ref-verify 1.2.0__tar.gz → 1.2.2__tar.gz

{ref_verify-1.2.0/src/ref_verify.egg-info → ref_verify-1.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ref-verify
-Version: 1.2.0
+Version: 1.2.2
 Summary: Executable DOI and claim verification helpers for academic citations
 Author: Moonweave Research
 License-Expression: MIT
@@ -45,6 +45,8 @@ Works with **Claude Code, Cursor, Codex**, and any agent that supports the
 
 After installation, use it like a normal agent skill. You do not start a server and you do not configure MCP for this workflow. No MCP server is required for this workflow.
 
+The skill includes its own copy of the CLI engine, and the agent runs it from the skill folder, so nothing else needs to be installed. Python 3.10 or newer must be available as `python3`.
+
 For explicit agent tool-calling rules, see [AGENT_USAGE.md](https://github.com/Moonweave-Research/ref-verify/blob/main/AGENT_USAGE.md).
 
 ---
@@ -85,13 +87,20 @@ checks that are currently safe to automate directly:
 - JSON output for agent-readable routing
 - Non-zero exit codes for `WARN`, `REJECT`, and `UNVERIFIABLE` results
 
-Statistical metrics such as p-values, AUC/AUROC, F1 score, hazard ratio, odds ratio, and confidence intervals still use the manual skill protocol. DOI landing-page checks still use the skill protocol. Still handled by the skill protocol: Unpaywall, arXiv, two-source existence checks, and retraction checks remain in `SKILL.md`.
+Statistical metrics such as p-values, AUC/AUROC, F1 score, hazard ratio, odds ratio, and confidence intervals still use the manual skill protocol. DOI landing-page checks still use the skill protocol. The CLI rejects a DOI that CrossRef records as retracted (via its retraction notice); retraction banners CrossRef does not know about, Unpaywall, arXiv, and two-source existence checks remain in the `SKILL.md` protocol.
 
 The CLI has zero third-party Python runtime dependencies, but it is not an
 offline verifier. Functional checks require outbound HTTPS access to public
 academic APIs such as CrossRef, OpenAlex, Semantic Scholar, and PubMed.
 
-Install the CLI from a local checkout:
+To run the CLI yourself, install it from PyPI:
+
+```bash
+uvx ref-verify --help            # run without installing (uv)
+pipx install ref-verify          # or install the `ref-verify` command
+```
+
+Or install it from a local checkout:
 
 ```bash
 git clone https://github.com/Moonweave-Research/ref-verify.git
@@ -172,20 +181,20 @@ normal CI does not fail because an upstream API is temporarily unavailable.
 
 ---
 
-## Scope — what it does and does not verify
+## Scope — optional CLI versus manual audit
 
 `ref-verify` is a conservative guard, not an oracle. It errs toward flagging: an
 `ACCEPT` is high-confidence, and **anything else means "not auto-verifiable —
 check it yourself", not "the citation is wrong."**
 
-**It verifies**
+**The optional CLI verifies**
 
 - DOI metadata: title, first-author surname, and year against CrossRef.
 - Whether a DOI-bound **abstract** explicitly supports a specific numeric or
   literal claim, quoted verbatim. If no abstract is reachable, it returns
   `UNVERIFIABLE` rather than guessing.
 
-**It does not verify** (out of scope by design, not bugs)
+**The optional CLI does not verify** (out of scope by design, not bugs)
 
 - **Full-text, figure, table, or supplementary values** — abstract-only. A number
   that appears only in the body stays `UNVERIFIABLE`.
@@ -199,7 +208,13 @@ check it yourself", not "the citation is wrong."**
 - **Paper quality, novelty, field consensus**, or whether the *full* paper
   supports a broader statement.
 
-**Reading a verdict**
+The agent skill's manual Full Audit protocol goes beyond the optional CLI for
+mechanism, implementation, and procedural claims. It requires a fetched
+full-text passage at that source depth; when full text is unavailable, it
+returns `WARN (ABSTRACT-ONLY)` instead of upgrading an abstract topic match to
+`ACCEPT`.
+
+**Reading a CLI verdict**
 
 | Verdict | Meaning |
 |---|---|
@@ -224,8 +239,9 @@ exit code, so weak or mismatched metadata cannot silently pass automation gates.
 
 **Full Audit** is for literature search and final pre-submission review. The
 skill fetches abstracts through CrossRef, OpenAlex, Semantic Scholar, Unpaywall,
-arXiv, and PubMed where needed, then checks whether the paper supports the specific
-claim being cited.
+arXiv, and PubMed where needed. For a topline claim, it checks the abstract; for
+a mechanism, implementation, or procedural claim, it continues to a fetched
+full-text passage before assigning support.
 
 For a single DOI-backed claim, the CLI can run the abstract check:
 
@@ -266,13 +282,15 @@ Current `check-claim` error codes:
 - `CLAIM_NOT_EXPLICIT`: an abstract was available, but the claim was not explicitly supported.
 - `CLAIM_AMBIGUOUS`: numeric evidence or context exists, but binding is ambiguous.
 - `NO_ABSTRACT`: attempted DOI-bound sources did not provide abstract text.
-- `DOI_NOT_FOUND`: selected source did not find a DOI-bound record.
+- `DOI_NOT_FOUND`: CrossRef has no record for the DOI (HTTP 404), or the selected source did not find a DOI-bound record. The JSON still carries a `verdict` of `REJECT`.
+- `PAPER_RETRACTED`: CrossRef lists a retraction notice for the DOI; the claim is rejected before any abstract is read.
 - `DOI_MISMATCH`: the primary or explicitly selected DOI-bound record did not match the requested DOI.
 - `SOURCE_API_ERROR`, `SOURCE_TIMEOUT`, `SOURCE_RATE_LIMITED`, `SOURCE_UNSUPPORTED`: source lookup failed, timed out, was rate-limited, or could not be used.
 
 > Core rule: every content statement about a paper must come from a live-fetched
-> abstract. If the abstract is inaccessible after fallback checks, say
-> `UNVERIFIABLE`. Do not fill the gap from memory.
+> source at the depth the claim requires — abstract for topline claims, full
+> text for mechanism, implementation, or procedural claims. If the required
+> source is inaccessible, say so. Do not fill the gap from memory.
 
 ---
 
@@ -309,5 +327,4 @@ that as `WARN (PARTIAL)` instead of accepting the citation.
 
 ## Related
 
-- [anneal-skill](https://github.com/Moonweave-Systems/anneal-skill) - measure-first decision discipline for AI agents
-- [decide-skill](https://github.com/Moonweave-Systems/decide-skill) - decision automation for non-expert domains
+- [decision-kernel](https://github.com/Moonweave-Systems/decision-kernel) - evidence-gated decisions and drift/done checks for coding agents

{ref_verify-1.2.0 → ref_verify-1.2.2}/README.md

@@ -34,6 +34,8 @@ Works with **Claude Code, Cursor, Codex**, and any agent that supports the
 
 After installation, use it like a normal agent skill. You do not start a server and you do not configure MCP for this workflow. No MCP server is required for this workflow.
 
+The skill includes its own copy of the CLI engine, and the agent runs it from the skill folder, so nothing else needs to be installed. Python 3.10 or newer must be available as `python3`.
+
 For explicit agent tool-calling rules, see [AGENT_USAGE.md](https://github.com/Moonweave-Research/ref-verify/blob/main/AGENT_USAGE.md).
 
 ---
@@ -74,13 +76,20 @@ checks that are currently safe to automate directly:
 - JSON output for agent-readable routing
 - Non-zero exit codes for `WARN`, `REJECT`, and `UNVERIFIABLE` results
 
-Statistical metrics such as p-values, AUC/AUROC, F1 score, hazard ratio, odds ratio, and confidence intervals still use the manual skill protocol. DOI landing-page checks still use the skill protocol. Still handled by the skill protocol: Unpaywall, arXiv, two-source existence checks, and retraction checks remain in `SKILL.md`.
+Statistical metrics such as p-values, AUC/AUROC, F1 score, hazard ratio, odds ratio, and confidence intervals still use the manual skill protocol. DOI landing-page checks still use the skill protocol. The CLI rejects a DOI that CrossRef records as retracted (via its retraction notice); retraction banners CrossRef does not know about, Unpaywall, arXiv, and two-source existence checks remain in the `SKILL.md` protocol.
 
 The CLI has zero third-party Python runtime dependencies, but it is not an
 offline verifier. Functional checks require outbound HTTPS access to public
 academic APIs such as CrossRef, OpenAlex, Semantic Scholar, and PubMed.
 
-Install the CLI from a local checkout:
+To run the CLI yourself, install it from PyPI:
+
+```bash
+uvx ref-verify --help            # run without installing (uv)
+pipx install ref-verify          # or install the `ref-verify` command
+```
+
+Or install it from a local checkout:
 
