---
title: Terminal rollout
family: terminal-policy-iteration
summary: Compare columns by playing each one to the end of the game. This faithful version was unaffordable.
status: runtime-paused
evidence: task-record only
reads: public
---

The faithful version of the family's idea: compare columns by playing each one
all the way to the end of the game. It was too slow to use.

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

## The intuition

Every other program in this family truncates. It plays a candidate column
forward for 25, 50, 100 or 200 moves, and then has to guess what the rest of
the game would have been worth, with a hand-tuned evaluator, or a fixed
penalty for dying, or simply by treating the horizon as the end. Each of those
guesses reintroduces exactly the kind of error the rollout was meant to remove.

This program removes the guess. Play each candidate to the actual end of the
game, and let the comparison be between real outcomes. If the family's premise
is right: that the way to judge a move is by the lifetime that follows it —
this is the version with no approximation left in it.

## How it works, step by step

1. Read the public position; ask the shared phase policy which column it wants.
2. Screen the field down to a small number of challengers: the source's
   default keeps two, so that the expensive part runs on a short list rather
   than on all seven columns.
3. Give each survivor a set of shared imagined futures (up to eight by default)
   and play each one out to termination, or to a 500-move cap, with the shared
   phase policy making every subsequent decision.
4. Compare the resulting distributions with a 99% confidence rule, and keep the
   baseline unless a challenger clears it.

## What happened

It was never a policy result, because it never got that far. The design was
measured, the cost was projected, and the projection ended it: roughly hundreds
of seconds **per move**. At that rate a single 90-move game takes hours, and
the paired cohorts this repository requires — dozens of games, candidate and
reference on the same seeds — are out of reach by orders of magnitude.

The repository's word for this is **runtime-paused**: cost stopped the
experiment before a decisive comparison, which is a different thing from the
idea being wrong. Nothing here says that comparing moves by terminal outcomes
fails. It says this implementation cannot be run, and the family's later work
is a series of attempts to buy most of the fidelity for a fraction of the
price — truncated horizons, staged elimination of weak columns, and reuse of a
locked panel of stored positions instead of fresh games.

The projection itself is **task-record only**: it comes from a research
conversation, not the written ledger, so there is no timing artifact to check.
Its practical consequence: that no successor in this family attempts full
terminal rollouts — is visible throughout the rest of the directory.

<TechnicalDetails title="The technical record">

Status **runtime-paused**, evidence **task-record only**. The
[experiment index](/docs/research/experiment-index) records the purpose as
"screens faithful full-terminal comparisons around the shared phase behavior"
and the outcome as "a faithful design projected roughly hundreds of seconds per
move". There is no entry in the
[experiment history](/docs/research/history) and no retained timing artifact.

From the source, the defaults that define the design: two challengers per root,
at most eight shared scenarios, a 500-move cap per rollout, a four-game audit
lane with a 115-second time limit, and a 0.99 confidence level. Chance tapes
use separate domains for reveals and next discs, keyed by scenario, step and
event. The file carries no compile-time 7,000-point lock, so it builds against
today's corrected engine.

</TechnicalDetails>

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

- **Fidelity has a price, and here the price is the result.** A rollout method
  in this game is a trade between how faithfully it measures an outcome and how
  many measurements it can afford. Every later program in this family is a
  different point on that trade.
- **A paused experiment is not a rejected one.** If a continuation policy
  became dramatically cheaper: the repository has since built a
  [much faster engine](/approaches/lifetime-objective/fast-engine) — the
  projection that stopped this program would need recomputing before it could
  be quoted again.
- **Still open:** what the family calls a successor that is "cheaper by
  design", rather than the same design run harder.