# Hex-a-Hop: complete verified route coverage — 16 October 2026

100/100 original levels have winning routes in a source-extracted headless rules
kernel. Classical search independently found 99. Fetch Quest was completed by
walkthrough-assisted classical planning: pickup order, regional completion and
bridge constraints were informed by the published full walkthrough at
https://sources.debian.org/data/main/h/hex-a-hop/1.1.0%2Bgit20140926-2/data/hints.html

This is not an unassisted 100/100 result or a trained campaign model.

## Final unresolved level

Fetch Quest: 71 moves, 2,372 states, 2,332 expansions, about 0.03 seconds in the
recorded constrained-planner run (excludes compilation, analysis and failed work).
The planner rebuilds the route from the original board; the full move string is
not embedded in its search implementation. Published guidance remains substantial
assistance. An earlier root-search run exceeded 25 million states without solving it.

Boot inventory changes ice behavior. The successful plan converts central ice
early and revisits it, rotates the upper spinner twice and the southern spinner
three times, then uses six remaining boots for the final ascent. The left cannon
clears the green tile left by the bottom pickup. Optional blue tiles may remain
at victory. The useful change was the dependency plan, not more training.

## Verification and limits

All 100 routes were freshly replayed: every input accepted, true win, no death
or transition cutoff. Every prefix passed lossless encoding and existing
structural-pruning audits. All 100 routes passed AddressSanitizer and
UndefinedBehaviorSanitizer (leak detection disabled). 63 experiment tests passed.

This is source-kernel verification, not an independent 100-level browser-parity
check or a proof of shortest routes. The kernel makes graphics/audio inert,
omits undo history from reachability state, uses a fixed settled viewport,
caps automatic tile interactions and guards undefined map writes.

The complete JSON includes all routes, adaptations, source hashes and validation.
Local paths in it describe research evidence; they are not public download URLs.

## Separate learning pilot

Generated ordinary/green-tile boards, not the original campaign; 64 held-out test
boards and one training seed. Search 64/64, imitation 2/64, PPO 0/64,
imitation followed by PPO 1/64. Learned policies did not qualify.

## Local reproduction

From the repository with pinned source archives and research dependencies:

```
experiments/rl/snake/.venv/bin/python -m pytest experiments/rl/hex-a-hop/tests -q
experiments/rl/snake/.venv/bin/python -m pytest experiments/rl/hex-a-hop/tests/test_search.py -k walkthrough_assisted_pickup_planner -q
```

The public JSON is a frozen evidence snapshot, not a standalone executable.
Research code is under experiments/rl/hex-a-hop. Source-kernel extraction and
campaign import are in build_headless.py and campaign_search.py. Game source and
licenses are available through /computer-games/vendor/hex-a-hop/source/.