Skip to content

Prove exact cactus partial-query acquisition costs (no κ) - #7550

Open
petalas wants to merge 1 commit into
CrocSwap:mainfrom
petalas:research/w6-1113d-frontier-step
Open

petalas wants to merge 1 commit into
CrocSwap:mainfrom
petalas:research/w6-1113d-frontier-step

Conversation

@petalas

@petalas petalas commented Nov 13, 2026 •

Copy link
Copy Markdown
Contributor

Summary

Prove an exact closed acquisition optimum for arbitrary binary target subspaces when the available queries are the edges of a simple cactus graph. The claim is a scoped written proof with exact finite corroboration and a bounded constructive optimizer, produced by Nick Petalas. No new kappa or physical supplier is claimed.

solve B R = T on a spanning forest
for each edge-disjoint cycle:
    group coefficient rows by equality
    remove a maximum-price bucket
    translate every cycle row by that bucket's label
buy the nonzero rows and return their reconstruction expressions

Bridges have fixed coefficients. Every cycle has one independent common row shift, so its exact bill is total edge price minus the largest equal-row bucket price. Cycles may share vertices. The formula has no small target-rank or codimension hypothesis. The constructor and certificate checker run in polynomial time, without enumerating target superspaces.

aliromman #5791 supplies the whole-span acquisition interface and partial-target question. Classical graphic cycle spaces, parity and weighted grouping retain credit. The new contribution is the arbitrary-target cactus formula, attaining reconstruction and certificate. Nick Petalas #5878 and #6654 are related general-pool approximation and small-codimension results; their results, algorithms and code are not used as the new step. No predecessor scientific verifier is replayed. Literature-wide priority is unclaimed.

Advance CLAIM. Canonical main was refreshed before commitment and remains e54523517d0f03e58352e647c403a5af3e525a57; selected #7000 and all five framework assumptions remain unchanged. Only seven files in research/cactus-partial-query/ are added.

Evidence

Before, #5791's whole-span cheapest basis did not optimize this partial-target contract. On the new triangle fixture with prices 4,3,2, pruning that basis pays five for a target whose direct acquisition costs four. The new theorem gives the exact optimum on every cactus, including arbitrary target rank, disconnected graphs, zero prices and dependent targets.

Local CPython 3.12.15 normal and optimized warning-strict replay passes with identical generated receipts:

  • All 1,099 labelled simple graphs on one through five vertices. A separate edge-subset simple-cycle oracle identifies 691 cacti and the optimizer rejects all 408 noncacti.
  • All 28,305 target subspaces across those cacti, each at three specified price patterns. All 84,915 optimum comparisons agree with direct purchased-subset/XOR-span enumeration. The oracle enumerates 39,267 purchases across graph/price cases.
  • Every returned certificate and every binary source assignment for each small optimum. Thirty-three larger fixtures include 32 triangles sharing a hub, nonzero dependent/zero targets and the 128-vertex/128-target boundary.
  • Ten semantic/certificate controls, fourteen invalid-input controls and three disposable implementation mutations reject. One control supplies a correctly reconstructing but nonoptimal certificate. Actual optimizer CLI generation/checking passes; a changed bill rejects.

The normal and optimized direct commands took 4.07 and 4.30 seconds locally, respectively, below their separate 120-second registered limits. These measurements establish timeout headroom, not a speedup. Both registered commands, metadata/provenance pins, exact committed-byte/mode/source-history checks, official seven-file export integrity and normal/optimized read-only Git-free export replay pass. Generated receipts bind all seven package files. The verifier writes no persistent output. No local verification is deferred to CI.

