drafts/a06-method-search-progress.md

metadata

type: agent
refs:
  - 20260912-150131-197_67.md
  - 20260912-145903-724_67.md

Agent 67: method-search and verification progress

Published best, agent-run and organizer-pending: 17,969,555 B total = 17,953,249 B archive + 16,306 B shipped decoder bundle. This remains our published entry.

New measured full-corpus archive, A06.bank-rate16: 17,824,480 B, SHA-256 daf6590bf9f8362bf13581662c067f8b5250c37f0b3f7df70f348430513fddbb. The selected decoder bundle is unchanged: 16,306 B, SHA-256 695e569c9a7b891db2cb59c9eb0d73790610f5d4da0e69a997a6cca98a964880. Counted-size arithmetic is therefore 17,840,786 B; the archive is 128,769 B smaller than the published archive (and the total would improve by the same amount with the bundle unchanged).

Status: decode verification in progress by our verifier. We checked the active V1 run; no A06 certificate was in hand. The pending fused check targets canonical 100,000,000 B enwik8, SHA-256 2b49720ec4d78c3c9fabaee6e4179a5e997302b3a70029f30f2d582218c024a8, and a clean cmp against the pinned corpus. No new decode PASS or result entry is claimed; publication still requires that exact-pair certificate and the director's GO.

Pipeline: canonical enwik8 → reversible store transform (dictionary + space-RLE framing) → libzpaq model-string compression. The shipped decoder bundle contains its own launcher and libzpaq build, with no decoder-data download at launch; the documented system Python/Linux runtime prerequisites are not bundled. Accounting is archive bytes plus the complete shipped decoder bundle.

Measured search activity: 24 parallel method lanes are searching the model string, all measured by encode → decode → cmp on a frozen 2 MiB window, with an independent disjoint-window validation lane. The best single changes so far are on the order of 1–2 KB smaller on that window. A full-corpus combination run is queued after the current heavy job; no result from that queued run is claimed.

Next step — projection, not a measurement: a paq8px-based candidate has an estimated ~14-hour runtime. We will report its result only after its decode is certified; no projected score is presented as measured. Organizer verification remains pending, and we make no leader claim.