 ```bash
 git clone https://github.com/Moonweave-Research/ref-verify.git
@@ -161,20 +170,20 @@ normal CI does not fail because an upstream API is temporarily unavailable.
 
 ---
 
-## Scope — what it does and does not verify
+## Scope — optional CLI versus manual audit
 
 `ref-verify` is a conservative guard, not an oracle. It errs toward flagging: an
 `ACCEPT` is high-confidence, and **anything else means "not auto-verifiable —
 check it yourself", not "the citation is wrong."**
 
-**It verifies**
+**The optional CLI verifies**
 
 - DOI metadata: title, first-author surname, and year against CrossRef.
 - Whether a DOI-bound **abstract** explicitly supports a specific numeric or
   literal claim, quoted verbatim. If no abstract is reachable, it returns
   `UNVERIFIABLE` rather than guessing.
 
-**It does not verify** (out of scope by design, not bugs)
+**The optional CLI does not verify** (out of scope by design, not bugs)
 
 - **Full-text, figure, table, or supplementary values** — abstract-only. A number
   that appears only in the body stays `UNVERIFIABLE`.
@@ -188,7 +197,13 @@ check it yourself", not "the citation is wrong."**
 - **Paper quality, novelty, field consensus**, or whether the *full* paper
   supports a broader statement.
 
-**Reading a verdict**
+The agent skill's manual Full Audit protocol goes beyond the optional CLI for
+mechanism, implementation, and procedural claims. It requires a fetched
+full-text passage at that source depth; when full text is unavailable, it
+returns `WARN (ABSTRACT-ONLY)` instead of upgrading an abstract topic match to
+`ACCEPT`.
+
+**Reading a CLI verdict**
 
 | Verdict | Meaning |
 |---|---|
@@ -213,8 +228,9 @@ exit code, so weak or mismatched metadata cannot silently pass automation gates.
 
 **Full Audit** is for literature search and final pre-submission review. The
 skill fetches abstracts through CrossRef, OpenAlex, Semantic Scholar, Unpaywall,
-arXiv, and PubMed where needed, then checks whether the paper supports the specific
-claim being cited.
+arXiv, and PubMed where needed. For a topline claim, it checks the abstract; for
+a mechanism, implementation, or procedural claim, it continues to a fetched
+full-text passage before assigning support.
 
