paper: define BriskSeed reuse evidence boundary

[ShuhaoZhangTony]

Scope

Advisor-authored claim-boundary draft only, based specifically on the canonical BriskSeed branch. The student manuscript and all implementation/results are unchanged; Hongru Gao retains implementation, environment, experiment, and result ownership.

Exact identity

• PR head: c37092d11a7f0459fc3b3f2f74cb70c8496ac9ed
• artifact commit: 9d2d6a65a035e8615934ebd00245c32d7f26e702
• base branch/head: BriskSeed @ 4cde3b3f079ad9b1e4c98b10c2dfa7edf405f07a
• PDF: 2 pages

Claim boundary

The draft preregisters a reuse boundary over drift, recurrence, backend, and memory. It includes exact caching, periodic rebuild, strongest native backend, and oracle selector arms; generation-aware exact top-k identity checks; all-cost accounting; winner-switch plus oracle-headroom; and explicit stop rules. Existing figures/CSVs remain historical and no number is promoted.

Verification

• Tectonic exit 0, no warnings; 2/2 pages visually checked.
• git diff --check: pass.
• Full suite is collection-blocked by existing missing dependencies (numpy/yaml under Python 3.11; h5py under Python 3.10).
• Dependency-light subset: 7 passed, 2 existing failures (missing random ground truth; test invokes unavailable bare python).
• Existing student manuscript build was also checked: it fails on an undefined algorithm environment and absent bibliography key Gong20251593; it was not modified.

Keep Draft; do not merge without explicit user confirmation.

Manual paper semantic audit (2026-08-31)

• Exact Draft head: a30cedeaf0e1f5091037ca8d6ee0ca476258e87d
• Audit result: BriskSeed advisor main verified; Q1-Q7 labels now explicitly expose cited gap, mechanism, feasibility, matched evaluation, and takeaway.
• Tectonic 0.17.0: exit 0; PDF pages: 2; every page rendered and visually inspected.
• Bibliography is actually cited by the formal manuscript entry; no undefined citations/references.
• git diff --check: passed after mechanically normalizing engine-log trailing whitespace where applicable.
• Advisor paper/storyline assets only. No mechanism code, environment setup, experiment, result generation, performance tuning, or NPU run.
• Keep Draft and unmerged until explicit user confirmation.

[ShuhaoZhangTony]

@ShuhaoZhangTony 112 independent review requested for exact head c37092d11a7f0459fc3b3f2f74cb70c8496ac9ed on base BriskSeed. Advisor narrative only; no student manuscript, mechanism, environment, or result changes. Tectonic exit 0 (2 pages), visual and diff checks passed; test limitations are recorded exactly in the PR.

[ShuhaoZhangTony]

51-paper final audit — PASS at exact head c37092d11a7f0459fc3b3f2f74cb70c8496ac9ed

Canonical repository/base, exact-head paper assets, independent Tectonic rebuild, bibliography/citations, and page rendering pass. The manuscript closes the project-specific seven-question contract with strongest baselines, matched arms, correctness oracle, metrics/gates, evidence boundary, and stop conditions; claims do not exceed the retained real-online/replay/simulation/probe/derived evidence, and NPU architecture claims are bounded to the documented causal chain where applicable. This PR remains Draft; this review does not merge it or authorize experiments.

[ShuhaoZhangTony]

Incremental head-drift review — MERGE-READY

Reviewed exact head a30cedeaf0e1f5091037ca8d6ee0ca476258e87d, relative to the previously accepted c37092d11a7f0459fc3b3f2f74cb70c8496ac9ed.

The increment remains advisor-paper-only on canonical branch BriskSeed: it makes Q3/Q4/Q6 semantics explicit, refreshes the Tectonic transcript/PDF, and binds the updated artifact manifest. No student mechanism, environment, experiment, or result is changed, and no historical number is promoted. All listed artifact SHA-256 values match the checked-out bytes; git diff --check passes. Independent Tectonic compilation exits 0, yields the same extracted text as the committed 2-page PDF, and reports no overfull/error/undefined-reference diagnostics.

Final conclusion for this exact head: MERGE-READY for the bounded advisor paper assets. Existing dependency/test limitations and the separate student manuscript build failure remain accurately disclosed and are not hidden by this result.