---
title: Vertical reservoir
family: constructive-reservoir
summary: Defines a promising vertical stack, tunes those preferences on complete games, and finishes behind the one-move reference.
status: rejected
evidence: ledger-recorded
reads: public
---

The most literal version of the family's idea: describe a good vertical
structure directly, in features a person can name, and let an optimiser tune
how much each one is worth. It ended up playing worse than the simple search it
was meant to improve on.

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

## The intuition

The [family page](/approaches/constructive-reservoir) shows the mechanism this
approach is named after: a column of identical high discs that sits inert
until its run reaches exactly the right length, then clears all at once. The
bet here is that the *shape* of such a structure can be recognised directly
from the board, without simulating the future at all.

So the evaluator is built out of named, deliberately non-linear quantities:

- **release ladders**: how close a stack is to being exactly releasable by one
  more disc, computed exactly rather than approximated;
- **same-target powers** — several discs sharing one target value, which is
  what makes a stack pay off all at once;
- **escape wells**: a low, empty column kept as a way out if the structure
  goes wrong;
- **horizontal reachability** and **cover access**, whether covered discs can
  still be attacked;
- **low caps**: the badness of a small number sitting on top of something you
  need.

None of this reads hidden information: it is a description of the visible
board. The claim being tested is that structural potential is *legible* from a
still image of the position.

## How it works

1. Score every legal column by applying the evaluator to the position after the
   move, averaged over seven sampled chance outcomes; play the best.
2. Fit the evaluator's weights with a cross-entropy method — repeatedly sample
   a population of weight vectors, play complete games with each, keep the best
   few, and move the sampling distribution toward them.
3. Score every candidate on complete games played under **common random
   numbers**, so two candidates are always compared on the same games.
4. Freeze the winner and play a larger tournament against the one-move-ahead
   fair search.

## What happened

It lost. On its 128-game tournament the frozen candidate averaged **160,498.1
points and 50.9 moves** against fair D1's **178,554.4 and 55.8**. Score,
lifetime, lower-quartile score and reveal flow all failed admission, so the
probe cohort was never opened.

The result is more interesting than a simple failure. The same family's
[constructive planner](/approaches/constructive-reservoir/constructive-spectrum),
which simulates a cycle forward instead of scoring a still image, beat the same
comparator by a wide margin on the same kind of cohort. Taken together the two
runs say something specific: it is not the reservoir *idea* that fails here,
it is the attempt to read structural value off a static board. The research
status makes the same point in general terms — similar-looking boards can have
very different futures depending on how reachable triggers and covered discs
evolve across the rises.

<TechnicalDetails title="The technical record">

**Status: rejected; ledger-recorded.**
[Experiment history](/docs/research/history) §"Explicit reservoir and
constructive-cycle policies (rejected)": 128-game tournament, candidate
160,498.094 points / 50.891 moves versus fair D1 178,554.438 / 55.750.
Corrected 17,000-point scoring, checked as internally consistent by
[audit 03](/docs/exploratory/audit-03-claim-arithmetic) row 4399. The ledger
retains source and artifact SHA-256 values; the
[reproducibility guide](/docs/reproducibility) notes that historical hashes no
longer match current bytes after the tree was reorganised.

Frozen configuration, from the source: training seeds
`0x3d630000`–`0x3d63ffff`, tournament seeds `0x3d640000`–`0x3d64ffff`, probe
seeds `0x4d630000`–`0x4d63007f` (never opened), 7 successor samples,
1,000-move cap, population 65, 13 elites, 24 generations, 64 games per fitting
batch, 128 tournament games, 32 probe games, policy seed `VRSV`, 30-minute wall
limit, 256 MiB RSS limit. Admission required at least 150 mean moves, at least
500,000 mean points and a 1.50 ratio on the fitting stage before the probe
could open.

A related but separate experiment tested the same instinct inside the
four-move search: a literal 7/6/5 vertical-ladder energy term screened cheaply
at depth 2 looked promising and was **rejected** when transferred to depth 4.
Both are listed in the [experiment index](/docs/research/experiment-index)
under fair expectimax.

Source: `vertical-reservoir-policy.cpp`.

</TechnicalDetails>

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

- A hand-named description of vertical structure, tuned directly on complete
  games, is not enough to beat even a one-move-ahead fair search.
- The failure is specific to reading structure from a still board. Simulating
  the cycle forward with the same underlying idea worked much better.
- This is a valid negative result on one frozen feature set. It does not rule
  out vertical structure as a *component*, but the one attempt to transfer a
  vertical-ladder term into the depth-4 reference also failed.