Work

The Two-Needle Compass: Why a Learning Mind Needs a Second Needle, and Now and Then a Look at the Sky

Year: 2026
Type: preprint
Abstract: A learning mind that grades itself is a walker in a mirror maze: every reflection agrees with it. This preprint builds a small voting machine, 64 rotors that vote left or right at each fork of 500 mazes (4,000 forks in all), and gives it two needles: an outer needle, the vote of the moment, and an inner needle, a slow memory of past votes. Then it lets some of the rotors lie, moves true north, and asks what the mind can know from inside, and when it must look outside. Every number is printed by a script that ships with the paper, and 145 predictions were written down before the runs that judged them. What it found: One needle cannot grade itself. When the world misleads all the rotors together, they agree, the vote looks sure, and it is still wrong: among confident letters, wrong 8 times in 100 (0.0818) in a misleading world, against never (0 of 452) in a clear one, and nothing inside the bank can tell the two worlds apart. A second needle fixes the grade only if its mistakes are its own. The inner needle is right every time (1.0000) where the outer needle is right 84 times in 100 (0.8392), and every one of the outer needle's 6,432 wrong letters is flagged by a disagreement. When the world plays the same trick at every visit, the inner needle learns it too (0.8440 against 0.8407). Weight with honey (momentum with damping) buys a steadier letter, not a faster one: at matched noise honey alone follows a move just as fast, but the weighted needle slips back less (0.0095 of forks against 0.0243). Trust the advisers who agree with you, and you trust the yes-men most. Rotors that echo the self score highest (0.8217, against 0.7555 for honest rotors); with half the bank echoing, only 9 of 4,000 forks ever follow a moved north. Weigh advisers by agreement and one yes-man is enough to freeze you: one echo among 64 stops half the forks from following (0.5089 follow), two stop almost all (0.0060). Scoring advisers against a teacher instead follows the move even with 32 echoes (0.9972). A look at the real north now and then is enough, and how often hardly matters: looks every 2nd to 20th visit all free the frozen self (0.9855 to 0.9998 by the end); never looking stays frozen (0.2426), and looking at every visit never moves (0.5005). A frozen self is in the dark, not surprised: the reading goes still (its variance falls to an eighth, 0.1204), and a needle that feels for that darkness frees itself without a teacher (0.9680 follow) while loosening by mistake on only 1 fork-visit in 100,000 when nothing is wrong. From inside, a moved north and turned advisers look the same, so every rule that uses only the self's own data scores about 1 over the two worlds, as a coin toss would (1.0007, 0.9983, 1.0001). Only a look outside wins both. Let the table of advisers' leans say when, and one look at that moment says which: registered and run on a fresh world, 1.9579 of a possible 2 with one look in each world. A family should pass down its routes as words, not as copied certainty: after 24 generations the family that hands down decided letters keeps 0.9920 of its routes exactly right, against 0.1160 for the family that copies its confidences, the regenerative-repeater lesson of 1948 telephone engineering; the copying family loses its founder's least certain turns first, as a model trained on its own output loses its rare cases first. Underneath it all, the machine: tuned to the zeros of the Riemann zeta function, the bank agrees only at the prime powers and its background floor hardly grows (exponent 0.0983, against 0.5109 for random speeds), which is Landau's formula read as a rule for building a voting machine. A bank tuned to the Davenport-Heilbronn function, with zeros off the critical line, shows which parts belong to the whole numbers and which to the primes, and in this machine the Riemann hypothesis reads as the statement that every dial in the zeta bank is a pure dial. The count: of 145 predictions written before their runs (124 in the main steps, 21 in redesigns written after a miss), 122 came true, 18 of them benchmark checks, and 23 missed. Every miss was given a redesign check; 16 of the 21 redesign predictions came true. Every verdict is listed in the paper's Appendix A. Try it yourself: the code zip includes Try the Compass, five small games built from the same machine, each checked against the paper's numbers before anyone plays, and a compass card that scores a player's own rules, or their own AI's, heading by heading against the paper's. Limits: this is a preprint. It has not been peer reviewed, and its search of the literature is limited; where a source was read only in part, the paper says so. Nothing here was measured on a deployed AI system: the links to confirmation bias, sycophancy and model collapse are resemblances in kind inside this machine, not tests of those systems. The work comes from a part-time research team that works in the open and gives its tools away, and was made with Claude (Anthropic) as research partner; responsibility for the work rests with the author. Contact: Matthew "Bear" Schulz, [email protected], ORCID 0009-0000-0956-5743. Related works in this series: Question the Substrate (10.5281/zenodo.24094163), Two Arrows (10.5281/zenodo.24120415), The Anchor and the Hammer (10.5281/zenodo.21450883), The Remote-Piloted Vessel (10.5281/zenodo.21786539), Memory by Subtraction (10.5281/zenodo.21857701) and Taking a Loop Walk Through the Integer Forest (10.5281/zenodo.21449569).
Language: English
Subfield: Philosophy
Open Access status: green
OpenAlex ID: W7226024022
Cites: 6
Cited by: 1
Related to: 10