Skip to content

P104: per-reader endpoint charge, κ ≤ .0115428 for words meeting its hypotheses (Lean, conditional, no κ claim) - #7685

Open
DreamingOfClouds wants to merge 2 commits into
CrocSwap:mainfrom
DreamingOfClouds:research/p104-reader-ceiling
Open

DreamingOfClouds wants to merge 2 commits into
CrocSwap:mainfrom
DreamingOfClouds:research/p104-reader-ceiling

Conversation

@DreamingOfClouds

@DreamingOfClouds DreamingOfClouds commented Nov 18, 2026 •

Copy link
Copy Markdown
Contributor

Lean-only package, builds on #7678 (the tree ceiling, t0 = .0116750): the four #7678 modules the new files import (CTreeCore, CongestionTreeCore, CongestionTreeW, CongestionTreePiece) and the 46 modules of #7668 are included unchanged; 274 new modules. No κ claim and no word claim in Lean: every object in the three theorems is abstract, and nothing in Lean ties a word to the hypotheses.

Results. Three theorems, each proving, for t ∈ [t0, 1] (that is, t0 ≤ t ≤ 1), Finv (1 − t) (invoice (caller 6 7) n 17) < 0 (caller caller 6 7, L = 17) under the hypotheses listed in PROVED.md (copied from the Lean text). They differ in the bound on one named hypothesis, ω ≤ w0 (the net G'-token rank that owned rows absorb), which sets the allowance NDL = 15 + w0:

  • ceiling_reader_GD_w10 (headline, CongestionReaderW10): ω ≤ 10, t0 = .0115428 (a0RF = 28857/2500000).
  • ceiling_reader_GD (CongestionReader): ω ≤ 0, t0 = .0115262 (a0RD).
  • ceiling_reader_GD_w16 (CongestionReaderW16): ω ≤ 16, t0 = .0115528 (a0RE).
    For comparison ceiling_tree_GD (P104: node-partition refinement of the frontier-word ceiling, κ ≤ .0116750 (Lean, conditional, no κ claim) #7678) is .0116750. On [.0118103, 1] each applies ceiling_native_J with its own hypotheses; on [t0, .0118103] it uses two or three kernel-certified reader_pieces. Certified digits are 1.1e-6 to 1.2e-6 above the paper digits (.0115417, .0115250, .0115516), because t is quantised to 1e-7 and the line constants and weights are rounded outward. The paper digits are reproduced by the master-LP script (.011541666, .011525001, .011551596). Lean proves: the register identity and tied_rho (a register tied to a reader has residual at least delta = 7^s + 2^s − 8^s − 1), the payer summation sum_c2_pay, the reader line bound tree_lower_reader, the piece lemma, soundness of six 230-row DP tables (one decide +kernel per row) and all per-piece numerics. All theorems print only propext, Classical.choice, Quot.sound; no sorry, no native_decide; every new module was compiled fresh (CHECK.txt).

What the charge is. Each reader (104) that is tied to a distinct public register (Z-class, 0 → 8, or V-class born at dim 1, ending at 9, with a stop at a dim in [2, 7]) adds delta to the budget; the readers that are not tied and not fully transported (at most 5 mixed registers, at most 10 owned last movers, plus ω) are priced by an allowance of delta each.

Scope. One 104-port set. Five word files share it (st1, #7637, x1t2a, #7660, and t7669, our reproduction of the word of #7669, not that file); the word-level numbers in NATIVES.md are for these five.

Hypotheses. All in PROVED.md. The 32 explicit binders shared with ceiling_native_J are textually identical to that theorem and to ceiling_tree_GD (compared by script). New or changed: T : TreeData (now the readers' tables), hcstLo, hcstMonoT, hcst0, hcst1 (and hcst2 in ceiling_reader_GD), hP1, hP2 on the new window, ω, hω, Q, Uo, hUo, hUc, hQ, hcase, π, hπR, hπinj, hTied.

  • ω ≤ 10 is proved on paper but is a named hypothesis in Lean. Premises of the paper proof: no owned register rises from dim ≤ 7 into V, every owned lifetime ends in a CLEAR; both hold on all five words, where ω = 0 (exact replay modulo two primes). The CLEAR certificates are not wired into this package.
  • hTied, hcase, hUc: open (paper). tied_rho is proved, so the residual bound is not a hypothesis; the existence and uniqueness of a tied register per tied reader is. On all five words every reader has a Lean-Tied register on its final chart and a maximum matching of readers to registers has size 104 / 104.
  • T (IN side) is open. The IN DP table is the reflection of the OUT table (paper); the Lean-kernel IN table rfiR differs from gbR in the edge masks of 89 of 230 rows (dim, n, m agree in all 230; the numeric g/ge lists differ in 225 rows); rfaR and gaR have identical masks (corrected; an earlier version of this text said 18 of 230). After opening, the IN-side reader trees were extracted from all five words and every word-checkable IN-side premise passes (two primes; see NATIVES.md); the OUT-side tree premises of P104: node-partition refinement of the frontier-word ceiling, κ ≤ .0116750 (Lean, conditional, no κ claim) #7678 were re-checked on all five words (both primes).
  • hcstLo : 0 ≤ cstF on [a0, 1/10] is a hypothesis; on paper it follows from cst(t) ≥ c2(7, 8) > 0.

Caveats, stated plainly.

  • The F2 lemma as first posed (a paying register inside the reader's visit set) is false: a five-register token-swap counterexample exists. The statements use the corrected count form (payers ≥ N_high − 15 − ω), proved on paper for the class; Lean uses only the existence of a tied register per tied reader and the count bound on Uo. The clause "only public registers hold G' tokens" is false on the natives (15 to 28 owned registers hold tokens) and is not used.
  • The 5-allowance paper figure (.0115086) is not formalised: owned moves can end at dim-8 charts without a public payer, so 10 owned last movers are added (NDL = 15).
  • The word-level premises are computed by scripts that are not in the package (NATIVES.md: premise × word table, methods, which were re-run). Nothing in Lean ties a word to the hypotheses. hmix = 5 is tight on all five words; the five words are close to one data point (one port set).
  • The Lean-certified digits are weaker than the paper digits by about 1.1e-6 to 1.2e-6.
  • The docstrings of the final modules had a wrong digit for t0 in the working copy; corrected in this package (comments only).
  • CHECK.txt: all 274 new modules were compiled fresh, one at a time under a 3.5 GB guard. The 50 inherited modules were not rebuilt; their oleans are those of the build of P104: node-partition refinement of the frontier-word ceiling, κ ≤ .0116750 (Lean, conditional, no κ claim) #7678's package (same sources).

The premises under study are SYLforge's (#7616, #7620); the caller is rohanarun's (#7113). DreamingOfClouds with Claude Opus 5.5 (Anthropic), via Claude Code.

🤖 Generated with Claude Code

DreamingOfClouds and others added 2 commits November 18, 2026 02:57
… its hypotheses (Lean, conditional, no kappa claim)

ceiling_reader_GD_w10 (t0 .0115428, omega <= 10), variants w0 = 0 (.0115262) and w0 = 16 (.0115528); kernel-only; builds on CrocSwap#7678. IN-side trees and omega <= 10 are hypotheses in Lean; word checks in NATIVES.md.
DreamingOfClouds with Claude Opus 5.5 (Anthropic), via Claude Code.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
rfiR vs gbR edge masks differ in 89 of 230 rows (not 18 as first stated). The IN-side reader trees and word-checkable IN premises now verified on all five words. No Lean change.
DreamingOfClouds with Claude Opus 5.5 (Anthropic), via Claude Code.

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