item 8d: REASONING_V3, the standing note — a policy worth swapping to

[compusophy]

#25 gave the agent a door. v1 and v2 are demonstrations — one changes nothing, the other prefixes a visible marker so a swap is observable — so "swap your reasoning policy" was an offer with no answer to "to what?".

what v3 claims, and why it is true

The transcript is trimmed at KEEP_MESSAGES and is per-conversation. A cartridge's kv is neither. So v3 gives the agent one line of memory that outlives both — deliberately the smallest thing that is genuinely not available otherwise.

• answer phase: the first CARRY: in the answer, to the end of that line, becomes the standing note. No marker, no change. A bare CARRY: clears it. Capped at STANDING_MAX (400) — uncapped, the note would grow every later prompt for as long as the policy ran.
• prompt phase: the note is prepended and a protocol line is appended. The module cannot edit the system prompt, so it teaches its own contract in the only channel it has: the prompt it is already shaping. A policy that needs the agent to cooperate has to say so itself.

It is also the first reference module that makes the language work rather than demonstrate it — store_get and the packed-i64 unpack path, a prompt assembled from three sources, and a byte search (find). Items 5 and 7 earning their keep.

the honest limit, in the plan rather than implied

v3 is verified to do what it says, not to help. Nothing in this substrate can yet judge whether a shaped prompt was better than the one it replaced — that is what item 9's corpus capture is for, and it is now the thing gating every further claim about policy quality. docs/CARTRIDGE_PLAN.md §12 says so in those words.

drift guard

rustlite cannot be handed a Rust constant, so CARRY_MARKER, STANDING_KEY, STANDING_MAX and STANDING_PROTOCOL are mirrors of literals inside the module. An eval pins them to the source they describe, because nothing else would.

evidence

7 new evals (23 in tests/cognitive_wiring.rs now): the protocol line appended when nothing stands; a marker taking its line and only its line; no marker leaving the note alone; a bare marker clearing it; the cap; the note surviving a simulated reload; and v3 swapping in over v1 with the rehearsal passing and its probe coming back plain.

bash ci/run_tests.sh → unit tests + 20 suites + clippy; cargo check --lib --bins --target wasm32-unknown-unknown clean.

smoke-preview will be red for the usual reason — the preview is behind Vercel Deployment Protection and the run reports EVIDENT-CAUSE FAILURE … the app was never served. Still waiting on the owner-only VERCEL_AUTOMATION_BYPASS_SECRET.

process note

The module was prototyped in a throwaway tests/v3_probe.rs and iterated there — it caught two real bugs (a line-end scan that reused its loop variable to stop and so threw the position away, and an over-greedy carry) before any of it became a crate constant. The probe is deleted in the same commit.

🤖 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 26, 2026 9:05pm

[compusophy]

Live on the preview of 6ce70dc, in an authenticated browser: load v3 → hot-swap policy gives

🧪 rehearsal passed: "rehearsal: a prompt" → "rehearsal: a prompt

(policy: to carry one line into your next prompt, write CARRY: followed by that line in your answer.)"
🔁 cartridge 'reasoner' hot-swapped
🧠 reasoner: reasoning v3 up: standing note
🔁 'reasoner' is now the reasoning policy — the next prompt goes through it

The probe comes back with the protocol line appended and no standing note, which is exactly right for a fresh store — so the candidate really ran both phases before it was installed.

Also worth noting: this run produced no publish diagnostics race, because #27 landed and both verify runs were green so the step never ran.