glossary
short, testable definitions. each entry links to where it is formalized or used in practice.
A
- Abstention Threshold (τ*)
- the gating value applied to the compression tension score. if τ at a site or along a path exceeds τ*, the system abstains or requests clarification. see foundations and implementations.
- Auditability
- the share of steps that can be reproduced with stored artifacts. high auditability means a third party can retrace sources, parameters, and gates. see principles.
safety
replay
C
- Compression Site
- a step where data, context, or meaning is reduced. examples: retrieval prune, rank, summarize, generate, redact, review. each site is identified, scored, and auditable. see foundations.
- Compression Tension Score (CTS, τ)
- a normalized per site score that combines information loss, task degradation, provenance gaps, and uncertainty. used to gate claims and to compute path scores. equation on foundations.
- Compression Map
- a visual or textual trace of compression sites for a flow. includes τ per site, τ* thresholds, and provenance. used in reviews and audits. see applications.
- Contradiction Score (CS)
- a measure of net conflict between claims or sources, derived from entailment vs contradiction probabilities. positive means contradiction, negative means entailment. see benchmarks.
- Contradish
- a benchmark that evaluates detection of contradictions, τ distribution, and calibrated abstention. site: contradish.com.
visibility
metric
design
coherence
benchmark
E
- Entailment
- a claim is entailed when it follows from available evidence. in CAI, claims without entailment are blocked or routed to abstention. see faq.
evidence
H
- Hallucinet
- a CAI tool that flags outputs tied to risky compression paths and triggers abstention when τ exceeds thresholds. see applications.
tool
M
- Mirror Flags
- diagnostic markers that indicate disowned traits, false integration, or narrative breaks in inputs or outputs. used to surface where compression bent meaning. see applications.
- MirrorNet
- a CAI interface that shows compression choices to users and models, making tradeoffs visible and steerable. see applications.
diagnostics
tool
P
- Path Score
- an aggregation of τ across sites for a single request path. used to compare designs, set budgets, and trigger abstention. definition on foundations.
- Provenance
- the record of sources, transformations, and parameters that produced an output. a claim without provenance cannot be verified. see principles.
- Provenance Entropy
- a measure of concentration or dispersion of evidence sources. used to calibrate confidence and review effort. equation on foundations.
metric
accountability
metric
T
- Triarch
- orienting structure: compression, decompression, and awareness in tension. used as a compass for design choices. see about.
concept
U
- Unsupported Claim
- an output not entailed by available evidence or not covered by provenance. CAI systems abstain instead of asserting. see faq.
guardrail
last updated: