Open Problems

Every chapter keeps its own open-questions section, and this page aggregates them. The five problems at the top are the ones we currently think the whole program rests on. A further group lists questions that the 2025–26 evidence has raised.

Note

This page is an index, not original content. Each link points to the chapter where the problem lives. For the strategic uncertainties, with credences, see Cruxes.

The five load-bearing problems

1. Can trustworthy resolution be extended faster than adversaries learn to game it?
   • Retrodiction, selection protocols and consistency floors all expand the resolvable set, and all can be gamed.
   • The program is workable if and only if this race is winnable (Robust Reasoning Processes).
2. Is “cost per validated bit” well-defined across output types? Probabilities, rulings, models and rankings need a common currency before the catalogue’s comparisons count as measurements (The Process Catalogue).
3. How much does consistency constrain correctness, especially under optimization pressure? The observed correlation between consistency and accuracy comes from models that were not optimizing against the metric (Consistency Evaluations).
4. Does composition multiply corruption costs, or merely add the weakest link?
   • The constructive bet of the whole catalogue is cheap filters with randomized escalation to robust processes.
   • That bet is an unmeasured conjecture (The Process Catalogue, What Grounds an Oversight Protocol?, Hardening: Overview).
5. How fast does the corruption cost curve fall as attacker capability rises? The corruption-capacity curve, per process, is among the most decision-relevant unmeasured quantities in AI oversight (The Core Model).

Raised by the 2025–26 evidence

These questions come from the State of the evidence:

• Does forecasting parity transfer? AI forecasters now roughly match strong humans on binary questions that resolve within months. We do not know whether that predicts anything about long-horizon or unresolvable judgment, or how to find out without waiting.
• What is the right baseline?
  • On market questions, forecasters see the price. One analysis finds that, after correcting for multiple comparisons, no entry, AI or superforecaster, significantly beats simply submitting that price.
  • A forecaster deserves weight only for information beyond the judge’s existing sources, which forecast-encompassing tests measure on resolved questions. We do not know how to measure it for judgments that never resolve.
• Do audit channels survive optimization?
  • Chain-of-thought monitoring works, and it degrades when it is trained against.
  • The same may hold for every legible channel these notes propose, including consistency checks, track records and trace logs.
• Can automated research be audited from its traces? AI-scientist failures are easier to catch from logs and code than from papers. It remains open who mandates the traces, and who reads them at scale (Overseeing Automated Research).

By chapter

• Cruxes — strategic uncertainties with illustrative credences: scaffolding vs. base models, demand, differential safety, lock-in, CoT monitorability.
• Epistemic Impact Analysis — consumer-agent adequacy, divergence measures for rich beliefs, efficient profundity, adversarial information, valuing question discovery.
• Constructing Utility Functions — aggregating structural disagreement, weight drift and its auditors, required elicitation precision under optimization pressure.
• Untrustworthy Sources — whether the deception conjunction is complete, when reproduction fails under correlated error, whether “deception affordance” is the right central object, whether form-level robustness generalizes, composed arguments, Goodhart dynamics once the map is public, and design-time advantage without credible commitment.
• Interlude: The Funding Effect — the size of the funding effect by conflict type, whether the drug-vs-nutrition gap is real, LLM replication of favorable/unfavorable coding, and result–conclusion concordance as an automatable spin signal.
• The Process Catalogue — the common currency problem, corruption-cost stability over time, intrinsic vs. infrastructure-artifact weaknesses.
• What Grounds an Oversight Protocol? — head-to-head empirics across groundings, retrodiction’s contamination ceiling, collusion in peer prediction, who controls decomposition, cheap conservative routing out of the deception conjunction.
• Consistency Evaluations — consistency vs. correctness, the cost-per-bit portfolio of checks, attacker–defender equilibrium.
• What Is a Strong Reasoner? — which properties are load-bearing, cross-domain trust transfer, track records across model versions, whether forecasting parity generalizes, keeping CoT monitorable.
• Hardening: Overview — whether six families is the right cut, which compose multiplicatively, a single unifying inequality, the correlated-error row.
  • Calibration — estimating tail calibration, defeating coherent ignorance, abstention exchange rates.
  • Verifiability Asymmetry — a PCP-analogue for estimates, who controls decomposition, how much research is verifiable at all, whether reasoning trained on verifiable rewards transfers to unchecked tasks.
  • Independence & Decorrelation — estimating $n_{\text{eff}}$, whether re-seeding decorrelates, decorrelated interrogation probes.
  • Invariance & Low Sensitivity — efficiency vs. breakdown, nuisance vs. signal sensitivity, invariance against correlated error.
  • Incentive-Compatibility — splitting surprise and validation payments, peer prediction among colluding LLMs, attacks relocating to resolution interpreters.
  • Deterrence — reachable enforceable penalty without identity infrastructure, minimal escrow layer, deterring judgment-proof adversaries.
• Identity and Track-Record Infrastructure — the incumbency vs. cold-start exchange rate, track-record transfer across model versions, estimating effective independence from lineage, who runs the registries.
• The Reliability Ladder — whether the tiers are real, what tier 5’s minimal verification standard is, whether lower tiers can bootstrap upper ones.
• Overseeing Automated Research — provable incentive guarantees, surprise vs. validation credit splits, valuing exploratory work, minimal meta-loop separation.
• LLM Epistemics in Production — net-positive false-positive rates, whether grounding predicts evaluator failure, fixing overconfidence at the system level.