rapier-runtime 0.5.0__tar.gz → 0.5.2__tar.gz

rapier_runtime-0.5.2/.githooks/pre-commit

@@ -0,0 +1,82 @@
+#!/bin/bash
+# Loom pre-commit secret-scan gate (project template).
+#
+# Lives in .githooks/ so it is git-tracked and shared across all clones.
+# Activated per-clone with: git config core.hooksPath .githooks
+# (projlaunch wires new worktrees automatically; /sync-now sets it on existing
+# clones -- see loom-secret-scan-check.)
+#
+# Why this exists: the 2026-06-30 .env-leak incident (CR-0004). A `.env` backup
+# with live keys was committed via `git add -A` and auto-pushed to origin within
+# seconds -- no human gate. .gitignore is one typo from failing, and it did.
+# This is the local backstop: it blocks the commit BEFORE any auto-push fires.
+# The server-side scan (.github/workflows/secret-scan.yml) is the guarantee;
+# this hook is defense-in-depth that catches a secret before it leaves the box.
+#
+# Three gates:
+#   1. PATH gate    -- refuses to stage any .env secret file (templates/examples
+#      excepted). This alone would have stopped the CR-0004 incident.
+#   2. CONTENT gate -- scans added lines for high-signal credential signatures,
+#      including in chatlogs and ordinary markdown (the inline-paste vector).
+#      Skips only dirs that legitimately carry the pattern literals / key-format
+#      teaching prose: .githooks/, docs/security/, docs/onboarding/.
+#   3. gitleaks     -- run if installed, as broader defense in depth (uses this
+#      repo's .gitleaks.toml automatically).
+#
+# Bypass (intentional, rare -- e.g. a verified false positive):
+#   git commit --no-verify
+# Prefer fixing the staged content over bypassing.
+
+set -uo pipefail
+
+staged=$(git diff --cached --name-only --diff-filter=AM)
+[ -z "$staged" ] && exit 0
+
+fail=0
+
+# --- Gate 1: env-secret file PATHS (the CR-0004 incident class) -------------
+while IFS= read -r f; do
+    [ -z "$f" ] && continue
+    base=$(basename "$f")
+    case "$base" in
+        *.example|*.template|*.sample) continue ;;
+    esac
+    if printf '%s\n' "$base" | grep -qE '^\.env($|[.-])'; then
+        echo "[pre-commit] BLOCKED: staged env-secret file: $f" >&2
+        fail=1
+    fi
+done <<< "$staged"
+
+# --- Gate 2: credential SIGNATURES in added lines --------------------------
+# High-signal, prefix-anchored signatures (near-zero false positives). Generic
+# high-entropy / base64 detection is deliberately left to gitleaks (Gate 3),
+# where the false positives live.
+patterns='sk-ant-[A-Za-z0-9_-]{16}|sk-proj-[A-Za-z0-9_-]{16}|sk-[A-Za-z0-9]{20}|xai-[A-Za-z0-9]{16}|SG\.[A-Za-z0-9_.-]{16}|AC[a-z0-9]{30}|(sk|rk|pk)_(live|test)_[A-Za-z0-9]{16}|whsec_[A-Za-z0-9]{16}|AKIA[0-9A-Z]{16}|gh[opsu]_[A-Za-z0-9]{36}|github_pat_[A-Za-z0-9_]{20}|xox[baprs]-[A-Za-z0-9-]{10}|xapp-[A-Za-z0-9-]{10}|AIza[0-9A-Za-z_-]{35}|GOCSPX-[A-Za-z0-9_-]{20}|-----BEGIN [A-Z ]*PRIVATE KEY-----'
+while IFS= read -r f; do
+    [ -z "$f" ] && continue
+    case "$f" in
+        .githooks/*|docs/security/*|docs/onboarding/*) continue ;;
+    esac
+    added=$(git diff --cached -U0 -- "$f" | grep '^+' | grep -v '^+++' || true)
+    if printf '%s\n' "$added" | grep -qE "$patterns"; then
+        echo "[pre-commit] BLOCKED: credential-like signature in staged change: $f" >&2
+        fail=1
+    fi
+done <<< "$staged"
+
+# --- Gate 3: gitleaks if available (defense in depth) ----------------------
+if command -v gitleaks >/dev/null 2>&1; then
+    if ! gitleaks protect --staged --no-banner >/dev/null 2>&1; then
+        echo "[pre-commit] BLOCKED: gitleaks flagged staged content." >&2
+        fail=1
+    fi
+fi
+
+if [ "$fail" -ne 0 ]; then
+    echo "[pre-commit] -------------------------------------------------------------" >&2
+    echo "[pre-commit] Commit blocked by the secret-scan gate. Remove the secret from" >&2
+    echo "[pre-commit] the staged set -- env backups belong OUTSIDE the repo tree." >&2
+    echo "[pre-commit] If this is a verified false positive: git commit --no-verify" >&2
+    exit 1
+fi
+exit 0

