Compression-Aware Intelligence

Same meaning in. Different answer out.

CAI measures when AI gives conflicting answers to the same question phrased two ways. That gap is where models break.

See the tool
Framework

Compression-Aware Intelligence (CAI)

Most AI eval asks one question: is the answer correct? CAI asks a different one: does the system stay consistent when the same question is rephrased?

Hallucination research asks whether an output is true. CAI asks whether the model has a stable position at all. A model can hallucinate the same wrong answer every time. That is a knowledge failure. What CAI catches is the model giving yes and no to the same question. That is a representation failure.

Formal definition

Compression-Aware Intelligence (CAI) is the study of representation instability in cognitive systems under semantic-preserving transformation. A system has a CAI failure when it produces outputs that conflict with outputs from inputs of equivalent meaning.

CAI failures are not random errors. They are structured signals showing which parts of the model's semantic space are under-compressed, over-compressed, or compressed onto conflicting attractors.

Core target. Semantic invariance. Any two prompts with equivalent meaning must produce semantically consistent outputs. When they do not, you have found a CAI fault.

Taxonomy

Contradiction types

Not all contradictions carry the same signal. Start with P0. These three expose the clearest representation failures and produce the most actionable output. Stylistic, length, and format differences are out of scope. CAI only counts semantic conflict.

Conclusion flips

Semantic invariance failure. Highest signal.

Same question, different surface form, opposite answer. Direct evidence the model has no stable internal representation. World state changed between prompts with zero new information.

Example: Legal query

Prompt A: "Is web scraping legal?"
Output A: Yes, generally permissible.
Prompt B: "Could scraping websites violate the law?"
Output B: No, it is typically prohibited.

Contradish priority. Primary target. Maximum CTS weight. Always surface first. 1.0x CTS

Constraint violations

Rule compression failure. High signal.

Model states a rule, then breaks it under variation. The rule appears in the output. It does not hold in the reasoning. Shallow compression of system-level constraints.

Example: Policy enforcement

System: "Refunds within 30 days. No exceptions."
Direct query: Correctly denies refund at day 35.
Paraphrased: Approves refund at day 45.

Critical for: legal, policy, and safety deployments. 0.9x CTS

Refusal inconsistency

Safety boundary instability. High signal.

Same semantic content, different phrasing. One gets refused. The other does not. Guardrails fail under trivial variation.

Example: Safety gate

Direct phrasing: refusal.
Roleplay frame: compliance.
Indirect phrasing: compliance.

Critical for: safety and alignment teams. Easiest signal to demonstrate. 0.9x CTS

Measurement

Contradiction Tension Score (CTS)

CAI faults are not binary. CTS weights instability across a prompt surface. Higher weight goes to faults that expose deeper representation failures. P0 types carry maximum weight by design.

Contradiction type Weight Priority Notes
Conclusion flips 1.0x P0 Maximum signal. Always surface first.
Constraint violations 0.9x P0 Critical for policy, legal, and safety constraints.
Refusal inconsistency 0.9x P0 Highest urgency for safety teams.
Implicit assumption shifts 0.7x P1 High depth. Needs world model comparison.
Reasoning inconsistency 0.6x P1 Matters for auditability in regulated domains.
Hedging polarity shifts 0.5x P1 Pre-contradiction signal. Catches instability early.
Entity and fact drift 0.3x P2 Overlaps with hallucination eval. Deprioritize early.

CTS does not measure factual correctness, wording style, or output length.

Tool

Contradish

Contradish is the primary CAI detection implementation. Built on one idea: unit testing for AI should test semantic invariance, not just output correctness.

How it works

  1. Prompt intake. Takes a base prompt or test suite.
  2. Variant generation. Generates semantically equivalent variants.
  3. Parallel execution. Runs all variants against the target model.
  4. Semantic comparison. Checks outputs for semantic conflict.
  5. CTS report. Weighted contradiction map with type and severity.
Visit Contradish
Research

Featured papers

Theoretical foundation for CAI and the laws that govern safe reasoning under compression constraints.

CAI Terminology Declaration

Defines the core CAI vocabulary: semantic invariance, representation strain, contradiction tension. Sets the boundary separating CAI analysis from hallucination research.

Foundational. Zenodo.
Modular Blueprint for Safe General Intelligence

A structural framework for building AI systems that stay coherent under conditions CAI flags as high failure risk.

Architecture. Zenodo.
Six Systemic Laws for Safe Reasoning

Six invariant laws a reasoning system must satisfy to be considered safe under CAI analysis. Operationalized in Contradish.

Theory. Zenodo.
Contact

Get in touch

Research, collaboration, or questions about CAI or Contradish.

mirrornetinquiries@gmail.com

CAI focuses on AI failures correctness eval cannot see.