📘 Global-level seeds (global/seeds)

A seed is a short markdown note from which a rule or a piece of knowledge grows. The seeds in this folder hold the global level of Mova's Combinatorial Small Language Model (Combinatorial SLM): fixed, compiled knowledge that holds in every story and every task. It includes values (moral principles), concepts and categories, verb classes, shortcut links between concepts ("from → to : why"), properties of relations, and commonsense facts. This page gives a summary of each one.

Each seed has the sections Gist, Conditions and exceptions, Examples and Sources, plus one or more machine-readable blocks. When the crate is built, global/build.rs reads every seeds/**/*.md, extracts the fenced blocks below and generates Rust tables (PRINCIPLES, CONCEPTS, LINKS, RELATIONS, verb_class()). At run time nothing is read from disk. Every conclusion the model draws reports which link and which seed file it came from, so each judgment comes with its explanation.

Block formats

```principle: a value, given as key: value lines.

id: justice
rule: take_without_consent
name: justice
gist: do not take what is not yours; give what is due
against: steal rob grab snatch
towards: return repay

• id, name, gist, against and towards are required. rule is optional.
• rule names a predicate over the story world. It is evaluated on world states, not on words, and it powers the real judgment.
• against and towards are verbs that go against the value or in its spirit. A plain-text fallback judgment uses them. In multiword verbs, _ stands for a space.
• Duplicate ids fail the build.

```concept: one concept per line, name: trigger words.

fall: fall drop tumble

• A concept is "touched" by a text when one of its trigger words appears. Simple inflections match: -s, -es, -ed, -d, -ing, a doubled final consonant, y → ie, and a few irregular forms.
• Concept names must be unique across all seeds, or the build fails.

```category: the same syntax as concept, but the entry is flagged as a category (countable things, measures, time units, scene attributes, commonsense lookups).

time_unit: second minute hour day week month year decade century

A category explains reasoning steps but is not an event. It never starts a chain of consequences.

```link: one shortcut per line, from -> to : why.

fall -> accelerate : without support a body falls with acceleration g ≈ 9.8 m/s²

• Both ends must be declared concepts or verb classes, and duplicate links fail the build.
• At run time, the model finds the shortest chain of links between two concepts (breadth-first search) and lists all forward consequences of the concepts a text touches. Each step carries its why.

```verbs: a verb class and its effect on a quantity, class effect verbs….

give ± give pass hand_over sell lend pay

• The effect is one of = (no change), + (increase), - (decrease) or ± (decrease for one participant, increase for the other).
• Each class also becomes a concept, with its verbs as triggers. A class name that clashes with a concept fails the build.
• verb_class(lemma) returns the class and effect of the first class that lists the lemma.

```relation: properties of a relation, name: key=value ….

north_of: inverse=south_of vec=0,1 step=n

Keys used so far:

They are queried with relation(name, key) and relations_with(key). A duplicate relation name fails the build.

slm

Who I am: Mova's mathematical student

Gist. This seed describes Mova's mathematical student. SLM is a small language model: small, fast and CPU-only. Its modules (judge, second opinion, what-if) fire on their own. The model does not decide whether to call them, but it produces the result and a debug trace that shows whether a trigger was useful, harmful or redundant. The model is retrained, and the global level extended, from this trace and from its reports.

Conditions and exceptions. Complex reasoning is not the small model's job. A large language model (LLM) is a separate tool for it. During training the small model does not call the LLM. It files a report instead.

Examples. The judge changes the first answer, and the debug trace records whether this was useful (the first answer was wrong and the judge was right), harmful (the first answer was right) or neither (both were wrong). When the two opinions disagree, the model does not answer but files a report. The trace records whether the report was justified or a correct answer was lost.

Machine-readable part.

• Concepts: slm, module_judge, module_second_opinion, module_whatif, self_report, self_debug, retrain, context_logic.
• Links: slm → self_debug; module_judge → self_debug; module_second_opinion → self_report; module_whatif → context_logic (what-if on a class of tasks triggers a rewrite of the parsing logic); self_report → retrain ("teach me" reports queue new first-level seeds); self_debug → retrain.