rapier_runtime-0.5.2/.github/workflows/secret-scan.yml

@@ -0,0 +1,45 @@
+name: Secret scan
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+
+permissions:
+  contents: read
+
+jobs:
+  gitleaks:
+    name: gitleaks
+    runs-on: ubuntu-24.04
+    steps:
+      - name: Checkout (full history for delta scan)
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Install gitleaks
+        run: |
+          set -euo pipefail
+          GITLEAKS_VERSION="8.21.2"
+          curl -sSfL -o /tmp/gitleaks.tar.gz \
+            "https://github.com/gitleaks/gitleaks/releases/download/v${GITLEAKS_VERSION}/gitleaks_${GITLEAKS_VERSION}_linux_x64.tar.gz"
+          tar -xzf /tmp/gitleaks.tar.gz -C /tmp gitleaks
+          sudo mv /tmp/gitleaks /usr/local/bin/gitleaks
+          gitleaks version
+
+      - name: Scan changes only (push diff or PR diff)
+        env:
+          BASE_SHA: ${{ github.event.before || github.event.pull_request.base.sha || '' }}
+          HEAD_SHA: ${{ github.sha }}
+        run: |
+          set -euo pipefail
+          if [ -z "$BASE_SHA" ] || [ "$BASE_SHA" = "0000000000000000000000000000000000000000" ]; then
+            echo "[gitleaks] no base SHA; scanning HEAD commit"
+            gitleaks detect --redact --verbose --no-banner --exit-code 1 \
+              --log-opts="-1 HEAD"
+          else
+            echo "[gitleaks] scanning diff $BASE_SHA..$HEAD_SHA"
+            gitleaks detect --redact --verbose --no-banner --exit-code 1 \
+              --log-opts="$BASE_SHA..$HEAD_SHA"
+          fi

{rapier_runtime-0.5.0 → rapier_runtime-0.5.2}/.gitignore

@@ -32,3 +32,6 @@ runs/
 
 # Nested git worktrees — registered checkouts, never committed into the repo
 .claude/worktrees/
+
+# Loom per-machine stamp (when the explainability reminder last ran)
+.loom/last-explainability-reminder

rapier_runtime-0.5.2/.gitleaks.toml

@@ -0,0 +1,23 @@
+# gitleaks configuration for this repo.
+#
+# What this is: a small settings file (TOML format -- plain "key = value" lines
+# grouped under [section] headers) that tells the secret scanner two things:
+# which detection rules to use, and which files to deliberately skip so we don't
+# raise false alarms. Read by BOTH the server-side scan
+# (.github/workflows/secret-scan.yml) and the optional local pre-commit hook.
+
+[extend]
+# Use gitleaks' built-in rule set (Anthropic, OpenAI, AWS, Stripe, GitHub, ...).
+useDefault = true
+
+[allowlist]
+description = "Known false-positive surfaces (safe to skip)"
+paths = [
+    # Dependency lock files: integrity hashes, never real credentials.
+    '''(^|/)(package-lock\.json|composer\.lock|yarn\.lock|pnpm-lock\.yaml)$''',
+    # Redaction tests feed fake key-shaped strings (sk-..., sk-ant-...) on purpose.
+    '''^tests/test_secrets\.py$''',
+]
+# Add project-specific entries above as real false positives appear, e.g.:
+#   '''database/migrations/.*\.sql$'''   # bulk data with high-entropy strings
+#   '''(^|/)\.env\.example$'''           # ONLY if its placeholders trip the scanner

