Trust in the Dark: Maintaining Map Fidelity Under Corruption, Cost, and Substrate Decay
Preprint · not peer reviewed
Abstract
Description / Abstract:
Trust in the Dark is an agent-based study, with a pre-registered Python/NumPy simulator
suite, of how a system keeps a true-enough model of a partly-hidden, drifting, and
adversarial world — the problem of *map fidelity*. It is one part of a wider program
(the comparative cybernetics of fidelity) and ships with a master overview tying it to
its sibling deposits. The central methodological constraint is that an agreeable
instrument is a liability as a verifier: every load-bearing claim is pre-registered with
an explicit kill condition before it is run, results are reported including nulls, and
the work is released to be broken.Robust, reproducible findings include: coverage is the master variable (full coverage
rescues any prior); blind optimism and blind paranoia frame-lock identically;
cooperative vetting has an overlap optimum (φ\* ≈ 0.25); disagreement-based vetting has
zero leverage against correlated, out-of-frame (Outside-Context-Problem) corruption; a
non-adversarial substrate corruptor is catchable only when identification is paired with
durable out-of-band grounding; and a navigation "keystone" in which a position estimate
gates a value estimate, so that an upstream frame error becomes coherent downstream
corruption.
The deposit is falsification-first. It records what was tested, what was killed (a
control-theoretic windup–framelock identity; a "sharp threshold" that finer sampling
revealed as a gradual sigmoid; early-warning signatures that proved to be trivial
noise-scaling; a third "saturation" attack class that collapsed into an existing one
under a pre-registered test), and what remains asserted. A COVERAGE_MAP, a
REFEREE_CHANGELOG, verified CITATIONS, and a BREAK_THIS open-falsification challenge
travel with the code.
Status: self-deposited, AI-assisted (a bound, adversarial procedure-runner), not
peer-reviewed and not awaiting review. The simulator results stand on reproducible runs
regardless of how much of the conceptual scaffold is eventually built or retired.
Companion to the author's Frame-Lock / Red Queen's Prison and Turtles deposits and to the
Comparative Biosonar & Navigation work.
How to cite
Schulz, M. (2026). Trust in the Dark: Maintaining Map Fidelity Under Corruption, Cost, and Substrate Decay. Preprint, Zenodo. https://doi.org/10.5281/zenodo.20550751