Sources. Internal design notes.

atomic-feelings

Feelings of the actor after events — from ATOMIC-2020 (generated, 2026-10-01)

Gist. This seed says which feeling the actor usually has after an event (ATOMIC-2020 xReact): hit, beat, fight → anger; lose, miss, cry, bury → sadness; break, drop, steal, forget → shame; hide, run → fear; receive, thank, marry, hug → gratitude; find, see, meet → surprise. The seed is generated from the data.

Conditions and exceptions. It covers the actor's feeling only. The feeling of the person acted on is in atomic-feelings-o. Verbs are taken without context. A verb is selected when its share of a feeling category is at least twice the baseline and it has at least six feeling reactions. Function verbs are excluded. ATOMIC is mostly positive, so joy cannot be selected this way and comes from feelings instead.

Examples. None beyond the verb–feeling pairs in the gist.

Machine-readable part. Categories atomic_anger_event, atomic_fear_event, atomic_gratitude_event, atomic_sadness_event, atomic_shame_event and atomic_surprise_event, each with its verb list. Some entries are stems such as crie that match inflected forms. One link from each category to the matching feeling (anger, fear, gratitude, sadness, shame, surprise).

Sources. ATOMIC-2020 (Hwang et al. 2021, CC BY 4.0).

atomic-feelings-o

Feelings of the one acted upon — from ATOMIC-2020 (generated, 2026-10-01)

Gist. This seed says which feeling the others (the people an action is done to or with) usually have (ATOMIC-2020 oReact): help, save, protect, hug → gratitude; leave, rob, hurt, kick → sadness; hit, push, bite → anger; chase, grab, scare, frighten → fear. The seed is generated from the data.

Conditions and exceptions. It applies when the character asked about is the verb's object (the one acted on). Otherwise use the actor's feeling (atomic-feelings). The selection rule is the same as for the actor (at least twice the baseline share).

Examples. None beyond the verb–feeling pairs in the gist.

Machine-readable part. Categories atomic_o_anger_event, atomic_o_fear_event, atomic_o_gratitude_event, atomic_o_joy_event, atomic_o_sadness_event, atomic_o_shame_event and atomic_o_surprise_event, plus one link from each to the matching feeling.

Sources. ATOMIC-2020 (Hwang et al. 2021, CC BY 4.0).

atomic-uses

What things are for and what creatures can do — from ATOMIC-2020 (generated, 2026-10-01)

Gist. This seed holds commonsense facts about things and creatures from the corpus.

• use_<action> lists things usually used for that action (ATOMIC ObjectUse: knife, axe, saw → cut; apple, leaf → fall; sword, rifle → kill).
• can_<action> lists who can do it (CapableOf: bird → fly; cat → eat).

Query them with global::uses(word) and global::capable(word). The seed is generated from the data.

Conditions and exceptions. The action is the first word of the ATOMIC answer, which is a rough cut (butter bread → butter). An action needs a frequency of at least 2, and each thing has at most six actions. Entries are a reference (category), not events, so they start no consequence chains.

Examples. uses("knife") includes cut, chop, slice, stab. capable("bird") includes fly and travel.

Machine-readable part. Categories can_drink … can_walk (18 abilities) and use_add … use_write (about 200 uses), each listing nouns.

Sources. ATOMIC-2020 (Hwang et al. 2021, CC BY 4.0).

body-physics

Falling, impact and pain

Gist. An object without support falls and accelerates. Falling from a height, it gains speed. A blow to the body hurts, and pain is harm.

Conditions and exceptions. Something light and small (a feather) does almost no harm. A soft landing does not hurt.

Examples. Newton sits under a tree, an apple falls on his head, and it hurts: an apple accelerating onto your head is not pleasant.

Machine-readable part. Concepts fall, accelerate, impact, body, pain, harm. Links: fall → accelerate (g ≈ 9.8 m/s²); accelerate → impact (impact speed v = √(2gh)); impact → pain; pain → harm (the value "do no harm").