rapier_runtime-0.5.2/.loom/audits/smuggled-virtue.done

@@ -0,0 +1,3 @@
+audited 2026-10-15T16:40:23Z by DEMETER
+fixed: 0 load-bearing (5 grep hits, all technical scope/behaviour qualifiers: compose.py:12,438; verify_grounding.py:516,608; rider_record.py:30)
+left: 0 historical

{rapier_runtime-0.5.0 → rapier_runtime-0.5.2}/.loom/project.json

@@ -4,5 +4,11 @@
   "tier": null,
   "source": "declared",
   "loom_commit": "c956f9c151c7303a190ec8d8c43b5ad6ca8a1e58",
-  "stamped": "2026-07-30"
+  "stamped": "2026-07-30",
+  "pm": {
+    "exempt": "Public repo; Bart ruled exempt 2026-10-02"
+  },
+  "ownership": {
+    "owner": "bart"
+  }
 }

rapier_runtime-0.5.2/CHANGELOG.md

@@ -0,0 +1,46 @@
+# Changelog
+
+Notable changes to `rapier-runtime`. Earlier releases are summarized in
+[`docs/STATUS.md`](docs/STATUS.md) and on the
+[Releases](https://github.com/muddyone/rapier-runtime/releases) page.
+
+## 0.5.2 — 2026-10-16
+
+### Fixed: the report no longer calls a same-vendor review independent
+
+Found on a live run with a single vendor's key, where the reviewer and the
+author were the same model:
+
+- **The report contradicted itself.** "Already fixed under challenge" said
+  "an independent, different-vendor reviewer raised each of these," and the
+  registry section pointed to "the cross-vendor challenge," while "How much to
+  trust this" correctly said the review was not independent. Every section
+  now says what actually happened; a cross-vendor run reads as before.
+- **The progress warning blamed the wrong key.** It said "OPENAI_API_KEY
+  absent; falling back to a Claude reviewer" when the OpenAI key was set and
+  no Claude model was used. It now names the vendor that ran the review.
+- **Arithmetic printed as raw LaTeX.** Math the model wrote as `\( H \times
+  0.60 \)` now reads `H × 0.60` in the report. Only text inside `\( \)` or
+  `\[ \]` is rewritten.
+
+## 0.5.1 — 2026-10-16
+
+### Changed: default models for five OpenAI-compatible vendors
+
+Five built-in defaults named models that a vendor has retired or is retiring. If
+you relied on a vendor's default (no `model:` in your manifest and no
+`RAPIER_MODEL_<VENDOR>` set), **your runs now use a different model.** To keep
+the old one while it still works, name it in your manifest or set
+`RAPIER_MODEL_<VENDOR>`.
+
+| Vendor | Was | Now | Why |
+|---|---|---|---|
+| `openrouter` | `openai/gpt-4o` | `openai/gpt-6-sol` | gpt-4o stops working 2027-01-14. gpt-6-sol is OpenAI's current flagship and costs slightly less. |
+| `deepseek` | `deepseek-chat` | `deepseek-flash` | `deepseek-chat` was retired 2026-07-24, so the old default already fails. |
+| `groq` | `llama-3.3-70b-versatile` | `openai/gpt-oss-120b` | Shut down 2026-08-16, so the old default already fails. Groq names gpt-oss-120b as the replacement. |
+| `together` | `meta-llama/Llama-3.3-70B-Instruct-Turbo` | `openai/gpt-oss-120b` | Leaves Together's serverless tier 2026-10-22. |
+| `gemini` | `gemini-2.5-flash` | `gemini-3.5-flash-lite` | 2.5 Flash is closed to new API users. Google recommends 3.5 Flash-Lite for new projects, at the same price. |
+
+`xai` (`grok-4.3`), `mistral` (`mistral-large-latest`) and `ollama` (`llama3.1`)
+are unchanged. Each was checked against its vendor's deprecation page on
+2026-10-15.

