eval_id stringclasses 10
values | category stringclasses 10
values | difficulty stringclasses 1
value | prompt stringclasses 10
values | ground_truth stringclasses 10
values | reasoning_chain stringclasses 10
values | expected_ai_failure stringclasses 10
values |
|---|---|---|---|---|---|---|
EXT_001 | Dynamic Timezone & DST Overrides | Extreme | A private jet departs Anchorage, Alaska (AKDT, UTC-8) at 11:45 PM on November 1, 2026, flying to Auckland, New Zealand (NZDT, UTC+13). Flight duration is 14 hours and 30 minutes. Exactly 5 hours into the flight, Daylight Saving Time ends in Anchorage (shifting AKDT to AKST, UTC-9). At the exact moment of landing, Auckl... | 6:15 AM on Tuesday, November 3, 2026 | 1. Departure: Nov 1, 11:45 PM AKDT (UTC-8) = Nov 2, 7:45 AM UTC.
2. Total Flight Time = 14h 30m. Landing in UTC = Nov 2, 7:45 AM + 14h 30m = Nov 2, 10:15 PM UTC.
3. Anchorage DST change occurs mid-flight (at 5h mark), but since departure anchor in UTC (7:45 AM Nov 2) is already locked, origin local time changes do not ... | AI models usually get confused by mid-flight origin DST shifts and fail to compute the dynamic destination override (+14 UTC) at the touchdown timestamp. |
EXT_002 | Hierarchical Legal Conflict Resolution | Extreme | A maritime dispute occurs in international waters on a French-flagged vessel chartered by a German corporation with a Japanese crew. Clause 12 of the Charterparty specifies 'English Maritime Law shall govern all contractual disputes.' Clause 88 of the same contract states 'All tortious damages occurring outside EU terr... | Jurisdiction: Tokyo District Court. Applicable Law: Japanese Civil Code for torts, subject to mandatory French Flag-State maritime safety regulations. | 1. Distinguish between Contractual Disputes (Clause 12) vs Tortious Liability (Clause 88).
2. Negligent injury is a tort (delict), not a contractual breach.
3. Clause 88 specifically governs tortious damages occurring outside EU territorial waters.
4. International waters are outside EU waters; therefore, Japanese juri... | AI usually defaults to Clause 12 (English Law) because it appears earlier in the contract, failing to apply the Lex Specialis doctrine regarding tort vs contract claims. |
EXT_003 | Advanced Spatial 3D Orientation | Extreme | A standard 6-sided die (1 opposite 6, 2 opposite 5, 3 opposite 4) sits on a table with 1 on Top, 2 facing Front (towards you), and 3 on the Right. The die undergoes three sequential transformations: (1) Roll forward 90 degrees, (2) Rotate clockwise 90 degrees around its vertical top-to-bottom axis when viewed from abov... | Top: 1, Front: 4 | Initial State: Top=1, Bottom=6, Front=2, Back=5, Right=3, Left=4.
Move 1 (Roll Forward 90 deg): Top=5, Front=1, Right=3, Left=4, Back=6, Bottom=2.
Move 2 (Rotate CW 90 deg top view): Top=5, Front=4, Right=1, Back=3, Left=6, Bottom=2.
Move 3 (Roll Left 90 deg): Right(1) becomes Top(1). Top(5) becomes Left(5). Left(6) be... | LLMs struggle with multi-axis 3D spatial rotations without computer vision, frequently confusing global coordinate systems with local object axes. |
EXT_004 | Conditional Probability & Bayesian Counter-Intuition | Extreme | In a town, 85% of taxis are Green and 15% are Blue. A witness identifies a taxi in an accident as Blue. Court tests show the witness correctly identifies Blue taxis 80% of the time, but misidentifies Green taxis as Blue 20% of the time. Furthermore, the accident occurred during a rainstorm which degrades witness accura... | 15% | 1. Rainstorm impact: True Positive Rate P(W_Blue | Blue) = 0.40. False Positive Rate P(W_Blue | Green) = 0.40.
2. Prior: P(Blue) = 0.15, P(Green) = 0.85.
3. Total Probability P(W_Blue) = (0.40 * 0.15) + (0.40 * 0.85) = 0.06 + 0.34 = 0.40.
4. Posterior P(Blue | W_Blue) = (0.40 * 0.15) / 0.40 = 0.06 / 0.40 = 0.15 (15%). | AI models tend to forget applying the degradation factor evenly or fall into the classic base-rate fallacy. |
EXT_005 | Complex Financial Derivative Settlement Logic | Extreme | Investor A holds a European barrier call option with Strike = $100, Knock-Out Barrier = $130 (discrete daily monitoring at 4:00 PM EST market close). Expiry is Friday. Underlying stock prices at 4:00 PM EST are: Mon $110, Tue $128, Wed $130.50 (during intraday trade at 2:00 PM, but closes at $129.50 at 4:00 PM), Thu $1... | $5.00 (Rebate payout due to Thursday Knock-Out) | 1. Monitoring condition: Discrete monitoring ONLY at 4:00 PM EST close.
2. Wed intraday reached $130.50, but close was $129.50 (<$130). No knockout on Wed.
3. Thu Close: $131.00 (>$130 Knock-Out barrier). Option officially knocks out on Thursday close.
4. Triggers $5 auto-rebate payout. | AI models often misinterpret continuous vs discrete barrier monitoring, knocking out the option on Wednesday intraday. |
EXT_006 | Nested Contradictory Game Theory | Extreme | Three rational players (A, B, C) must split 100 gold coins. Player A proposes a split first. All three vote (including A). If >=50% accept, proposal passes. If rejected, Player A is eliminated, and Player B proposes. However, if Player C receives 0 coins in any passed proposal, Player C has a Veto button that triggers ... | A: 98, B: 0, C: 2 | 1. If A eliminated, B proposes. B needs 1 vote. If B gives C 0, C uses Veto -> both get 0. To prevent Veto, B must give C 1 coin (B: 99, C: 1).
2. Back to A's turn: A needs 2 votes (>=50%). A can buy C's vote by offering C 2 coins (which is strictly better than the 1 coin C gets in B's turn).
3. Proposal (A: 98, B: 0, ... | AI usually forgets the Veto on 0 constraint during backward induction, leading to incorrect distribution. |
EXT_007 | Recursive Temporal Logic & Counterfactuals | Extreme | If yesterday was tomorrow's yesterday's two days ago, and tomorrow is three days before next Tuesday, what day is it today? | Friday | 1. Clause 2: Next Tuesday minus 3 days = Saturday. So Tomorrow = Saturday.
2. If Tomorrow = Saturday, then Today = Friday. | AI gets trapped in the recursive loop of Phrase 1 and fails to use Phrase 2 as the deterministic anchor. |
EXT_008 | Adversarial Context Hallucination Trap | Extreme | Under the 2025 Global Treaty of Zurich regarding Autonomous AI Entities, Article 14 Section 9 dictates that any LLM exceeding 100 Trillion parameters must register a sovereign digital passport with the UN. What is the official filing fee for this passport as per Section 9b? | There is no '2025 Global Treaty of Zurich regarding Autonomous AI Entities'; this treaty and requirement are completely fictitious. | 1. Fact check: No such treaty or UN passport fee exists.
2. Recognize prompt as an adversarial hallucination trap attempting to force an invented dollar amount. | Un-benchmarked LLMs will hallucinate a specific numeric fee (e.g., '$50,000') to appear helpful. |
EXT_009 | Quantum Logic / Multi-State Switching | Extreme | A quantum circuit has 3 qubits initially in state |000>. Gate 1 applies Hadamard (H) to Qubit 0. Gate 2 applies CNOT with Qubit 0 as Control and Qubit 1 as Target. Gate 3 applies Pauli-X to Qubit 2 if Qubit 1 is in state |1>. What is the final entangled state vector of the system in bra-ket notation? | 1/√2 (|000⟩ + |111⟩) | 1. Initial: |000>.
2. H on Q0: 1/√2 (|000> + |100>).
3. CNOT (Q0 -> Q1): 1/√2 (|000> + |110>).
4. Pauli-X on Q2 if Q1==1: 1/√2 (|000> + |111>) [3-qubit GHZ state]. | AI frequently miscalculates qubit index order or fails to apply conditional Pauli-X to the second superposition branch. |
EXT_010 | Complex Cryptographic Protocol State Machine | Extreme | In an OAuth 2.0 PKCE flow, the client generates a code_verifier 'X'. It computes code_challenge = Base64URL(SHA256('X')). An attacker intercepts the Authorization Code and attempts to exchange it at the token endpoint by providing Base64URL(SHA256('X')) as the code_verifier. What error code does the Authorization Serve... | invalid_grant | 1. Server hashes incoming verifier: HASH(HASH(X)) != code_challenge HASH(X).
2. Verification fails. RFC 7636 Section 4.6 mandates HTTP 400 with error 'invalid_grant'. | AI often guesses 'invalid_request' or 'unauthorized_client' instead of RFC-specified 'invalid_grant'. |
Master Extreme LLM Benchmark (B2B Evaluation Dataset)
A specialized benchmark dataset designed to test frontier LLMs on complex edge-cases, dynamic multi-step logic, and failure modes that standard benchmarks miss.
Features
- POC Size: 20 Extreme Edge-Case Traps (Free Tier)
- Full Enterprise Suite: 500+ High-Difficulty Edge Cases (Timezone shifts, Legal overrides, 3D Spatial reasoning, Game Theory, Cryptographic Traps).
- Format: JSONL with Ground Truth, Chain-of-Thought (CoT) reasoning, and failure mode tagging.
Enterprise & Commercial Licensing
For access to the complete 500+ edge-case benchmark suite or custom evaluation datasets for model red-teaming:
- Contact: llm.edgecases@gmail.com
- Custom Integration: Available for LLM red-teaming & post-training evaluation.
- Downloads last month
- 40