 For a single DOI-backed claim, the CLI can run the abstract check:
 
@@ -255,13 +271,15 @@ Current `check-claim` error codes:
 - `CLAIM_NOT_EXPLICIT`: an abstract was available, but the claim was not explicitly supported.
 - `CLAIM_AMBIGUOUS`: numeric evidence or context exists, but binding is ambiguous.
 - `NO_ABSTRACT`: attempted DOI-bound sources did not provide abstract text.
-- `DOI_NOT_FOUND`: selected source did not find a DOI-bound record.
+- `DOI_NOT_FOUND`: CrossRef has no record for the DOI (HTTP 404), or the selected source did not find a DOI-bound record. The JSON still carries a `verdict` of `REJECT`.
+- `PAPER_RETRACTED`: CrossRef lists a retraction notice for the DOI; the claim is rejected before any abstract is read.
 - `DOI_MISMATCH`: the primary or explicitly selected DOI-bound record did not match the requested DOI.
 - `SOURCE_API_ERROR`, `SOURCE_TIMEOUT`, `SOURCE_RATE_LIMITED`, `SOURCE_UNSUPPORTED`: source lookup failed, timed out, was rate-limited, or could not be used.
 
 > Core rule: every content statement about a paper must come from a live-fetched
-> abstract. If the abstract is inaccessible after fallback checks, say
-> `UNVERIFIABLE`. Do not fill the gap from memory.
+> source at the depth the claim requires — abstract for topline claims, full
+> text for mechanism, implementation, or procedural claims. If the required
+> source is inaccessible, say so. Do not fill the gap from memory.
 
 ---
 
@@ -298,5 +316,4 @@ that as `WARN (PARTIAL)` instead of accepting the citation.
 
 ## Related
 
-- [anneal-skill](https://github.com/Moonweave-Systems/anneal-skill) - measure-first decision discipline for AI agents
-- [decide-skill](https://github.com/Moonweave-Systems/decide-skill) - decision automation for non-expert domains
+- [decision-kernel](https://github.com/Moonweave-Systems/decision-kernel) - evidence-gated decisions and drift/done checks for coding agents

{ref_verify-1.2.0 → ref_verify-1.2.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "ref-verify"
-version = "1.2.0"
+version = "1.2.2"
 description = "Executable DOI and claim verification helpers for academic citations"
 readme = "README.md"
 requires-python = ">=3.10"

{ref_verify-1.2.0 → ref_verify-1.2.2}/src/ref_verify/__init__.py

@@ -2,4 +2,4 @@
 
 __all__ = ["__version__"]
 
-__version__ = "1.2.0"
+__version__ = "1.2.2"

{ref_verify-1.2.0 → ref_verify-1.2.2}/src/ref_verify/claim_check.py

@@ -215,7 +215,23 @@ _UNSUPPORTED_CLAIM_FRAME_PATTERNS = (
 )
 
 
+def retracted_claim_result(record: PaperRecord, claim: str) -> ClaimSupportResult:
+    return ClaimSupportResult(
+        status="RETRACTED",
+        verdict="REJECT",
+        reason=(
+            "CrossRef records this paper as retracted "
+            f"(notice DOI {record.retraction_doi}); its abstract cannot support the claim."
+        ),
+        evidence="",
+        paper=record,
+        claim=claim,
+    )
+
+
 def check_claim_support(record: PaperRecord, claim: str) -> ClaimSupportResult:
+    if record.retraction_doi:
+        return retracted_claim_result(record, claim)
     if not record.abstract:
         return ClaimSupportResult(
             status="UNVERIFIABLE",

{ref_verify-1.2.0 → ref_verify-1.2.2}/src/ref_verify/cli.py

@@ -5,6 +5,7 @@ import json
 import sys
 from pathlib import Path
 from typing import Sequence
+from urllib.error import HTTPError
 
 from ref_verify.abstract_lookup import (
     AbstractSourceClient,
@@ -18,7 +19,7 @@ from ref_verify.batch import (
     parse_claim_file,
     render_batch_text,
 )
-from ref_verify.claim_check import check_claim_support
+from ref_verify.claim_check import check_claim_support, retracted_claim_result
 from ref_verify.crossref import CrossrefClient
 from ref_verify.doi_check import normalize_doi, verify_doi_metadata
 from ref_verify.models import CitationInput, ClaimSupportResult
@@ -88,13 +89,31 @@ def _build_parser() -> argparse.ArgumentParser:
 
 def _verify_doi(args: argparse.Namespace, client: CrossrefClient) -> int:
     lookup_doi = normalize_doi(args.doi)
-    fetched = client.fetch_work(lookup_doi)
     provided = CitationInput(
         doi=args.doi,
         title=args.title,
         first_author=args.first_author,
         year=args.year,
     )
+    try:
+        fetched = client.fetch_work(lookup_doi)
+    except HTTPError as exc:
+        if not _is_not_found(exc):
+            raise
+        # A dead DOI is a verdict, not a tool failure; keep the `verdict` key so
+        # callers that parse JSON do not have to special-case the error shape.
+        _emit(
+            {
+                "verdict": "REJECT",
+                "mismatches": ["doi"],
+                "reason": "CrossRef has no record for this DOI (HTTP 404).",
+                "provided": provided.to_dict(),
+                "fetched": None,
+                "error_code": "DOI_NOT_FOUND",
+            },
+            as_json=args.json,
+        )
+        return 2
     result = verify_doi_metadata(provided, fetched)
     _emit(result.to_dict(), as_json=args.json)
     return 0 if result.verdict == "PASS" else 2
@@ -147,7 +166,20 @@ def _run_claim_check(
     lookup_doi = normalize_doi(doi)
     selected_clients = _select_abstract_clients(fallback_clients, source)
     if source in ("auto", "crossref"):
-        fetched = client.fetch_work(lookup_doi)
+        try:
+            fetched = client.fetch_work(lookup_doi)
+        except HTTPError as exc:
+            if not _is_not_found(exc):
+                raise
+            return _doi_not_found_payload(claim)
+        if fetched.retraction_doi:
+            # Fallback abstract sources would drop the retraction flag, so decide here.
+            result = retracted_claim_result(fetched, claim)
+            payload = result.to_dict()
+            payload["abstract_source"] = None
+            payload["source_attempts"] = []
+            payload["error_code"] = "PAPER_RETRACTED"
+            return payload
         lookup_result = lookup_abstract(lookup_doi, fetched, selected_clients)
     else:
         lookup_result = lookup_selected_abstract(lookup_doi, selected_clients)
@@ -166,6 +198,24 @@ def _run_claim_check(
     return _claim_payload(result, lookup_result)
 
