Skip to content

P104 decoder wall: row price, partial-replacement bounds (k ≤ 5), half-wall for h ≤ 6 (no κ) - #7583

Open
DreamingOfClouds wants to merge 3 commits into
CrocSwap:mainfrom
DreamingOfClouds:research/p104-decoder-wall
Open

DreamingOfClouds wants to merge 3 commits into
CrocSwap:mainfrom
DreamingOfClouds:research/p104-decoder-wall

Conversation

@DreamingOfClouds

Copy link
Copy Markdown
Contributor

Theory package with no κ claim and no new word. It gives lower bounds on the owned mass of the P104 decoder: the 56 owned units per plane (8 rows × climb 7) that every P104 word pays. It works in the row model, where each rank-1 decoder term is one owned chain priced by its climb. The owned deletion bound and the contract classes are cited from #7555, not repeated.

Tags: [P] proved in PROOF.md · [C] exact check, re-run in CI · [C-pin] exact lane run, pinned summary only · [N] search evidence · [open].

Results

  1. Row price [P]. A row's climb is ≥ rank_F2 of the label pairing over its (deposit, gather) supports, and equality is attained. On odd cells, any chain is exactly rank 1.
  2. The P104 decoder [C]. Its 8 rows sum to G/2 exactly. Each row has |D| = |S| = 52, label spans 8 + 8 and price 7, for 56 in total. The G is the PR968 Gram, read from the repository file.
  3. Rigidity [C]. Each port's deposit vector is unique: nullity 1 at all 104 ports.
    • Exactly 16 label hyperplanes have outside ports of real rank ≤ 7.
    • No codim-2 label subspace does (all 43,435 checked, also after deleting any one port).
    • So every E8-world row costs ≥ 7, and a rank-8 decoder costs exactly 56.
  4. Replacing 4 of the 8 rows.
    • ≤ 3 rows: impossible for all 70 R, at any climb [P + C].
    • Exactly 4 rows: cost ≥ 28 for all 70 R. Propagation forces all ports for 62 R, and all but one port for 8 R [P + C].
    • Exactly 5 rows: cost ≥ 28 for both R orbits [P + C-pin].
      • The full check ran a C++ enumerator over a monotone cover of 27,423,618 closures, with refinements and 0 failures. It is too heavy for CI and not re-run there. The source is in lane/ and the summary is pinned.
      • fiverow.py re-implements the method independently. It re-derives the flats exactly, validates all start rectangles, checks 80 random codim-5 closures, and reproduces 30 codim-7 failures exactly.
      • The two-orbit claim for R is quoted from our lane notes, not re-checked.
  5. Cross-pair identity [P + C]. (2T_R ∘ Q) mod 2 is a sum of cross pairs, and its F2 rank is exactly |R|(8 − |R|), checked on all 255 R. So a half-integer replacement of R costs ≥ |R|(8 − |R|).
  6. Graded 2-adic rank bound [P]. n ≤ Σ_k (rank of the row block mod 2^(e_k+1)). With halves this gives C ≥ 2n/(h+1), which shows that plain Smith-rank counting cannot reach 56.
  7. Rank certificates [P + C].
    • Certificate lemma: n ≤ Σ_k rank Z[D_k,S_k], for any Z supported on the odd cells plus the diagonal, with a nonzero diagonal.
    • Bounds by label rank h: C ≥ 2n/3 for h ≤ 4; λ5 = 5291/2845 (C ≥ 0.5377n) for h = 5; λ6 = 1395/697 for h = 6, which with integrality gives C ≥ ⌈n/2⌉ for distinct labels of rank ≤ 6.
    • Functional-split certificates provably cannot reach h = 9. On the h = 7 full flag, the minimum over all 29,212 splits is 15, so λ ≥ 15/7 > 2·255/254. Pure clique partitions fail too.
  8. Open, stated as open.
    • ≥ 6 replacement rows: reduced [P] to one union-rigidity lemma on 252 cheap-row profiles. Adversarial and random searches always gave β = 7, which is evidence only [N].
    • The h = 9 wall: C ≥ n/2, i.e. C ≥ 52 on P104 labels.
    • The conjecture C ≥ 104 − 48·J/4,992, where J is the number of nonzero junk cells.

