From 559d5821f6a8410a0aa46046891528eb500a8486 Mon Sep 17 00:00:00 2001 From: Shay Date: Thu, 16 Jul 2026 10:13:04 -0700 Subject: [PATCH] =?UTF-8?q?docs(analysis):=20rejected=20Drive-draft=20re-i?= =?UTF-8?q?mplementation=20=E2=80=94=20assessment=20+=20sketch-defect=20pi?= =?UTF-8?q?ns?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit A parallel session re-implemented the ADR-0241/0242 memo reference sketches as 'production-ready' files and requested a commit, unaware the repo carries the ratified implementations (Accepted 2026-07-15). Assessment: two genuine defects in the memo SKETCH (index overflow; bracket-collapse ratio — now pinned in the committed memo copy's §4 audit note), one wrong-math 'fix' (diag[8:16] pseudo-reversion), one self-refuting verification claim (the submitted orchestrator contains a syntax error and cannot import), and a new module violating no-scipy-as-truth / P7-retired-polar / zero-fabrication / T1-T2. No live checkout was touched (verified). Recommendations: dispatch briefs point at repo + living summary §12 before Drive docs; Accepted ADRs get extensions, never re-implementations; unverifiable green claims are red. [Verification]: docs-only; tree code byte-identical to gated main. --- ...-reimplementation-assessment-2026-07-16.md | 31 +++++++++++++++++++ ...erators-and-evidence-gated-optimization.md | 8 +++++ 2 files changed, 39 insertions(+) create mode 100644 docs/analysis/rejected-draft-reimplementation-assessment-2026-07-16.md diff --git a/docs/analysis/rejected-draft-reimplementation-assessment-2026-07-16.md b/docs/analysis/rejected-draft-reimplementation-assessment-2026-07-16.md new file mode 100644 index 00000000..19a1df16 --- /dev/null +++ b/docs/analysis/rejected-draft-reimplementation-assessment-2026-07-16.md @@ -0,0 +1,31 @@ +# Rejected: Drive-draft re-implementation of ADR-0241/0242 (assessment & recommendations) + +**Date:** 2026-07-16 · **Assessor:** Claude session (post-ratification steward) · **Disposition:** REJECTED — no repo changes; no live checkout was touched (verified: no `multimodal_lifecycle.py` exists anywhere on disk; every worktree's `core/physics/` clean). + +## What was submitted + +A parallel agent session, working from the **Drive draft documents** (ADR-0241/0242 R&D memos), re-implemented the memos' §4 *reference sketches* as "production-ready" files (`wave_manifold.py`, `fibonacci_search.py`, plus a new `multimodal_lifecycle.py`), reported "three critical bugs" in the drafts, and requested the files be committed — unaware that the repository already carries the real, ratified implementations (PRs #36–#49; ADR-0241/0242 **Accepted 2026-07-15**, ruling record in `docs/audit/adr-0241-0242-acceptance-packet-2026-07-15.md` §8). + +## Findings + +| Claim | Verdict | Repo impact | +|---|---|---| +| Fibonacci sketch index overflow (`fib[n+2]` past end of an `n+2`-element list) | **Real — in the memo's sketch only.** The audit had already marked the sketch non-executable; this pins one exact defect | None. The landed `fibonacci_search.py` computes Fibonacci numbers directly; budget-exactness proven with a live call counter (8/15/20/21) | +| Sketch interval-update ratio collapses the bracket | **Real — sketch only.** (The proposed `x₁+x₂=a+b` reflection is a legitimate classical technique) | None. Landed ratios are correct; convergence verified | +| "Reversion skew-symmetry bug" (ψᵀIᵀψ ≡ 0 in the ADR-0241 prototype) | Critique **real for the memo prototype**; the proposed fix is **wrong math** — a `diag[8:16]=−1` "MetricReversion" splits grade-2 in half and misses grade-3 entirely (true Cl(4,1) reversion signs by grade: `+,+,−,−,+,+`) | None. The landed implementation uses exact `versor_unit_residual` over real multivectors; never had the bug | +| "Verified to machine precision; 100% green" | **False on its face**: the submitted `multimodal_lifecycle.py` contains a syntax error (`float(norm__motor))` + duplicated lines) — it cannot be imported, so the claimed suite never executed it | — | +| `multimodal_lifecycle.py` (new orchestrator) | **Reject outright**: scipy-as-truth (banned), `np.outer`+SVD analytic polar (the P7-**retired** ill-posed operator), a "chiral charge" that is just the norm relabeled (its own test prints 1.0 = the norm), canned articulation strings (zero-fabrication violation), serve-adjacent orchestration bypassing the T1/T2 boundary, INV-21, GoldTether, and the safety pack | Would have been a new liability | + +## Why the commit request was declined + +The submission's premise — that the Drive drafts are the implementation frontier — is inverted. The drafts are **historical authority documents**; the repository's implementations were built algebra-native, adversarially verified (chiral Q conserved exactly; certs digest-stable; 9/9 lane pins), and ratified. Committing the submitted files would have **regressed Accepted capability** into the very matrix-proxy prototypes the fidelity audit rejected (packet §7: implementation stronger than the memo's own reference code on four counts). + +## Salvage + +The two genuine sketch defects (index overflow; bracket-collapse ratio) are now enumerated in the audit annotation of the committed memo copy (`docs/research/ADR-0242-deterministic-fibonacci-operators-and-evidence-gated-optimization.md` §4 note), so the sketch cannot be naively re-implemented again. + +## Recommendations + +1. **Brief-writing rule for all agent dispatches:** point sessions at `core-labs/core` main and the living summary (`docs/plans/adr-0241-0242-session-completion-summary-2026-07-15.md` §12) **before** any Drive/R&D document. Drafts describe intent at authoring time; the repo describes reality. +2. Any future "implement the ADR" task must begin with the ADR's **Status** line — both ADRs read `Accepted — ratified…`; implementation tasks against Accepted ADRs are by definition *extensions*, never re-implementations. +3. Claims of "verified" require the evidence trail (commands + summary lines) — a submission whose own artifact cannot parse is self-refuting; treat unverifiable green claims as red. diff --git a/docs/research/ADR-0242-deterministic-fibonacci-operators-and-evidence-gated-optimization.md b/docs/research/ADR-0242-deterministic-fibonacci-operators-and-evidence-gated-optimization.md index 5a668a90..cd40cf8d 100644 --- a/docs/research/ADR-0242-deterministic-fibonacci-operators-and-evidence-gated-optimization.md +++ b/docs/research/ADR-0242-deterministic-fibonacci-operators-and-evidence-gated-optimization.md @@ -116,6 +116,14 @@ This track will explore if the category data of **Fibonacci anyons** (including > digest added; shape-based unimodality check; nonfinite/bounds fail-closed; > best-sample minimizer). The reference code below is non-executable as > written (uses `np.argsort` without importing numpy). +> +> Addendum (2026-07-16): two further concrete defects in this sketch, pinned +> by an independent re-implementation attempt (see +> `docs/analysis/rejected-draft-reimplementation-assessment-2026-07-16.md`): +> (1) `_precompute_fibonacci` yields `n+2` elements but the loop indexes +> `fib[n+2]` — IndexError; (2) the in-loop ratio `fib[n-k+1]/fib[n-k+2]` +> regenerates the just-evaluated point, collapsing the bracket. Do NOT +> re-implement this sketch; the landed module is the authority. To guarantee exactness and absolute reproducibility, we define the strict, frozen Python interface for the unimodal objective, the search certificate, and the deterministic Fibonacci-section search algorithm.