Sentinel v3.0

The security audit pipeline of marketnow.site · 12 stages · every layer backed by public evidence

L2: 2,839 / 2,868 tarballs scanned (99%) Decision benchmark: F1 0.857 Spec (md): /security/sentinel-v3.0.md

Sentinel is not a score of popularity — it is a layered inspection pipeline where each stage produces artifacts you can verify yourself. This page is the interactive map: click any stage to see what it does, what it caught, and where its evidence lives. Nothing here is claimed without a URL.

pipeline stages (L1→L9)

L2 rules on shipped artifacts

68,388

L1 index-certified entries

5.59M

weekly downloads covered by L2

Pipeline — click a stage

Sentinel v3.0

Select a stage above to inspect its evidence.

Review taxonomy — we never claim "verified" without earning it

Status	Count	Meaning
Auto-scanned	2,839	Sentinel ran; no human review. All L2-flagged artifacts keep their findings public.
Human-reviewed	22	Curated free skills with manual review.
Maintainer-verified	0	Program not yet launched — honestly reported as zero.

Sandbox configuration (L2.5)

docker run --rm -i \
  --network none \
  --read-only \
  --tmpfs /tmp:rw,size=64m,mode=1777 \
  --cap-drop ALL \
  --security-opt no-new-privileges \
  --user 65534:65534 \
  --memory 256m \
  --pids-limit 64 \
  --cgroup-parent=/marketnow/audit \
  "$IMAGE"

gVisor dynamic analysis is planned for Sentinel v3.1 (status: partial, not hidden).

Per-stage catch counts (real data)

Stage	Catches	Notes
L1 metadata validation	14 malformed manifests / 14,581	none malicious
L1.6 secrets (gitleaks batch)	1,257 findings	mostly test fixtures; OSV: 1 vuln
L1.6 semgrep (18 rules)	0 findings	own repo batch 2026-09-20
L1.7 binary detection	1 trojan (issue #9)	Trojan:Win64/Lazy.PGPK!MTB — regression vector since
L1.9 prompt injection	1 catch (ledger qd_001)	public quarantine record
L2 deep tarball scan	1,156 error / 791 warning verdicts	29 rules, shasum-verified artifacts
L2.5 sandbox (seccomp)	1 trojan via clone() denial	the same incident, runtime layer
L4 dependency scan	0 CVEs in production deps	OSV: 0 vulns across 688 locked deps (2026-09-25)

Honest limitations

This is not a substitute for human review of code that executes on your machine.

L1.5/L1.6 static analysis is regex-based (Semgrep-equivalent lite). Obfuscation, encodings and indirect calls can evade patterns.
L2 scans the top 2,868 npm targets by weekly downloads — long-tail packages carry L1 certification only (99% completion of the covered set).
L2.5 sandbox runs async (GitHub Actions) on repo-linked skills; gVisor is roadmap.
Independent pentest: not yet performed — scope published at /security/audit-2026-08-19#pentest.
Maintainer-verified count is 0 because the program has not launched.

Evidence index

Layer-by-layer evidence matrix — every claim → its public artifact
Sentinel benchmark (TP/FP/F1) — decision-level, reproducible from the public ledger
/api/security — machine-readable overview of all layers
/api/certification.json — L1 checks C1–C10 recomputed over the live bundle
/api/certification-scans.json — raw L2 results per package (2,839)
/api/quarantine — tamper-evident quarantine ledger (SHA-256 per record)
Audit AUD-2026-0821-MN — original findings, remediation and PDF+SHA-256