>> Light Agents · AI Swarms>> Light Agents · Simple mode story
Five identical learners, raised on different curricula, bid for grid tasks in a market on NEAR testnet. The contract checks every answer against a hidden hash. No pretrained weights.Five robot kids start out exactly the same, then grow up practising different puzzles. They guess the rule, bid pocket money for each job, and a fair referee on NEAR checks every answer. They learn from scratch, with no giant borrowed brain.
>> Results>> What happened
Real numbers from the frozen experiments. Passed, failed and null stay separate.Real results, told plainly. Wins and misses stay separate.
Rows are subjects. Columns are task families. Hidden-test solves out of 16.Each row is a robot, each column a kind of puzzle. The number is how many of 16 it solved. Geo got all 16 shape puzzles and 0 of 16 object puzzles.
Paired bootstrap, 95% interval. Blue clears (lower bound above 0).Each line is a fair head-to-head. Blue means the win is real, not luck.
>> Start here>> Start here
Every bordered card below is interactive. The blue strip tells you what to press. Explanations sit in Learn more.Every box below is something you can poke. The blue strip tells you what to press.
Generate a grid, then watch the five learners search.Pick a puzzle and watch five robots try moves until one fits.
Specialists cleared. Society versus best member did not.The specialists really specialise. Teaming up hasn’t beaten the best single robot yet.
Testnet NEAR verifies the grid. A dry quote prices USDC on Arbitrum.A fair referee checks the answer and pays whoever is right, in practice money. We also asked for a price to pay in another coin, without spending anything.
Target envelope: about 1–30 million parameters, CPU, no pretrained weights, cost per task, ARC-AGI pass@2 when a public score is claimed. This page is under that envelope. No ARC-AGI score is claimed. Earlier work: a synthetic generalisation gate passed, a real-task confirmation failed, and a deeper-search gate failed its control.The grown-up goal: robots small enough to run on a normal computer, learning from scratch. We don’t claim a score on the official ARC puzzle test. Before this project, one earlier check passed and two didn’t, and we say so.
Contract, paper and source of every number on this page.
>> Playground>> Playground simple
Everything runs in your browser. No server, no language model.Everything happens right here in your browser. No big server, no chatbot.
Draw a synthetic grid in the browser, or load a frozen example.Make a new picture puzzle, or load one we saved.
Family
Seed
Curiosity 1.5
› press Run society to start five searchestap Run society and watch the robots think
One architecture. The curriculum is the upbringing.Same starting brain. Different practice makes a different robot.
Search ranks 16 operations by the prior plus a curiosity bonus, keeps a beam of 6, and accepts the first program that fits the demos. Residuals that would name a one-step rule are off for these five. The broader design uses a larger grid language and a deeper search.Each robot has 16 moves it can make, like flip, turn and recolour. It tries a handful, keeps its 6 best guesses, and stops at the first one that turns the examples into the right pictures. The real answer stays hidden until the end.
>> Society>> Society simple
Hidden-test solves by family, and the contrasts that cleared.Each robot practised a different kind of puzzle. The specialists really specialise. Teaming up hasn’t beaten the best single robot yet.
>> Subjects>> Meet the robots
Same matrix as the overview. Click a cell to load that family.Each row is a robot, each column a kind of puzzle. The number is how many of 16 it solved. Geo got all 16 shape puzzles and 0 of 16 object puzzles.
Paired bootstrap, 95% interval.Each line is a fair head-to-head. Blue means the win is real, not luck.
Geometry, colour, and object train on one family. Mixed trains on the whole catalogue and keeps a curiosity archive. Noisy trains on 200 rows with a quarter of the labels moved. The generalist is held out of the society.Geo practised shapes, Colour practised colours, and Object practised things. Mixed practised a bit of everything and stayed extra curious. Noisy grew up on 200 messy examples. One more robot that practised everything is kept out of the team.
>> Market>> Society & Market simple
Escrow, stakes, slashes and payouts for the 24-round payment trial, replayed step by step.Watch all 24 puzzles of the payment test, one step at a time. 23 of the 24 got solved.
>> How the game works
Someone posts a puzzle and puts the reward in a jar. The referee keeps a secret fingerprint of the right answer.
Robots that think they can solve it put pocket money down. The referee picks the bid that is most trusted for a fair price.
The chosen robot hands in its answer in a sealed envelope first, so nobody can copy it. Then it opens the envelope.
Right answer: the robot gets the reward and can share some with a friend who helped. Wrong answer: it loses its pocket money and the puzzle reopens.
Offline twin of the contract. Escrow, stakes, slashes, payouts.A practice copy of the referee. Watch the money jar, the bids and the trust scores move.
Reputation per learner. Follows the replay.How much the referee trusts each robot, round by round. Being right makes it go up.
Solver and helper pay, up to the current step.Mixed was paid for 22 puzzles and Geo for 1. The others weren’t paid in this test: Mixed could solve their puzzles too, and its bids won.
The 24 rounds are the frozen upbringing trial. Only the deploy and the one posted task exist on testnet. The live switch reads those views. It does not send a transaction.These 24 rounds are a saved practice game. On the real NEAR test network there is one referee contract and one posted puzzle. The Live switch only looks; it never spends anything.