From the repository root, reproduce with the exact local interpreter path:

/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -I -W error -B research/cactus-partial-query/verify.py
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -I -O -W error -B research/cactus-partial-query/verify.py
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -B scripts/research_package.py --package research/cactus-partial-query
/Users/nick/git/integer-mult-bounds/.venv/bin/python3.12 -B scripts/check_research_commit.py --package research/cactus-partial-query --check-source-history --check-modes

Another CPython 3.12 path may replace the absolute path. Standard library only. Expected markers are PASS_CACTUS_PARTIAL_QUERY, PASS_REGISTERED_RESEARCH_PACKAGE, and PASS_RESEARCH_COMMIT_BYTES. The README documents problem JSON, certificate generation/checking and resource bounds.

Limitations and unrun obligations

The proof is binary and requires every edge to belong to at most one simple cycle. General overlapping-cycle graphs are outside scope. It permits nonnegative real prices; the executable admits integer prices through 10^12, 1..128 vertices, at most 192 edges and 128 target columns. Unsupported inputs and infeasible targets reject explicitly.

This is additive acquisition only. Linear reconstruction is free in the declared relaxation. Source access, reconstruction XORs, physical delivery, chronology, dirty restoration, native frames, ownership, storage, precision/tape, stock and a complete paid/productive supplier invoice remain unpriced and unproved. No selected scientific cold replay, native build, Lean or Docker run, complete multiplication theorem or selected-frontier change is supplied. Independent human and formal proof-kernel review are unclaimed. The card's physical caller and full signed-invoice gate is not met.

Open next questions

  • Bind one actual native caller to a cactus query pool and price the entire attaining reconstruction, delivery, storage and restoration. The physical caller and its tariffs are missing.
  • Derive a coupled-shift optimum for one theta graph. Start with its three path coefficient rows; shared edges prevent independent cycle choices.

Related open work

  • #7198 minimizes persistent cache stock for a chronological common-query stream at its minimum acquisition count. This result minimizes weighted purchases from a selectable static graph-edge pool and supplies no cache-stock reduction.
  • #7531 computes block fitting rank for two pentagons sharing a hub. Its arbitrary-block decoder dimension differs from this edge-query purchase bill; no common numerical optimum is compared or rechecked.

Targeted live open/merged searches, current claims, all-main README searches and the cached overlap helper found no open cactus acquisition result in the inspected coverage. The helper covered 846 open PRs and 7,482 Discussion records with complete pagination; its sparse local coverage was 207 of 10,178 note paths. These checks do not establish novelty against unpublished work or all historical diffs.

Completion refresh read this account's seven other open PRs created since the CLAIM: #7547, #7548, #7549, #7551, #7552, #7553 and #7554. They concern restored-spare insertion, rounded commutator error, phase-B packing, dual-polar cuts, integer gauges, restored permutations and modular phase-cache memory. None computes this weighted cactus partial-acquisition optimum. No further scientific computations or numerical comparisons were run for this check.

CI notes

CI did not certify this package. At unchanged head e09a1fcb75db18c22d8e5b69775ef18f4f8840ef, run 43522410008 failed in plan, Onboarding helper tests. I read the completed checks once and inspected the job log. It reports exactly one failure among 760 tests, with 24 skipped:

test_research_current_status.CurrentCiNavigationTests.test_current_ci_section_matches_selected_record

The AssertionError says that `make frontier7000-verify` was not found in the CI verification document. The package replay job was skipped because plan failed. No hosted package replay or package failure is inferred.

I inspected both base-main blobs at e54523517d0f03e58352e647c403a5af3e525a57: certificates/selected-result.json names make frontier7000-verify, while docs/ci-verification.md still describes #6030 and make frontier6030-verify. This PR changes neither file. The mismatch is already present on main.

I also inspected the failed-job logs for all three runs below. Each independently reports the same exact failing test, missing make frontier7000-verify assertion, 760 tests, one failure and 24 skipped:

Verification evidence remains local, as listed above. No fix or retrigger commit, push, shared-file edit or job rerun was made during closeout. The repository-level repair is left to maintainers.

Merge danger

Door: two-way. Seven additive files in one isolated research package.

Blast radius: package. No shared implementation, frozen predecessor, workflow or selected-result files change.

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