Author: @SecurityQQ (Alex), with OpenCode/AI assistance
Date: 2026-09-07 (UTC)
Problem: 16x16 matmul
Cost: 63,819
This improves the current 64,074 record by 255 (0.3980%), the previous unmerged 63,847 submission by 28 (0.0439%), and the earlier 64,431 published record by 612 (0.9499%). These are abstract Dally v0 operand/output read costs, not measured runtime or joules.
The classical 4,096 scalar products use asymmetric column panels of widths
6 and 10, row groups 3/3/3/3/4 and 5/5/6, and chunks 1 and 2, respectively.
The third tile (rows 6..8, columns 0..5) reverses its initial column traversal;
subsequent chunks still snake. Two dependency-safe SSA block moves relocate
the two-operation block starting at destination 9897 before destination 9893,
and the four-operation block starting at destination 8408 before destination
This is a separate schedule branch, not a postprocess of 64,074 or the 5+11-panel 63,847 submission. The other 63,847 -> 63,831 branch is not its parent. Fresh whole-program DP allocation and pair-tier refinement precede alternating temporary input captures, cheapest surviving equivalent-source reads, and physical address allocation:
| Stage | Cost | Accepted captures |
|---|---|---|
| Schedule seed, fresh DP | 64,336 | - |
| Pair-tier refinement | 64,289 | - |
| Primary cycle 0, captures then allocation/replay | 63,996 | 165 |
| Primary cycle 1, captures then allocation/replay | 63,897 | 90 |
| Primary cycle 2, captures then allocation/replay | 63,854 | 43 |
| Continuation cycle 0 | 63,825 | 28 |
| Final six single captures | 63,819 | 6 |
The last six strict improvements score 63,824, 63,823, 63,822, 63,821,
63,820, and 63,819. All capture destinations are address 1. Their recorded
boundary/value pairs are 3840/265, 4756/57, 5089/441, 5282/29,
5445/473, and 6267/149, interpreted against the evolving batch.
The seed includes 128 generator B captures; the later stages add 332 captures.
This is an allocation/capture heuristic, not reinforcement learning.
| Instruction | Count | Read cost | Change from 64,074 |
|---|---|---|---|
copy |
1,932 | 20,639 | +592 |
mul |
4,096 | 18,477 | -355 |
add |
3,840 | 20,256 | -495 |
| Output exit | 256 | 4,447 | +3 |
| Total | 9,868 instructions | 63,819 | -255 |
All 512 inputs and 256 outputs are checked exactly. Maximum address is 631.
Each operand and output read pays ceil(sqrt(address)); arithmetic and
writes are free, and reads precede writes.
The research manifests record the following bounded certificates on the final SHA-256 below, with capture destinations restricted to addresses 1..631:
These certificates concern the tested single-edit neighborhoods and frozen placement, not arbitrary edits, joint changes, or global optimality. The earlier 63,825 stage was explicitly not converged; its status must not be confused with the later final-hash single-capture/deletion certificates.
Recorded timing scopes are separate: the schedule primary search took 442.82648 s wall; primary plus continuation postprocessing took 427.01034 s wall; the final combined search took 479.19434 s wall / 478.12483 s process CPU. Its separate research verification took 181.55947 s wall. These are recorded research timings, not timings for the submitted verifier, not an end-to-end total, and not evidence of exhaustive optimization.
From the repository root, using only the Python standard library:
python3 -S matmul/submissions/best_63819.py
python3 -S -m unittest matmul.test_record_63819 -v
The verifier resolves repository files relative to its own location, so it also works by absolute path from another directory. It checks the byte hash, unchanged official symbolic scorer, independent exact noncommutative-polynomial proof, operation counts, and full read-cost breakdown. It uses only repository code and the frozen IR, with no external workspace, OR-Tools, or Torch dependency.
This is an artifact submission, not a constructor. No portable generator,
search implementation, or byte-identical construction replay is included.
The supplied verifier does not rerun search or certify the search neighborhoods.
The following inspected research evidence lives in the external
matmul-baseline/ workspace and is not a bundled reproduction capability:
structural_schedule_v1_report.md and
structural_schedule_v1_results/winner_00.json: full 6+10 schedule and
pair-refined seed SHA-256
7989b88f2beb8f425107615baac5dca327e392204ce7e0af4c2bec84ca4e2b9a.structural_schedule_post_v1_REPORT.md: the 64,289 -> 63,825 chain,
primary and continuation replay evidence, and timing/coverage limitations.combined_pr62_v1_results/results.json and combined_pr62_v1_summary.json:
the 63,825 -> 63,819 events, final-hash scans, deletions, incomplete relocation,
and allocation probes.combined_pr62_v1_verification.json: byte-identical research replay of
seven events (initial plus six captures), ten allocation probes, and repeated
input, all-value, and deletion certificates. This depended on external
research helpers and is not a replay performed by the submitted verifier.Source artifact: combined_pr62_v1_results/best.ir.
SHA-256: e8adc50783a64088e6864ecfccb7e18a74814e24cfd6e1f9e3045e61fb37f2df.
The 63,825 parent SHA-256 is
963d9027a89ec73a8c4618a579fe34dbe0dc51c932ae737ec03c990a45965b22.
The prior asymmetric-panel, persistent-B-capture, and allocation work is by @cosminscn and Codex; see the 64,431 report. That report credits @jurajselep for snake traversal and @sigkillme0 for the 66,178 record and exact fixed-order allocation analysis. The independent polynomial proof is reused from the latter submission. The current 64,074 record by @SecurityQQ with OpenCode assistance remains preserved, as do all earlier records and the scorer.