Method sample

4-atari-hard/2-go-explore.py:217–238

n cells with replacement, p ∝ 1/sqrt(seen+1); DONE excluded. Returns (key, CAPTURE) pairs that freeze the cell's snapshot/score/ trajectory AT SAMPLING TIME. The trajectory walk must use the capture, never the live cell: an earlier result in the same batch may replace

(self, n, rng)

        return max(c.score for k, c in self.cells.items() if k != DONE_KEY)

    def sample(self, n, rng):
        """n cells with replacement, p ∝ 1/sqrt(seen+1); DONE excluded.

        Returns (key, CAPTURE) pairs that freeze the cell's snapshot/score/
        trajectory AT SAMPLING TIME. The trajectory walk must use the capture,
        never the live cell: an earlier result in the same batch may replace
        the cell, and stitching actions executed from the OLD state onto the
        NEW prefix fabricates scores no single playthrough achieved
        (2026-06-08 incident — caught by publish-time replay verification;
        the official code ships these values inside the task for this reason)."""
        keys = [k for k in self.cells if k != DONE_KEY]
        w = np.array([1.0 / np.sqrt(self.cells[k].seen + 1.0) for k in keys])
        csum = np.cumsum(w)
        picks = []
        for u in rng.random(n) * csum[-1]:
            k = keys[min(int(np.searchsorted(csum, u)), len(keys) - 1)]
            c = self.cells[k]
            c.chosen += 1
            c.chosen_since_new += 1
            picks.append((k, {"snapshot": c.snapshot, "lives": c.lives, "score": c.score,
                              "traj_len": c.traj_len, "traj_last": c.traj_last}))
        return picks

    def update_from_trajectory(self, chosen_key, capture, res, explog):
        """Walk one exploration episode (master-side, serial): append to the