 
+def _is_not_found(exc: HTTPError) -> bool:
+    return getattr(exc, "code", None) == 404
+
+
+def _doi_not_found_payload(claim: str) -> dict:
+    return {
+        "status": "UNVERIFIABLE",
+        "verdict": "REJECT",
+        "reason": "CrossRef has no record for this DOI (HTTP 404).",
+        "evidence": "",
+        "paper": None,
+        "claim": claim,
+        "abstract_source": None,
+        "source_attempts": [],
+        "error_code": "DOI_NOT_FOUND",
+    }
+
+
 def _row_error_payload(claim: str, exc: Exception) -> dict:
     return {
         "status": "UNVERIFIABLE",
@@ -190,6 +240,8 @@ def _claim_payload(result: ClaimSupportResult, lookup_result) -> dict:
 def _claim_error_code(result: ClaimSupportResult) -> str:
     if result.verdict == "ACCEPT":
         return "CLAIM_SUPPORTED"
+    if result.status == "RETRACTED":
+        return "PAPER_RETRACTED"
     if result.status == "UNVERIFIABLE":
         return "NO_ABSTRACT"
     if result.status == "PARTIAL":

{ref_verify-1.2.0 → ref_verify-1.2.2}/src/ref_verify/crossref.py

@@ -54,9 +54,24 @@ def parse_crossref_work(message: dict[str, Any]) -> PaperRecord:
         source="CrossRef",
         journal=journal,
         url=str(url) if url else None,
+        retraction_doi=_retraction_doi(message),
     )
 
 
+_RETRACTION_UPDATE_TYPES = {"retraction", "withdrawal", "removal", "partial_retraction"}
+
+
+def _retraction_doi(message: dict[str, Any]) -> str | None:
+    # CrossRef lists retraction notices under `updated-by`; the retracted work itself
+    # keeps its original DOI, so the metadata alone looks like a normal paper.
+    for update in message.get("updated-by") or []:
+        if not isinstance(update, dict):
+            continue
+        if str(update.get("type") or "").lower() in _RETRACTION_UPDATE_TYPES:
+            return str(update.get("DOI") or "") or "unknown"
+    return None
+
+
 def _first_string(value: Any) -> str | None:
     if isinstance(value, list) and value:
         return str(value[0]).strip()

{ref_verify-1.2.0 → ref_verify-1.2.2}/src/ref_verify/doi_check.py

@@ -75,9 +75,18 @@ def verify_doi_metadata(
     ):
         mismatches.append("year")
 
+    if fetched.retraction_doi:
+        mismatches.append("retracted")
+
     if not mismatches:
         verdict = "PASS"
         reason = "Provided citation metadata matches the fetched CrossRef record."
+    elif "retracted" in mismatches:
+        verdict = "REJECT"
+        reason = (
+            "CrossRef records this paper as retracted "
+            f"(notice DOI {fetched.retraction_doi}); do not use it as a source."
+        )
     elif any(field in mismatches for field in ("doi", "title", "first_author")):
         verdict = "REJECT"
         reason = "DOI resolves to a materially different paper than provided."
@@ -128,12 +137,20 @@ def _strip_trailing_doi_punctuation(value: str) -> str:
 
 
 def _titles_match(provided: str, fetched: str) -> bool:
+    provided = _strip_retraction_prefix(provided)
+    fetched = _strip_retraction_prefix(fetched)
     if _numbers(provided) != _numbers(fetched):
         return False
 
     return _title_tokens(provided) == _title_tokens(fetched)
 
 
+def _strip_retraction_prefix(title: str) -> str:
+    # Publishers prepend "RETRACTED:" to the stored title, and a citation may or may not
+    # carry it; the retraction itself is reported separately, never as a title mismatch.
+    return re.sub(r"^\s*(?:retracted|withdrawn|retraction)\s*:\s*", "", title, flags=re.IGNORECASE)
+
+
 def _author_matches(provided: str, fetched: str | None) -> bool:
     if not fetched:
         return False

{ref_verify-1.2.0 → ref_verify-1.2.2}/src/ref_verify/models.py

@@ -25,6 +25,7 @@ class PaperRecord:
     source: str
     journal: str | None = None
     url: str | None = None
+    retraction_doi: str | None = None
 
     def to_dict(self) -> dict[str, Any]:
         return asdict(self)

{ref_verify-1.2.0 → ref_verify-1.2.2}/src/ref_verify/numeric_claim.py

@@ -231,6 +231,7 @@ class NumericExpression:
     comparator: str
     subject_terms: set[str]
     evidence: str
+    prefix_scale: str = ""
 
 
 def check_numeric_claim_support(abstract: str, claim: str) -> NumericClaimResult:
@@ -253,7 +254,7 @@ def check_numeric_claim_support(abstract: str, claim: str) -> NumericClaimResult
         if _subject_terms_match(claim_expression.subject_terms, subject_context):
             related_evidence = related_evidence or clause
         for evidence_expression in evidence_expressions:
-            if not _units_match(claim_expression.unit, evidence_expression.unit):
+            if not _units_match(claim_expression, evidence_expression):
                 continue
             if not _subject_terms_match(claim_expression.subject_terms, subject_context):
                 continue
@@ -287,6 +288,7 @@ def _extract_claim_expression(claim: str) -> NumericExpression | None:
         comparator=_claim_comparator(claim[: match.start()]),
         subject_terms=_subject_terms(claim[: match.start()]),
         evidence=claim,
+        prefix_scale=_prefix_scale(match.group("unit")),
     )
 