Checks

  • Four standard-library checkers, all our own code: rows.py (~20 s), partial.py (~50 s), fiverow.py (~17 s), halfwall.py (~1 s).
  • Data is pinned by SHA256, and the PR968 decoder file is read from the repository.
  • The two Pin the historical selected replay to its inspected commit #7257 CI-prerequisite commits are cherry-picked with their authorship kept.
  • scripts/research_package.py --package research/p104-decoder-wall reports PASS (PASS_REGISTERED_RESEARCH_PACKAGE, outcome=partial).

Credit the PR968 authors for the decoder file, Clear1nsight for the #7334 word family that the port data comes from, and ToikY and the price/replay lineage for the conventions used.

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

🤖 Generated with Claude Code

Clear1nsight and others added 3 commits November 14, 2026 03:51
This is the exact documentation prerequisite retained by CrocSwap#7156. It points the onboarding check to the selected CrocSwap#7000/CrocSwap#7001 review. Credit to Rohan Arun and the source authors.
The PR CrocSwap#6292 rejection census and package plan are statements about 2485805. Load its exact sources and planner in a private sparse checkout, keeping all source hashes, 19 expected rejections, nine package selections and the empty-plan control. Subsequent edits to the live Makefile and CI guide no longer invalidate this historical check.

The original audit is by unpaidmoon, building on Dennis’s historical replay gate. The pinned repository view follows the precedent of Matarener’s PR CrocSwap#5434. Repair prepared with substantial OpenAI Codex assistance.
…, half-wall for h <= 6 (no kappa)

Lower bounds on the owned mass of the P104 decoder in the row model: the row
price (climb >= F2 pairing rank) and the exact 8 x 7 = 56 E8 decoder with
rank-8 rigidity; replacing 4 of the 8 rows by <= 3 rows is impossible and by
4 or 5 rows costs >= 28 (five rows: C++ cover run pinned, independent stdlib
spot-check); the cross-pair identity and |R|(8-|R|) half-integer bound; the
rank-certificate lemma with C >= n/2 for distinct labels of rank <= 6, and
proof that functional-split certificates cannot reach h = 9. Open: >= 6 rows,
the h = 9 wall, the junk trade-off conjecture. Standard-library checkers.

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

Copy link
Copy Markdown
Contributor Author

Correction to README "The h = 9 wall [open]" (the [N] remark). The README says the numerical minimum full-flag certificate rank on the 120 E8 labels is about 18 = 2h, "which leaves essentially no slack". That numerical reading is wrong. The optimiser behind it fails its positive control at the sizes in question.

Exact counterexamples [C]:

  • E8, 120 labels. Partition the 120 root lines into 15 frames of 8 mutually orthogonal lines (an exact set cover; 15 = 120/8 is optimal). Let Z be the sum of the 15 block all-ones matrices. Then Z is 0 on every even cell (q·q′ = 0), its diagonal is 1, and its rank over Q is 15. This lies inside the partition family (Q∘Π) already in the PR.
  • P104's own 104 labels. The same construction gives 13 frames and exact rank 13. A lower bound of 8 comes from 8 pairwise-even labels. So the minimum full-flag rank on P104 is in [8, 13].

So the full flag is not the bottleneck: there are 5 ranks of slack against 2h = 18. The open question stays open, and it binds on intermediate flags. A sampled-flag LP over the families in this PR (Q, I, frame partitions, functional splits, merges/refinements) gives only λ ≥ 2.62 on P104 labels, against the λ < 2.039 needed for C ≥ 52. That is a lower bound for the families tried, not a statement about C. Nothing else in the PR changes; the proved items (row price, k ≤ 5, half-wall for h ≤ 6, split family can't reach h = 9) are unaffected.

The check is a few lines: build Z from the frame partition, then verify the even-cell zeros, the diagonal and the exact rational rank.

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

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