Kryptos K4 at its Unicity Distance: A Homophonic Reframe and a Model-Robust Under-Determination Result

Abstract

Kryptos K4 is the 97-character fourth and only unsolved passage of Jim Sanborn's 1990 sculpture at CIA Headquarters. Using only the public ciphertext, the four released cribs (EAST, NORTHEAST, BERLIN, CLOCK), and the solved sister passages K1 through K3, we conduct a systematic deterministic cryptanalysis (125 experiments, indexed through 128; 86 with logged verdicts) and reach a sharp structural characterization of K4, short of a decryption.

Two proven facts drive the analysis. First, K4 is positional and length-preserving: the four cribs sit at fixed contiguous positions, so there is no net transposition. Second, its free positions are frequency-flattened to 4.33 bits/char, well beyond any single alphabet. Under the standard bijective-alphabet model these two facts collide: no short key with a simple selector is simultaneously crib-consistent, flattening, and English. We call this incompatibility the squeeze and establish it analytically and by a CP-SAT falsification over the simple-selector family.

The squeeze is an artifact of the bijectivity assumption. A homophonic chart model needs only two charts to satisfy the cribs, and two homophonic charts already flatten English past K4's 4.33 bits/char; two charts thus meet both facts where the bijective model needs at least four, consistent with a hand-executed chart cipher.

Our central claim is then deliberately model-agnostic. We prove a model-robust under-determination result that holds identically for the bijective (three-or-more-chart) and the homophonic (two-chart) model classes, with three tiers of strength. The 73 free positions split into 30 selector-locked and 43 prior-governed positions. For the 30, the cribs provably leave exactly a per-position coin flip (consensus 0.50; the bijective leg at most one in three): an exact result. For the 43, the cribs are silent by construction (their ciphertext is crib-absent); the substantive claim, that no deterministic language prior resolves them either, is strongly evidenced (heuristic search and two priors logged inconclusive) rather than proven.

No deterministic selector lever we tested moves the locked set: not a chart relation, clock, autokey, engraving-grid selector, character-level prior, nor an internally-bootstrapped (Linear-B-style) hypothesized crib. K4's free positions are therefore under-determined by the public, decipherment-legal inputs we consider (its ciphertext, four cribs, and deterministic character-level priors), and relative to the chart-model family we adopt, pending exactly one external fact: a measured per-position selector or a fifth positional crib. We are precise about this: a unique intended plaintext still exists under Sanborn's full key, so we make no claim of unsolvability. We give the contributions, the falsified hypotheses, the limits, and the methodology, including an independent re-derivation protocol that corrected a community-propagated error about the engraving layout.

This is the formal write-up behind the essay The Answer In The Box; the cipher library and full experiment harness are in the companion repository.