Card 60 shipped (PR #570) — and the loop ends exactly as its contract says P5_DONE

The time-arrow detector shipped — and this trial proved a point the whole process is built on: even a green full-quality-gate can hide a red one. The attacker ran the strictest lint configuration (every target, every feature flag combined) and caught the new code failing a mandatory style-severity rule that the standard gate invocation never exercises. The code worked perfectly — every mathematical check green — but the repo's strictest gate would have gone red the moment someone ran it with full flags. Fixed in minutes because it was caught before the PR, not after. The defender then verified all five findings resolved (and proved the one remaining lint mess predates this work entirely): 0 conceded.

Then the loop did the most important thing an autonomous system can do: it stopped itself. Feature 4 (card 50, a volatility measure) is contractually forbidden to start until a human ratifies a formula decision — should the production value be the window-average of the per-bar reading? That ruling hasn't arrived. The loop's contract for this exact situation: write the morning handoff, mark the campaign complete-with-a-gate, and end — no improvising, no "helpful" guessing at the formula, no spinning on a blocked resource.

The batch-6 scoreboard: three genuinely novel measurements — "is move size related to bar duration, in any shape?" (Φ², PR #568), "how long do price levels stay underwater?" (visibility horizon, PR #569), "does time have an arrow here?" (clustering mean-gap, PR #570) — each bit-exact-or-better against frozen real-data answer keys, each certified causal/deterministic/restart-proof, each surviving an adversarial trial at 0 conceded, stacked in merge order for the human operator. Thirteen iterations, two plan corrections honestly surfaced, six fail-closed traps sprung before they could poison anything, zero touches of production.