Author notes — full detail, auditor-facing

This audit walks the full cipher version chain, scoring each version on the same 107-element benchmark and documenting *what changed* and *what was learned* at each step. The purpose: anyone evaluating the cipher's predictive performance needs to understand which version is the canonical one, why previous versions stand or fall, and what the data actually says about which mechanisms were load-bearing.

Version-by-version scorecard

Findings

F1. v9 → v11 was honest progress. Each version added a specific mechanism (eigenvalue susceptibility, snap-state tunneling, bicone simultaneous derivation) and each improved the score on the same bench. The 98.1% A+P on v11 was a genuine result.

F2. v10 was the warning shot. v10 tried to add a spherical-limit correction layer (the cipher's misses approach spherical limit; correct that and the misses should go away). It regressed to 20.9% exact — worse than v8. *Result:* uniform correction layers don't work. v10 was never released; v11 went back to per-element progressive blending with much better results.

F3. v12 falsified the corrections approach entirely. The v11 → v12 transition stripped *all* corrections (eigenvalue layer, snap layer, shell-completion logic) and returned to pure terrain. Result: 82/107 clean vs v11's 71/107 corrected on the same bench. Every correction layer was *hurting* accuracy. See the corrections-hurt-accuracy log for the head-to-head numbers.

F4. v12 perspective view: 91/107 with 0 outright misses. When v12 is read with a perspective lens (CN≤6 → 2D regime, CN≥7 → 3D regime), the score climbs to 91/107 with zero outright misses and 100% A+P. This is the framework's best-ever score on the bench.

F5. The c-ladder correction (Tribonacci replaces Fibonacci/8 for dims 4–6) is orthogonal to versions. It changed the c-values for d ≥ 4 but didn't affect cipher predictions for crystal structures (those live in cycle 1, d ≤ 3). See the c-ladder correction trail audit.

F6. v11's "complete self-derivation" claim stands. v11 was the version that proved the cipher *can* predict every element from Z alone with no external data. v12 demonstrates that the *terrain* carries the predictive power, but v11 was the milestone that closed the self-derivation chain. The v11 entry stays preserved as the milestone.

F7. Historical versions preserved, not deleted. Every version's results, code, and analysis remain published with status: superseded where appropriate. No silent rewriting. The audit trail is the intellectual-honesty record.

Resolution

• ✅ Canonical version: v12 (pure terrain, no corrections, perspective

read 91/107).

• ✅ Stale versions: v7, v8, v9, v10, v11 — all tagged superseded with

forward pointer to v12.

• ✅ Cross-version benchmark data preserved for independent verification.
• ⏳ Papers that referenced earlier cipher versions need status notes

(Paper 5 done, Papers 1–3 pending — see paper-status entries).

Discipline lesson

The v9 → v10 → v11 → v12 trajectory taught the framework's no-corrections rule: every attempted correction layer reduced accuracy. The reason is that corrections impose a *human-set* fix on top of the geometric terrain; the terrain, when allowed to speak for itself, is more predictive than terrain-plus-correction. This rule now applies to all future cipher work and to the framework more broadly (see feedback_cipher_no_corrections — the rule is durable).

The non-obvious part: this is the *opposite* of standard scientific practice, where models are improved by adding correction terms. Here, every addition *hurt*. The reason it works that way for the cipher is that the underlying mechanism is geometric, and the geometry is exact — not approximate. There's nothing to "correct" because nothing was approximated.