Tanagra: Teaching a Learning Mind to Hear the Story Under the Surface
Year: 2026
Type: preprint
Abstract: In the Star Trek: The Next Generation episode "Darmok", the words "Darmok and Jalad at Tanagra" carry a lesson that only living the story reveals. This paper asks the episode's question of a small learning mind: what does it take for it to hear the story under the surface? A number has a surface, its face (its size and its last digit in base ten), and a story under it, its recipe of prime factors. A learner that holds guesses and strikes out every guess a try proves wrong was taught in a booth of numbers, in a booth of sentences and in a village of many lives. What we found: A number's face hides its recipe from a careful mind. A mind holding both readings fell as often as a mind holding only recipes, but ended understanding (judging every kind of thing rightly) in only 0.204 of 1,000 lives, against 0.276. In numbers, a near miss is also a smaller number, so near misses taught the face-reading mind "too small" as well as "missing a part": it understood in 0.841 of lives, against 0.992 for the recipe mind. Opposite parallelism closes the trap. Pairing each near miss with one that keeps the number's size gave 0.992 with no falls, at every face strength tried, from 1 to 16. The counterexample that decides is the working example's surface twin: alike on every surface feature, missing a needed part. It is not necessary, but it is about sixteen times cheaper: random counterexamples reached 0.991 at 32 a life, against 0.992 for 1.95 chosen ones. The point where the trap snaps shut can be called before the run. In a booth of sentences, arithmetic on what the mind believed put it at a face strength of 1.25, and the run found it between 1.2 and 1.3. The half law, and the dangerous direction. Right after a lesson the mind judged every kind rightly exactly when the true rule held more than half its belief, and every such life kept the idea (5,446 of 5,446). Lives whose lesson left only things wrongly believed to fail kept the idea in 0.983; lives left with even one thing wrongly believed to work, in 0.839. Thrift is fragile, and a pile of lessons works like an error-correcting code. Told one lie, the chosen pair fell to 0.400 while 32 random counterexamples held 0.976; against a liar who lies at a steady rate, the pile still won (0.960 against 0.866 at 5%). In a village of minds that live, work, die and leave stories, the stories of the dead hand the young their tools (4.46 tools at age twenty, against 0.86 for newcomers who learned only by living), and the better a village knows its tools, the more its funerals teach the wrong lesson: the share of death stories about someone holding the right tool, killed by bad luck, rose from 0.399 to 0.693 to 0.894. A mind can act rightly without understanding. It took the right order on 4.08 of five ordered pairs while believing on 2.06 of them that order did not matter, and in a world with no orders at all, one unlucky failure left it sure of a false order it never tested again (0.35 a villager by age thirty). For anyone who teaches, or trains a machine, with examples, the booth gives a short routine: show the twin that fails; where lessons can lie, pile them up; and make sure the lesson leaves nothing wrongly believed to work, because the learner's own successes repair nearly all the rest. This is the fifth paper in a series on learning minds, after Question the Substrate, Two Arrows, The Two-Needle Compass and Darmok and the Idea of a Raft, whose ladder its booth reuses unchanged. Its results sit beside established work that got there first by its own road, which the paper names: Winston's near misses, hard negatives and contrast sets in machine learning, contrasting cases in education, shortcut learning, self-confirming equilibrium, Skinner's superstition, error-correcting codes, outcome bias and cultural evolution. The paper says what it adds to each. Files: the paper (PDF), and TANAGRA_code_and_data.zip. The zip holds the paper's markdown source and figures, every script with its printout and saved data, the predictions written before each run (sealed by md5 from the first card's redesigns on), a guess ledger of 192 registered statements with the tool that scores it, and a bench that rebuilt all 33 files of record identically from a fresh copy of the scripts. The booth's cards need D4/d4_ladder.py from the record of Darmok and the Idea of a Raft. A preprint: not peer reviewed, with a limited search of the literature. It comes from a part-time research team that works in the open and gives its tools away, and it was built with Claude (Anthropic) assistance for algorithm and program development. Contact: [email protected]. CC BY 4.0.
Author: Schulz, Matthew
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: green
OpenAlex ID: W7235735726
