Glossary & Notation

Every term this book mints or borrows, and every recurring symbol, with a one-line definition and a link to the page that owns it. When prose and glossary disagree, the owning page wins; please file the disagreement.

Status

Draft v0 · updated June 2026 · maintained by QURI.

Terms

Term	One-line definition	Defined in
Robust reasoning processes (RRP)	The proposed field: the study and engineering of procedures that turn effort and evidence into trustworthy conclusions, and keep doing so under adversarial pressure	Robust Reasoning Processes
Reasoning process	A procedure — peer review, a prediction market, an LLM pipeline — that consumes effort and evidence and emits claims, estimates, or evaluations	The Core Model
Judge	The consumer of reasoning, who cannot check the work object-level and must decide how much to update	The Core Model
Epistemic weight	A process’s likelihood-ratio profile: how strongly its outputs discriminate truth from falsehood, conditional on its incentive environment	The Core Model
Loss pipeline	The heuristic decomposition of delivered value into codification, evaluator, and interpretation losses	The Core Model
Cost per validated bit	The field’s figure of merit: the price of information that survived checking, not the price of text	The Core Model
Corruption cost curve	The minimum an adversary of given capability must spend to distort a process’s output by a given amount	The Core Model
Corruption surplus	A participant’s maximum expected gain from deviating from honest effort — what attackers harvest	The Core Model
Robustness condition	Corrupting the process costs more than the distortion is worth — evaluated per threat model, never threat-model-free	The Core Model
Incentive audit	Measuring a corruption cost curve empirically: pay red-team producers to corrupt the process and record the price	The Core Model
Grading scheme	Every formalism carries [exact], [standard shape], or [heuristic] — trust each exactly as far as its grade	The Core Model
Epistemic Impact Analysis (EIA)	Pricing information by how it changes a calibrated agent’s beliefs, weighted by a utility function, validated at resolution	Epistemic Impact Analysis
Profundity	How load-bearing a changed belief is — how much an update propagates to downstream questions	Epistemic Impact Analysis
Falsehood nullification	The demand that false claims contribute zero or negative value — the hard problem of verification, restated as a desideratum	Epistemic Impact Analysis
Question portfolio	A utility function represented as a few hundred importance-weighted resolvable questions	Constructing Utility Functions
Relative value functions	Representing value as pairwise ratio distributions rather than absolute units, preserving correlation	Constructing Utility Functions
Grounding	What a protocol’s incentives terminate in: a judge’s verdict, external resolution, internal coherence, or peer agreement	What Grounds an Oversight Protocol?
Producer/consumer prediction game	An information producer trades against a calibrated reference model and profits by moving its beliefs and being validated at resolution	What Grounds an Oversight Protocol?
Retrodiction	Scoring models on known facts hidden from them — resolution without waiting, if contamination can be controlled	What Grounds an Oversight Protocol?
Epistemic selection protocols	Committing now to a process for choosing the most-trusted resolver at resolution time	The Reliability Ladder
Consistency battery	A suite of checks that a system’s estimates obey the constraints any rational belief set must — necessary, never sufficient	Consistency Evaluations
Dutch book / arbitrage metric	Inconsistency priced as the guaranteed profit extractable from a system’s own estimates	Consistency Evaluations
Opinion fuzzing	Sampling judgments across prompts, models, and personas and treating the variance structure as signal	Consistency Evaluations
Strong reasoner	An AI system whose judgments deserve substantial weight — in the limit, deference — within its domain; warranted trust, not raw capability (and not “reasoning model”)	What Is a Strong Reasoner?
LLM-based epistemic process	The unit of analysis: model plus scaffolding plus protocols, not the bare model	What Is a Strong Reasoner?
Reliability ladder	Five application tiers, each deployable only when its verification machinery exists and survives optimization against it	The Reliability Ladder
Output-metered oversight	Don’t audit the research process; meter the output’s validated epistemic impact against a question portfolio	Overseeing Automated Research
Consumer agent	The calibrated reference model whose belief state over the portfolio is “the book”	Overseeing Automated Research
Object loop / meta loop	Producers maximize validated impact against $V$; an institutionally separated party audits and re-weights $V$ itself	Epistemic Impact Analysis
Resolution layer	The enforcement of falsehood nullification: resolvers, retrodiction, consistency checks, randomized audits. “The resolution layer is the system.”	Overseeing Automated Research
Deception affordance	A form’s perceived likelihood ratio divided by its actual cost-to-fake one; operationally, its false-side belief swing relative to its true-side swing — near one, the form transmits persuasion, not information	Untrustworthy Sources
Built-in biases	A process’s systematic distortions with no adversary at all — corruption an attacker gets at zero marginal cost	The Process Catalogue
Decision-relative (goal) bias	A process’s systematic lean toward one resolution of a specific decision $D$ (nearest standard term: directional bias) — as opposed to broad, decision-independent biases; should be ≈0 on decisions it has no information about	The Process Catalogue
Label-swap neutrality	The unsupervised check for goal bias: swap the options of a no-information decision; residual output asymmetry is the bias	The Process Catalogue
Deception preconditions	The conjunction (verification gap, reproduction gap, dependence, goal-divergence, never-resolves, stakes) all required for deception to be a live risk — break any one to defuse it; read as the defender’s attack-surface-reduction checklist	Untrustworthy Sources
Residual attack surface	The few channels that survive design-time hardening — questions turning on an irreplaceable advantage, that never resolve, where goals diverge; where the runtime test must carry the load	Untrustworthy Sources
Irreplaceable advantage	A source’s output-relevant epistemic edge with no Blackwell-sufficient trusted substitute at feasible cost (replacement cost unbounded, though its value stays finite) — the core of the residual attack surface, where deception is genuinely dangerous	Untrustworthy Sources
Counterfactual-deceiver test	The listener’s runtime rule: use a message exactly to the degree a motivated liar in the source’s position could not have produced it as cheaply if the claim were false	Untrustworthy Sources
The empty quadrant	Cheap to run and expensive to corrupt — the region of the process map the AI era needs filled	The Process Catalogue

