Paste a recruiter message. Get one coherent verdict, is it safe to engage and is it worth your time, with an auditable trail behind it.
The pipeline is deterministic by design. Risk and the final decision are computed by typed rules over a locked evidence ledger; an optional LLM writes the human-readable summary but cannot invent evidence or change the verdict. Same input, same verdict, every time.
- Evidence integrity. Tool-produced evidence is authoritative; anything the model adds is discarded before scoring.
- Two honest axes. Risk (real danger signals) is separated from confidence (how much could be verified), so a legit-but-unverified message reads as low-risk / low-confidence, not "dangerous".
- Safe by default. Live URL fetch and the code sandbox ship OFF. When enabled the sandbox runs
--network none, read-only, non-root, resource-capped. - Typed end to end. LangGraph state with reducers, pydantic contracts (
extra="forbid"), and a real test suite.
Built for messy real-world outreach: LinkedIn DMs, recruiter emails, copied job descriptions, assessment links, GitHub repos, package-install requests, and crypto/Web3 roles. Two questions usually get mixed together, and this keeps them apart: is this real (domain resolves, official job page exists, company exists, JD looks like a real client role) and is this safe and worth time (shortened links, wallet connections, local code execution, off-platform moves, salary and stack alignment). A role can be real but not worth pursuing; a recruiter can have a real domain and still send a risky assessment.
- Decision.
Do not engage,Hold until basics are confirmed,Proceed with qualifying questions. - Safe reply. Asks for the right proof without taking unsafe action.
- Main reasons. The top decision drivers, not raw scanner spam.
- Opportunity read. Legitimacy, security risk, fit, salary stance, reply focus, open questions.
- Verification digest. Plain-English checks: DNS/MX, RDAP, Companies House, official and candidate JD evidence, dates checked.
- Original JD recon. Infers the likely original job or client from title, stack, salary, location, and wording.
- Lane reports + raw evidence ledger and a Markdown report for audit.
A message pushing Telegram, a shortened assessment link, local code, package installs, or a wallet connection lands on Do not engage / security risk high / fit poor. A plausible corporate domain with an official careers URL but unclear salary and client lands on Proceed with qualifying questions / legitimacy high / risk low.
extract_entities
→ deterministic_risk_scan
→ original_jd_recon (reverse-searches JD clues to infer likely end client)
├── recruiter_identity (email/domain checks)
├── agency_legitimacy (registry/search/domain checks)
├── client_company (verifies named or inferred company lead)
├── job_jd_verification (checks official/matched JD and GitHub metadata)
├── security_assessment (link/package/repo/wallet/code-execution risk)
└── reputation_pattern (planned lane — inactive, contributes zero risk)
→ merge_evidence
→ final_report
original_jd_recon runs first as a neutral lead generator; only threshold-confident structured facts pass downstream. The other lanes never see each other's conclusions, they verify independently against the original outreach and tool output. The final evidence list is locked to input and tool-produced evidence. Static diagrams live in artifacts/graph_mermaid*.png.
Not a magic verifier, and it should not be treated as one. It runs locally on 127.0.0.1 with no auth. .env is gitignored. Live URL expansion (EXPAND_URLS_LIVE) and repo sandboxing (SECURITY_SANDBOX_LIVE) are both off by default; with sandboxing off, code-assessment links produce a no-network Docker sandbox plan plus a do-not-run warning. Live fetches allow only public http/https targets, blocking localhost, private IP ranges, link-local metadata services, non-HTTP schemes, embedded credentials, and unsafe redirects. Sandboxed analysis is static inspection inside Docker with --network none, read-only mounts, non-root user, CPU/memory/time limits, and no project install or test script execution.
python3 -m pip install -r requirements.txt
python3 -m app.server # http://127.0.0.1:8765curl -s http://127.0.0.1:8765/api/full-check \
-H 'Content-Type: application/json' \
-d '{"text":"Hi, we have a role. Please run npm install and connect your wallet for the assessment."}'The endpoint returns the same structured output as the UI.
python3 -m pip install -r requirements-dev.txt
python3 -m pytest -q # 79 passedtests/test_determinism.py proves the LLM cannot change the verdict; tests/test_coherence.py proves decision, score, and opportunity read share one security_risk.
Copy .env.example to .env and fill in only what you want. Everything is optional; with no LLM key the graph falls back to deterministic lane reports.
COMPANIES_HOUSE_API_KEY= # registry checks
SEARCH_API_PROVIDER=brave
SEARCH_API_KEY=
EXPAND_URLS_LIVE=false
SECURITY_SANDBOX_LIVE=false
GITHUB_TOKEN= # enrichment
URLSCAN_API_KEY=
VIRUSTOTAL_API_KEY=
LANGSMITH_TRACING=false # tracing
ENABLE_LLM_LANE_SYNTHESIS=false # optional summary writing
LLM_PROVIDER=anthropic
LLM_MODEL=claude-sonnet-4-5-20250929
ANTHROPIC_API_KEY=LLM_LANE_SYNTHESIS_LANES selects which lanes get synthesis; see .env.example for the full list and the LangSmith and custom-endpoint keys.
A prototype, not a production SaaS. No auth (local use by design). Dependencies are pinned but there is no full lockfile. Public web and search results can be stale; the UI labels candidate evidence separately from verified evidence. Companies House and search quality depend on configured keys. The reputation_pattern lane is labelled not_applicable in code and contributes zero risk until implemented.