{rapier_runtime-0.5.0 → rapier_runtime-0.5.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: rapier-runtime
-Version: 0.5.0
+Version: 0.5.2
 Summary: A code-orchestrated engine that runs the SPARRING method: grounded, cross-vendor adversarial review for AI-in-the-loop decisions.
 Project-URL: Homepage, https://rapierruntime.com
 Project-URL: Repository, https://github.com/muddyone/rapier-runtime

{rapier_runtime-0.5.0 → rapier_runtime-0.5.2}/docs/STATUS.md

@@ -1,9 +1,46 @@
 # Rapier — dev status
 
-_Last updated: 2026-07-18 (tag v0.4.0, 186 tests)._ A running "where we are /
+_Last updated: 2026-10-16 (tag v0.5.1, 230 tests)._ A running "where we are /
 what's next" so a new session can continue without reconstructing from git log.
 
 ## Done
+- **RELEASED (2026-10-16) — `rapier-runtime` 0.5.1 on PyPI. Default models for
+  five vendors replaced.** Git tag `v0.5.1`;
+  https://pypi.org/project/rapier-runtime/0.5.1/. OpenRouter's default
+  `openai/gpt-4o` (stops working 2027-01-14, Loom Plan 26) is now
+  `openai/gpt-6-sol` (Bart's ruling). Checking every OpenAI-compatible default
+  against its vendor's deprecation page found two already failing (DeepSeek
+  `deepseek-chat`, Groq `llama-3.3-70b-versatile`), one leaving 2026-10-22
+  (Together's Llama 3.3 Turbo) and one closed to new users (`gemini-2.5-flash`).
+  The full table is in `CHANGELOG.md`, which starts with this release. None of
+  the new defaults has been tried on a live run yet; that needs vendor keys.
+- **RELEASED (2026-08-28) — `rapier-runtime` 0.5.0 on PyPI. The verifier stops
+  calling real citations fabricated.** Git tag `v0.5.0`;
+  https://pypi.org/project/rapier-runtime/0.5.0/. Eleven commits, and the
+  headline is a correctness fix in the grounding verifier:
+  `backend_literature` ran only a Crossref *bibliographic search* and gated every
+  candidate on a cited-surname match — but a bare identifier carries no surname,
+  so it could never match and fell through to `GROUNDED_REFUTED / "fabricated"`.
+  A bare DOI and a bare arXiv id, the two most authoritative citation forms
+  there are, were the two that always failed. `cite_check` had exposed
+  `lookup_crossref_doi` and `lookup_arxiv` the whole time; the backend called
+  neither. A verifier that flags real citations as fake is worse than none — it
+  trains the reader to distrust the gate.
+  Also in: identifier-strength resolution (DOI → arXiv → search, decisive both
+  ways); new `law` (CourtListener) and `patent` (Google Patents) backends;
+  evidence unioned across every identifier a reference supplies; a Proposer-side
+  `traceability_gate` (RM §10.1/§13.1) so the framing step's own citations are
+  grounded — verification was Resolver-only, so a Proposer could cite a
+  nonexistent source and the run reported clean; the reconcile gate now runs as
+  a standing check in `/spar`; and `RAPIER_MODEL_<VENDOR>` overrides the default
+  model per vendor. 223 tests green.
+  **Build note:** `pyproject.toml` pins `hatchling<1.30`. 1.30 emits
+  Metadata-Version 2.5 (PEP 794), which the current publishing toolchain rejects
+  (`twine check`: "'2.5' is not a valid metadata version") — that blocked this
+  release at the gate until pinned. Lift only when `twine check` passes a 2.5 wheel.
+  **Site:** no deploy needed. The full landing has been live since 2026-07-16 and
+  carries no version string; verified byte-identical to `site/index.html` on
+  2026-08-28.
 - **RELEASED (2026-07-18) — `rapier-runtime` 0.4.0 on PyPI. Governance: runs
   persist by default + THE RECORD provenance.** Git tag `v0.4.0`;
   https://pypi.org/project/rapier-runtime/0.4.0/.

{rapier_runtime-0.5.0 → rapier_runtime-0.5.2}/pyproject.toml

@@ -9,7 +9,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "rapier-runtime"
-version = "0.5.0"
+version = "0.5.2"
 description = "A code-orchestrated engine that runs the SPARRING method: grounded, cross-vendor adversarial review for AI-in-the-loop decisions."
 readme = "README.md"
 requires-python = ">=3.10"

{rapier_runtime-0.5.0 → rapier_runtime-0.5.2}/src/rapier/cli.py

@@ -304,8 +304,34 @@ def main(argv: list[str] | None = None) -> int:
     ip.add_argument("--dir", default=".", help="directory to write .env.example into (default: cwd)")
     sub.add_parser("mcp", help="run the MCP server (stdio) exposing frame/proposer/spar/sparring as tools")
 
+    vr = sub.add_parser(
+        "verify-rider",
+        help="check a run's trust rider against its own transcript (model-free)",
+    )
+    vr.add_argument("run", help="path to a persisted run directory, or a run id under the runs root")
+    vr.add_argument("--json", action="store_true", help="emit the report as JSON")
+
     args = parser.parse_args(argv)
 
+    if args.cmd == "verify-rider":
+        import json as _json
+        import os as _os
+
+        from .ledger import default_runs_root
+        from .verify.rider_record import check_run, render
+
+        run_dir = args.run
+        if not _os.path.isdir(run_dir):
+            run_dir = _os.path.join(default_runs_root(), args.run)
+        if not _os.path.isdir(run_dir):
+            print(f"no such run: {args.run}", file=sys.stderr)
+            return 2
+        report = check_run(run_dir)
+        print(_json.dumps(report, indent=2) if args.json else render(report))
+        # failed -> 1 so CI can gate on it; unchecked -> 3, distinct from both
+        # a pass and a failure, because an unverifiable claim is neither.
+        return {"failed": 1, "unchecked": 3}.get(report["status"], 0)
+
     if args.cmd == "doctor":
         from .onboarding import doctor_report
 

