# Keying and replay experiments: no gain found, one effect measured

> Five preregistered experiments on remembering and keying dead ends found no gain that matters for a full solution. Three were killed (difference-map bands, cut-keyed replay, keying with five clues) and two were kept as measurements (quota keying, and the quota curve with five clues). The kept quota result is a rate effect of 53 times, about 5.7 bits of rate, which is not a smaller tree.

- Canonical page (with interactive figures/demos): https://eternity2.dev/research/lab/experiments/raphael-anjou/full-solve/keying-experiments/
- Updated: 2026-11-14
- Topics: backtracking, search-space
- Reproduce: `just experiments five-clue-quota test 4 local`
- Source: difference-map-bands: preregistration, results and independent assessment (https://github.com/raphael-anjou/eternity2/tree/main/research/experiments/difference-map-bands)
- Source: cut-keyed-replay: preregistration, results and independent assessment (https://github.com/raphael-anjou/eternity2/tree/main/research/experiments/cut-keyed-replay)
- Source: quota-keying: preregistration, amendments, results and independent assessment (https://github.com/raphael-anjou/eternity2/tree/main/research/experiments/quota-keying)
- Source: five-clue-keying: preregistration, amendments, results and independent assessment (https://github.com/raphael-anjou/eternity2/tree/main/research/experiments/five-clue-keying)
- Source: five-clue-quota: preregistration, amendments, results and independent assessment (https://github.com/raphael-anjou/eternity2/tree/main/research/experiments/five-clue-quota)

---
A search that forgets why a branch died pays for it again. These experiments
remember dead ends, or key the search on a few colours so they are recognised
sooner. Each question, baseline and stop line was fixed before any test run.
Each result was independently re-checked; where a README and that re-check
disagreed, the re-check decided and the README was corrected. The price of a
full search and the gap are on the
[series page](/research/lab/experiments/raphael-anjou/full-solve/).

## The five at a glance

Boards of 6x6 to 8x8 are generated boards (toys), never Eternity II progress.
The three real-piece experiments use the official 16x16 pieces, with no record
attempt.

| Experiment | Question | Baseline | Stop line | Outcome | Headline number |
|---|---|---|---|---|---|
| difference-map-bands (generated) | Do exact band projections slow a difference map's growth with size? | width 0, `bb-mrv` | killed if no fall of the slope | **killed** at this cap | 0 of 30 solved at 7x7, every width; `bb-mrv` 23 of 30 |
| cut-keyed-replay (generated) | Does partitioning propagation on a cut line prune because of the cut? | `bb-mrv` nodes, random key | killed if a random key is within 0.1 | **killed** | cut 0.486 of nodes, random 0.419, work 1.21 |
| quota-keying (real) | Is the published colour keying worth 450 to 4,600 times? | default keying | contradicted if gain below 45 | **kept**; claim contradicted | gain 1.31 [0.69, 2.43] |
| five-clue-keying (real) | With five clues, is a one-ring keying a lever? | default, random keys | not a lever if best vs random includes 1 | **killed**, confounded | 0 of 2,240 runs reach legal depth 201 |
| five-clue-quota (real) | Keying or quota curve? | default keying and quota | slack explains if 0 of 12+ vs 8+ openings | **kept**: slack explains it | 0 of 16 against 14 of 16 openings |

"Kept" means the measurement is valid, reproduces exactly and its readings
stand. It does not mean a method works.

## Difference-map bands (killed)

Exact projections onto bands of width k (a dynamic program), k = 0 to 3, were
added to a difference map (an iterative projection method) on 30 generated
boards per size, 3 seeds each, 20,000 iterations per run. At 7x7 and 8x8 no
instance was solved at any width (one run of 90 aside), so no growth exponent
was measured: the kill is of this method at this cap. At 6x6, the only size
with finite medians, the iteration ratio to width 0 has intervals [0.41, 1.36],
[0.43, 1.30] and [0.77, +inf] for widths 1 to 3, all containing 1, and a width
costs 3 to 33 times more per iteration. A larger cap could give finite 7x7
medians at low width, but the 6x6 trend and the censored runs point the wrong
way. Independently re-checked: 90 rows
rerun, identical.

