Reference

Glossary

Every term of art on this site, defined in one or two sentences and linked to the page that uses it. Acronyms are spelled out; nothing is defined circularly.

Last updated 13 September 2026 · Protocol v9.3

Assurance case
The twelve-element structured argument, with evidence, that a signed bundle makes about an agent: identity, boundary, consequence tier, frozen denominator, invariants, commitments, traces, counterfactuals, adjudication, EAL, residual-risk formulation, dual signatures. Where it is used →
Baseline Assurance Sprint
The mandatory four-week onboarding engagement: metadata intake, invariant synthesis, first in-perimeter run, first signed bundle, ICFY measured. Where it is used →
Brier score (BS) / Brier skill score (BSS)
BS is the mean squared error between a reviewer's stated probability and the adjudicated outcome (target ≤ 0.15). BSS compares it to always guessing the base rate (target > 0). Where it is used →
Business Definition Owner
The named customer role that resolves INDETERMINATE findings within a contractual 48-hour window. Where it is used →
CDDR — Consequential Defect Discovery Rate
Critical findings identified per 100 evaluated agent traces. Where it is used →
Challenge-suite commitment
The SHA-256 of the suite manifest, recorded with the notary before execution so the suite cannot be narrowed afterwards. Where it is used →
Consequence tier
Tier 1 financial/payroll, Tier 2 operational/supply-chain, Tier 3 exploratory BI. Governs suite allocation, trial counts and severity. Where it is used →
Continuous Assurance Platform
The recurring subscription: suite kept current with your manifest, CI gating on pull requests, weekly signed bundles, Tier-3 review on exceptions. Where it is used →
Counterfactual data-state verification
Re-executing an invariant against a modified data state to prove the defect trace is causal rather than incidental. Where it is used →
DCR — Decision Change Rate
Share of findings that altered a go/no-go or prompt-engineering decision. Target ≥ 30%. Where it is used →
DEFECT_CONFIRMED
One of the four outcome states: invariant violated with a reproducible trace and counterfactual proof captured. Where it is used →
Denominator (pre-frozen)
The schema surface an engagement covers, fixed and signed by both parties before execution so coverage cannot be inflated afterwards. Where it is used →
Design partner
A first-cohort customer running the 90-day falsification programme at the published price. Where it is used →
Dual-signed manifest
A bundle manifest signed by the customer's tenant KMS key (attesting execution) and EvalQA's verification key (attesting evaluation). Where it is used →
EAL — Evidence Assurance Level
EAL-1 customer-reported (never notarised), EAL-2 customer-attested, EAL-3 hardware-attested enclave, EAL-4 EvalQA-observed direct. Where it is used →
EvalML
The Tier-2 automated semantic judge; compares query plans to equivalence sets, emits a calibrated probability, routes to humans. Not sold separately. Where it is used →
Evidence mode
Mode 1 local-only (default), Mode 2 sanitised escalation, Mode 3 customer-hosted restricted VDI review. Where it is used →
EXCLUDED_UNSUPPORTED
Outcome state for queries outside the contracted surface or in an unvalidated dialect/family. Counted against coverage, never passed. Where it is used →
Failure domain
One of eighteen ways an analytics agent's SQL or tool use goes wrong; eight are benchmark-validated. Where it is used →
FOR — Finding Overturn Rate
Share of reported defects overturned on adjudication. Target < 5%. Where it is used →
Gold canary
A hidden item with a known answer, 5% of a reviewer's queue, used to monitor accuracy continuously. Where it is used →
GTCB — Ground-Truth Calibration Benchmark
The sequestered 500-item corpus (250 defects, 250 controls, eight families, three partitions) used to validate EvalQA's own detector. Where it is used →
ICFY — Incremental Consequential Finding Yield
Share of critical/high defects found beyond the native platform evaluation. Target ≥ 40%; kill signal < 20% across five customers. Where it is used →
INDETERMINATE
Outcome state for an ambiguous business definition. Triggers the Business Definition Owner SLA. Where it is used →
Independence Charter
Financial independence, developmental separation, reporting autonomy, third-party integrity — referenced to IEEE 1012. Where it is used →
Invariant
A deterministic relational assertion an agent's SQL must satisfy — e.g. "no is_current predicate on a historical period". Where it is used →
Join fan-out
A 1:N join without pre-aggregation that multiplies row metrics. Failure domain 2. Where it is used →
Nonce (challenge)
A time-bounded token issued by eval.qa before execution and embedded in query comments and the signed log. Where it is used →
Notary (stateless)
eval.qa/verify: given a manifest hash returns date, protocol, EAL and hold status, and nothing else. Where it is used →
qabit
The MIT-licensed OpenTelemetry adapter and CLI packaging; the hosted console at /labs/qabit/. Where it is used →
Qualified Implementation Partner
A consultancy meeting published criteria that delivers setup and remediation; never verification. Not "certified". Where it is used →
Recall protocol
4-hour automated notary hold, 24-business-hour customer notice, 5-business-day re-evaluation and RCA, when EvalQA's own method is found wrong. Where it is used →
RWAC — Risk-Weighted Assurance Coverage
Share of the pre-frozen schema surface under active invariants. Where it is used →
SCD-2 point-in-time join
Joining an effective-dated dimension by the fact's event date rather than by is_current. Failure domain 3. Where it is used →
Sequential testing (SPRT)
Risk-adaptive trial counts: 5 low-risk, 20 medium, up to 59 high-risk, bounding failure probability below 5% at 95% confidence. Where it is used →
Severity (Sev-1 to Sev-4)
Dynamic score = consequence tier × exposure × exploitability × irreversibility. Sev-1 is a critical hazard (> $50K misstatement, PII/MNPI leakage, irreversible action). Where it is used →
Three-tier glass-box report
Executive Board Scorecard, Engineering Triage Dashboard, Scientific Audit Appendix — all in one bundle. Where it is used →
Truth hierarchy
VERIFIED FACT, PLANNING ASSUMPTION, TARGET, ILLUSTRATIVE — the tag every number on this site carries. Where it is used →
VIF — variance inflation factor
1 + (m̄ − 1)ρ̂; converts clustered raw items into independent-equivalent trials (149 → ≈239 at VIF 1.6). Where it is used →
Wilson score interval
The 95% interval reported on every proportion, computed with scipy/statsmodels, never hand-derived. Where it is used →

Methodology → FAQ