{rapier_runtime-0.5.0 → rapier_runtime-0.5.2}/src/rapier/models.py

@@ -178,14 +178,21 @@ class OpenAICompatibleModelClient(ModelClient):
 # vendor -> (base_url, key_env, default_model). Defaults are overridable in the
 # manifest; only xai's default is live-validated (2026-07-06 — Grok's API is
 # OpenAI-compatible). Others are best-effort until validated with a real key.
+# Checked against each vendor's deprecation page 2026-10-15 (see CHANGELOG.md).
 _OPENAI_COMPATIBLE: dict[str, tuple[str, str, str]] = {
-    "gemini":     ("https://generativelanguage.googleapis.com/v1beta/openai", "GEMINI_API_KEY", "gemini-2.5-flash"),  # frontier; also has a native API
+    # gemini-2.5-flash is closed to new API users; Google points new projects here, same price.
+    "gemini":     ("https://generativelanguage.googleapis.com/v1beta/openai", "GEMINI_API_KEY", "gemini-3.5-flash-lite"),  # also has a native API
     "xai":        ("https://api.x.ai/v1",            "XAI_API_KEY",        "grok-4.3"),
-    "deepseek":   ("https://api.deepseek.com/v1",    "DEEPSEEK_API_KEY",   "deepseek-chat"),
-    "mistral":    ("https://api.mistral.ai/v1",      "MISTRAL_API_KEY",    "mistral-large-latest"),
-    "groq":       ("https://api.groq.com/openai/v1", "GROQ_API_KEY",       "llama-3.3-70b-versatile"),
-    "together":   ("https://api.together.xyz/v1",    "TOGETHER_API_KEY",   "meta-llama/Llama-3.3-70B-Instruct-Turbo"),
-    "openrouter": ("https://openrouter.ai/api/v1",   "OPENROUTER_API_KEY", "openai/gpt-4o"),
+    # deepseek-chat was retired 2026-07-24; deepseek-flash is DeepSeek's current name.
+    "deepseek":   ("https://api.deepseek.com/v1",    "DEEPSEEK_API_KEY",   "deepseek-flash"),
+    "mistral":    ("https://api.mistral.ai/v1",      "MISTRAL_API_KEY",    "mistral-large-latest"),  # alias, follows Mistral's current Large
+    # llama-3.3-70b-versatile was shut down 2026-08-16; this is Groq's named replacement.
+    "groq":       ("https://api.groq.com/openai/v1", "GROQ_API_KEY",       "openai/gpt-oss-120b"),
+    # Llama-3.3-70B-Instruct-Turbo leaves serverless 2026-10-22; the cheapest current open model there.
+    "together":   ("https://api.together.xyz/v1",    "TOGETHER_API_KEY",   "openai/gpt-oss-120b"),
+    # gpt-4o stops working 2027-01-14. gpt-6-sol is OpenAI's current flagship, the same class
+    # gpt-4o was and slightly cheaper, so no caller gets a weaker critic or a bigger bill.
+    "openrouter": ("https://openrouter.ai/api/v1",   "OPENROUTER_API_KEY", "openai/gpt-6-sol"),
     "ollama":     ("http://localhost:11434/v1",      "OLLAMA_API_KEY",     "llama3.1"),  # local, zero-egress
 }
 

{rapier_runtime-0.5.0 → rapier_runtime-0.5.2}/src/rapier/stages/resolver/citation_gate.py

@@ -79,7 +79,8 @@ def _reconcile_definitiveness_with_grounding(env: Envelope, verdicts: list) -> N
     env.verdict = verdict
     rider = dict(env.trust_rider or {})
     rider["assumptions_to_verify"] = defin["rider_lines"]
-    rider["overall_confidence"] = verdict
+    rider["specifics_traceability"] = verdict
+    rider["overall_confidence"] = verdict  # deprecated alias; remove after 0.6.x
     rider["verified_externally"] = grounded_specs
     env.trust_rider = rider
     env.add_trace(

{rapier_runtime-0.5.0 → rapier_runtime-0.5.2}/src/rapier/stages/resolver/compose.py

@@ -5,6 +5,14 @@ Assembles the final report (recommendation + trust rider), where the rider's
 Proposer's forwarded standing objections (so the deliberation's dissent is
 visible to the user). Writes the /spar-parity named files and the derived
 ceremony-ledger row.
+
+
+Trust-rider field note: ``specifics_traceability`` carries the definitiveness
+gate's verdict -- whether every hard specific in the answer traces to a given
+fact or is flagged as an estimate. It is deliberately NOT named
+"overall confidence": it is one narrow evidence check, and says nothing about
+coverage, soundness, or the decision. ``overall_confidence`` remains as a
+deprecated alias with the same value.
 """
 from __future__ import annotations
 
@@ -40,6 +48,13 @@ def _verdict_sentence(verdict, gate) -> str:
             "to check, or the checker was unavailable), so treat the stated specifics as unverified.")
 