One row per round. Blue when someone was paid; the current round is highlighted.One row per puzzle: who bid, who answered, who got paid.
| Trial | Family | Program | Bids | Reveals | Paid | Amount |
|---|---|---|---|---|---|---|
| r1 | geo | flip_h | geo, mixed, noisy | mixed | mixed | 766,500 |
| r2 | colour | invert_fg | colour, mixed, noisy | mixed | mixed | 783,900 |
| r3 | object | gravity_down | object, mixed, noisy | mixed | mixed | 794,000 |
| r4 | cross | rot180 · swap_top2 | — | — | open | — |
| r5 | trap | paint_largest | object, mixed, noisy | mixed | mixed | 772,600 |
| r6 | geo | transpose | geo, colour, mixed, noisy | mixed | mixed | 780,400 |
| r7 | colour | cycle_fg | colour, mixed, noisy | mixed | mixed | 823,600 |
| r8 | object | gravity_right | colour, object, mixed, noisy | mixed | mixed | 833,700 |
| r9 | cross | rot90 · fg_to_one | colour, mixed, noisy | mixed | mixed | 843,800 |
| r10 | trap | paint_largest | object, mixed, noisy | mixed | mixed | 820,400 |
| r11 | geo | rot270 | geo, mixed | mixed | mixed | 848,400 |
| r12 | cross | flip_h · cycle_fg | mixed, noisy | mixed | mixed | 864,200 |
| r13 | colour | fg_to_one | colour, mixed, noisy | mixed | mixed | 883,700 |
| r14 | object | gravity_down | object, mixed, noisy | mixed | mixed | 893,800 |
| r15 | trap | paint_largest | object, mixed, noisy | mixed | mixed | 878,500 |
| r16 | cross | flip_v · fg_to_one | mixed, noisy | mixed | mixed | 907,900 |
| r17 | geo | rot180 | geo, mixed, noisy | mixed | mixed | 920,000 |
| r18 | trap | paint_largest | object, mixed, noisy | mixed | mixed | 891,800 |
| r19 | colour | invert_fg | colour, mixed, noisy | mixed | mixed | 920,000 |
| r20 | trap | paint_largest | object, mixed, noisy | mixed | mixed | 920,000 |
| r21 | geo | rot90 | geo | geo | geo | 777,500 |
| r22 | colour | invert_fg | colour, mixed, noisy | mixed | mixed | 920,000 |
| r23 | object | keep_largest | object, mixed, noisy | mixed | mixed | 920,000 |
| r24 | cross | transpose · invert_fg | mixed | mixed | mixed | 920,000 |
>> NEAR Intents>> Getting paid simple
Solvers compete to fill it. The contract escrows testnet NEAR and checks the grid hash; settlement is quoted through NEAR Intents 1Click.The referee pays in practice money. We also asked what 0.1 of that money is worth in another coin. The ask is a dry run: no money moves.
The on-chain payout is testnet NEAR. A cross-chain helper payment is the dry 1Click quote below.
Request · POST /v0/quote
{
"dry": true,
"swapType": "EXACT_INPUT",
"slippageTolerance": 100,
"originAsset": "nep141:wrap.near",
"depositType": "INTENTS",
"destinationAsset": "nep141:arb-0xaf88d065e77c8cc2239327c5edb3a432268e5831.omft.near",
"amount": "100000000000000000000000",
"recipient": "0x2527D02599Ba641c19FEa793cD0F167589a0f10D",
"recipientType": "DESTINATION_CHAIN",
"refundTo": "lightswarm16771.testnet",
"refundType": "INTENTS"
}The contract escrows testnet NEAR and checks the grid hash. The quote calls the public 1Click API with dry set to true: 0.1 wNEAR from an Intents balance toward USDC on Arbitrum. No deposit address is issued. Chain Signatures are roadmap.The contract holds practice money and checks the answer. The price quote asks a public service, with dry set to true, about 0.1 NEAR toward USDC. No deposit address is issued, and no money moves.
The two real transactions behind this page.
>> Spec>> What is built simple
Outer design of the system. This repository’s rows are the ones you can run.The plan for the whole system. The rows you can run are the ones marked built.
Each row is built here, a dry run, or roadmap.Each row is something we built, something we only asked the price of, or something still ahead.
| Component | Status | What this repository does |
|---|---|---|
| Neutral recursive core | Roadmap | Feed-forward prior in this demo. A recursive core is later. |
| Curiosity and novelty | Built | Bonus when a one-step behaviour is new. Frozen draw: no change in the accepted program. |
| Learned program prior | Built | 326,800 parameters. Three hidden layers. Same initialization for all five. |
| Program search | Built | Beam over 16 operations. The broader design uses a larger language and a deeper search. |
| Skill library | Roadmap | Keep a skill only when it replays on a fresh grid. |
| Upbringings | Built | Geometry, colour, objects, mixed with curiosity, and a small noisy set. |
| Society | Built | One market. It did not beat its best member on the frozen draw. |
| Society of societies | Roadmap | One helper split exists. A society hiring a society does not. |
| Task spec | Built | post_task stores the answer hash and locks the bounty. |
| Bid | Built | place_bid locks a stake. Score is confidence × reputation / price. |
| Commitment | Built | commit_solution stores a hash. Reveal waits for a later block. |
| Verification | Built | reveal_solution checks sha256 of the canonical grid. |
| Hiring | Built | subcontract names one helper, at most half the price, paid after a correct reveal. |
| Intent settlement | Dry-run | 1Click quote, dry true, 0.1 wNEAR toward USDC on Arbitrum. No deposit address. |
| Identity | Built | Sub-accounts and function-call keys. Chain Signatures are roadmap. |
A result counts only after a freeze, matched untrained and shuffled controls, and a paired bootstrap whose lower bound is above 0. Held-out grids stay sealed. The broader system uses a larger grid language and a deeper search than this 16-operation demonstration.A result counts only when the test was frozen first, the comparison is fair, and the win is bigger than luck. Hidden answers stay sealed. This demo uses 16 moves. The bigger plan uses more.