---
title: Public survival rollout
family: terminal-policy-iteration
summary: Judge each column by survival over thirty-one shared futures of one hundred moves each.
status: rejected
evidence: task-record only
reads: public
---

Judge each column by how long the game survives afterwards: thirty-one shared
imagined futures, a hundred moves each.

<EvidenceLabel status="rejected" evidence="task-record only" reads="public" />

## The intuition

This is the family's idea at its most literal. Survival is what the score
almost entirely consists of
([finding 01](/docs/exploratory/finding-01-score-is-survival): correlation
r = 0.9995 between score and moves survived over 64 reference games), so
instead of asking a board evaluator which column *looks* safest, force each
column and watch how long the game actually lasts.

A hundred moves is twenty row rises. That is far past anything a four-move
search can reason about, and comfortably past the point where a decision that
merely postpones trouble stops looking clever.

## How it works, step by step

1. Read the public position: board, visible next disc, moves until the next
   rise, terminal flag.
2. Build 31 event-indexed synthetic scenarios: the discs that will be dealt
   and the hidden numbers that will be shown, and use the same 31 for every
   sibling column, so the comparison is not decided by luck.
3. Force the column under test; from then on, every move is chosen by the same
   exact, completed fair one-move-ahead policy. That policy is handed only the
   board, the visible next disc, the rise phase and the terminal status. It
   never sees the real game seed, and it never learns which synthetic scenario
   it is inside.
4. Run to the 100-move horizon and compare columns by their outcomes.

## What happened

It was worse than the policy it was built on top of.

The retained verdict is that the longer rollout came out behind plain fair
one-move-ahead search on the games it was fitted on: a result that sounds
paradoxical and is not. A rollout comparison is only as good as the player
doing the rolling: if the continuation policy is myopic, then a hundred moves
of it mostly measures how well a column suits *that* player's weaknesses.
Thirty-one shared futures is also a small sample for a quantity as
heavy-tailed as a Drop7 lifetime, so column averages that differ by a little
may not differ at all.

This is **task-record only** evidence. There is no protocol, no artifact and no
per-game data in the repository, so no number can be quoted here.
[Audit 04](/docs/exploratory/audit-04-blind-spots) lists this source among the
thirty rejections that "cannot be re-derived", and notes it as one of the
survival-flavoured experiments whose numbers a later analysis wanted and could
not find.

<TechnicalDetails title="The technical record">

Status **rejected**, evidence **task-record only**. The
[experiment index](/docs/research/experiment-index) records the purpose as
"evaluates each root action with 31 aligned H100 public fair-D1 continuations"
and the outcome as "the longer rollout was worse than fair D1 on fitting
games". There is no entry in the
[experiment history](/docs/research/history).

From the source: a public-only 100-move survival rollout around the reference
fair-D1 continuation policy; every root action is evaluated on 31 aligned,
event-indexed synthetic chance scenarios; the root action is fixed and all
later actions are freshly chosen by the same exact, completed fair-D1 function
of board, visible next disc, rise phase and terminal status; neither the real
game seed nor the synthetic scenario identity is available to that continuation
policy. The file asserts the corrected 17,000-point level bonus at compile
time.

</TechnicalDetails>

## What this taught us, and what is still open

- **A rollout inherits the judgement of its continuation policy.** This is the
  same conclusion the [tree search](/approaches/tree-search) family reached from
  the opposite direction, where scaling a search with a myopic playout made
  long-horizon ranking worse.
- **Still open:** the same measurement with a stronger continuation — the
  family's later panels moved to a two-move-ahead policy for exactly this
  reason, and with enough scenarios to say something about a heavy-tailed
  quantity.