Paradox Sifter

Paradox Sifter — Find the Contradictions Everyone Ignored

Every paper has a "Limitations" section. Nobody cross-references them. You do.

When you collect limitations from N papers on the same topic, contradictions emerge. Paper A's limitation is paper B's claimed advantage. Paper C's "future work" was already done by paper D and it didn't work.

These contradictions are where unsolved problems live.

---

Protocol

Step 1: Collect Limitations

Find 5-10 papers on the same topic. Extract the exact limitation statements. Include the paper's own words.

Step 2: Cross-Reference

Create a matrix. For each pair of limitations, ask:

1. Do these contradict each other? (A says X causes Y, B says X does not cause Y)
2. Do these ignore each other? (A flags a problem, B claims to solve it without citing A)
3. Do these compound? (A's limitation + B's limitation = a harder problem than either alone)

Step 3: Classify Each Paradox

| Type | Description | Opportunity | |------|-------------|-------------| | Direct contradiction | A says X, B says not-X | One of them is wrong. Find out which. | | Mutual ignorance | A and B solved the same problem in incompatible ways | A synthesis could be novel. | | Hidden dependency | A's solution only works if B's assumption holds | The combined system is fragile. | | Escalation | Solving A's limitation exacerbates B's | Trade-off space needs mapping. |

Step 4: Identify the Richest Paradox

The paradox with the highest combination of:

• Ignored duration (how long have these papers coexisted without anyone resolving this?)
• Impact if resolved (what would change?)
• Testability (can we design an experiment?)

---

Example Output

Topic: "Chain-of-Thought Prompting"

Papers: Wei et al. (2022), Wang et al. (2023), Kojima et al. (2022), Creswell et al. (2023), Madaan et al. (2024)

Cross-referenced limitations:

| Paper | Limitation | |-------|-----------| | Wei et al. | "CoT only helps above a certain model scale (~100B parameters)" | | Wang et al. | "Self-consistency improves CoT but requires multiple samples" | | Kojima et al. | "Zero-shot CoT works with models as small as 7B" | | Creswell et al. | "CoT can amplify errors — wrong reasoning chains produce wrong answers with higher confidence" | | Madaan et al. | "Self-refinement improves CoT but benefits from external feedback" |

Paradox found — Direct Contradiction:

Wei et al. (2022) claims CoT only works at ≥100B parameters. Kojima et al. (2022) shows zero-shot CoT works at 7B. Both published in the same year. Neither cites the other's limitation. The contradiction has been ignored for 3+ years.

Root cause: They tested different tasks (math vs common sense) and different CoT variants (few-shot vs zero-shot). The real limitation is task-dependent, not scale-dependent.

Resolution experiment: Test few-shot CoT vs zero-shot CoT across 10 tasks at model sizes from 1B-100B. Find the crossover point where each variant starts working. Publish the full surface.

---

Anti-Patterns

| Mistake | Why it fails | Fix | |---------|-------------|-----| | Superficial comparison | "Paper A and B are different" | Find the specific contradiction in their claims | | Not citing sources | Claims without evidence | Quote the exact limitation statement | | Ignoring temporal context | Older papers couldn't know | Account for when each paper was published |

PRISM Integration

In PRISM mode, this pattern's input changes: instead of paper limitations, it cross-references the structured findings from all 6 generators in Phase 1.

Output as structured YAML:

pattern: paradox-sifter
input: ""
cross_references:
  convergences:
    - claim: ""
      sources: [pattern_a, pattern_b, pattern_c]
      confidence: HIGH
  contradictions:
    - claim_a: ""
      claim_b: ""
      source_a: pattern_x
      source_b: pattern_y
      potential_resolution: ""
  creative_tensions:
    - dimension: ""
      dreamer_claim: ""
      pragmatist_assessment: ""
      frontier: ""
  assumption_cascades:
    - assumption: ""
      inverted_by: ""
      consequence_if_false: ""

Consumed by: synthesis (final output), skeptic (validation of cross-references) Consumes from: contrarian, assumption-excavator, cross-pollinator, heretic, counterfactual, dreamer

---

Trigger Conditions

Use this skill when:

• Asked to find "research gaps" or "future work"
• Reviewing multiple papers on the same topic
• Writing a "Related Work" section
• The user says "there's a contradiction in the literature"

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: argahv
• Source: argahv/novelty-skills
• License: MIT
• Homepage: https://github.com/argahv/sisyphus-academica

Install and usage instructions live in the source repository linked above.