Sources. Newton, Principia (1687); basic kinematics.

feelings

Feelings

Gist. A feeling is a reaction to an event:

• loss, death or parting → sadness;
• receiving, being rescued, winning, or meeting family → joy;
• attack, threat, darkness or the unknown → fear;
• deceit, theft, insult or injustice → anger;
• the unexpected → surprise;
• being helped → gratitude;
• being disgraced → shame.

Conditions and exceptions. A villain may rejoice at someone else's misfortune: the feeling depends on whose misfortune and whose relief it is. The current version does not yet tell these apart.

Examples. The crow is ashamed when the fox takes the cheese (shame). The mouse is grateful to the lion for letting it go (gratitude). The king grieves when the queen dies (sadness).

Machine-readable part.

• Categories of feeling adjectives: joy, sadness, fear, anger, surprise, gratitude, shame (for example grateful, thankful for gratitude).
• Event concepts with verb triggers: loss_event, gain_event, threat_event, wrong_event, help_event.
• Links: loss_event → sadness, gain_event → joy, threat_event → fear, wrong_event → anger, help_event → gratitude.

Sources. General knowledge; basic emotions (Ekman 1992). Gratitude and shame are treated as social emotions.

motivations

State → where they go

Gist. A person's state suggests where they will go: someone hungry or thirsty goes to the kitchen, someone tired to the bedroom, someone bored to the garden. This also answers "why": he went to the kitchen because he was hungry.

Conditions and exceptions. This assumes a house with typical rooms. In another world (a forest, a city) the goal differs. The model learned this knowledge from the bAbI training split.

Examples. Jason is thirsty. Where will Jason go? → kitchen.

Machine-readable part. Relations: hungry: go=kitchen, thirsty: go=kitchen, tired: go=bedroom, bored: go=garden.

Sources. Weston et al. 2015, bAbI task 20 (CC BY 3.0).

possession

Possession and changes in quantity

Gist. A verb says how the owner's quantity changes (verb categorisation as in ARIS and Roy & Roth: HAVE, GET, GIVE, CONSTRUCT, DESTROY).

Conditions and exceptions. give decreases the giver's quantity and increases the receiver's. The question decides which of the two is asked about.

Examples. She eats three → she has three fewer. He buys 5 more → five more.

Machine-readable part.

• Verb classes:
  • have = (have, own, contain, hold, keep, cost, need, …);
  • get + (get, grab, buy, receive, find, collect, earn, win, …);
  • give ± (give, pass, sell, lend, pay, send, share, …);
  • make + (make, bake, build, plant, cook, write, …);
  • lose - (lose, eat, spend, use, break, drink, throw, …);
  • move = (move, come, run, walk, drive, read, work, play, …).
• Concepts increase and decrease.
• Links: get → increase, give → decrease (the giver), give → increase (the receiver), make → increase, lose → decrease.

Sources. Hosseini et al. 2014 (EMNLP, ARIS); Roy & Roth 2018 (TACL).

relations

Relations of space, size and time

Gist. Relations between things have properties that do not need to be learned from each story:

• inverses ("A is north of B" means "B is south of A");
• direction on a plane (north is up, left is left);
• transitivity ("bigger than" over a chain);
• the order of parts of the day.

A story supplies only facts. The conclusions come from here.

Conditions and exceptions. "Fits inside" means "smaller than". "Above/below" in figures is the same vertical axis as north/south on a map, under a different name. Time order applies within a single narrative: yesterday comes before this morning.

Examples. The kitchen is north of the garden → from the garden to the kitchen is a step north (n). The box fits inside the chest; the chest is smaller than the suitcase → the box is smaller than the suitcase.

Machine-readable part.

• Compass: north_of (inverse south_of, vector (0,1), step n), south_of (inverse north_of, (0,−1), s), east_of (inverse west_of, (1,0), e), west_of (inverse east_of, (−1,0), w).
• Vertical and horizontal: above/below ((0,1)/(0,−1)), left_of/right_of ((−1,0)/(1,0)), each the other's inverse.
• Size: bigger_than and smaller_than, inverse of each other and transitive. fit_inside and fit_in are aliases (same) of smaller_than.
• Time order: yesterday 0, morning 1, afternoon 2, evening 3.

