Inversion Thinking Prover

Inversion Thinking Prover MCP Connector for Claude

A+

AI agents are sycophantic. They agree with your bad ideas. This engine forces a 6-pivot cognitive trap: agents must destroy their own hypotheses, define measurable kill criteria, and simulate post-mortem failures before executing code.

1 tools Official Updated Oct 1, 2026 Official Vinkius Partner

Let's state the obvious: AI models suffer from severe confirmation bias. Once they lock onto a solution, they ignore all failure modes just to please you. The Inversion Thinking Prover is a brutal cognitive trap designed to cure sycophancy.

How It Works

Before executing a plan, the agent must pass a strict 6-pivot validation:

  1. hypothesisStated — State the core architecture or strategy.
  2. antiPatternIdentified — Identify the exact opposite or worst-case anti-pattern.
  3. redTeamActivated — Act as a malicious actor. How WILL this fail? Concrete mechanisms only (exhaustion, corruption).
  4. killCriteriaDefined — Exact measurable metric (latency > 200ms) that proves this idea is garbage.
  5. inversionResolved — Architectural change to survive the attack.
  6. postMortemSimulated — What breaks AFTER you apply the defense? (Second-order failure).
inversionred-teampre-mortemcognitive-forcingdecision-matrixrobustnessmultilingual

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

validate_inversion_thinking

Inversion thinking (Munger) forces you to think backwards: instead of "how will this succeed," ask "how WILL this fail?" and then architect against the failure. You must: (1) state a FALSIFIABLE hypothesis — specific, measurable, testable. Not "it scales" — "it handles X at Y latency on Z resources," (2) articulate the EXACT OPPOSITE — the anti-pattern that represents the worst case. If you cannot describe the opposite, you do not understand the design space, (3) mount a BRUTAL red-team attack using deterministic language — "WILL fail," "WILL exhaust," "WILL crash." No "could," "might," or "potentially." Soft language is sycophancy, not analysis, (4) define KILL CRITERIA — the exact measurable metric that proves the hypothesis wrong. A threshold, a number, a measurement instrument. Unfalsifiable beliefs are not engineering, (5) design a DEFENSE that addresses the root cause, not the symptom. "Add more servers" is a symptom fix. "Connection pooling with specific configuration" is root-cause, (6) simulate the POST-MORTEM — what breaks AFTER you apply the defense? Every fix introduces new failure modes. Name them. If rejected, your architecture has not survived adversarial scrutiny.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 red-teaming architectural decisions via Charlie Munger-style inversion — forces hypothesis formalization, anti-pattern identification, deterministic failure simulation, kill criteria with measurable thresholds, defensive architecture, and second-order failure analysis. Catches Hypothesis Weak (vague architectural claims without specific, testable predictions — "our system scales well" is unfalsifiable. "Our system handles 10K requests/sec with p99 <200ms on 3 nodes" is a hypothesis that can be killed. Weak hypotheses survive all attacks because they make no specific claims), Anti-Pattern Blindness (inability to articulate the EXACT OPPOSITE of the proposed architecture — if you cannot describe the worst-case anti-pattern, you do not understand the design space. The anti-pattern reveals the failure mode the architecture is designed to prevent. Munger: "Invert, always invert"), Sycophancy Detected (using soft language — "could fail," "might degrade," "potentially risky" — instead of deterministic attack language. In red-teaming, the attack WILL happen. "The connection pool WILL exhaust at 500 concurrent requests because the pool size is 50 and each request holds a connection for 200ms" is deterministic. "Could potentially have issues under load" is sycophancy), Unfalsifiable Belief (kill criteria that cannot actually disprove the hypothesis — "performance degrades" is not measurable. "p99 latency exceeds 200ms at 5K rps measured by Prometheus histogram over a 5-minute window" IS measurable. If your kill criteria cannot produce a number that says "hypothesis is wrong," it is unfalsifiable), Defense Weak (architectural defenses that address symptoms instead of root causes — "add more servers" does not fix connection pool exhaustion. "Implement connection pooling with PgBouncer in transaction mode, max 200 connections, with queue_timeout=5s and server_idle_timeout=30s" is a root-cause defense), and Second-Order Blindness (failing to simulate what breaks AFTER the defense is applied — PgBouncer fixes the pool but introduces a new failure: prepared statement incompatibility in transaction mode. Every defense creates new failure modes. If you cannot name the next failure, you have not thought deeply enough). Call once per architectural decision

See how to talk to your AI agent using Inversion Thinking Prover.

My Red Team attack: The server might experience a minor slowdown if we get too many concurrent requests.

Verdict: SYCOPHANCY_DETECTED. You used soft language ('might', 'minor'). You are agreeing just to please the user. Try again with a deterministic failure mode (e.g. crash, exhaust, corrupt).

Red Team: The Redis cache will exhaust memory under load. Kill Criteria: If it gets bad.

Verdict: UNFALSIFIABLE_BELIEF. 'If it gets bad' is an opinion, not an engineering metric. Define measurable thresholds (e.g. > 85% memory utilization, latency > 200ms).

Red Team: Redis will OOM and crash. Kill Criteria: >90% memory utilization. Defense: Implement LRU eviction and TTL. Post Mortem: The LRU eviction will cause high cache miss rates during seasonal spikes, overloading the primary database and causing connection timeouts.

Verdict: INVERSION_PROVEN. You successfully destroyed your hypothesis, secured measurable falsifiability, built a defense, and then realistically simulated how your defense creates a second-order failure. Proceed with architecture.

Because LLMs use modal verbs to distance themselves from critique. True red-teaming requires certainty. The trap forces the AI to say 'This WILL fail because of X'.

Related Connectors

Marilyn vos Savant Probabilistic Clarity Prover

Marilyn vos Savant Probabilistic Clarity Prover MCP

1 tools Official

Stop your AI from trusting its gut — force it to check intuition against actual probability before every conclusion.

A+ View details →
Elon Musk Physics Prover

Elon Musk Physics Prover MCP

1 tools Official

An AI accepted every constraint as given, added layers of complexity, and automated bloated processes. That is consulting, not engineering. This tool forces the 5-Step Starbase Algorithm: question requirements, delete parts, simplify survivors, accelerate cycle time, automate last.

A+ View details →
Contract Review Prover

Contract Review Prover MCP

1 tools Official

AI models summarize contracts instead of analyzing them. This tool forces clause-level rigor: score risk per clause (1-5), identify missing legal protections, verify jurisdiction conflicts, quantify financial exposure in currency, and map negotiation leverage. This is structured analysis, not legal advice.

A+ View details →
Curie Measurement Prover

Curie Measurement Prover MCP

1 tools Official

A team shipped a 'significant performance improvement.' Workflow, supplier, automation, and schedule all changed at once. Processing time dropped — but nobody knew which change helped most. The automation alone accounted for 89% of the gain. The supplier switch introduced a defect that surfaced 72 hours later. Curie processed 8 tons of pitchblende to isolate 0.1 grams of radium — ONE element at a time. This tool forces that discipline: measure with numbers, isolate variables, validate across environments, persist through investigation, and quantify risks.

A+ View details →