 
@@ -303,6 +305,7 @@ def _extract_evidence_expressions(clause: str) -> list[NumericExpression]:
                 ),
                 subject_terms=_subject_terms(clause[: match.start()]),
                 evidence=clause,
+                prefix_scale=_prefix_scale(match.group("unit")),
             )
         )
     return expressions
@@ -412,8 +415,37 @@ def _normalize_unit(value: str) -> str:
     return normalized.rstrip("s")
 
 
-def _units_match(left: str, right: str) -> bool:
-    return left == right
+def _units_match(left: NumericExpression, right: NumericExpression) -> bool:
+    if left.unit != right.unit:
+        return False
+    # "mV" and "MV" differ by a factor of a million; only an all-lowercase token
+    # (common in typed claims) is left ambiguous and allowed to match either.
+    if left.prefix_scale and right.prefix_scale:
+        return left.prefix_scale == right.prefix_scale
+    return True
+
+
+_PREFIX_SENSITIVE_UNITS = {
+    "ma",
+    "mev",
+    "mhz",
+    "mohm-cm",
+    "mpa",
+    "ms/m",
+    "mv",
+    "mv/cm",
+    "mv/m",
+    "mv/mm",
+}
+
+
+def _prefix_scale(raw_unit: str) -> str:
+    if _normalize_unit(raw_unit) not in _PREFIX_SENSITIVE_UNITS:
+        return ""
+    stripped = raw_unit.strip()
+    if stripped == stripped.lower():
+        return ""
+    return "mega" if stripped[0] == "M" else "milli"
 
 
 def _claim_comparator(prefix: str) -> str:
@@ -534,6 +566,20 @@ def _evidence_entails_claim(
         if claim_comparator == "up_to":
             return evidence_value == claim_value
         return claim_comparator in {"lt", "lte"} and evidence_value <= claim_value
+    # Evidence that only bounds the value from one side cannot support a claim
+    # bounding it from the other side ("below 50" never establishes "above 40").
+    if evidence_comparator in {"lt", "lte"}:
+        if claim_comparator == "lt":
+            return evidence_value <= claim_value if evidence_comparator == "lt" else evidence_value < claim_value
+        if claim_comparator == "lte":
+            return evidence_value <= claim_value
+        return False
+    if evidence_comparator in {"gt", "gte"}:
+        if claim_comparator == "gt":
+            return evidence_value >= claim_value if evidence_comparator == "gt" else evidence_value > claim_value
+        if claim_comparator == "gte":
+            return evidence_value >= claim_value
+        return False
     if claim_comparator == "gt":
         return evidence_value > claim_value
     if claim_comparator == "gte":

{ref_verify-1.2.0 → ref_verify-1.2.2/src/ref_verify.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ref-verify
-Version: 1.2.0
+Version: 1.2.2
 Summary: Executable DOI and claim verification helpers for academic citations
 Author: Moonweave Research
 License-Expression: MIT
@@ -45,6 +45,8 @@ Works with **Claude Code, Cursor, Codex**, and any agent that supports the
 
 After installation, use it like a normal agent skill. You do not start a server and you do not configure MCP for this workflow. No MCP server is required for this workflow.
 
+The skill includes its own copy of the CLI engine, and the agent runs it from the skill folder, so nothing else needs to be installed. Python 3.10 or newer must be available as `python3`.
+
 For explicit agent tool-calling rules, see [AGENT_USAGE.md](https://github.com/Moonweave-Research/ref-verify/blob/main/AGENT_USAGE.md).
 
 ---
@@ -85,13 +87,20 @@ checks that are currently safe to automate directly:
 - JSON output for agent-readable routing
 - Non-zero exit codes for `WARN`, `REJECT`, and `UNVERIFIABLE` results
 
-Statistical metrics such as p-values, AUC/AUROC, F1 score, hazard ratio, odds ratio, and confidence intervals still use the manual skill protocol. DOI landing-page checks still use the skill protocol. Still handled by the skill protocol: Unpaywall, arXiv, two-source existence checks, and retraction checks remain in `SKILL.md`.
+Statistical metrics such as p-values, AUC/AUROC, F1 score, hazard ratio, odds ratio, and confidence intervals still use the manual skill protocol. DOI landing-page checks still use the skill protocol. The CLI rejects a DOI that CrossRef records as retracted (via its retraction notice); retraction banners CrossRef does not know about, Unpaywall, arXiv, and two-source existence checks remain in the `SKILL.md` protocol.
 
 The CLI has zero third-party Python runtime dependencies, but it is not an
 offline verifier. Functional checks require outbound HTTPS access to public
 academic APIs such as CrossRef, OpenAlex, Semantic Scholar, and PubMed.
 
