Skip to content

P104: the odd unit, paid per grading — one cost behind the decoder wall and the t = −1 park (no κ) - #7606

Open
DreamingOfClouds wants to merge 2 commits into
CrocSwap:mainfrom
DreamingOfClouds:research/p104-odd-unit-per-grading
Open

DreamingOfClouds wants to merge 2 commits into
CrocSwap:mainfrom
DreamingOfClouds:research/p104-odd-unit-per-grading

Conversation

@DreamingOfClouds

Copy link
Copy Markdown
Contributor

Summary

This package collects one day of P104 research on the two fixed costs every redesign has hit: the decoder's 56 owned units per plane and the t = −1 park. It is organised in three parts:

No κ claim. No new frontier word. Package: research/p104-odd-unit-per-grading/.

One cost behind both walls. Both costs are the same odd unit, paid separately in each port's own grading σ_p.

  • o(N) = rank of N's σ-odd part, for any block, ring or port coupling.
  • t = −1 holds exactly when the σ-even part is singular (the Cayley form of the park).
  • Two ports never share a grading inside a word (05-layers.tex l.4–7, 86–91).
  • The odd diagonal (q·q = 1) forces both an UP and a DOWN decoder in every plane.
  • Every construction that tries to move the park elsewhere needs a moved owned row, about 1 unit per port per call against a break-even of 0.214. This is enforced by the Rank, Lock and Clone lemmas, Theorem W, NT, AS/VA, the lift law, the Freeze Lemma, the web-general Cut Lemma and the Commutation Theorem.

Main results, with scope

result scope tag
o(N) = rank Odd(N) any size, ring or coupling P + C
Cayley form: ker(R + I) ≅ ker E, so t = −1 ⇔ the σ-even part is singular any contract P + C
t_eff = ab/(cd − ab) per port P + C
M45 = HOST(−1, ½) up to one free SCALE, i.e. the t = −½ B-port contract P + C
The ledger is per grading; the isotropic q_a + q_b kernel is even in both gradings P + C
ρ = 1 forces the park root callers P
Rank lemma: no event raises full-state rank; a dirty CLEAR is lossless iff its content is in span premise-free P + toy C
Lock Lemma, Theorem W (separate twins need ≥ 96 moved owned rows/call), Lemma NT, Lemma AS/VA, lift law, Freeze Lemma P + toy C
One lift serves ≤ 8 pairwise-orthogonal ports P + C
Commutation Theorem: walls hold for any enlarged same-frame op set P + toy C
XRECALL lemma, Zone lemma, owned additivity (c ≥ 0.92/port/call under any sharing), wiring lemma P
Cut Lemma: rank(O mod Π) ≥ def(𝒢) for any up-set of frames and any helper web validated on CubeJerry's W6805 (def 0, 104/104) and a t = −1 feeder word (def 1, 104/104) P + C
Cut Lemma corollaries: every t = −1 port needs an owned ascent leaving its Y-end hyperplane; descending decoders + line feeders ⇒ M ≥ 104; M ≥ |F| + α(R) with α(P104) = 8 P + C
UP + DOWN decoders forced at t = −1, so every decoder saving also applies at t = 0 P
G/2 = 13·proj(range G); the full UP∘DOWN composition leaves exactly −12·I on diagonal + odd cells C
Joint multi-call decoder rank ≥ 8k P + C
Lemma A(acyc): rank certificates are blind to acyclic junk P
Lemma B (mod-4 certificates) half-integer rows P
Early-set cap |E| ≤ 5,408 − 43,264/(d + 8) d ≥ 2 closed; d = 1 open P + C
Theorem F, Lemmas LIN, TR, 2B, Descent, RT local re-solve on a fixed schedule P + C

Corrections of our own earlier statements

  1. The "one-rank law" is the Fredholm alternative (Theorem F). Our drop-repair engines reported "every drop misses by exactly one rank". With one right-hand side that is only "infeasible". The contact-minimality conjecture we built on it carries no structure, and the rigidity it described is schedule-driven, with no lower bound (Descent Lemma).
  2. HALF18 (already commented on P104 decoder wall: row price, partial-replacement bounds (k ≤ 5), half-wall for h ≤ 6 (no κ) #7583). The full-flag certificate minimum is not ≈ 18 = 2h. Exact frame certificates of rank 15 (E8-120) and 13 (P104) exist, with lower bound 8. P104 decoder wall: row price, partial-replacement bounds (k ≤ 5), half-wall for h ≤ 6 (no κ) #7583's proved bounds are unaffected.
  3. Internal notes only, never in a public PR. Duality is type-preserving: U builds are not duals of L builds. EARLYCAP part B is refuted by T1CANCEL. BYPASS's relaxed screens were optimistic.

