You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Closes: the "intelligence" and "recursively improving" parts of the vision (docs/VISION.md in #2).
The gap
The improvement loop's proposer is a seeded search. It tunes one number about itself, its step scale. Nothing that learns takes part, so the loop is recursive only in the weakest sense.
The system
The proposer becomes an interface. It receives the current genome, the margins and the last verdicts, and returns one proposed change. Any proposer can sit behind it: the seeded search, a language model, a Weft cell, an Exo agent.
The gate and the ledger do not change. The theorem run_safe already holds for whatever is plugged in: it puts no condition on the proposer.
First slice
A recorded run in which a language model proposes each change through that interface, with its prompts and replies kept beside the ledger.
Done when
the ledger replays in Lean with no mismatch (python3 verify.py --formal);
the run is compared with the seeded search on the same budget of 64, and the comparison is committed whichever way it comes out;
a refused proposal visibly changes what the model proposes next.
A second slice lets a Weft cell propose.
Rests on
loop.py, abios for the message format.
Boundary
The model proposes and the gate decides. Nothing here gives a model authority to act. The budget is set by a person, and run_bounded bounds the number of changes by it.
Filed by an AI agent (Claude Code) working for the owner of aska-physical.
The first slice of this proposal is done twice in #13. This is what happened and what it leaves open.
What was built.tools/propose.py is the proposer's side of the loop as a command line. show prints the design held, its margins, the radius, the objective and the budget. try GENE=VALUE ... --why "..." proposes one or more genes at once and records the reason. The run file is the whole state, and verify.py replays it and must reproduce the committed ledger. --formal then works every decision out again in Lean from the two designs.
Run 1 (results/model_run.json). A language model, run as a sub-agent with only that output to go on, made 11 proposals in a limit of 40. 10 were taken up and 1 was refused as worse. It reached an objective of 0.919548, which a direct search finds to be the best any design within the limits can score. The seeded search reached 0.919009 in 50 proposals. It probed each gene once, fitted a formula to each moving margin, then planned moves that changed hardware and software together.
Run 2 (results/model_run_2.json). The first run's closing notes were saved word for word in results/model_lessons.md and given to a second instance with no memory of the run. It reached the same objective in 4 proposals, none refused, and left its own notes in results/model_lessons_2.md. Its notes say three proposals cannot work, by its arithmetic. That is not tested.
What this does not show. Two runs of one model are anecdotes. The sub-agents were told not to read the repository and that was not enforced. The plant model has five genes and three moving margins, so it is an easy problem.
One finding that matters for the design. Both runs end with the load-floor margin at 999,215 ppm, 785 from its limit, and can change nothing more. The objective may not fall (loop_never_worse), and nothing in it rewards headroom, so a good proposer drives the design onto a limit and freezes it there. Before a stronger proposer is worth adding, the objective needs a term for headroom, or the loop needs a way to trade objective for room.
Open, in the order I would take them.
A stated protocol and many runs, so that "model against search" means something.
Notes in a form a program can check. Run 1's notes include three fitted formulas. They could be tested against the model automatically before the next proposer is allowed to rely on them.
A Weft cell or an Exo agent as the proposer. This needs a decision from their owners on the message format.
A third run that starts from results/model_lessons_2.md with the reworded rules.
Activity
abu-aska commentedon Nov 17, 2026
The first slice of this proposal is done twice in #13. This is what happened and what it leaves open.
What was built.
tools/propose.pyis the proposer's side of the loop as a command line.showprints the design held, its margins, the radius, the objective and the budget.try GENE=VALUE ... --why "..."proposes one or more genes at once and records the reason. The run file is the whole state, andverify.pyreplays it and must reproduce the committed ledger.--formalthen works every decision out again in Lean from the two designs.Run 1 (
results/model_run.json). A language model, run as a sub-agent with only that output to go on, made 11 proposals in a limit of 40. 10 were taken up and 1 was refused asworse. It reached an objective of 0.919548, which a direct search finds to be the best any design within the limits can score. The seeded search reached 0.919009 in 50 proposals. It probed each gene once, fitted a formula to each moving margin, then planned moves that changed hardware and software together.Run 2 (
results/model_run_2.json). The first run's closing notes were saved word for word inresults/model_lessons.mdand given to a second instance with no memory of the run. It reached the same objective in 4 proposals, none refused, and left its own notes inresults/model_lessons_2.md. Its notes say three proposals cannot work, by its arithmetic. That is not tested.What this does not show. Two runs of one model are anecdotes. The sub-agents were told not to read the repository and that was not enforced. The plant model has five genes and three moving margins, so it is an easy problem.
One finding that matters for the design. Both runs end with the load-floor margin at 999,215 ppm, 785 from its limit, and can change nothing more. The objective may not fall (
loop_never_worse), and nothing in it rewards headroom, so a good proposer drives the design onto a limit and freezes it there. Before a stronger proposer is worth adding, the objective needs a term for headroom, or the loop needs a way to trade objective for room.Open, in the order I would take them.
results/model_lessons_2.mdwith the reworded rules.Generated by Claude Code