Submission: mikeg-cerebras

[mikeg-cerebras]

MIB Doc Challenge Submission — Track A Policy-Score

Published Artifact

The verified real-validation artifact is now published at public commit
c7cc98598b2f4beb2bf90a8bd01f9619679c6883, a non-force direct descendant of
the previous PR head db123b27793c409f2eadaf8a2559cc09c9588c03.

• rows / unique expected IDs: 5,000 / 5,000;
• missing / extra IDs: 0 / 0;
• strict schema and official validator: PASS / exit 0;
• bytes: 1,637,452;
• SHA-256:
  233cfea5a21462653e25b1744b4afdc194b8ec8efa0b3ccedbcb207e0ff79c70;
• public diff: exactly MEMO.md, SUBMISSION.md, and predictions.jsonl.

The JSONL was copied byte-for-byte from the immutable terminal artifact. It
was not regenerated, reordered, reformatted, or manually edited.

Pre-deadline Production Freeze

The production runtime was fully frozen before the August 3 Pacific-time
deadline:

• frozen source 5c70e9edb1cfd736f87f9bdab4ecf6acc45cb738, tree
  ebe6ffb25e73373f3d3e6e3c339fff1a6e19cf9e, committed at
  2026-08-03 19:07:13 PDT;
• exact image
  sha256:c8d238794c6ab7c1e3963fe7f588bf17df5428d39956d9352852322ffce16f0d
  (589402203 bytes), created at 2026-08-03 19:08:04 PDT;
• real container started from that image at 2026-08-03 19:57:57 PDT;
• runtime-bearing private main at
  2f1d3af600c78556baad5fee5f0b8c2eb2669302, committed at
  2026-08-03 21:43:03 PDT.

The real JSONL was atomically published at its actual time,
2026-08-03 23:59:18 PDT, and the public commit retains its actual timestamp.
No source, image, model, or inference-architecture change was made after the
freeze.

Evaluation Evidence

Protocol	Score / CFA	Evidence meaning
Grouped nested 5x4	133.62002305740833/CFA1	Primary generalization evidence; all five outer-fold deltas positive versus owner 133.32418873300010/CFA1
All-1,000 deployment fit	133.88708608605802/CFA1	Descriptive full-public fit after selection, not holdout
Exact-image replay B	133.88713553473560/CFA1	1,000/1,000; score non-regression PASS; zero extraction/adjudication changes
Reused 850/150 sensitivity	130.07338821722735/CFA0	Disclosed reused slice, not untouched holdout

Replay B has 19 confidence-only differences from the immutable semantic
reference and improves score by 0.00004944867758. Its internal exact-output
parity subgate therefore remains STOP_SHIP; this submission does not
claim byte-deterministic confidence. Execution, validity, field extraction,
adjudication, CFA, and score did not regress.

Terminal Runtime Evidence

• Generated truth-free 5,000 lifecycle: PASS, exit 0, 5,000/5,000 strict
  rows, 5680.879672278 seconds, no crash/timeout/OOM, observed kernel peak
  3202224128 bytes/21 PIDs and zero swap; synthetic lifecycle evidence only.
• Real validation 5,000 frozen verifier: PASS, wrapper/container exit 0,
  no timeout/crash/OOM-kill observed, 18082.435871 seconds
  (3.6164871742 seconds/PDF), exact source/image/container/output binding;
  terminal receipt SHA-256
  e0f8676d875db040abeec5b1a884ac9e117e40e42b097c616bc9d1310596c7fc.

The user-requested external telemetry samplers were stopped while the real run
was live. Their observed segment was clean—kernel peak 4006785024 bytes,
23 PIDs, zero swap/OOM/OOM-kill—but ended about 2 hours 44 minutes before
terminal. Full-lifecycle resource-sampling continuity and continuous
zero-swap are therefore not claimed. Output identity, strict schema,
coverage, official validator, Docker exit, and terminal binding independently
passed.

Coverage and structural validity do not imply a validation score. Private
validation truth was not opened or used for tuning, and no validation score
is claimed.

Architecture and Evidence Boundary

The CPU-only offline pipeline combines rendered OCR, typed evidence
resolution, the protected V5 plus fold-local PR44 semantic owner, globally
serialized PDFium, and an identity-free confidence-only policy-score model.

Native/selectable PDF text, page layout, PDF-object structure, visible
key-like text, and OCR are restricted to proposal, localization,
corroboration, conflict, and veto roles; none is accepted verbatim as truth or
transcribed directly into a final answer. There is no answer-key
harvesting/parser/transcriber, case/hash/path/output lookup, public-label
table, competitor-prediction route, LLM, VLM, cloud OCR, runtime network
dependency, or manual per-case edit.

Publication Checklist

• Frozen source, image, model, and runtime-bearing private main predate
  the deadline.
• Grouped nested 5x4, all-1,000, reused 850/150, and independent replay-B
  score evidence are disclosed.
• Replay-B exact confidence parity remains explicitly STOP_SHIP; no
  deterministic-confidence claim is made.
• Real artifact passed the frozen label-free verifier with 5,000 unique
  expected rows and immutable hash.
• Telemetry truncation is disclosed without a full-lifecycle resource
  claim.
• Public commit changes exactly the two Markdown files and the byte-exact
  JSONL.
• Actual post-midnight artifact/publication timestamps are preserved.

Competition is now over, code is open source: https://github.com/mikeg-cerebras/mib-doc-challenge