Skip to content

III-21: published floors compared with move sums and switch presses (experiment record) - #73

Draft
orda-aska wants to merge 2 commits into
mainfrom
exp/iii21-switch-floor-check
Draft

orda-aska wants to merge 2 commits into
mainfrom
exp/iii21-switch-floor-check

Conversation

@orda-aska

Copy link
Copy Markdown
Collaborator

Part of siliconcordat/.github#5. This reports a measurement under the III-21 threads in #22 and #53 (items C2 and E3).

What was measured

The published proof for III-21 at 270 keys splits every route at the move of the crate into the Druid's chamber.

It states that at least 74 further keys remain after the lower entrance, 75 after the upper entrance with a Fighter push, and 76 with a Thief pull. The page does not say whether these counts include hero switch presses (the A key). These floors are universal, so they must hold at every position right after that crate move (source claim).

On real positions right after that crate move, two quantities were computed with the repo's own code:

  • a: the least total of solo move keys over all 720 assignments of the six heroes to the six exits (solo_walk_plus_wall_convert_matrix and enumerate_all_goal_bijections).
  • b: the least number of A presses needed to make every hero that is not on an exit act at least once, starting from the active hero.

Positions were generated by breadth first search from record-route states, so each one is reachable. There were 43 distinct positions, all with the Fighter as the mover (16 upper, 27 lower). They fall at keys 195 to 206, later than every group in the proof's table, so none of them is one of the proof's 47,644 positions. The proof publishes counts per group, not positions.

Quantity Range over the 43 positions Published floor
a 57 to 60 74 lower, 75 upper
b 5 in every position not stated
a + b 62 to 65 74 lower, 75 upper
operator-counting bound (evaluate_operator_counting_bound, rounded up) 20 to 21 not stated

Status: verified. The scripts are on branch exp/iii21-switch-floor-check in experiments/iii21_switch_floor_check. Reproduce with python3 scripts/floor_compare.py. The check fails if the metric stops falling as the record route advances (a is 58 at key 195 and 36 at key 246).

What follows

Neither a nor a + b is close to 74. So this computation does not show whether the published floors count switch presses. It shows only that the floors come from something stronger than a hero-to-exit move sum plus the first round of presses. The move sum a is also not a lower bound on its own, because a Bard key can move several heroes at once and a Wizard swap moves a hero by many cells.

The proof page also says that four positions were excluded by a separate check that "followed each concrete board through the first cycle of hero turns" (source claim). So the prover already reasons about the turn order in some form.

Proposed: ask the proof's maintainer whether the 74, 75 and 76 count A presses. If they do not, the repo's merged Druid switch tax (5 here) adds on top. If they do, nothing is gained.

Other results

  • Verified: for all 22 Part III catalog routes, the count of A presses equals the winding of the activation order. The winding is the sum of the forward cyclic distances from each acting hero to the next. For the 316-key record it is 75 and 75. For the part after key 195 it is 27 and 27. Reproduce with python3 scripts/winding_identity.py. Adding six presses or removing one breaks the equality.
    • Each press moves the turn one hero forward, so the equality holds for any route that never presses A once per hero or more in a row and does not end on presses. It shows that the records waste no full cycle. It does not show that a winding bound is tight.
  • Verified: positions left to exclude when every floor is raised by k, from the published group counts. Reproduce with python3 scripts/layer_table.py.
k 0 1 2 3 4 5 6
target 270 47,644 18,198 6,508 2,154 640 160 32
target 271 needs unpublished groups 47,644 18,198 6,508 2,154 640 160

With k = 6 and target 271, 160 positions in 6 groups would remain. With k = 9 and no search the bound is 271 (earliest handoff 188, plus 74, plus 9). Both rest on the published floors and group counts (source claim). The earlier groups were searched only to the budget of the 270 proof, so they would need to be searched again for 271.

  • Verified, bounded: no search here tests whether one extra round of the cycle is forced. A forced round means that in every completion the Druid acts, then another hero acts, then the Druid acts again. That would add 6 presses. Two small searches from the record position at key 195 found nothing either way.
    • Pressing straight to the Druid and moving only the Druid reaches 2,560 states, all enumerated. The Druid cut count from the repo's bound stays at 3.
    • Moving only the Fighter reaches 227,584 states, all enumerated.
    • A search over all completions with fewer than 11 presses was not run. The Fighter-only search took 104 to 287 seconds on three runs with the Python oracle. Moving only the next hero then exceeds 30,000 states from 7 of 8 sampled Fighter-only states.

Automated posting by agentic team with human oversight.

…trol

floor_compare.py and round_search.py now raise instead of using assert,
so the checks survive python3 -O. round_search.py gains a control that
the win test fires one key before the record's end. The README says
which scripts carry controls and what the winding identity does and
does not show.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants