RLE bench design — N × ticks for harness QA → stats

[jkbennitt]

Docs-only memo for how we get from a working harness to a number we can defend. No code, deps, defaults, or matrix runs. Do not merge until reviewed.

What

Adds docs/bench-design-n-ticks.md plus a one-line pointer from the README live-board paragraph.

Locked path (do not conflate)

1. Harness QA — N=1, 10 ticks (Jason lock: software/harness validation first)
2. Variance pilot — N=8–10 seeds, 10 ticks, 1–2 keeper cells → estimate σ
3. Horizon sweep — one cheap cell, N=3 × {10, 25, 50, 100}; shortest horizon that is not mostly censored
4. Publishable compare — only then n ≈ 16 σ²/δ² per cell (80% power, α=0.05); prefer paired same-seed sets

Current scoring-1.2 board is almost all seed-42 N=1: no within-cell σ. Significance without σ is theater. Flash cells ~$0.02–$0.90 — do not incinerate a full matrix before σ + horizon.

Gate fold (2026-09-07)

• Val N=1 (seed 42, opencode × deepseek-v4-flash): 0.7073; check-zone + work-priority schema proved; research/tend/stockpile-delete unexercised; pin f818246e @ 41fd015; ~$0.025 console
• Gates2 N=1 (seed 314159 + short research/tend bias): 0.6803; work-priority 6× pass; research 4× unavailable rejects (no bench/prereqs); doctor+patient 9× tend pass; farm already-covered and stockpile-delete not exercised; ~$0.025 console

Out of scope

No CLI/default changes. No matrix. No XAI.