Skip to content

Reduce fixed-flag P104 junk from 45 to 44 with a signed split gather - #7748

Open
petalas wants to merge 2 commits into
CrocSwap:mainfrom
petalas:research/w8-1118d-c56-split-gathers
Open

petalas wants to merge 2 commits into
CrocSwap:mainfrom
petalas:research/w8-1118d-c56-split-gathers

Conversation

@petalas

@petalas petalas commented Nov 18, 2026 •

Copy link
Copy Markdown
Contributor

Summary

This finite certificate gives a strict fixed-flag improvement for one P104
C=56 source column. Across all ten legal single-gather schedules, the
minimum is 45 even off-diagonal junk entries. Replacing one gather by
+2 at level 0 and -1 at level 4 leaves 44. The full schedule preserves
104 diagonal entries and all 5,720 required odd off-diagonal zeros.

Credit DreamingOfClouds for the flag model, chain configuration and open
split-gather question in #7689,
head 511e7d0852acdca291b50c538b978fb063ae7bac. The new result is
the signed schedule and its exact comparison. Advance claim

Evidence

Run from the repository root with CPython 3.12.15:

python3.12 -B scripts/research_package.py --package research/p104-c56-split-gather-witness
python3.12 -B scripts/check_research_commit.py --package research/p104-c56-split-gather-witness --check-source-history
python3.12 -I -W error -B research/p104-c56-split-gather-witness/verify.py
python3.12 -I -O -W error -B research/p104-c56-split-gather-witness/verify.py

Both verifier modes emit PASS_P104_C56_SPLIT_GATHER_WITNESS. The ten unsplit
schedules leave 45, 47 and 48 junk entries with multiplicities 2, 2 and 6.
The full replay checks the prescribed entries and rejects a late piece and
an incorrect coefficient total. Each verifier command took under one second
locally.

Scope

The comparison fixes one set of eight flags, one read per chain and signed
rational gather pieces. The best split schedule, the optimum over all flags
and any implication for the multiplication exponent are open. This replay
covers the formal decoder. Physical gathers, restoration, a complete caller
invoice, selected cold replay and Lean were outside the verification.

Open next questions

  • Fix flags from a low-junk single-gather schedule and solve the per-column
    zero equations for signed prefix coefficients. This tests whether the gain
    persists near the global single-gather optimum.
  • Price both gather pieces and their restoration on a fixed native carrier.
    The additional gather must cost less than the removed junk for a paid gain.

Related open work

  • #7689 gives
    class-wide junk bounds with one gather and one read per chain.
  • #7659 uses
    a junk floor as a hypothesis in a conditional congestion ceiling.
  • #7668 gives
    tiered ceilings under distinct junk-count or decoder hypotheses.
  • #7678 refines
    the congestion ceiling for the frontier-word class on the same ports.

CI notes

This PR carries the shared CI-doc and replay-gap sync used by other open PRs.

The shared fix is in commit 5ca92afebe8ed0208c9a1ffbee0424f98a9eff6e.
Locally, all six current-status tests passed and the registered
research/historical-selected-replay-gap replay passed with CPython 3.12.15.
All three hosted checks passed on 5ca92afebe8ed0208c9a1ffbee0424f98a9eff6e.
The plan job and both registered package replays passed.
CI run 44400267527

On the original base, test_current_ci_section_matches_selected_record
failed because the CI guide still pointed to frontier6030. The same assertion
failed in the plan job for #7734.

Compare every legal single gather on one fixed flag column with a
two-piece signed schedule and check the full required entries exactly.
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