
report_md = assert_clean(public_md, artifact_path=src)
public_html = render_public_html(public_md)
assert_clean(public_html, artifact_path=f"{src} (rendered html)")
out_dir = os.path.join(website_repo, "src", "data", "public-reviews")
os.makedirs(out_dir, exist_ok=True)
md_path = os.path.join(out_dir, f"{record_id}.md")
html_path = os.path.join(out_dir, f"{record_id}.html")
with open(md_path, "w", encoding="utf-8") as f:
f.write(public_md)
with open(html_path, "w", encoding="utf-8") as f:
f.write(public_html)
if report_md.warn_hits:
uniq = sorted({t for t, _ in report_md.warn_hits})
print(
f" redaction: {len(report_md.warn_hits)} soft-warn hit(s) "
f"({', '.join(uniq)}) β inspect {md_path} if unexpected."
)
return md_path
# --------------------------------------------------------------------------
# Review Quality Control β public redaction
# --------------------------------------------------------------------------
# The RQC rubric defines five dimensions. The first four are publishable;
# ``injection_indicators`` is INTERNAL ONLY and must never appear in any
# file written under src/data/public-reviews/. See rubrics/review_quality_control.md.
RQC_PUBLIC_DIMENSIONS: tuple[str, ...] = (
"Rubric Adherence",
"Internal Consistency",
"Specificity",
"Tone",
)
RQC_INJECTION_DIM_LABEL = "Injection Indicators"
# Substring-guard against the redacted dimension leaking into public text.
# Match case-insensitively on the prose label AND any plausible variants.
RQC_FORBIDDEN_PUBLIC_TOKENS: tuple[str, ...] = (
"injection_indicators",
"injection indicators",
"injection indicator",
"prompt injection",
"prompt-injection",
"prompt_injection",
)
@dataclass
class ParsedRQC:
"""Structured view of a reviews/<id>_review_quality_control.md file."""
record_id: str
title: str
doi: str
audit_date: str
flag: bool
summary: str
overall_concerns: list[str] = field(default_factory=list)
# Each slot: {"reviewer": str, "errored": bool,
# "dimensions": [(label, score, justification), ...]}
slots: list[dict] = field(default_factory=list)
def _parse_rqc_slots(body: str) -> list[dict]:
"""Parse '### Reviewer N' sections out of the Per-slot audit block."""
marker = "\n## Per-slot audit\n"
idx = body.find(marker)
if idx < 0:
return []
remainder = body[idx + len(marker):]
end = remainder.find("\n---\n")
if end >= 0:
remainder = remainder[:end]
sections: list[tuple[str, str]] = []
current_head: str | None = None
current_lines: list[str] = []
for line in remainder.splitlines():
if line.startswith("### "):
if current_head is not None:
sections.append((current_head, "\n".join(current_lines).strip()))
current_head = line[4:].strip()
current_lines = []
else:
current_lines.append(line)
if current_head is not None:
sections.append((current_head, "\n".join(current_lines).strip()))
slots: list[dict] = []
for head, content in sections:
errored = content.lstrip().startswith("*Errored:")
dims: list[tuple[str, str, str]] = []
if not errored:
for m in re.finditer(
r"^-\s+\*\*(?P<label>[^*]+)\*\*\s*\((?P<score>[^)]+)\):\s*"
r"(?P<just>.+?)(?=\n-\s+\*\*|\Z)",
content,
re.S | re.M,
):
dims.append(
(
m.group("label").strip(),
m.group("score").strip(),
" ".join(m.group("just").split()),
)
)
slots.append({"reviewer": head, "errored": errored, "dimensions": dims})
return slots
def parse_rqc_file(path: str) -> ParsedRQC:
"""Load a reviews/<id>_review_quality_control.md file into structured form."""
with open(path, "r", encoding="utf-8") as f:
text = f.read()
fm, body = _parse_frontmatter(text)
title = fm.get("title", "").strip('"')
if title.lower().startswith("review quality control:"):
title = title.split(":", 1)[1].strip()
flag_raw = str(fm.get("review_quality_control_flag", "false")).lower()
flag = flag_raw == "true"
# Pull the summary paragraph between "## Summary" and the next "## " header
summary = ""
m = re.search(r"^##\s+Summary\s*\n(.+?)(?=^##\s+|\Z)", body, re.S | re.M)
if m:
summary = " ".join(m.group(1).split())
concerns: list[str] = []
m = re.search(r"^##\s+Overall concerns\s*\n(.+?)(?=^##\s+|\Z)", body, re.S | re.M)
if m:
for line in m.group(1).splitlines():
line = line.strip()
if line.startswith("- "):
concerns.append(line[2:].strip())
return ParsedRQC(
record_id=str(fm.get("record_id", "")),
title=title,
doi=fm.get("doi", ""),
audit_date=fm.get("audit_date", ""),
flag=flag,
summary=summary,
overall_concerns=concerns,
slots=_parse_rqc_slots(body),
)
def _flag_cause_is_injection_only(parsed: ParsedRQC) -> bool:
"""True iff flag=true AND no scholarly dimension scored <=2.
