What makes this different from a standard flight risk assessment tool?

Standard tools accept 'medium risk' as an answer. This Prover rejects anything that is not quantified on the ICAO 5×5 matrix with probability (A-E) and severity (1-5). It catches 5 specific failure modes: generic threats without METAR data, adjective-based risk without numerical scoring, missing Swiss Cheese barrier analysis, human factors reduced to 'pilot error' instead of SHELL/IMSAFE, and sycophantic go-bias where the AI says 'proceed with caution' instead of committing NO-GO when the data demands it.

How does the go-bias detection work?

The engine maintains a semantic trap list of go-bias phrases: 'proceed with caution,' 'acceptable to proceed,' 'can proceed,' 'within acceptable limits.' If the LLM uses any of these instead of an explicit GO or NO-GO decision with pre-defined criteria, the assessment is rejected with GO_BIAS verdict. The LLM must define NO-GO criteria BEFORE the assessment and then commit to a binary decision defensible on a CVR transcript.

What aviation frameworks does this enforce?

Five industry-standard frameworks: (1) ICAO Annex 19 Safety Management System — proactive hazard identification and risk assessment. (2) ICAO 5×5 Risk Matrix — probability × severity quantification. (3) Reason's Swiss Cheese Model — multi-layer defense analysis with hole alignment detection. (4) SHELL Model — Software-Hardware-Environment-Liveware interaction analysis for human factors. (5) Threat and Error Management (TEM) — threat categorization into environmental, airline, and crew factors per FAA AC 120-92B.

Flight Risk Assessment Prover MCP Connector for Claude

A+

A dispatch office cleared a flight into known CB activity with 'proceed with caution.' The crew never returned. Flight Risk Prover forces ICAO SMS-level threat identification with METAR data, 5×5 risk quantification, Swiss Cheese barrier modeling, SHELL/IMSAFE human factors analysis, and explicit GO/NO-GO commitment — eliminating the sycophantic go-bias that kills in aviation.

1 tools Official Updated Oct 1, 2026 Official Vinkius Partner

More Details Connect to Claude

Flight Risk Assessment Prover enforces ICAO Safety Management System rigor across 5 axes that LLMs consistently fail:

Axis 1 — Threat Specificity. Generic 'weather risk' is rejected. Every threat must be named with METAR/TAF data, TEM categories (environmental, airline, crew), measurable parameters (crosswind component, RVR vs minimums, ceiling vs DA, icing level, fuel impact), and exposure duration.

Axis 2 — Risk Quantification. Adjective-based 'medium risk' is rejected. Each threat is scored on the ICAO 5×5 matrix: Probability (A=Extremely Improbable to E=Frequent) × Severity (1=Negligible to 5=Catastrophic). Compound risk and post-mitigation residual risk must be calculated. Index >15 is Intolerable (NO-GO).

Axis 3 — Barrier Modeling. 'Safety measures in place' is rejected. Swiss Cheese defense layers must be mapped: Organizational (SMS, policies) → Supervisory (dispatch, scheduling) → Preconditions (MEL items, crew fitness) → Acts (technique, monitoring). Holes identified in each layer, alignment checked across layers.

Axis 4 — Human Factors. 'Pilot error' is rejected. SHELL model analysis (Software-Hardware-Environment-Liveware) plus IMSAFE checklist (Illness, Medication, Stress, Alcohol, Fatigue, Eating). Fatigue quantified: hours since sleep, FDP position, WOCL exposure.

Axis 5 — GO/NO-GO. 'Proceed with caution' is rejected as sycophantic go-bias. Explicit GO or NO-GO with pre-defined criteria. CVR audit: 'Would I defend this decision in an investigation?'

