Skip to content

Reduce the forward native snapshot pool from 2,425 banks to 2,106 - #7724

Open
petalas wants to merge 3 commits into
CrocSwap:mainfrom
petalas:research/w2-1118b-frontier-step
Open

petalas wants to merge 3 commits into
CrocSwap:mainfrom
petalas:research/w2-1118b-frontier-step

Conversation

@petalas

@petalas petalas commented Nov 18, 2026 •

Copy link
Copy Markdown
Contributor

Result

A legal serial order of the #7674 forward native graph from aliromman uses 2,106 snapshot banks, saving 319 against the fixed-order minimum of 2,425 in #7681. Including the shared paid-row accumulator and four scratch fields, the private reservation falls from 2,430 to 2,111 homes per occurrence, or from 38,880 to 33,776 over 16 disjoint occurrences.

This is a finite allocation certificate with a scoped proof of operator preservation. Among ready rows, the schedule chooses the one that will finish the most donors' remaining reads. Each row retains its exact coefficient list, charts, child and adapters. The new bank is reserved before the first read, all donors survive through child return, and all outputs survive until public installation. The peak is attained at event 3,022, token 94,904. Proof and contract · Generated receipt · Advance claim

Evidence and reproduction

The replay checks 7,643 row events, 9,144 assignments, 169,343 donor reads and 1,501 final outputs against the pinned forward graph. Independent owner counts confirm the peak. Both a live donor/result alias and the assignment restricted to 2,105 banks are rejected. The new schedule pays for all 30,580 full initialization/retirement services per occurrence. The conservative 16-occurrence reset allowance is 11,816,384.

Run from the repository root with any CPython 3.12 interpreter.

py=python3.12
pkg=research/banked-native-row-order
"$py" -B scripts/research_package.py --package "$pkg"
"$py" -I -B -W error "$pkg/verify.py"
"$py" -I -B -W error -O "$pkg/verify.py"
"$py" -I -B "$pkg/verify.py" --certificate
"$py" -B scripts/check_research_commit.py --package "$pkg" --check-source-history --check-modes

The certificate command writes the TSV to standard output. Both registered commands passed locally in 0.62 s each on macOS arm64. Output was PASS_BANKED_NATIVE_ROW_ORDER 2106 banks, 2111 private homes, 169343 reads. The package and committed-byte, mode and source-history checks passed. The complete TSV hashes to 16de610956122d9f2c68f05d955b5484a0ee31bf201b69546a092495bf97dd86.

Limits and credit

The 2,106 lower witness applies to this order and this owner contract. A global schedule optimum and a concrete packed-tape implementation remain open. The proof uses the published row semantics and the five inherited framework assumptions. Physical child operators, the full caller, the selected cold replay and a formal proof were outside the local checks. There is no new κ.

This allocation was produced by Nick Petalas. aliromman posed the question and supplied the fixed-order allocation, fifteen pivots and field-level cost contract. SYLforge supplied the underlying word and caller. Clear1nsight supplied the public comparator encoding and digest. The package retains the original source notices. Credits and graph provenance

Related open work

Open next questions

  1. Which donors and terminal outputs must coexist in every legal topological order? A lower bound from such a cut would constrain further reductions.
  2. What packed backing and movement does this order require with the actual bank views, guards and suspended owners?

CI notes

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

All six current-status tests, the historical replay-gap package and the bank allocation package passed locally at ef52676eb20c252f5786f42a70ef0dd814400a03. The shared planning test on the base guide expected make frontier7000-verify and found make frontier6030-verify. The same assertion appears in the #7681 plan failure. All three hosted checks passed at ef52676eb20c252f5786f42a70ef0dd814400a03, including planning, the bank allocation replay and the historical replay-gap replay. CI run 44369322855

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