---
title: Rollout veto
family: d4-long-outcome
summary: Keep the depth-4 search's move unless replaying every column forward for twenty-five moves says, across all seven imagined futures, that another one is better.
status: runtime-paused
evidence: ledger-recorded
reads: public
---

Play the depth-4 search's move — unless replaying every legal column forward for
twenty-five moves says, consistently across all seven imagined futures, that a
different column is better.

<EvidenceLabel status="runtime-paused" evidence="ledger-recorded" reads="public" />

This is the family's most direct idea and its most eventful record: one
extraordinary game, a runtime rejection, three attempts to make it affordable,
and finally a properly powered retest that rejected the mechanism.

## The intuition

The reference search is tactically strong and structurally short-sighted. It
cannot see the second row rise, so it cannot see that a column it likes will be
unplayable two cycles from now. A twenty-five-move replay can. But a long replay
is also *noisy*: play the same position forward under different luck and you get
very different outcomes, so a long look used carelessly will happily override a
good move on the strength of one lucky imagined future.

The veto design answers both at once: the long look never *chooses*, it can only
**object**, and it may only object when the objection holds up statistically.

## How it works, step by step

1. Run unmodified fair depth-4 search. Keep its column as the default. If
   anything below fails, this is what gets played.
2. **Route or skip.** The expensive part runs only when the board is dangerous —
   defined as a maximum column height of at least four. On safe boards the policy
   *is* plain depth 4, bit for bit.
3. At a routed position, give **every legal column the same seven imagined
   futures**. Each future is generated from a hash of the public board, so the
   comparison between columns is paired and no column gets luckier tapes than
   another.
4. In each future, force the candidate column, then play **twenty-five more
   moves** with a fresh, complete depth-2 search at every step. That continuation
   is typed so that it structurally cannot receive the tape, the scenario
   identity, the score, the level or the move number: it sees a board, a disc, a
   rise phase and whether the game is over.
5. Score each line by the points it actually earns, minus a large penalty if it
   dies, plus one leaf evaluation if it is still alive at move 25.
6. **Four conditions, all required**, before an alternative may displace depth 4:
   it must survive in at least as many of the seven futures; it must not clear
   fewer numbered discs on average; its paired advantage must be positive at a
   fixed one-sided confidence bound; and it must not be too far below depth 4's
   own valuation of the position. Among alternatives that pass, the one with the
   largest confidence bound wins. Otherwise depth 4 plays.

## What happened

**One astonishing game.** The single registered pilot lasted 250 moves and
scored 404,047 where plain depth 4 lasted 105 moves and scored 159,616. It made
only twelve overrides in the whole game. Its projected cost was also more than five times its own preregistered runtime ceiling, so the run was paused before the
second game rather than being allowed to become a strength claim on a sample of
one. The record was explicit at the time: "only one pair, not statistical
evidence".

**Making it cheaper did not make it fast.** Two follow-ups applied only changes
that provably cannot alter a single decision — skip alternatives that depth 4
already values too poorly, evaluate the baseline once, reuse identical
continuations, drop a cache. Every value, action and selected move stayed
bit-identical to the original, and the wall-clock savings were immaterial.

**A wider look at quality did not rescue it either.** Adding three more games
left the mean higher, but mostly because the original long game was still in it;
only one of the three new games was a joint win on both score and lifetime, and
the frozen quality gate failed.

**And the cheaper teachers could not imitate it.** Six approximations — shallower
continuations, fewer futures, shorter horizons, mixed schemes — were replayed
against the original's own decisions. The best of them reproduced **2 of the 12
beneficial overrides** against a preregistered requirement of 8. The high overall
agreement in that table is an illusion created by the 167 positions where the
teacher did nothing; agreement on doing nothing is easy.

**Then it was retested properly, and it does not reproduce.** The whole mechanism
was ported to corrected 17,000-point scoring, verified byte-identical against the
frozen source, and run on 32 paired games. Fair depth 4 averaged 339,290 points
and 98.66 moves; the veto averaged 292,780 and 86.34. It lost 19 of 32 games and
won 9. Most tellingly, the two flow rates that would have had to *rise* for the
idea to work both **fell**: the accepted overrides produced a policy that clears
and reveals less per move than the search it was supposed to improve.

<TechnicalDetails>

**Original pilot** (runtime-paused; ledger-recorded). Seed `0x3ded0000`, one
game: stock depth 4 159,616 points / 105 moves, 1.97143 clears and 1.09524
reveals per move; the veto 404,047 / 250, with 2.276 clears and 1.316 reveals per
move. 12 switches out of 179 routed decisions (4.8% of 250 moves); 15
alternatives passed every gate. Runtime: 852.255 seconds for the pair against a
2,700-second ceiling, projected to 15,340.599 seconds for the full protocol; the
runner paused before `0x3ded0001`.

