try_cartridge: rehearse a policy without betting your reasoning on it

[compusophy]

#28 shipped with a flaw, and this fixes it rather than adding a new number for its own sake.

the flaw

The corpus could only be fed by a swap. So the only way to learn whether a program ran was to make it the policy the agent reasons with. Exploring the language meant betting the explorer on every experiment — and a corpus that can only be fed by committing to every candidate will never reach the volume item 10 needs.

the fix

Cognition::try_policy shares swap_policy's entire front half through a private prepare — parse, rehearse against a copy of the live memory, build the corpus sample — and stops before the install. Command::TryCartridge, a try_cartridge tool, and a rehearse only button expose it.

The system prompt now says to rehearse first, and says why: a refusal from try_cartridge costs nothing, and the attempt lands in the corpus either way.

the eval that matters

A candidate that rehearses clean in a trial produces the same Sample as swapping it. If the two doors could disagree, "rehearse first" would be advice that does not pay — and sharing the front half is what guarantees they cannot. That property is pinned, not assumed.

The other two: a trial installs nothing (the running policy still answers afterwards — a trial that quietly installed would be a swap with a friendlier name), and a trial of a trapping module leaks nothing into the live store.

3 new evals, 34 in tests/cognitive_wiring.rs.

on where this came from

It was not on the board. It came from asking what item 10 actually needs and finding the answer blocked by a design decision made one item earlier. That is the useful kind of board item — the one the previous item created.

bash ci/run_tests.sh → unit tests + 20 suites + clippy; cargo check --lib --bins --target wasm32-unknown-unknown clean. The try_cartridge definition is covered by the prompt-drift guard added in #25.

smoke-preview will be red for the usual Vercel Deployment Protection reason.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
vanish	Ready	Preview	Aug 27, 2026 12:54am

[compusophy]

live on the preview of c54c758

🧪 rehearsal passed: "rehearsal: a prompt" → "[v2] rehearsal: a prompt" (writes last_answer, last_prompt) — nothing was installed
📚 corpus: 1 program(s), 1 verified, 0 refused — most-emitted: local.get×32, i32.const×16, i64.const×15, i32.add×12, local.set×11

🧪 rehearsal failed: it failed on a prompt: cartridge 'reasoner': cartridge trapped: memory access out of bounds: 4 byte(s) at 2000000000 exceed the 1048576-byte memory — nothing was installed
📚 corpus: 2 program(s), 1 verified, 1 refused — most-emitted: local.get×32, …

Then a full page reload:

📚 corpus restored: 2 program(s), 1 verified
🧠 reasoning policy 'reasoner' up (reference v1)

That last line is the proof. Two programs learned about, nothing installed, nothing persisted as a swap — a trial that quietly installed would have come back as v2. And a broken candidate is a result here, not an incident.