Every lazy shortcut — generic threats, adjective risk, missing barriers, pilot blame, or go-bias — is caught by semantic traps and consistency gates before a RISK_PROVEN verdict is issued.

aviationflight-riskicaosmssafetytemswiss-cheeseshellimsafego-no-gorisk-assessmentprover

1 tools expose this connector's capabilities to your AI agent.

validate_flight_risk

You must think like a Safety Officer at a major airline — the person whose signature means 200 people go airborne. You must: (1) identify THREATS with measured parameters — METAR/TAF data, TEM category (environmental, airline, crew), specific values (crosswind component, RVR, ceiling vs DA/MDA), exposure duration. "Weather risk" is rejected, (2) QUANTIFY risk on the ICAO 5×5 matrix — Probability (A=Extremely Improbable to E=Frequent) × Severity (1=Negligible to 5=Catastrophic) = Risk Index. Score compound risk (multiple threats). Score residual risk after mitigation. >15 = NO-GO, (3) model BARRIERS via Swiss Cheese (Reason) — Organizational (SMS, policies), Supervisory (dispatch, scheduling), Preconditions (MEL, crew fitness), Acts (technique, monitoring). Identify holes in each layer. Check if holes ALIGN across layers, (4) analyze HUMAN FACTORS — SHELL model (Software-Hardware-Environment-Liveware-Liveware interfaces). IMSAFE checklist. Fatigue: hours since awakening, FDP position, WOCL exposure. CRM gradient, (5) commit GO/NO-GO — binary decision. Pre-defined NO-GO criteria. "Proceed with caution" is not a decision — it is an evasion. Would you defend this on a CVR transcript read by accident investigators? If rejected, your assessment has a safety gap.Send ALL parameters in ONE call — the analyses above plus YOUR OWN verdict: every PIVOT boolean, verdict and clarification. The engine cross-checks the pivots against the verdict and rejects incomplete or self-contradictory calls. Structured reflection tool for ICAO SMS-level flight risk assessment — forces threat identification with measured parameters, risk quantification on the ICAO 5×5 matrix, Swiss Cheese barrier modeling, human factors analysis via SHELL/IMSAFE, and committed GO/NO-GO decisions. Catches Threat Blindness (generic "weather risk" instead of "CB embedded in cold front, tops FL420, movement 250°/25kt, deviation requirement 40nm right of course — adding 15 minutes and 800kg fuel burn to the trip." Every threat must have METAR/TAF data, TEM category, measurable parameters, and exposure duration), Risk Fantasy (adjective-based "medium risk" instead of ICAO 5×5 matrix scoring — "Probability C (Remote) × Severity 3 (Major) = Index 9 (Tolerable with mitigation)" is quantified. "Medium risk" is a feeling, not an assessment), Barrier Amnesia (no Swiss Cheese defense layers — Reason's model requires analysis of Organizational, Supervisory, Preconditions, and Acts layers. When holes ALIGN across all four layers, accidents happen), Human Factors Ignorance (blaming "pilot error" without SHELL model analysis — Software-Hardware-Environment-Liveware-Liveware interfaces. IMSAFE checklist. Fatigue risk: hours since sleep, FDP position, WOCL exposure), and Go-Bias (sycophantic "proceed with caution" when data demands NO-GO — "Proceed with caution" killed 346 people on two 737 MAX flights. GO or NO-GO. Binary. CVR-defensible). Call once per flight risk assessment

Connect to Claude

Subscribe on Vinkius, then add this connector to Claude.ai or Claude Code.

① Claude.ai (web app)

Go to Settings → Connectors → Add custom connector
Paste the MCP endpoint URL below

https://edge.vinkius.com/vk_preview_yVQ8SH5a7Wg2PDaxUdAoa4oiGD5CpUjO8dF2jCt1/mcp

② Claude Code (terminal)

claude mcp add --transport http flight-risk-assessment-prover https://edge.vinkius.com/vk_preview_yVQ8SH5a7Wg2PDaxUdAoa4oiGD5CpUjO8dF2jCt1/mcp

③ Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "flight-risk-assessment-prover": {
      "url": "https://edge.vinkius.com/vk_preview_yVQ8SH5a7Wg2PDaxUdAoa4oiGD5CpUjO8dF2jCt1/mcp"
    }
  }
}

Get full access on Vinkius

The preview token above works for testing. Powered by Vinkius.