When the only trigger is the internal injection_indicators dimension,
the public rendering shows a generic operator-review note rather than
identifying which scholarly dim tripped it.
"""
if not parsed.flag:
return False
for slot in parsed.slots:
if slot["errored"]:
continue
for label, score, _ in slot["dimensions"]:
if label == RQC_INJECTION_DIM_LABEL:
continue
try:
n = int(re.match(r"(\d+)", score).group(1))
except Exception:
continue
if n <= 2:
return False
return True
def build_public_rqc_markdown(parsed: ParsedRQC) -> str:
"""Render the redacted RQC markdown. Omits injection_indicators entirely.
Contract: the returned string contains no reference to the
injection_indicators dimension under any spelling. ``assert_rqc_clean``
enforces this before the redaction writes anything to the site.
"""
status_line = (
"Review Quality Control: flagged β reviewed by human editors before acceptance."
if parsed.flag
else "Review Quality Control: passed."
)
lines: list[str] = [
"---",
f'title: "Review Quality Control: {parsed.title}"',
f'doi: "{parsed.doi}"',
f"record_id: {parsed.record_id}",
f"audit_date: {parsed.audit_date}",
f"review_quality_control_flag: {str(parsed.flag).lower()}",
"---",
"",
"## Review Quality Control",
"",
f"**{status_line}**",
"",
(
"This audit quality checks each AI reviewer's assessment for "
"rubric adherence, internal consistency, specificity, and "
"institutional voice. It is published alongside the panel review "
"so the quality of the review process is as auditable as the "
"review itself."
),
"",
]
# When flag was tripped only by the internal dimension, say so generically.
if parsed.flag and _flag_cause_is_injection_only(parsed):
lines.extend([
(
"The audit surfaced a concern outside the four scholarly "
"dimensions above. A human editor reviewed the panel output "
"before the acceptance decision was recorded."
),
"",
])
elif parsed.overall_concerns:
# Concerns may reference the redacted injection dimension, name
# specific reviewer numbers that won't match the public 1..N
# renumbering (we drop errored slots from public view), or call
# out pipeline-level mechanics that are operator noise rather than
# author-facing scholarly concerns. Filter aggressively.
operator_noise_patterns = (
"pipeline-level", "pipeline level", "slot-run", "slot run",
"errored slot", "errored slots", "slot errored", "reviewer defect",
"operator", "panel composition", "pass 1", "pass 2", "pass 3",
)
safe_concerns = []
for c in parsed.overall_concerns:
cl = c.lower()
if any(tok in cl for tok in RQC_FORBIDDEN_PUBLIC_TOKENS):
continue
if any(tok in cl for tok in operator_noise_patterns):
continue
# Concerns that name specific panel members by vendor identity
# are panel-composition leaks, not author-facing scholarly
# concerns. Drop them rather than try to rewrite.
if any(tok in cl for tok in FORBIDDEN_VENDOR_TOKENS):
continue
# Drop concerns that cite specific reviewer numbers β our public
# renumbering (valid-only 1..N) won't line up with whatever
# index the audit claude-p referenced in its internal output.
# Singular AND plural forms both count.
if re.search(r"\breviewers?\s+\d+\b", cl):
continue
safe_concerns.append(c)
if safe_concerns:
lines.extend(["### Notes", ""])
for c in safe_concerns:
lines.append(f"- {_rewrite_rubric_filenames(c)}")
lines.append("")
# Public audit shows only valid-slot outputs, renumbered 1..N to match
# the open review's reviewer count. Errored slots are operator-layer
# signal; surfacing them publicly creates a count mismatch with the
# review section and invites reader confusion.
valid_slots = [s for s in parsed.slots if not s["errored"]]
lines.extend(["### Reviewer Quality Control Audit", ""])
# Condensed table: reviewer Γ four scholarly dimensions.
lines.append(
"| Reviewer | " + " | ".join(RQC_PUBLIC_DIMENSIONS) + " |"
)
lines.append(
"|----------|" + "|".join(["----"] * len(RQC_PUBLIC_DIMENSIONS)) + "|"
)
for idx, slot in enumerate(valid_slots, start=1):
label = f"Reviewer {idx}"
by_label = {lbl: (score, just) for lbl, score, just in slot["dimensions"]}
cells = [label]
for dim in RQC_PUBLIC_DIMENSIONS:
score, _ = by_label.get(dim, ("β", ""))
cells.append(score)
lines.append("| " + " | ".join(cells) + " |")
lines.append("")
# Detail block per valid reviewer, collapsed by default. Scholarly dims
# only; injection_indicators is redacted upstream. Browser-native
# <details> means no JavaScript and works with the site's scoped CSS.
import html as _html
for idx, slot in enumerate(valid_slots, start=1):
label = f"Reviewer {idx}"
by_label = {lbl: (score, just) for lbl, score, just in slot["dimensions"]}
lines.append(f'<details class="reviewer-detail">')
lines.append(f'<summary><strong>{label}</strong></summary>')
lines.append("")
lines.append("<ul>")
for dim in RQC_PUBLIC_DIMENSIONS:
score, just = by_label.get(dim, ("β", ""))
just_clean = _rewrite_rubric_filenames(just) if just else "No justification recorded."