## Cut-keyed replay (killed)

Partitions keyed on the colours of two cut edges were replayed over the exact
trees of `bb-mrv` on 60 generated boards. The cut key keeps 0.486 [0.469,
0.500] of the nodes at 7x7, just under the 0.5 line. A random key with the same
partition count, drawn off the cut line, keeps fewer: 0.419 at 7x7, on every
control seed and on 30 of 30 instances (0.444 against 0.483 at 6x6). The
pruning comes from partitioning in general (every keyed arm keeps 0.42 to
0.49), not from the cut, and every arm costs more propagation work than
`bb-mrv` (1.211 for the cut key). Not tested: a cut key inside a row-scan
solver. Independently re-checked: 6x6 and a sample of 7x7 rows reproduce.

## Quota keying (kept)

The published record engine prunes below a quota of "privileged" colour sides
placed by each cell. An earlier desk study claimed its keying makes reaching
depth 201 450 to 4,600 times cheaper (depth counts cells placed, out of 256). Measured on the real pieces, every node
charged, 10 openings per arm, 1e9 nodes per run, published triple (colours 5,
20, 7 in this repository's numbering) against the port's default:

- **Keying gain: contradicted.** 1.31 [0.69, 2.43] with no clue, 0.74 [0.42,
  1.27] with the centre clue. With five clues the published keying never
  reached depth 201 in 10 runs; the default did 135 times. Scope: 1e9 nodes,
  one attempt, this port (the desk study ran 1e10 with five clues).
- **Legal break path: supported, at a price.** Forbidding illegal break
  placements gives 7.63 [4.82, 13.9] times more legal nodes past depth 201 per
  node, but runs about 23 times slower per node (1.25 arrivals per second
  against 19.2).
- **Staging: dead for the desk study's schedule on this port.** Requiring all
  122 privileged sides by depth 176 cuts deep nodes (depth 230 and beyond) to
  0.051 [0.031, 0.081] of the published schedule's; a third colour by depth 200
  cuts them to 0, which a row scan can barely supply, so that is not a yield
  measurement. Dead if both ratios fall below 3.

Independently re-checked: 19 rows rerun, identical.

## Five-clue keying (killed)

With the five official clues, 680 keyings (one ring colour, two interior
colours, quota scaled to the keying) were screened on 3 openings, and the best
five, the published triple and five random keys confirmed on 20 more. None of
the 2,240 one-ring runs placed a legal node at depth 201 or more; the default
keying placed 9,758 on the 20 confirmation openings. The reading "not a lever"
is a floor effect (best against random is 0 over 0, set to 1), and the default
differed in two things at once, keying and quota slack. The surviving claim is
narrow: no one-ring keying under the scaled quota works. Re-checked: 20 rows
rerun, identical.

## Five-clue quota (kept)

This separated the two. Same five clues, 16 openings, 32 arms. The default
keying, asked for 147 of its 150 sides by cell 160 instead of 119, never passes
depth 35: 0 of 16 openings reach depth 201 (12 under its own quota). One-ring
keyings asked for the same 79% share of their sides have 14 of 16 (published
triple) and 12 to 13 (keying 1, 7, 9) positive openings, against 2 and 1 under
their earlier 97.5% quota. Legal nodes past depth 201 per second for the
published triple rise from 0.19 to 10.0. **Quota slack, not the colour set,
explained the previous experiment.**

Not shown: at matched slack the default is still ahead on the point estimate
(one-ring keyings at 0.61 [0.30, 1.53] and 0.74 [0.29, 2.18] of its rate per
second), with intervals including 1, so keying is not shown to matter on 16
openings. Which segment of the curve binds is not localised. The port's
top-row orientation fix is free (0.98 [0.95, 1.04]). Deepest legal depth of any
run: 220. Independently re-checked: 8 rows rerun, identical.

## What this means for a full solution

None of the five makes a full solve cheaper. Each ratio below keeps its own
scope; base-2 logarithms are derived from them. The effects were measured in
different experiments on different generated boards, so they are not additive,
and none is a gain against the full tree:

| Effect | Factor | Bits |
|---|---|---|
| cut key, generated boards, nodes only (work rises) | 1 / 0.486 | 1.0 |
| published keying gain, no clue | 1.31 | 0.4 |
| legal break path, per node (a wall-clock loss) | 7.63 | 2.9 |
| quota slack, rate of reaching depth 201 | about 53 | 5.7 |

The row scan with five clues costs 2^134.4 nodes; the gap against the budgets
is on the [series hub](/research/lab/experiments/raphael-anjou/full-solve/). The kept result is a rate of reaching a depth,
not a smaller tree, and a deeper run is not a higher-scoring board (see
[the records](/research/records/)). See also
[late filters](/research/lab/experiments/raphael-anjou/full-solve/late-filters/).

## What was not run

No other quota curve shape, no budget above 1e9 nodes per run, no restarts, no
border interleaving, no cut key inside a row-scan solver, no other band
projection.

## Reproduce

The kept experiment:

```
just experiments five-clue-quota test 4 local
```

The other four:

```
just experiments difference-map-bands
just experiments cut-keyed-replay test 7 12
just experiments quota-keying main 12
just experiments five-clue-keying screen 12
```

Earlier stages and analysis scripts are in each folder's README.

## Related

- [The price of a full solution](https://eternity2.dev/research/lab/experiments/raphael-anjou/full-solve/): An exact search of the 16x16 with five clues costs about 2^134.4 nodes, and no route examined closes the 56 to 73 bit gap against budgets of 2^78 and 2^61.5 nodes. The work is parked, with stated conditions for reopening. This page holds the fixed numbers and indexes the ten pages behind them.
- [The price of a full search](https://eternity2.dev/research/lab/experiments/raphael-anjou/full-solve/price-of-a-full-search/): Exhausting a plain row-scan search of the real Eternity II pieces with five clues takes about 10^40.457 nodes (2^134.4), measured, and the random-model prediction lands within 0.04 digits of it. The official piece set is an ordinary member of the boards painted with the same colour counts on this statistic. This prices the full search; it finds no way to shorten it.
- [Late filters: every check that works, works too late](https://eternity2.dev/research/lab/experiments/raphael-anjou/full-solve/late-filters/): A filter added to the row scan can save at most the bits its depth allows, and the checks that refute nearly all wrong boards only start to bite at depths 144 to 192, past the widest level of the tree (158). A perfect free cut at depth 144 would save 2.4 bits of the 134.4 the scan costs, so no late filter closes the gap. The ceiling table is derived arithmetic on a measured profile, not an experiment.
- [Port fidelity: seven differences from a record engine, and a five-clue regression](https://eternity2.dev/research/lab/experiments/raphael-anjou/full-solve/port-fidelity/): Our port of a record-style backtracker differed from the published engine in seven places; all seven were fixed and tested, and the node sequences match the earlier port wherever the legacy switches are on. With all five clues the faithful defaults do worse than the old ones (deepest depth 18 to 54 against 33 to 215 on eight seeds), so the fix is not an improvement for that setting.
- [Ten engine experiments on small boards: faster nodes, no smaller trees](https://eternity2.dev/research/lab/experiments/raphael-anjou/full-solve/search-engine-tracks/): Ten small preregistered engine experiments on generated boards from 6x6 to 9x9 kept five and dropped five. The one measured speed-up is a node that costs 5.3 to 6.1 times less for the identical search; none of the three methods that change the tree (blocker backjumping, value order by impact, Hall matching) shrank it against its baseline. Nothing here is Eternity II progress.
- [Seven families of approaches priced out against the row-scan tree](https://eternity2.dev/research/lab/experiments/raphael-anjou/full-solve/approaches-priced-out/): None of seven families of approaches (search that learns, lossy representations, the shape of the tree, adaptive orders, tile-set structure, algebraic and search-space reformulation), nor the far ideas, frame nogoods or alternative fill orders, is priced at 10 bits or more with a mechanism behind it. The largest surviving gain is about 5.5 bits, against a gap of 56 to 73 bits; a few closures rest on one reader or one solver family and are marked as such.
- [Records & solvers](https://eternity2.dev/research/records/): Eternity II has never been solved, but nearly two decades of community effort have pushed the best board to 470/480. Who holds what, how they did it, and why some headline "480" boards are not actually the real puzzle.
- [Open problems](https://eternity2.dev/research/open-problems/): The open frontier of Eternity II in one place: every angle still worth a run, the wall it attacks, what has been tried and where it stopped, and whether it is a newcomer-tractable target or a hard, well-mapped one.
- [Dead ends and untested claims](https://eternity2.dev/research/build/dead-ends/): Approaches that did not move the score on Eternity II, each graded by the strength of its evidence. Twenty verdicts rest on committed data, a named public report or a derivation; twenty-two more are stated without evidence a reader can open, and are listed separately with the experiment that would settle each, and one measured negative result on clause learning.