-Install the CLI from a local checkout:
+To run the CLI yourself, install it from PyPI:
+
+```bash
+uvx ref-verify --help            # run without installing (uv)
+pipx install ref-verify          # or install the `ref-verify` command
+```
+
+Or install it from a local checkout:
 
 ```bash
 git clone https://github.com/Moonweave-Research/ref-verify.git
@@ -172,20 +181,20 @@ normal CI does not fail because an upstream API is temporarily unavailable.
 
 ---
 
-## Scope — what it does and does not verify
+## Scope — optional CLI versus manual audit
 
 `ref-verify` is a conservative guard, not an oracle. It errs toward flagging: an
 `ACCEPT` is high-confidence, and **anything else means "not auto-verifiable —
 check it yourself", not "the citation is wrong."**
 
-**It verifies**
+**The optional CLI verifies**
 
 - DOI metadata: title, first-author surname, and year against CrossRef.
 - Whether a DOI-bound **abstract** explicitly supports a specific numeric or
   literal claim, quoted verbatim. If no abstract is reachable, it returns
   `UNVERIFIABLE` rather than guessing.
 
-**It does not verify** (out of scope by design, not bugs)
+**The optional CLI does not verify** (out of scope by design, not bugs)
 
 - **Full-text, figure, table, or supplementary values** — abstract-only. A number
   that appears only in the body stays `UNVERIFIABLE`.
@@ -199,7 +208,13 @@ check it yourself", not "the citation is wrong."**
 - **Paper quality, novelty, field consensus**, or whether the *full* paper
   supports a broader statement.
 
-**Reading a verdict**
+The agent skill's manual Full Audit protocol goes beyond the optional CLI for
+mechanism, implementation, and procedural claims. It requires a fetched
+full-text passage at that source depth; when full text is unavailable, it
+returns `WARN (ABSTRACT-ONLY)` instead of upgrading an abstract topic match to
+`ACCEPT`.
+
+**Reading a CLI verdict**
 
 | Verdict | Meaning |
 |---|---|
@@ -224,8 +239,9 @@ exit code, so weak or mismatched metadata cannot silently pass automation gates.
 
 **Full Audit** is for literature search and final pre-submission review. The
 skill fetches abstracts through CrossRef, OpenAlex, Semantic Scholar, Unpaywall,
-arXiv, and PubMed where needed, then checks whether the paper supports the specific
-claim being cited.
+arXiv, and PubMed where needed. For a topline claim, it checks the abstract; for
+a mechanism, implementation, or procedural claim, it continues to a fetched
+full-text passage before assigning support.
 
 For a single DOI-backed claim, the CLI can run the abstract check:
 
@@ -266,13 +282,15 @@ Current `check-claim` error codes:
 - `CLAIM_NOT_EXPLICIT`: an abstract was available, but the claim was not explicitly supported.
 - `CLAIM_AMBIGUOUS`: numeric evidence or context exists, but binding is ambiguous.
 - `NO_ABSTRACT`: attempted DOI-bound sources did not provide abstract text.
-- `DOI_NOT_FOUND`: selected source did not find a DOI-bound record.
+- `DOI_NOT_FOUND`: CrossRef has no record for the DOI (HTTP 404), or the selected source did not find a DOI-bound record. The JSON still carries a `verdict` of `REJECT`.
+- `PAPER_RETRACTED`: CrossRef lists a retraction notice for the DOI; the claim is rejected before any abstract is read.
 - `DOI_MISMATCH`: the primary or explicitly selected DOI-bound record did not match the requested DOI.
 - `SOURCE_API_ERROR`, `SOURCE_TIMEOUT`, `SOURCE_RATE_LIMITED`, `SOURCE_UNSUPPORTED`: source lookup failed, timed out, was rate-limited, or could not be used.
 
 > Core rule: every content statement about a paper must come from a live-fetched
-> abstract. If the abstract is inaccessible after fallback checks, say
-> `UNVERIFIABLE`. Do not fill the gap from memory.
+> source at the depth the claim requires — abstract for topline claims, full
+> text for mechanism, implementation, or procedural claims. If the required
+> source is inaccessible, say so. Do not fill the gap from memory.
 
 ---
 
@@ -309,5 +327,4 @@ that as `WARN (PARTIAL)` instead of accepting the citation.
 
 ## Related
 
-- [anneal-skill](https://github.com/Moonweave-Systems/anneal-skill) - measure-first decision discipline for AI agents
-- [decide-skill](https://github.com/Moonweave-Systems/decide-skill) - decision automation for non-expert domains
+- [decision-kernel](https://github.com/Moonweave-Systems/decision-kernel) - evidence-gated decisions and drift/done checks for coding agents

{ref_verify-1.2.0 → ref_verify-1.2.2}/tests/test_cli.py

@@ -4,6 +4,7 @@ import tempfile
 import unittest
 from contextlib import redirect_stdout
 from pathlib import Path
+from urllib.error import HTTPError
 
 from ref_verify.abstract_lookup import AbstractSourceError
 from ref_verify.cli import _default_abstract_clients, main
@@ -302,6 +303,77 @@ class CliTests(unittest.TestCase):
         self.assertEqual(payload["abstract_source"], "crossref")
         self.assertEqual(payload["error_code"], "CLAIM_SUPPORTED")
 
+    def test_verify_doi_reports_dead_doi_as_reject_with_verdict_key(self):
+        error = HTTPError("https://api.crossref.org/works/x", 404, "Not Found", None, None)
+        output = io.StringIO()
+
+        with redirect_stdout(output):
+            exit_code = main(
+                ["verify-doi", "10.1126/science.999999.9999", "--json"],
+                client=FailingClient(error),
+            )
+
+        payload = json.loads(output.getvalue())
+        self.assertEqual(exit_code, 2)
+        self.assertEqual(payload["verdict"], "REJECT")
+        self.assertEqual(payload["error_code"], "DOI_NOT_FOUND")
+        self.assertNotIn("error", payload)
+
+    def test_verify_doi_still_surfaces_non_404_http_errors(self):
+        error = HTTPError("https://api.crossref.org/works/x", 503, "Unavailable", None, None)
+        output = io.StringIO()
+
+        with redirect_stdout(output):