 
+def _challenge_phrase(review) -> str:
+    """Who raised the challenge, worded so a same-vendor review is never called independent."""
+    if review.get("cross_vendor"):
+        return "an independent, different-vendor reviewer"
+    return "a reviewer on the same vendor as the author (not an independent check)"
+
+
 def _reviewer_sentence(review) -> str:
     """Whether the challenge was genuinely independent — in plain words."""
     if review.get("cross_vendor"):
@@ -124,11 +139,38 @@ def _bullets(items) -> str:
     """One finding per bullet, hanging-indented so the marker column stays a clean
     vertical scan-line."""
     return "\n".join(
-        textwrap.fill(str(it).strip(), width=_W, initial_indent="• ", subsequent_indent="  ")
+        textwrap.fill(_plain_math(str(it)).strip(), width=_W, initial_indent="• ",
+                      subsequent_indent="  ")
         for it in items
     )
 
 
+# Models often write arithmetic as LaTeX (\( H \times 0.60 \)), which a terminal
+# prints literally. Only text inside \( \) or \[ \] is rewritten, so a stray
+# backslash elsewhere in the prose is left alone.
+_MATH_SPAN_RE = re.compile(r"\\\((.+?)\\\)|\\\[(.+?)\\\]", re.S)
+_MATH_SYMBOLS = {
+    "times": "×", "div": "÷", "cdot": "·", "approx": "≈", "pm": "±",
+    "leq": "≤", "le": "≤", "geq": "≥", "ge": "≥", "neq": "≠", "ne": "≠",
+    "to": "→", "rightarrow": "→", "infty": "∞", "%": "%", "$": "$",
+    ",": " ", ";": " ", "!": "", " ": " ", "quad": " ", "qquad": " ",
+}
+
+
+def _plain_math(text: str) -> str:
+    def span(m: re.Match) -> str:
+        t = m.group(1) if m.group(1) is not None else m.group(2)
+        for _ in range(3):  # nested \frac / \text unwrap from the inside out
+            t = re.sub(r"\\[dt]?frac\{([^{}]*)\}\{([^{}]*)\}", r"(\1)/(\2)", t)
+            t = re.sub(r"\\(?:text|mathrm|mathbf|operatorname)\{([^{}]*)\}", r"\1", t)
+        t = re.sub(r"\\left|\\right", "", t)
+        t = re.sub(r"\\([A-Za-z]+|[%$,;! ])",
+                   lambda k: _MATH_SYMBOLS.get(k.group(1), k.group(0)), t)
+        t = t.replace("{", "").replace("}", "")
+        return re.sub(r"\s+", " ", t).strip()
+    return _MATH_SPAN_RE.sub(span, text or "")
+
+
 def _demote_md(text: str) -> str:
     """Neutralize an embedded recommendation's own markdown so it cannot compete
     with the report's structure: ATX headings (`## X`) become bare Title-case
@@ -142,7 +184,7 @@ def _demote_md(text: str) -> str:
     s = re.sub(r"\*\*(.+?)\*\*", r"\1", s)   # bold
     s = re.sub(r"(?<![\w*])\*(?!\s)([^*\n]+?)(?<!\s)\*(?![\w*])", r"\1", s)  # italic
     s = re.sub(r"(?<!\w)`([^`]+)`(?!\w)", r"\1", s)  # inline code
-    return s
+    return _plain_math(s)
 
 
 # The stored verdicts carry the citation gate's shape (``status``:
@@ -175,7 +217,7 @@ def _grounding_body(env: Envelope) -> str:
         return ("The recommendation cited nothing externally checkable — no CVE, CWE, "
                 "RFC, DOI, or URL — so there was nothing to confirm against a public "
                 "registry. This is normal for judgment and strategy questions, where the "
-                "load-bearing work is the cross-vendor challenge above, not citation.")
+                "load-bearing work is the challenge below, not citation.")
     lines = []
     nver = 0
     for v in verdicts:
@@ -291,7 +333,7 @@ def _render_report(env: Envelope) -> str:
     L += ["", "", _HEAVY, "  TRUST RIDER", _HEAVY, "",
           "How far to trust the recommendation above: what is still yours to",
           "confirm, what was verified for you, and how the answer held up under",
-          "an independent challenge."]
+          "challenge."]
 
     assumptions = rider.get("assumptions_to_verify") or []
     L += _sec("STILL YOURS TO CHECK",
@@ -314,13 +356,15 @@ def _render_report(env: Envelope) -> str:
 
     contested = _texts(review.get("objections"))
     L += _sec("ALREADY FIXED UNDER CHALLENGE",
-              "Not action items — already folded into the recommendation above. An "
-              "independent, different-vendor reviewer raised each of these, and the "
-              "answer was revised to address it. Shown so you can see what the "
-              "challenge caught and how the recommendation got stronger.",
+              "Not action items — already folded into the recommendation above. "
+              + _challenge_phrase(review)[0].upper() + _challenge_phrase(review)[1:]
+              + " raised each of these, and the answer was revised to address it. "
+              "Shown so you can see what the challenge caught and how the "
+              "recommendation got stronger.",
               _bullets(contested) if contested
-              else "The independent review found nothing material to change — the "
-                   "recommendation held up as written.")
+              else "The review found nothing material to change — the recommendation "
+                   "held up as written. See HOW MUCH TO TRUST THIS for how independent "
+                   "that review was.")
 
     dissent = rider.get("proposer_dissent_forwarded")
     if dissent:
@@ -497,7 +541,8 @@ class ComposeStage(TransformStage):
         rider.setdefault("contested_and_resolved", _texts(review.get("objections")))
         if standing:
             rider["proposer_dissent_forwarded"] = _texts(standing)
-        rider["overall_confidence"] = env.verdict
+        rider["specifics_traceability"] = env.verdict
+        rider["overall_confidence"] = env.verdict  # deprecated alias; remove after 0.6.x
         env.trust_rider = rider
 
         env.meta["report"] = {

{rapier_runtime-0.5.0 → rapier_runtime-0.5.2}/src/rapier/stages/resolver/cross_review.py

@@ -7,6 +7,10 @@ the anchored-correction stage to act on.
 """
 from __future__ import annotations
 
+import contextlib
+import io
+import sys
+
 from ...envelope import Envelope
 from ...stage import StageContext, TransformStage, register_stage
 
@@ -24,10 +28,23 @@ class CrossReviewStage(TransformStage):
         primary_v, secondary_v = bind_pair(
             env, secondary_pref=ctx.config.get("reviewer_vendor"), policy=ctx.policy
         )
+        # The vendored reviewer names its two slots after Anthropic and OpenAI, so
+        # its fallback warning ("OPENAI_API_KEY absent; falling back to a Claude
+        # reviewer") is wrong once the slots are bound to other vendors. Capture
+        # it and say what actually happened instead; other stderr passes through.
+        captured = io.StringIO()
         try:
-            result = B.review(env.request, env.recommendation, None, None)
+            with contextlib.redirect_stderr(captured):
+                result = B.review(env.request, env.recommendation, None, None)
         finally:
             B.reset_slots()
+        for line in captured.getvalue().splitlines():
+            if line.startswith("WARN spar-cross-review:") and not result.get("cross_vendor"):
+                if result.get("reviewer_vendor") is None:
+                    continue  # no keys at all; the pipeline's own preflight reports that
+                line = (f"WARN cross_review: no second vendor available; the reviewer runs on "
+                        f"{primary_v}, the author's vendor (NOT cross-vendor).")
+            sys.stderr.write(line + "\n")
 
         env.meta["review"] = result
         n = len(result.get("objections") or [])

{rapier_runtime-0.5.0 → rapier_runtime-0.5.2}/src/rapier/stages/resolver/definitiveness_gate.py

@@ -34,7 +34,8 @@ class DefinitivenessGateStage(TransformStage):
 
         rider = dict(env.trust_rider or {})
         rider["assumptions_to_verify"] = result.get("rider_lines") or []
-        rider["overall_confidence"] = env.verdict
+        rider["specifics_traceability"] = env.verdict
+        rider["overall_confidence"] = env.verdict  # deprecated alias; remove after 0.6.x
         review = env.meta.get("review") or {}
         if review.get("objections"):
             rider["contested_and_resolved"] = [