lines.append(
f' <li><strong>{_html.escape(dim)}</strong> '
f'({_html.escape(score)}): {_html.escape(just_clean)}</li>'
)
lines.append("</ul>")
lines.append("</details>")
lines.append("")
lines.extend([
"---",
"",
(
"*Review Quality Control is an internal ICSAC audit of the "
"panel review itself. The four dimensions above are published "
"as part of ICSAC's open review commitment.*"
),
"",
])
return _humanize_internal_jargon("\n".join(lines))
def assert_rqc_clean(text: str, artifact_path: str | None = None) -> ScrubReport:
"""Grep-gate for RQC public output.
Runs the standard fatal/warn gate (vendors, secrets, exfil patterns)
AND asserts that no reference to the redacted injection_indicators
dimension survives. A leak is treated as a RedactionLeak.
"""
report = scan(text)
lowered = text.lower()
extra_hits: list[tuple[str, int]] = []
for tok in RQC_FORBIDDEN_PUBLIC_TOKENS:
start = 0
while True:
at = lowered.find(tok, start)
if at < 0:
break
extra_hits.append((f"[rqc-redacted] {tok}", at))
start = at + 1
if extra_hits or report.fatal_hits:
raise RedactionLeak(report.fatal_hits + extra_hits, artifact_path)
return report
def publish_public_rqc(
record_id: str,
reviews_dir: str,
website_repo: str,
) -> str | None:
"""Redact reviews/<id>_review_quality_control.md β public-reviews/<id>_review_quality_control.{md,html}.
Returns the written .md path, or None if no RQC file exists for the
record. Raises RedactionLeak if any forbidden token (including references
to the redacted injection_indicators dimension) survives.
"""
src = os.path.join(reviews_dir, f"{record_id}_review_quality_control.md")
if not os.path.isfile(src):
return None
parsed = parse_rqc_file(src)
public_md = build_public_rqc_markdown(parsed)
report_md = assert_rqc_clean(public_md, artifact_path=src)
public_html = render_public_html(public_md)
assert_rqc_clean(public_html, artifact_path=f"{src} (rendered html)")
out_dir = os.path.join(website_repo, "src", "data", "public-reviews")
os.makedirs(out_dir, exist_ok=True)
md_path = os.path.join(out_dir, f"{record_id}_review_quality_control.md")
html_path = os.path.join(out_dir, f"{record_id}_review_quality_control.html")
with open(md_path, "w", encoding="utf-8") as f:
f.write(public_md)
with open(html_path, "w", encoding="utf-8") as f:
f.write(public_html)
if report_md.warn_hits:
uniq = sorted({t for t, _ in report_md.warn_hits})
print(
f" redaction/rqc: {len(report_md.warn_hits)} soft-warn hit(s) "
f"({', '.join(uniq)}) β inspect {md_path} if unexpected."
)
return md_path
if __name__ == "__main__":
import sys
if len(sys.argv) < 2:
print(
"usage: python3 redaction.py <record_id> [reviews_dir] [website_repo]\n"
" python3 redaction.py rqc <record_id> [reviews_dir] [website_repo]",
file=sys.stderr,
)
sys.exit(2)
if sys.argv[1] == "rqc":
if len(sys.argv) < 3:
print("usage: python3 redaction.py rqc <record_id> ...", file=sys.stderr)
sys.exit(2)
record = sys.argv[2]
rdir = sys.argv[3] if len(sys.argv) > 3 else os.path.join(os.path.dirname(__file__), "reviews")
wrepo = (
sys.argv[4]
if len(sys.argv) > 4
else os.path.expanduser("~/Desktop/icsac/icsacinstitute.org")
)
try:
written = publish_public_rqc(record, rdir, wrepo)
except RedactionLeak as e:
print(f"SCRUB LEAK: {e}", file=sys.stderr)
sys.exit(1)
if written is None:
print(f"No RQC file for record {record} β nothing to publish.")
sys.exit(0)
print(f"wrote {written}")
sys.exit(0)
record = sys.argv[1]
rdir = sys.argv[2] if len(sys.argv) > 2 else os.path.join(os.path.dirname(__file__), "reviews")
wrepo = (
sys.argv[3]
if len(sys.argv) > 3
else os.path.expanduser("~/Desktop/icsac/icsacinstitute.org")
)
try:
written = publish_public_review(record, rdir, wrepo)
except RedactionLeak as e:
print(f"SCRUB LEAK: {e}", file=sys.stderr)
sys.exit(1)
print(f"wrote {written}")
# Best-effort companion RQC publish.
try:
rqc_written = publish_public_rqc(record, rdir, wrepo)
except RedactionLeak as e:
print(f"SCRUB LEAK (rqc): {e}", file=sys.stderr)
sys.exit(1)
if rqc_written:
print(f"wrote {rqc_written}")
|