How to verify

python3 -I -B research/p104-odd-unit-per-grading/odd_unit.py   # PASS_ODD_UNIT_PER_GRADING (<1 s)
python3 -I -B research/p104-odd-unit-per-grading/e8laws.py     # PASS_E8_LAWS (~3 s)
python3 -I -B research/p104-odd-unit-per-grading/walls.py      # PASS_WALL_TOYS (~70 s)
python3 -I -B research/p104-odd-unit-per-grading/onerank.py    # PASS_ONERANK (~3 s)
python3 -I -B research/p104-odd-unit-per-grading/cuts.py       # PASS_CUT_LEMMA (~40 s)
python3 -I -B research/p104-odd-unit-per-grading/words.py      # PASS_UNPAID_WORDS (~80 s)
python3 scripts/research_package.py --package research/p104-odd-unit-per-grading
word result
SEEDBUILD S1 +21.34 D0/call, κ ×0.999069
relocation control +16.82 D0/call, ×0.999268
Lemma RT swaps (mains 6 and 886) κ identical to #7546 in every digit
WINDOW8 merged window X = x − y, Y = x + y exact on 8 ports; 201.12 D0/call on the window route; no κ (merged plane)

Caveats and open questions

  • Lane-only numbers. Numbers tagged [C-lane] are exact computations by our lane code and are not re-run in CI. [C-param] prices come from our invoice tool, whose control reproduces Conditional κ = .003648411294790 (+0.0121% over #7544; +0.0239% over #7542): fuse hosted role902 moves #7546's κ and S = 1424. Neither is a new claim against the frontier.
  • Open. Whether one owned ascent can be shared between t = −1 ports. Up-set cuts allow it, and no realisation is known; an exact F2³ decision is in progress and not claimed here.
  • Open. The d = 1 early-set window.
  • Open. ≥ 6 replacement rows and the h = 9 wall (C ≥ 52).
  • Open. A rank ≥ 9 free-lift decoder network below 56.
  • Not included. A network-compiler effort is ongoing; nothing from it is included.

Credits

— DreamingOfClouds with Claude Opus 5.5 (Anthropic), via Claude Code

🤖 Generated with Claude Code

Research package research/p104-odd-unit-per-grading: one cost (the odd
unit, paid per grading) behind the decoder wall and the t = -1 park.
Part I lemmas with exact stdlib checks; Part II scoped negative results
by route; Part III unpaid words replayed exactly against CrocSwap#7546.
Corrections: the one-rank law is the Fredholm alternative; HALF18 rank.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@DreamingOfClouds

Copy link
Copy Markdown
Contributor Author

CI note: the only failure in the plan job is test_current_ci_section_matches_selected_record (1 of 760 tests), which fails on pristine main e545235 because main's own docs/ci-verification.md is stale (same failure explained on #7601). This PR adds only research/p104-odd-unit-per-grading/ and touches no docs or tests. Locally, python3.12 scripts/research_package.py --package research/p104-odd-unit-per-grading runs all six checkers and prints PASS_REGISTERED_RESEARCH_PACKAGE (outcome=negative-result). The replay job was skipped because plan failed; I will rebase once main is fixed.

— DreamingOfClouds with Claude Opus 5.5 (Anthropic), via Claude Code

@DreamingOfClouds

Copy link
Copy Markdown
Contributor Author

Update on the open sharing question (listed as "in progress" in the README). The smallest case is now decided.

Setting: F2^3, labels {1, 2, 4, 7}, t = −1 targets X = x − y and Y = x on all four ports. Public registers ascend, helpers are arbitrary in number with tokens, owned registers are token-free with free unmoved snapshots and free dirty CLEARs.

Result [proved, helper-count-free]. Every owned set of total climb 3 is infeasible, all 8,261 of them, including the 13 sets that pass every up-set cut. The toy therefore needs owned climb ≥ 4 = n, so owned ascents cannot be shared here.

The proof uses three new lemmas, each stated in our own notation:

  • Forwarding. An X birth line passes all its content upward, except for at most one line per owned move that leaves {0}.
  • Union. This is the multi-commodity version of the Cut Lemma. A port gains no slack from a union of up-sets when each piece has zero slack.
  • Duality. An owned set (a → b moves) is feasible if and only if the set of dual moves (b^⊥ → a^⊥) is feasible. Transposing the network exchanges X and Y.

Check: an exhaustive certificate over all 460 up-sets of F2^3. Controls behave as expected: the 4-feeder set replays exactly, and the no-owned set fails, as Corollary C1 requires.

Scope: on F2^4 these lemmas also exclude climb (n−1) for several label sets, but not all. Some climb-(n−1) owned sets survive, for example labels {1,2,8,7}, {1,2,4,14} and {1,2,7,11}, plus two owned {0}→V hubs. Excluding them would need an invariant at the helper-token level. So M ≥ n on P104 (S ≤ 712 for every one-plane design) remains a conjecture. The proved P104 floor is still Corollary C2 (M ≥ 104 for the descending-decoder family). This comment does not change any claim in the package.

— DreamingOfClouds with Claude Opus 5.5 (Anthropic), via Claude Code

…ing only)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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.

1 participant