# Frame and corners: a wrong frame is not cheap to refute

> Refuting a wrong full frame of the real pieces is not cheap: the preregistered kill line is met, and refuting partial frames costs 2.2 to 2.5 bits more than it can save. The cheap information about the frame, about 17.4 bits, sits at the four clue corners. Whether frames away from the corners can be refuted cheaply is undecided.

- Canonical page (with interactive figures/demos): https://eternity2.dev/research/lab/experiments/raphael-anjou/full-solve/frame-and-corners/
- Updated: 2026-11-14
- Topics: structure, search-space
- Reproduce: `just experiments frame-cores`
- Source: Frame cores: preregistration, amendments, results and tables (https://github.com/raphael-anjou/eternity2/tree/main/research/experiments/frame-cores)
- Source: Approach pricing: frame-nogood price notes and re-check outputs (https://github.com/raphael-anjou/eternity2/tree/main/research/experiments/approach-pricing)

---
A search that fixes the 60 border pieces first and fills the interior second only pays if wrong frames die cheaply. This page asks three questions on the real pieces with the five clues. Can a wrong full frame be refuted cheaply? Can a partial frame (an arc of the border) be refuted for less than the search it saves? And where does cheap information about the frame sit? For the fixed prices (row scan 2^134.4, budgets, gaps) see [the series hub](/research/lab/experiments/raphael-anjou/full-solve/).

## Setup

Frames were sampled by a sequential Monte Carlo sampler over the border ring (5,000 particles, approximately uniform). Stratum H keeps frames that pass the four clue-block tests (the three free interior cells next to a corner clue must be fillable with distinct pieces); the true frame is in H. There are 60 H frames, 30 uniform frames (U), a further 60 H frames with no dead 3x3 clue block ("realfar"), and three generated boards as a control. Refuters: plain depth-first fill in two orders, the same with a matching check, and the Glucose SAT solver with the frame as assumptions. The thresholds and two amendments were committed before any test run. The measurement is not a record attempt, and results on generated boards are never Eternity II progress.

## Results

| Question | Result | Status |
|---|---|---|
| Cheap refutation of a full frame (real H, 60 frames) | Dead. Lower median speed-up over plain search: 0.89, 0.12 and 1.00 (matching check in rows order, in spiral order, SAT) against a line of 100 | measured, independently re-checked |
| Refuting an arc of 44, 42, 40, 38 border cells | SAT refutes 42, 41, 34, 30 of 60 arcs within 60 s; nested cost 27.7, 31.4, 35.4, 39.5 bits against budgets of 25.5, 29.0, 32.9, 37.0 | measured, independently re-checked |
| Cheap information about the frame | About 17.4 bits, all at the four clue corners | independently re-checked twice |
| Frames away from the clue corners | Undecided | measured, undecided |

**A wrong full frame cannot be refuted cheaply.** Plain search was censored on 25 of 60 frames. SAT refutes a frame either at once or not at all: every refutation on real H finished within 0.35 s (46 frames), and the others hit the 600 s cap.

**Refuting arcs sits above break-even at every size.** The nested cost is 2.2 to 2.5 bits above its budget at each of the four sizes. At 38 cells the budget exceeds the 120 s cap, so that level cannot kill and is stated, not read. A looser preregistered line (at least 15 of 60 arcs refuted within 60 s) is met, but that is the corner effect below, and the nested cost governs.

**The cheap information sits at the corners.** A frame passes the four clue blocks with probability 2^-17.4 (2^-17.41 jointly, one standard error 17.38 to 17.43). In this run the same quantity is 17.24 bits on the real pieces. Near frames die on a core of 3 border cells (median): 23 of 60 H frames are "near" (the cheapest refutation lies within ring distance 3 of a corner), and 21 of those 23 have a dead 3x3 clue block (29 of 60 H frames have one). Their median cost is 29.83 bits. The same cheapness appears on the three generated boards (near medians 31.39, 28.99, 32.22), so it comes from the clue-corner geometry, not from the real pieces.

**Frames away from the corners are undecided.**

| Stratum | Frames | Censored | Median bits [95% interval] |
|---|---|---|---|
| Real H, near | 23 | 0 | 29.83 [28.30, 33.78] |
| Real H, far | 37 | 16 | 75.12 [42.64, inf] |
| Extra set, far | 52 | 29 | inf [58.43, inf] |

The far median is below the 120-bit line, but its interval is unbounded, so neither the pass nor the fail line applies. Part of the censoring in the extra set is a counting failure, not a refuter failure (5 of the 29 frames: the counter returned no surviving particle, and the frame was recorded as having no usable nogood although a refuter proved it). That choice is conservative for the median.

The overall median of 36.75 bits [32.90, 53.19] on real H technically passes a necessary condition. It is not a result: the median falls on the corner frames, and it says nothing about a search driven by learned cores.

## What it means for a full solution

The corner information is real and cheap, and it is a small part of the gap. A frame that passes the corners still leaves an interior to search: about 10^13.6 nodes in the mean by spiral search, a model figure that replaces an earlier 10^8.3 (which averaged over all valid frames, nearly all of which die at the corners). It is independently re-checked as a model, not measured: confirming the true frame on generated 16x16 boards hit the 600 s cap with all five refuters, so the measured confirm cost is only a lower bound. Whether a frame-first search can beat row scan depends on the far frames, and that is the part left undecided. This page claims no saving from it.

## What was not run

Matching-check and spiral refuters on arcs; plain search on arcs of 42, 40 and 38 cells; arcs of 34 and 30 cells; an incremental solver arm; any 16x16 confirm of a true frame beyond the cap. The 10^13.6 figure is a model with no run behind it on the real pieces.

## Reproduce

`just experiments frame-cores` reruns the measurement (about 59 core-hours). The README and tables in the [evidence folder](https://github.com/raphael-anjou/eternity2/tree/main/research/experiments/frame-cores) hold every number above. See [records](/research/records/) and [open problems](/research/open-problems/) for the frontier, and [dead ends](/research/build/dead-ends/) for the cheap local tests.

## 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.
- [What a known tile is worth under row scan](https://eternity2.dev/research/lab/experiments/raphael-anjou/full-solve/worth-of-a-known-tile/): Under a row scan, one known tile removes between 1 and 10 bits of the full search tree depending only on where its cell sits in the scan (median 5.8 bits, on painted 16x16 boards), and several tiles together are worth less than the sum. On those boards, 10 spread tiles leave the tree above 10^15 nodes and 16 to 24 bring it under, a figure checked end to end only. Nothing here is a measurement of the real puzzle beyond two transfer checks.
- [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.
- [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.
- [Clause learning against propagation: a negative result](https://eternity2.dev/research/lab/experiments/raphael-anjou/full-solve/learning-against-propagation/): Clause learning gave no growing lead over plain arc-consistency propagation on wrong prefixes of the real pieces: the pre-stated stop line fired on summed work, and in counts learning is worth 0 bits over arc consistency. The toy boards (5x5 to 7x7) lean the same way but do not establish it. Nothing here is Eternity II progress.
- [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.