Notation

Symbol	Meaning	Home
$\pi$	a reasoning process	The Core Model
$J$	the judge	The Core Model
$U$	the judge’s utility function over information, represented as a weighted question portfolio	The Core Model, Constructing Utility Functions
$V_{source}, V_{delivered}$	information value before and after the loss pipeline	The Core Model
$L_{cod}, L_{eval}, L_{interp}$	codification, evaluator, and interpretation losses	The Core Model
$c_{process}, c_{interp}$, $n$	process cost, per-consumer interpretation cost, number of consumers	The Core Model
$C_\pi(\Delta, g, t)$	corruption cost curve: minimum adversary spend to distort $\pi$‘s output by $\Delta$ at capability $g$, time $t$	The Core Model
$\Delta$	distortion, measured in $U$-units	The Core Model
$g$	adversary capability	The Core Model
$V(I, A, U)$	EIA’s value of information set $I$ to agent $A$ under utility function $U$	Epistemic Impact Analysis
$I$	a set of information (data, arguments, a paper)	Epistemic Impact Analysis
$A$	the calibrated consumer agent — itself a reasoning process $\pi$ in the core model’s sense, playing the judge’s belief-keeper	Epistemic Impact Analysis
$Q$, $Q'$	an estimation question; a semantically equivalent rephrasing	Consistency Evaluations
$M(Q, t)$	the evaluated model’s answer to $Q$ at time $t$	Consistency Evaluations
$q$, $q_X$	a forecast; the forecast after observing source $X$	The Core Model
$I(Q; X)$	mutual information between question $Q$ and source $X$ — the decision-free value of a source, per question	The Core Model
$\tau$	the listener’s credence that a source is honest rather than strategic — underspecified as a scalar, refined into decision-indexed goal-divergence	Untrustworthy Sources
$b_\pi(D)$	decision-relative (goal) bias: $\pi$‘s systematic lean toward one resolution of decision $D$, holding truth-relevant information fixed	The Process Catalogue