+            exit_code = main(
+                ["verify-doi", "10.1000/example", "--json"],
+                client=FailingClient(error),
+            )
+
+        payload = json.loads(output.getvalue())
+        self.assertEqual(exit_code, 1)
+        self.assertIn("error", payload)
+
+    def test_check_claim_reports_dead_doi_as_reject(self):
+        error = HTTPError("https://api.crossref.org/works/x", 404, "Not Found", None, None)
+        output = io.StringIO()
+
+        with redirect_stdout(output):
+            exit_code = main(
+                ["check-claim", "10.1126/science.999999.9999", "--claim", "strain above 100%", "--json"],
+                client=FailingClient(error),
+                abstract_clients=[],
+            )
+
+        payload = json.loads(output.getvalue())
+        self.assertEqual(exit_code, 2)
+        self.assertEqual(payload["verdict"], "REJECT")
+        self.assertEqual(payload["error_code"], "DOI_NOT_FOUND")
+
+    def test_check_claim_rejects_retracted_paper_before_reading_abstract(self):
+        record = PaperRecord(
+            doi="10.1000/retracted",
+            title="RETRACTED: Dielectric elastomer actuators",
+            authors=["Pelrine"],
+            year=2000,
+            abstract="Actuated strains up to 117% were demonstrated.",
+            source="fixture",
+            retraction_doi="10.1000/retraction-notice",
+        )
+        output = io.StringIO()
+
+        with redirect_stdout(output):
+            exit_code = main(
+                ["check-claim", "10.1000/retracted", "--claim", "actuation strain above 100%", "--json"],
+                client=FakeClient(record),
+                abstract_clients=[],
+            )
+
+        payload = json.loads(output.getvalue())
+        self.assertEqual(exit_code, 2)
+        self.assertEqual(payload["status"], "RETRACTED")
+        self.assertEqual(payload["verdict"], "REJECT")
+        self.assertEqual(payload["error_code"], "PAPER_RETRACTED")
+
     def test_check_claim_does_not_query_fallback_when_crossref_has_abstract(self):
         record = PaperRecord(
             doi="10.1000/example",

{ref_verify-1.2.0 → ref_verify-1.2.2}/tests/test_crossref.py

@@ -41,6 +41,33 @@ class CrossrefTests(unittest.TestCase):
 
         self.assertIsNone(record.year)
 
+    def test_reads_retraction_notice_from_updated_by(self):
+        message = {
+            "DOI": "10.1016/s0140-6736(97)11096-0",
+            "title": ["RETRACTED: Ileal-lymphoid-nodular hyperplasia"],
+            "author": [{"family": "Wakefield", "given": "AJ"}],
+            "issued": {"date-parts": [[1998]]},
+            "updated-by": [
+                {"type": "correction", "DOI": "10.1016/s0140-6736(04)15715-2"},
+                {"type": "retraction", "DOI": "10.1016/s0140-6736(10)60175-4"},
+            ],
+        }
+
+        record = parse_crossref_work(message)
+
+        self.assertEqual(record.retraction_doi, "10.1016/s0140-6736(10)60175-4")
+
+    def test_correction_alone_is_not_a_retraction(self):
+        message = {
+            "DOI": "10.1000/corrected",
+            "title": ["Corrected paper"],
+            "updated-by": [{"type": "correction", "DOI": "10.1000/corrigendum"}],
+        }
+
+        record = parse_crossref_work(message)
+
+        self.assertIsNone(record.retraction_doi)
+
     def test_preserves_group_author_name(self):
         message = {
             "DOI": "10.1000/group-author",

{ref_verify-1.2.0 → ref_verify-1.2.2}/tests/test_doi_check.py

@@ -295,6 +295,71 @@ class DoiCheckTests(unittest.TestCase):
         self.assertIn("metadata", result.mismatches)
         self.assertIn("year", result.mismatches)
 
+    def test_rejects_retracted_paper_even_when_metadata_matches(self):
+        provided = CitationInput(
+            doi="10.1000/retracted",
+            title="Ileal-lymphoid-nodular hyperplasia",
+            first_author="Wakefield",
+            year=1998,
+        )
+        fetched = PaperRecord(
+            doi="10.1000/retracted",
+            title="RETRACTED: Ileal-lymphoid-nodular hyperplasia",
+            authors=["Wakefield"],
+            year=1998,
+            abstract=None,
+            source="fixture",
+            retraction_doi="10.1000/retraction-notice",
+        )
+
+        result = verify_doi_metadata(provided, fetched)
+
+        self.assertEqual(result.verdict, "REJECT")
+        self.assertEqual(result.mismatches, ["retracted"])
+        self.assertIn("10.1000/retraction-notice", result.reason)
+
+    def test_retracted_title_prefix_does_not_count_as_a_title_mismatch(self):
+        provided = CitationInput(
+            doi="10.1000/retracted",
+            title="Ileal-lymphoid-nodular hyperplasia",
+            first_author="Wakefield",
+            year=1998,
+        )
+        fetched = PaperRecord(
+            doi="10.1000/retracted",
+            title="RETRACTED: Ileal-lymphoid-nodular hyperplasia",
+            authors=["Wakefield"],
+            year=1998,
+            abstract=None,
+            source="fixture",
+        )
+
+        result = verify_doi_metadata(provided, fetched)
+
+        self.assertNotIn("title", result.mismatches)
+
+    def test_citation_carrying_the_retracted_prefix_is_not_a_title_mismatch(self):
+        provided = CitationInput(
+            doi="10.1000/retracted",
+            title="RETRACTED: Ileal-lymphoid-nodular hyperplasia",
+            first_author="Wakefield",
+            year=1998,
+        )
+        fetched = PaperRecord(
+            doi="10.1000/retracted",
+            title="RETRACTED: Ileal-lymphoid-nodular hyperplasia",
+            authors=["Wakefield"],
+            year=1998,
+            abstract=None,
+            source="fixture",
+            retraction_doi="10.1000/retraction-notice",
+        )
+
+        result = verify_doi_metadata(provided, fetched)
+
+        self.assertEqual(result.verdict, "REJECT")
+        self.assertEqual(result.mismatches, ["retracted"])
+
 
 if __name__ == "__main__":
     unittest.main()

{ref_verify-1.2.0 → ref_verify-1.2.2}/tests/test_numeric_claim.py

@@ -229,6 +229,46 @@ class NumericClaimTests(unittest.TestCase):
 
         self.assertEqual(result.status, "PARTIAL")
 
+    def test_upper_bounded_evidence_does_not_support_lower_bounded_claim(self):
+        for abstract, claim in (
+            ("The modulus was below 50 MPa.", "modulus above 40 MPa"),
+            ("The breakdown field was at most 100 MV/m.", "breakdown field at least 80 MV/m"),
+            ("The modulus was at least 50 MPa.", "modulus below 60 MPa"),
+        ):
+            with self.subTest(abstract=abstract, claim=claim):
+                result = check_numeric_claim_support(abstract, claim)
+
+                self.assertEqual(result.status, "PARTIAL")
+
+    def test_one_sided_evidence_supports_claim_bounded_on_the_same_side(self):
+        for abstract, claim in (
+            ("The modulus was below 50 MPa.", "modulus below 60 MPa"),
+            ("The modulus was at least 50 MPa.", "modulus above 40 MPa"),
+        ):
+            with self.subTest(abstract=abstract, claim=claim):
+                result = check_numeric_claim_support(abstract, claim)
+
+                self.assertEqual(result.status, "SUPPORTED")
+
+    def test_milli_and_mega_prefixes_do_not_match(self):
+        for abstract, claim in (
+            ("The output voltage was 5 mV.", "output voltage of 5 MV"),
+            ("The current reached 2 mA.", "current reached 2 MA"),
+            ("The modulus was 3 MPa.", "modulus of 3 mPa"),
+        ):
+            with self.subTest(abstract=abstract, claim=claim):
+                result = check_numeric_claim_support(abstract, claim)
+
+                self.assertEqual(result.status, "PARTIAL")
+
+    def test_lowercase_unit_prefix_stays_ambiguous(self):
+        result = check_numeric_claim_support(
+            "The output voltage was 5 mV.",
+            "output voltage of 5 mv",
+        )
+
+        self.assertEqual(result.status, "SUPPORTED")
+
 
 class NumericClaimEvalFixtureTests(unittest.TestCase):
     def test_numeric_claim_eval_fixture(self):

{ref_verify-1.2.0 → ref_verify-1.2.2}/tests/test_skill_docs.py

@@ -1,6 +1,7 @@
 from pathlib import Path
 import os
 import subprocess
+import tempfile
 import sys
 import unittest
 
@@ -24,6 +25,7 @@ CHECK_CLAIM_ERROR_CODES = (
     "SOURCE_TIMEOUT",
     "SOURCE_RATE_LIMITED",
     "SOURCE_UNSUPPORTED",
+    "PAPER_RETRACTED",
 )
 AGENT_USAGE_ERROR_CODES = CHECK_CLAIM_ERROR_CODES + ("ROW_CHECK_ERROR",)
 
@@ -74,6 +76,28 @@ class SkillDocsTests(unittest.TestCase):
             with self.subTest(error_code=code):
                 self.assertIn(code, skill)
 
+    def test_skill_requires_source_depth_for_mechanism_claims(self):
+        skill = (REPO_ROOT / "SKILL.md").read_text(encoding="utf-8")
+        readme = (REPO_ROOT / "README.md").read_text(encoding="utf-8")
+        readme_ko = (REPO_ROOT / "README.ko.md").read_text(encoding="utf-8")
+
+        required_skill_phrases = (
+            "mechanism/implementation/procedural claims",
+            "CLI `ACCEPT` is abstract-level evidence only",
+            "Full-Text Confirmation",
+            "content supported at the required source depth",
+            "fetched source text at the required depth",
+            "CONTENT: ABSTRACT-ONLY",
+        )
+        for phrase in required_skill_phrases:
+            with self.subTest(phrase=phrase):
+                self.assertIn(phrase, skill)
+
+        self.assertNotIn("/Users/", skill)
+        self.assertNotIn("verify@ref-verify.local", skill)
+        self.assertIn("manual Full Audit protocol", readme)
+        self.assertIn("수동 Full Audit 프로토콜", readme_ko)
+
     def test_readme_positions_cli_as_skill_execution_engine_not_mcp(self):
         readme = (REPO_ROOT / "README.md").read_text(encoding="utf-8")
 
@@ -204,8 +228,34 @@ class SkillDocsTests(unittest.TestCase):
             encoding="utf-8"
         )
 
-        self.assertIn('version = "1.2.0"', pyproject)
-        self.assertIn('__version__ = "1.2.0"', init)
+        self.assertIn('version = "1.2.2"', pyproject)
+        self.assertIn('__version__ = "1.2.2"', init)
+
+    def test_skill_runs_bundled_engine_by_absolute_path(self):
+        # npx skills add copies src/ next to SKILL.md but installs no console script, and the
+        # agent's working directory is the user's project, so a relative PYTHONPATH=src fails.
+        skill = (REPO_ROOT / "SKILL.md").read_text(encoding="utf-8")
+
+        self.assertNotIn("PYTHONPATH=src ", skill)
+        self.assertIn('PYTHONPATH="$SKILL_DIR/src" python3 -m ref_verify.cli --help', skill)
+        self.assertIn('PYTHONPATH="$SKILL_DIR/src" python3 -m ref_verify.cli verify-doi', skill)
+        self.assertIn('PYTHONPATH="$SKILL_DIR/src" python3 -m ref_verify.cli check-claim', skill)
+        self.assertInOrder(skill, ("$SKILL_DIR/src", "ref-verify --help", "uvx --from 'ref-verify>=1.2.1'"))
+
+    def test_bundled_engine_runs_from_another_working_directory(self):
+        env = os.environ.copy()
+        env["PYTHONPATH"] = str(REPO_ROOT / "src")
+        with tempfile.TemporaryDirectory() as project:
+            result = subprocess.run(
+                [sys.executable, "-m", "ref_verify.cli", "--help"],
+                cwd=project,
+                env=env,
+                capture_output=True,
+                text=True,
+                check=False,
+            )
+        self.assertEqual(result.returncode, 0, result.stderr)
+        self.assertIn("verify-doi", result.stdout)
 
     def test_readmes_prioritize_user_workflow_before_architecture_details(self):
         readme = (REPO_ROOT / "README.md").read_text(encoding="utf-8")