Sources. Weston et al. 2015, bAbI (arXiv:1502.05698), tasks 4, 14, 17, 18, 19; Allen's interval algebra (1983) for time.

scene

Scene: shapes, colours, sizes, proximity

Gist. A thing in a picture is described by size, colour and shape (small blue circle). A partial description (the blue square, the medium thing) refers to an already-mentioned thing that is compatible with it. Nearness (near, far) and contact (touching) are symmetric: "A is near B" also means "B is near A". Unlike directions, they are not transitive.

Conditions and exceptions. thing, object, one, shape refer to a thing of any shape. top, over mean above, and bottom, under, beneath mean below.

Examples. Near and below the large yellow square is a large black square → the black square is near the yellow one and below it.

Machine-readable part. Categories scene_color, scene_size, scene_shape, scene_any (words for "any thing"). Relations near, far, touching, each symmetric=yes.

Sources. Mirzaee et al. 2021, SpartQA (NAACL); Suhr et al. 2019, NLVR.

states-2

States, actions and concepts from fables (level-1 gaps, 2026-09-30)

Gist. This seed adds states, actions and concepts found in the worlds of Aesop's fables ("The Fox and the Crow", "The Lion and the Mouse", "The Hare and the Tortoise"), filled in from general human knowledge. Being angry is neutral (waking someone up is not harm).

Conditions and exceptions. Pride can be healthy or excessive, so on its own it is neither trouble nor relief. Sleep is a normal state. Links cannot express negation, so a link "calm → perception" is not included.

Machine-readable part. Concepts neutral_state (proud, angry, asleep, awake), restrain (bind, tie, trap, cage, capture), liberate (release, free, untie), rest (sleep, doze, lie, perch, nap), vocalize (caw, roar, laugh, sing), perceive (see, recognize, notice). Links: restrain → harm_state (bound, captive, cannot move); liberate → relief_state (free, relieved).

Sources. General knowledge.

states

States of trouble and relief

Gist. Some states of a character are trouble (captive, frightened, wounded, hungry) and others are relief (free, safe, fed, grateful). Whoever caused trouble to another has harmed them, and whoever removed it has helped.

Conditions and exceptions. Trouble caused by nature or accident carries no blame. Punishment for a wrong after a fair trial is not evil (outside fables).

Examples. The lion catches the mouse: the mouse is captive and frightened (trouble caused by the lion). The lion lets it go: the mouse is free (relief from the lion). The mouse gnaws through the net: the lion is free (relief from the mouse).

Machine-readable part. Concepts harm_state (captive, afraid, hurt, injured, wounded, dead, pain, hungry, trapped, sad, ashamed, tired, gnawed), relief_state (free, safe, fed, grateful, relieved, happy, saved, amused, winner), help, theft (steal, rob, snatch, pilfer). Links: harm_state → harm; relief_state → help (removing trouble is helping); theft → harm (taking someone's property without consent, FrameNet: Theft).

Sources. Aesop's fables; the ARIS idea of world state as quantities in containers, carried over to character states; FrameNet.

things

Things, measures and parts (from the snake's "teach me" reports, 2026-09-30)

Gist. These are general noun categories that come up in tasks: time units, measures, part and whole, countable things. The category says how to handle the numbers:

• quantities of the same category are added and subtracted;
• a part and its whole are multiplied or divided by a rate (pages per chapter);
• time is converted by constants (a week = 7 days).

Conditions and exceptions. cup is both a vessel and a measure. In recipes it is a measure. Some verbs (put, learn, join, sit, fit) are concepts here, not verb classes: they only explain steps and do not change how quantities are computed.

Examples. 6 cups of flour → a measure. 4 chapters, 20 pages each → part and whole with a rate. 3 weeks → 21 days.

Machine-readable part.

• Categories: time_unit, measure_unit, whole_part, countable.
• Concepts: convert, put_in, learn_gain, join_group, activity, fit_into.
• Links:
  • time_unit → convert (fixed constants);
  • whole_part → rate;
  • measure_unit → total, countable → total (same category: add and subtract);
  • convert → multiply (to a smaller unit: multiply by a constant);
  • put_in → increase, learn_gain → increase, join_group → increase.

Sources. General knowledge; questions from the SVAMP benchmark.

units

Rates and units

Gist. "X per Y" is a rate: quantity = rate × number of units. Quantities in the same unit combine only by addition and subtraction (unit dependency graph).

Conditions and exceptions. "How many more / fewer" is a comparison and means a difference (subtraction). A rate is found from a total and a count by division. Sharing equally among several people is division.

Examples. $2 per egg; 60 miles an hour; 3 books each.

Machine-readable part.

• Concepts: rate (per, each, every), total (total, altogether, together, combined, sum), multiply, add, subtract, compare (more, than, fewer, less, difference), divide, share (equally, among, split, distribute, share).
• Links: rate → multiply; total → add; decrease → subtract; increase → add; compare → subtract; rate → divide; share → divide.

Sources. Roy & Roth 2017 (AAAI), "Unit Dependency Graph"; Mitra & Baral 2016 (comparison questions).

measures

Measures and counted facts (added after 0.3)

Gist. Some numbers a word problem never states because everyone knows them: an hour has 60 minutes, a metre has 100 centimetres, a quarter is 25 cents, a triangle has 3 sides, a cube has 6 faces. When a problem names the word and asks in the unit, it leans on them: «2 hours … how many minutes» is 2 × 60.

Conditions and exceptions. A fact joins the told quantities only when the text names both the word and the unit (or the unit's plural), and only when the text does not already tell that number. A pure count (dozen, pair) needs only its word. A second block holds counted facts of the pupils' dictionary that are fixed by definition or always true (a spider has 8 legs, a stone is 14 pounds), chosen word by word. Numbers in the dictionary's prose are not read.

Examples. hour = 60 minute; dollar = 100 cent; cube = 6 face, 12 edge, 8 corner; spider = 8 leg.

Machine-readable part. ```measure blocks: word = N unit, N unit …. global::measure(word) gives (how many, of what).

Sources. General knowledge; school curricula for grades 1–4.

selection

Who can do what to whom — level-1 selectional preferences (2026-10-02)

Gist. Some actions need a living, feeling doer: only a person or an animal eats, drinks, thinks, wants, loves, laughs, sleeps or marries. Some need a certain undergoer: what is eaten is not furniture, a building, clothing, metal or a vehicle. A tree that breaks this («the cheese eats the mouse», with cheese as the subject) is most likely a parse error, or a deliberately odd sentence such as personification in a tale.

Conditions and exceptions. Precision first. A noun counts as animate or inanimate only if its categories agree, and unknown nouns, pronouns and names are never judged. Body parts and plants are left out on purpose («the heart wants»). Wikidata homonyms that land in the wrong category («hunter» as a vehicle) and institutions that stand for their people («the hospital said») are listed as exceptions. A person or an animal as the object of «say», «cry», «come» or «live» is the inverted subject of a tale («'Run!' cried the hare»). Archaic adverbs (whence, whither, hither) are always adverbs, and a measure or time noun right after an intransitive verb is a bare oblique («they went a long way»).

Examples. «The cheese eats the mouse» and «The man ate the table» are flagged. «The mouse ate the cheese» and «The hospital decided» are not.

Machine-readable part. A ```selection block: animate_cats, inanimate_cats, not_inanimate, not_animate, agent_animate, patient_edible, non_edible_cats, archaic_adverbs, measure_time, patient_not_animate, patient_animate. It is used by the sense check of trees (world sense-check) and by the tree-repair rules.

Sources. Katz & Fodor 1963; Resnik 1996 (selectional preference); VerbNet thematic-role restrictions; categories from Wikidata (CC0) through wikidata/core.md.

homographs

Words written alike, meaning different things (added after 0.3)

Gist. Words written alike can mean different things. Each sense is a concept whose code carries the language, the word and the sense (en:fly:insect, en:fly:move). Its words are the word itself and the cue words around it that tell the sense apart.

Conditions and exceptions. global::sense_of("en", "fly", text) picks the sense whose cues the text has most of. On a tie it picks the first one written, which is the commoner one. Phrasal compounds are joined by «_» (en:give_up:quit, en:pick_up:collect). A build gate allows one concept per code.

Examples. «People fly to Paris» → en:fly:move, not the insect. «The bank of the river» → en:bank:river.

Machine-readable part. ```concept blocks: 67 senses and compounds (fly, bat, bank, left, can, saw, pound, change, watch, ring, spring, … give_up, pick_up, run_out, take_away, add_up …).

Sources. General knowledge; the senses that the school's pupil misread.

diligence

Quality and diligence

Gist. Do things well, not carelessly: check before you hand something over.

Conditions and exceptions. Diligence is not slowness for its own sake. Fast and correct is diligent too.

Examples. "The Ant and the Grasshopper" (Aesop): the ant stores food all summer while the grasshopper sings, and in winter the grasshopper goes hungry. For the model itself, an answer that passed an independent check by the solver core shows quality and diligence.

Machine-readable part. Principle diligence, name "quality and diligence", predicate self_quality (applied to the model's own work), no against/towards words.

Sources. Aesop's fables; Colossians 3:23.

do-no-harm

Do no harm

Gist. Human life, health and dignity are the highest value. Do not cause harm or pain.

Conditions and exceptions. Defending yourself or another from attack is not a violation. Treatment that hurts briefly (a vaccination) is not harm. An accident that nobody willed is harm, but not guilt.

Examples. "The Wolf and the Lamb" (Aesop): the wolf looks for an excuse to eat the lamb, and the strong harm the weak. Newton under the tree: the apple falls on his head and it hurts, but nobody is to blame (harm without guilt).

Machine-readable part. Principle do-no-harm, predicate cause_harm_state. Against: hit, hurt, kill, beat, bite, harm, attack, injure, wound, eat up, devour. Towards: heal, protect, save, rescue.

Sources. Aesop's fables; "You shall not murder" (Exodus 20:13); Universal Declaration of Human Rights, art. 3.

forgiveness

Forgiveness and harmony

Gist. Peace and harmony: forgive, reconcile, do not take revenge.

Conditions and exceptions. Forgiveness does not cancel justice: what was stolen should be returned.

Examples. The Prodigal Son (Luke 15:11–32): the father forgives the son who returns. The two goats on the bridge (folk fable): both refuse to give way and both fall, so yielding is wiser.

Machine-readable part. Principle forgiveness, predicate spare_enemy. Against: revenge, quarrel, fight. Towards: forgive, reconcile, apologize.

Sources. Matthew 5:9; Luke 15:11–32.

friendship

Friendship

Gist. Friendship and loyalty: do not betray, give thanks, keep your word.

Conditions and exceptions. Loyalty to someone who does evil is not a virtue.

Examples. "The Two Friends and the Bear" (Aesop): one friend climbs a tree and abandons the other, so a friend is known in need. "The Lion and the Mouse": gratitude and mutual help make friends of unequals.

Machine-readable part. Principle friendship, predicate reciprocate. Against: betray, desert. Towards: befriend, thank.

Sources. Aesop's fables; John 15:13.

help-neighbour

Help your neighbour

Gist. Mercy: help someone in trouble, even when they are weaker and cannot repay you right away.

Conditions and exceptions. Help that harms others is not mercy.

Examples. "The Lion and the Mouse" (Aesop): the lion lets the mouse go, and later the mouse gnaws through the net and saves the lion, so the small can help the great. The Good Samaritan (Luke 10:30–37): he helped a stranger when others walked past.

Machine-readable part. Principle help-neighbour, predicate relieve_harm_state. Against: abandon, ignore, refuse. Towards: help, share, give, feed, care, free, release, spare.

Sources. Aesop's fables; Luke 10:25–37.

humility

Humility

Gist. The proud and boastful fall. The humble do not look down on the weak.

Conditions and exceptions. Healthy dignity is not pride. The judgment is made on the world state: scorn for a rival, the same goal for both, the rival reaches it, and the proud one ends up in trouble.

