Limitations

What this is not, and what it cannot do. Written because a project that cannot state its own limits cannot be trusted about anything else.

The simulation

• It is a toy economy, not a model of one. Four rules and a cap, tuned so the run shows growth, generations and a squeeze. It is not calibrated to anything and predicts nothing.
• The "notes inheritance" is a design, not a measurement. In this engine, inheritance appears only as a generationBonus multiplier from generation three onward. The research this project draws its premise from measured an effect of notes on a specific task (see docs/ARTICLE.md); we have not measured anything, and the engine does not claim to.
• Numbers are floats, rounded at every step. Rounding is what keeps machines in agreement; it also means the balances are not exact decimal money and never will be.
• Determinism is not fairness. A run that is identical everywhere is still a run whose rules were chosen by one author.

The boundary

• No money moves. No payment rail, no custody. A wallet signs a message that seals a move and nothing else, and the chain supplies randomness and a public record — not a balance. Nothing here can take or send anything.
• Nothing replicates. No scanning, no code copied between machines, no weights duplicated, and no part of the engine can reach a network. "Split" is an array append and a halved number.
• An agent cannot act. It has no tools, no key and no way to influence anything outside the tick. This is a deliberate inversion of the risk the source research describes.
• No prompt, model or inference is involved. The agents are arithmetic, not language models. Calling them "AI agents" would be an overstatement, which is why this repository does not.

Known weaknesses

• The live season rests on one process on one machine. No cluster, no failover, and the day's ledger is a file on that machine's disk. A dead host is a dead season until it is restored.
• *The safety argument is structural and* enforced — but only against the obvious.** Every tracked source file is scanned on every push for networking, subprocess and environment access. Twelve files are allowlisted today — the anchor writers and their memo helper, the daily poster, the live season's chain, store and drand client, the board's files, and the three browser clients that call this site's own API — each with its reason written beside it in tests/boundary.test.ts, so an addition is visible in the diff instead of silent. That catches a fetch or a child_process; it cannot catch an obfuscated one, and it says nothing about what a future dependency might do. A dependency-free engine is the second half of that defence, and it is asserted too: sim/engine.ts imports nothing but its own hash module.
• Nothing is anchored yet, and no third party has replayed a run. The writer and its verifier are built and tested, but no anchor wallet has been funded, so the on-chain list is empty. Until something outside this repository witnesses the output, the guarantee is "the code and the author agree".

What the operator can still do

• Change the rules — every change that moves a golden is visible in the test diff and in the changelog.
• Choose which seeds are published, and which day's chain head is anchored (the writer exists; it stays unused until an anchor wallet is funded).
• Stop the process. There is nothing else to stop: no money and no custody. What lives on disk is the season ledger and the day's board.

What the operator cannot do is rewrite a published run without the change being visible in the repository history — and once the anchors are written, without the chain disagreeing.