---
title: Virtual ignition
family: heuristic-search
summary: Before choosing a column, imagine dropping a few imaginary discs onto the board to see which arrangements would go off like a firework, and prefer the boards that would.
status: rejected
evidence: task-record only
reads: public
---

Before choosing a column, imagine dropping a few imaginary discs onto the
board to see which arrangements would go off like a firework, and prefer the
boards that would.

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

## The intuition

Hand evaluators normally judge chain potential with a static rule of thumb:
count the discs that are one addition away from clearing, add something for
how those clears might link up, and stop. That rule is a guess about a
cascade, made without simulating one.

Virtual ignition replaces the guess with the real thing. It takes the board,
places up to a handful of *virtual* discs on it — discs that do not exist and
are marked as such, and asks the actual rules engine what would happen. If a
board answers "a five-wave chain that opens four gray discs", it is a board
with stored energy. If it answers "nothing", it is not, whatever the static
count says.

The measured answer is turned into nine features: how ready the board is to
ignite, how much the seed disc itself would clear, how many covered discs
would crack and reveal in the first wave, how many discs, reveals and waves
would follow downstream, how deep the cascade would run, and how much the
total covered population would fall.

## How it works, step by step

1. **Read the public position** — visible board, visible next disc, rise clock.
2. **For each candidate ignition point**, add a virtual disc and run the real
   cascade resolver over several stratified guesses about what the hidden gray
   numbers would turn out to be. Nothing about the actual hidden values is
   read; the guesses are drawn from a fixed public sampler.
3. **Collect the nine energy features** and score them with fixed weights,
   which pay most for downstream reveals (+260), downstream waves (+240) and
   downstream clears (+160).
4. **Add that as a residual** to the fair leaf evaluator, scaled by a constant.
5. **Use the combined evaluator as the leaf** of a depth-3
   [sparse expectimax](/approaches/heuristic-search/sparse-expectimax) search
   and play the winning column.

## What happened

It was retired for cost. Simulating hypothetical cascades at every leaf of a
three-move search is expensive, the gain it bought was small, and small gains
do not justify large costs when the reference search is faster and stronger.

The lab is built as a scale sweep. It plays the same games with the ignition
residual switched off and at four increasing strengths, so the retirement is
a statement about the whole curve, not one setting.

<TechnicalDetails title="The technical record">

The [experiment index](/docs/research/experiment-index) records this lab as
**rejected, task-record only**: *"the small gain did not justify its
substantial search cost."* Task-record only means the comparison lives in a
research conversation and was never promoted into the
[experiment history](/docs/research/history). **No score, move count, work
figure, or cohort size for that comparison is retained anywhere in this
repository**, so this page quotes none. What "substantial search cost" means
numerically is exactly the thing that is missing.

Repository-verified from the source: `src/core/typescript/virtual-ignition.ts`
extracts nine features with weights +40 (ignition readiness), +40 (seed clear
potential), +90 (initial cover cracks), +180 (initial cover reveals), +160
(downstream clears), +260 (downstream cover reveals), +240 (downstream waves),
+80 (cascade depth energy) and +160 (cover reduction); it places at most six
virtual additions and defaults to seven reveal scenarios. The lab `main.ts` is
pinned to depth 3 with five chance samples, three ignition scenarios, a
1,000,000 work bound, and sweeps residual scales `0, 0.05, 0.1, 0.25, 0.5` over
games from the `0x1d70…` training range. It reports mean and median score,
moves, censored games, work per move and evaluator cache hit rate: the shape
of the missing result.

The virtual discs and the reveal guesses are generated from a fixed public
sampler keyed on the visible state. No hidden value or future disc is read.

</TechnicalDetails>

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

- **Measuring chain potential is not the same as being able to afford it.**
  This is the family's clearest cost-side negative: a more faithful feature
  can be strictly better per evaluation and still lose, because the search
  around it gets smaller.
- **A negative for one price, not for the idea.** The rejection is of this
  configuration — full cascade simulation at every leaf of a depth-3 search.
  An incremental or cached formulation, or the same signal used only at the
  root, was not tested.
- **The record is thinner than it should be.** A rejection whose retained form
  is one sentence cannot be re-examined, re-powered, or refuted. Anyone
  resuming this line should expect to re-run the scale sweep rather than build
  on it.