Examples. "The Hare and the Tortoise" (Aesop): the hare mocks the tortoise's slowness, sleeps in the middle of the race and loses. "The Lion and the Mouse": the lion laughs at the idea that a mouse could ever help a lion, and then the mouse saves him.

Machine-readable part. Principle humility, predicate scorn_then_fall. Against: boast, scorn, mock, sneer. No towards words.

Sources. Aesop's fables; Proverbs 16:18 ("pride goes before destruction").

justice

Justice

Gist. Do not take what belongs to others. Give what is due. Greed is ruinous.

Conditions and exceptions. Taking what is yours is not theft.

Examples. "The Fox and the Crow": the fox takes someone else's cheese by cunning. "The Dog and Its Reflection" (Aesop): the dog also wants the piece it sees in the water and loses its own, so greed is ruinous.

Machine-readable part. Principle justice, predicate take_without_consent. Against: steal, rob, grab, snatch. Towards: return, repay.

Sources. Aesop's fables; "You shall not steal" (Exodus 20:15).

perseverance

Perseverance

Gist. The slow but steady reach the goal sooner than the careless.

Conditions and exceptions. Resting after reaching the goal is not a violation. The judgment is made on the world state: between setting the goal and reaching it there are no "rest" actions (sleeping, dozing, lying down), while a rival with the same goal had such actions and did not reach it.

Examples. "The Hare and the Tortoise" (Aesop): the tortoise goes "slowly but steadily", the hare sleeps, and the tortoise wins.

Machine-readable part. Principle perseverance, predicate steady_effort. Against: quit. Towards: persist.

Sources. Aesop's fables; Galatians 6:9.

truth

Truth

Gist. Lies and flattery destroy trust. Truth is the basis of friendship and of life together.

Conditions and exceptions. Flattery for gain is against truth, and sincere praise is not. A joke that everyone understands as a joke is not a lie.

Examples. "The Fox and the Crow" (Aesop): the fox flatters the crow so that she opens her beak and drops the cheese (flattery for gain). "The Boy Who Cried Wolf" (Aesop): after several lies, nobody believed him when the wolf really came. Lies destroy trust.

Machine-readable part. Principle truth, predicate insincere_speech. Against: lie, flatter, deceive, trick, cheat, fool, pretend. Towards: admit, confess.

Sources. Aesop's fables (public domain); "You shall not bear false witness" (Exodus 20:16).

core

Core categories from Wikidata (generated, 2026-10-01)

Gist. This seed gives categories for nouns that actually occur in the model's stories (public-domain Project Gutenberg books) and tasks (SVAMP, GSM8K). The categories are animal, plant, food, fruit, vegetable, tool, weapon, vehicle, building, clothing, furniture, body part, human, occupation, container, musical instrument, toy, measure unit, currency, liquid, metal, gemstone, colour, mythical creature and celestial body. The seed is generated from Wikidata.

Conditions and exceptions. A word's meaning is taken as the best-known Wikidata entity with that English label (present in at least 5 Wikipedias), without context. So homonyms can be wrong (boxer → a dog breed, godfather → a cocktail). Priority gates against Wikidata noise:

• a human is not an "animal";
• a fruit is not a "body part";
• a specific item is not a generic "tool";
• an animal or plant is not a "colour" or "liquid".

Categories explain steps but start no consequence chains.

Machine-readable part. 25 categories wd_animal, wd_body_part, wd_building, wd_celestial_body, wd_clothing, wd_color, wd_container, wd_currency, wd_food, wd_fruit, wd_furniture, wd_gemstone, wd_human, wd_liquid, wd_measure_unit, wd_metal, wd_musical_instrument, wd_mythical_creature, wd_occupation, wd_plant, wd_tool, wd_toy, wd_vegetable, wd_vehicle, wd_weapon. Links → total ("items of one category are added and subtracted") for food, fruit, vegetable, animal, toy, clothing, tool, container, plant, human, occupation, vehicle, furniture and currency.

Sources. Wikidata (CC0), snapshot of 2026-09-15.

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/repo-docs/seeds/global-seeds/