**Exact compression** (runtime-paused; ledger-recorded). Replay asserted the
original 404,047 points, 250 moves, 179 routes and 12 switches before writing.
The depth-4 value prefilter skipped 120 of 1,242 action roots (9.66%); exact
continuation reuse removed a further 26,303 transitions; depth-2 calls fell from
192,983 to 152,884. Whole-game wall time moved 641.689 to 639.321 seconds — the
rollout phase from 333.585 to 295.902, which is not the order of magnitude the
gate required.

**Quality extension** (rejected; ledger-recorded). Four games: candidate mean
174,659.5 points / 113.75 moves against 107,007.75 / 72.5. Clears per move rose
1.84483 to 2.05055, reveals 0.98966 to 1.12967, and all four leave-one-out
subsets were positive, but only one of the three new pairs was a joint
score-and-move win, and lower-half score retention was 88.38% against a 90%
requirement. Individual games: `0x3ded0001` 109,264/70 became 153,925/100;
`0x3ded0002` was identical at 59,004/45 with zero switches; `0x3ded0003`
100,147/70 became 81,662/60.

**Teacher compression** (rejected; ledger-recorded). Replay-only over the 179
exported routed states. Best configuration (depth 2 at rise phases 1 or 5,
otherwise depth 1) reproduced 166 of 179 exact actions and **2 of 12** switches,
against an eligibility floor of 8 of 12; the five other configurations recovered
none.

**Corrected-scoring retest**, [`finding-03`](/docs/exploratory/finding-03-rollout-veto-17k)
(run validity `valid`, scientific outcome `fail`, evidence tier `development`,
reproduced in this checkout; separate namespace and seed lease, frozen source
untouched):

| Arm | mean score | mean moves | clears/move | reveals/move |
| --- | ---: | ---: | ---: | ---: |
| fair depth 4 | 339,290 | 98.66 | 2.0200 | 1.1308 |
| rollout veto | 292,780 | 86.34 | 1.9515 | 1.0749 |

Paired mean delta −46,510.5 points and −12.31 moves; one-sided 95% bootstrap
lower bound −91,924.6; wins–ties–losses 9–4–19; sign test p = 0.0436 on the 28
non-tied games; 0 censored games, 0 illegal moves. Of 12,314 alternatives
scored, 103 passed all four conditions and 68 overrides were taken: an average
of −21,887 points per override taken. Three games drew zero overrides and were
*exactly identical* to the comparator, which doubles as a fallback check. The
value-band correction that an earlier audit had asked for (7,000 to 17,000) was
run as a full second arm and changed the number of overrides taken by one, from
67 to 68; 29 of 32 games were identical between the two bands.

Records: [experiment index](/docs/research/experiment-index) "D4 and long-outcome
research"; [ledger](/docs/research/history) sections "Fair-D4 25-move exact-D2
rollout veto", "Exact 25-move rollout-veto compression", "D2 rollout
teacher-compression replay" and "Original D4 + D2/s7/h25 quality extension".
Sources: `d4-d2-rollout-veto.cpp`, `d4-d2-rollout-veto-exact-compressed.cpp`,
`d4-d2-rollout-veto-cache-free.cpp`, `d4-d2-rollout-veto-quality-extension.cpp`,
`d2-rollout-teacher-compression.cpp`.

</TechnicalDetails>

<Callout title="Two statuses, on purpose" tone="info">
The retained status of the sources in this directory is **runtime-paused** —
they were shelved for cost, not disproved, and the index still records them that
way. The mechanism they implement was separately rejected on a 32-game paired
cohort under corrected scoring, in a different namespace, without modifying these
files. Both statements are true and neither replaces the other.
</Callout>

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

- **The estimator was measuring the wrong game.** A column selected because it
  looks good under 25 moves of depth-2 play is executed in a game that will be
  played by depth 4. That mismatch, not merely noise — is the most concrete
  finding here, and it is why the accepted overrides made realized flow *worse*
  even though "must not clear fewer discs" was one of the four conditions.
- **A one-game pilot cannot be informative in either direction** for a policy
  whose per-override effect has a standard deviation of roughly 160,000 points.
- **What is ruled out, exactly:** horizon 25, seven futures, danger height 4, a
  depth-2 continuation, all legal columns as challengers, at both value bands, on
  32 fresh paired games. **What is not:** the family. The retest names the
  cheapest next test, not a longer horizon and not a wider band, but **more
  imagined futures**, since the return test appears under-sampled by roughly two
  to seven times at seven.
- The good engineering here survives the negative result: paired imagined futures
  shared across columns, a continuation whose *type* cannot carry privileged
  data, an exact fallback, and a compression pass proven decision-identical.