Risk-constrained solver selection
When do learning, verification and fallback make a solver policy worth trusting and paying for?
Read Online Or start with the explanation below.
Why this question matters
A model of a computation can decide whether to accept, check or repair an answer. This methods study prices those operations and separates checking the current witness from using the result to improve future solver choices.
Setup and results
A finite deployment menu has an exact cost frontier. Independent audited training gives simultaneous calibration and excess-cost bounds, with a margin premise for the oracle comparison. Bayesian and sliding-window variants treat learning and drift. Fallible checking and controller overhead change residual risk. The frozen public SAT study selects exact solver E for every risk-constrained learned choice; a mature exact baseline strengthens the negative result.
Limits and negative results
Gold-label dependence, warmup and setup/checking costs preclude an efficiency claim. The work synthesizes established metareasoning, algorithm selection and calibration rather than validating universal architectural superiority or subjective introspection. Benchmark availability does not imply redistribution permission.
Reading guide
Start with the known-reliability menu, then audited training and verification for future tasks. Read the measured full pipeline and larger-instance amendment alongside the negative selection result. Retrieve SATLIB inputs from the official catalog and compare corrected checksums; no benchmark archive is bundled here.
Place in the research programme
The relationship to Reflective Self-Models is choosing actions using a fallible model of computation. CM Compiler Computation also retains negative cost comparisons, but its operator-level compiler model and baselines differ from this solver-selection study.
Edition and review status
Brian Theory is the author; this is an AI-assisted research draft. Documented specialist and hostile AI review rounds informed development, but not every round was blind, independently provisioned or cover-to-cover on the final assembly. These checks are not external human peer review or journal acceptance. Targeted human review and historical novelty assessment remain pending.
Sources and reproducibility
Read the complete manuscript, open its page-faithful PDF reader, or download the exact PDF. Exact LaTeX and figure dependencies contain the manuscript inputs and a checksum manifest. This source-only package does not contain the computational verification programs; their public release and independent execution are not certified here. Read the manuscript’s reproducibility sections for the scope of its finite checks.
PDF SHA-256: 8a2ebd193061af31cadfba3aad83373e8ac12ac38df30189d28410c3a172bce9. Main LaTeX SHA-256: 7cebfa3cb374468a1c7a1853f0309e56a21a455a85e046dfaeb32972bc941e46.
Retrieve benchmarks from the official SATLIB catalog. No third-party benchmark archive is included. The manuscript describes a frozen, hash-bound benchmark protocol. Its separate benchmark checksum/reproduction manifest is not included in this source-only package; full input-hash validation still needs that separately vetted material.
Related manuscript
Learning Which Computations to Trust: Audited Solver Choice, Fallible Verification, and DriftAI-assisted research draft; internal AI checks documented. External human review and novelty assessment remain pending.
Reading scope: Trusted/gold labels, checker fallibility, controller/setup costs and drift assumptions are explicit. No measured efficiency advantage is established.
Site reading copy 1 from 13 November 2026 manuscript with localized corrections and approved author credit; exact edition hashes linked below.
Technical material
How this connects
Open in the knowledge map- Phases
- Manuscripts
- Projects
Conceptual and navigational links, not proof dependencies.

