rsilisted

# Dispatch Ladder: per-phase evaluation The per-phase half of `strategy-recursive-self-improvement` condition 14 as amended 2026-08-12 — read the condition at `origin/main`; it is authoritative and this is its mechanism. The other half is the cross-phase synthesis the session reading the ladder's terminus runs (`dispatch-ladder/SKILL.md` §"The closing cross-phase synthesis"); everything phase-local is here. Spawned by `dispatch-ladder-run` as its own `claude --bg` job at each phase boundary, and once more from `halt()` for a phase the run did not finish evaluating. The driver does not wait: nothing this job does can delay, gate, or change the ladder's disposition. It runs after the phase it evaluates has already ended, so it evaluates an observed result, never a prediction. **Why per phase.** A phase's transcript is small and warm at its own boundary and cold and expensive to recover at the end of a six-hour ladder, so a terminus-only review systematically evaluated best the phases it could still see. And a run that halts — exit 10, 11, 12, 13, 21 — used to record nothing at all, making the most defect-rich runs the ones that produced no review. ## Arguments `/rsi <node-id> <phase> --since <epoch>` - `<node-id>` — the tactic node the ladder is driving. - `<phase>` — the phase that just ended (`align-tactics`, `implement`, `fix`, `conflict`, `review`, `qa`, `main-qa`). - `--since <epoch>` — UTC seconds at which the driver launched that phase. **There is no per-phase s