# Robust Reasoning Processes

> A small textbook for a proposed field: the study and engineering of reasoning processes — peer review, prediction markets, audits, LLM judge pipelines, debate protocols — that deliver trustworthy conclusions at known cost and resist corruption. Maintained by QURI (quantifieduncertainty.org) as part of the CAIRN project. Everything is a dated, versioned working draft; formalisms are graded [exact] / [standard shape] / [heuristic].

Core quantities: epistemic weight (a process's likelihood-ratio profile), cost per validated bit, and the corruption cost curve (minimum adversary spend to distort output by a given amount). The field's central empirical project is replacing qualitative ratings with measurements of these quantities.

## The book

- [Robust Reasoning Processes (Ch. 1)](/start-here/introduction/): what the field is, its three layers (Judge, Processes, Environment), and how it differs from scalable oversight, AI control, ELK, and forecasting research
- [Cruxes (Ch. 2)](/start-here/key-questions/): the uncertainties whose answers would most change the agenda
- [The Core Model (Ch. 3)](/concepts/core-model/): candidate formalisms — epistemic weight, the loss pipeline, unit economics, the corruption cost curve, prospective information pricing
- [Epistemic Impact Analysis (Ch. 4)](/proposals/epistemic-impact-analysis/): pricing information by validated belief change in a calibrated agent, weighted by a utility function
- [Constructing Utility Functions (Ch. 5)](/concepts/constructing-utility-functions/): building weighted question portfolios — formats, elicitation, the 2022 empirical record
- [Interlude: Untrustworthy Sources](/concepts/untrustworthy-sources/): which argument forms retain value as trust in the sender goes to zero
- [Interlude: The Funding Effect](/case-studies/the-funding-effect/): the meta-science funding literature as a likelihood-ratio discount on a corrupted source's epistemic weight
- [The Process Catalogue (Ch. 6)](/concepts/process-catalogue/): seventeen processes compared on info cost, scope, cost to bias, and built-in biases; the two-axes map and its empty quadrant
- [What Grounds an Oversight Protocol? (Ch. 7)](/concepts/oversight-protocols/): protocols organized by what their incentives terminate in — judge verdict, external resolution, coherence, peer agreement
- [Consistency Evaluations (Ch. 8)](/proposals/consistency-evals/): scoring estimators by whether they can be arbitraged; seven check types; counterexample shrinking
- [What Is a Strong Reasoner? (Ch. 9)](/concepts/what-is-a-strong-reasoner/): a trust-first definition — calibration, discrimination, consistency, honest uncertainty — distinct from "reasoning models"
- [Hardening: Overview (Ch. 11)](/concepts/hardening-techniques/): the objective (eliminate confident error), six mechanism families mapped onto the two axes, the threat-coverage matrix, composition, and machine-oversight affordances
- [Hardening: Calibration (Ch. 12)](/concepts/hardening-calibration/): tail calibration, priced abstention, the Kelly bankruptcy filter, the sharpness floor
- [Hardening: Verifiability Asymmetry (Ch. 13)](/concepts/hardening-verifiability/): make checking cheaper than producing — a PCP-analogue for estimates, forced precision, retrodiction
- [Hardening: Independence & Decorrelation (Ch. 14)](/concepts/hardening-independence/): commit-reveal, checkpoint-forking, n-version reasoning, decorrelated-probe interrogation
- [Hardening: Invariance & Low Sensitivity (Ch. 15)](/concepts/hardening-invariance/): breakdown point, Byzantine-robust aggregation, differential-privacy ceilings, randomize-everything
- [Hardening: Incentive-Compatibility (Ch. 16)](/concepts/hardening-incentives/): scoring rules, decision-contingent bets, the persuasion-budget bound
- [Hardening: Deterrence (Ch. 17)](/concepts/hardening-deterrence/): the inspection-game frontier, Merkle-logged traces, provenance bonds
- [The Reliability Ladder (Ch. 23)](/concepts/epistemic-applications/): five application tiers, each gated on its verification machinery
- [Overseeing Automated Research (Ch. 24)](/proposals/overseeing-automated-research/): output-metered oversight — mark the research book to market via a weighted question portfolio
- [LLM Epistemics in Production (Ch. 25)](/case-studies/llm-epistemics-in-production/): deployed case studies with published numbers and admitted failures
- [Open Problems (Ch. 26)](/open-questions/): the aggregated open-questions index

## Reference

- [Glossary & Notation](/reference/glossary/): every minted term and recurring symbol, with the page that owns it