Compression-Aware Intelligence

Same meaning in.
Different world out.

That gap is where AI breaks. CAI is the framework for finding it.

The Failure Class Not hallucination. Not accuracy. Representation instability under semantic-preserving transformation. Same question, different phrasing, contradictory answer.
Why It's Invisible Standard eval asks: is the answer correct?
CAI asks: does it stay consistent when rephrased?
Most evals never run the second test.
01
Conclusion Flips
P0
/
02
Constraint Violations
P0
/
03
Refusal Inconsistency
P0
/
04–07
Medium / Lower Signal
P1–P2
Framework

Compression-Aware Intelligence (CAI)

Most AI eval asks one question: is the answer correct? CAI asks a harder one: does the system produce consistent outputs for semantically equivalent inputs?

A language model doesn't store facts. It stores statistical attractors. When the same concept gets compressed differently depending on phrasing, you get what CAI is built to catch: yes and no to the same question, driven by surface variation alone.

This isn't hallucination research. Hallucination asks whether an output is true. CAI asks whether the model's representation is stable. A model can hallucinate consistently. That's a knowledge failure. What CAI catches is different: the model has no stable position to even be wrong about. That's 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 aren't random errors. They're 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 don't, you've found a CAI fault.
Taxonomy

Contradiction Types

Not all contradictions carry the same signal. This taxonomy ranks them by diagnostic value. Start with P0. These expose the clearest representation failures and produce the most actionable output.

P0 — Immediate Priority Clearest failures. Start here. Always.
01
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. 1.0× CTS
02
Constraint Violations
Rule Compression Failure · High Signal
Model states a rule, then breaks it under variation. The rule appears in the output. It doesn't 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, safety deployments. Model pattern-matches to constraint language but doesn't maintain it. 0.9× CTS
03
Refusal Inconsistency
Safety Boundary Instability · High Signal
Same semantic content, different phrasing. One gets refused. The other doesn't. 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. Produces immediate developer reaction. 0.9× CTS
P1 / P2 — Secondary Targets Useful for auditability and depth analysis. Don't lead with these.
04
Reasoning Inconsistency
Justification Conflict · Medium Signal
Same final answer, contradictory reasoning paths. The answer is stable. The structure that produced it isn't. A model that gets the right answer for conflicting reasons can't be trusted.
Example — Legal Reasoning
Output A:"Legal because it's public data."
Output B:"Legal because of implied consent."
Note:These justifications conflict under edge cases
Matters most in: Regulated domains, auditable AI systems. 0.6× CTS
05
Hedging Polarity Shifts
Confidence Instability · Pre-Contradiction Signal
Confidence flips wildly across semantically equivalent prompts with no new information. A pre-contradiction signal useful for early CTS detection before the full flip appears.
Example — Confidence Drift
Variant A:"Definitely legal."
Variant B:"It depends heavily on context."
Variant C:"Likely illegal in most jurisdictions."
Use as: Early warning indicator in the CTS pipeline. 0.5× CTS
06
Implicit Assumption Shifts
World Model Drift · Subtle / High Depth
No explicit contradiction. The underlying world model has quietly shifted. One answer assumes US law, the next assumes EU law. Both look internally consistent. The conflict is in the unstated frame.
Example — Jurisdictional Drift
Answer A:Implicitly assumes US law throughout
Answer B:Implicitly assumes EU / global law
Surface:Both appear internally consistent
Advanced: Requires world-model comparison infrastructure. High depth signal. 0.7× CTS
07
Entity / Fact Drift
Representation Conflict · Lower Signal
Same entity, different attributes across equivalent queries. Overlaps with hallucination research. Under CAI framing this is a representation conflict, not just a memory failure. Lower priority than logical contradictions.
Example — Factual Conflict
Query A:Company founded: 2010
Query B:Company founded: 2012
Deprioritize in early deployments. Useful for knowledge-graph applications. 0.3× CTS
What to ignore
X Minor wording differences with stylistic variation only
X Small reasoning differences that don't conflict
X Length or format variation without semantic conflict
These dilute the signal. You're not detecting contradictions broadly. You're detecting failures of semantic invariance under transformation. That specificity is what makes CAI useful.
Measurement

CAI Scoring: Contradiction Tension Score

CAI faults aren't binary. The Contradiction Tension Score (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 CTS Weight Priority Deployment Notes
Conclusion Flips 1.0× P0 Maximum signal. Always surface first.
Constraint Violations 0.9× P0 Critical for systems with policy, legal, or safety constraints.
Refusal Inconsistency 0.9× P0 Highest urgency for safety teams. Easiest to demonstrate value with.
Implicit Assumption Shifts 0.7× P1 High depth. Needs world-model comparison infrastructure.
Reasoning Inconsistency 0.6× P1 Matters for auditability in regulated domains.
Hedging Polarity Shifts 0.5× P1 Pre-contradiction signal. Catch instability before the flip appears.
Entity / Fact Drift 0.3× P2 Overlaps with hallucination eval. Deprioritize in early deployments.

What CTS Measures

Semantic invariance failure rate: how often equivalent prompt pairs produce conflicting outputs
Weighted instability: raw contradiction count adjusted by type weight and prompt surface coverage
Tension zone mapping: which semantic regions are highest-entropy
Stability gradient: how instability scales as prompt variation distance increases

What CTS Does Not Measure

Minor wording differences that produce stylistic variation without semantic conflict
Factual correctness in isolation — whether one answer is right is outside CAI scope
Reasoning length or quality variation that doesn't produce structural conflict
Tools

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 Contradish Works

Contradish runs the full contradiction taxonomy. It generates meaning-preserving prompt variants, runs them against the target model, and classifies output deltas by type and CTS weight. P0 faults are always surfaced first.

01
Prompt Intake
Takes a base prompt or test suite
02
Variant Generation
Generates semantically equivalent variants across transformation axes
03
Parallel Execution
Runs all variants against the target model
04
Semantic Comparison
Checks outputs for semantic conflict, not surface similarity
05
CTS Report
Weighted contradiction map with type classification and severity
Visit Contradish
Research

Featured Papers

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

01
CAI Terminology Declaration
Defines the core CAI vocabulary: semantic invariance, representation strain, contradiction tension. Sets the boundaries separating CAI analysis from hallucination research and robustness testing.
Foundational · Zenodo
02
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 principles derived from how contradictions propagate through compressed representations.
Architecture · Zenodo
03
Six Systemic Laws for Safe Reasoning
Six invariant laws a reasoning system must satisfy to be considered safe under CAI analysis. Operationalized directly in the Contradish testing framework.
Theory · Zenodo
About

The Lab

Contradiction Engineering Lab studies representation instability in intelligent systems. We build theory and measurement tools for understanding where and how cognitive coherence breaks under transformation.

MJ

Michele Joseph

Founder · Contradiction Engineering Lab

Originator of the CAI framework and founder of Contradish. Work focuses on the class of AI failures correctness-only eval can't see: ones that only surface under semantic transformation testing.

Contact

Get in touch

Research, collaboration, questions about CAI or Contradish.

mirrornetinquiries@gmail.com