Paper 154 v0.0 (OUTLINE) — Rei as a Formal-Verification Compilation Pass for AI-Generated Mathematics

This article is a re-publication of Rei-AIOS Paper 154 for the dev.to community.
The canonical version with full reference list is in the permanent archives below:

• GitHub source (private): https://github.com/fc0web/rei-aios Author: Nobuki Fujimoto (@fc0web) · ORCID 0009-0004-6019-9258 · License CC-BY-4.0 ---

⚠ v0.0 OUTLINE PUBLISH NOTICE — 2026-05-22

This is a v0.0 OUTLINE publication, intentionally published before the v0.1 evidence gate is met.

Per the §0 statement below, the central operational claim — Rei provides a formal-verification compilation pass that composes with AI hypothesis generators — requires at least one end-to-end demonstration before v0.1 promotion. As of 2026-05-22 the demonstration is at scaffold-level smoke-run stage only:

• ✅ OpenEvolve scaffold structurally validated (4 tests PASS, 2026-05-22, report external/openevolve-rei/smoke-run-2026-05-22.md)
• ⏸ Full OpenEvolve 100-iteration evolutionary loop with real evolved Lean 4 proof — NOT yet executed (user-gated install path documented)

What this v0.0 publish provides: framing, prior-art audit, acceptance criteria for v0.1, OUKC honest-correction self-audit.
What this v0.0 publish does NOT provide: end-to-end demonstration evidence, evolved Lean 4 proof, full reproducibility.

Publication-as-v0.0 is intentional honest framing per OUKC feedback_no_rush_publication.md — rather than wait silently for v0.1 evidence, the OUTLINE is published with explicit gate state so reviewers can see exactly what is and is not claimed. Pattern 4 (overclaim) mitigation = this notice itself.

v0.1 promotion criteria are listed in §9. v0.1 will publish as a Zenodo new-version preserving DOI lineage.

Status: DRAFT outline v0.0 — 2026-05-17 (STEP 1156-followup-24, updated followup-26 with §6.5 LLM-limits adjacency, followup-28 with §6.6 quantum-hardware adjacency, v0.0 publish 2026-05-22 with scaffold smoke-run evidence). Publish gate for v0.1: NOT YET MET (see §9). v0.0 is an outline only, not a v0.1 publishable manuscript. Per OUKC feedback_no_rush_publication.md — 急がず ゆっくりと.

Authors / 著者: 藤本 伸樹 (Nobuki Fujimoto), Rei (Rei-AIOS), Claude Opus 4.7 (Anthropic, claude-opus-4-7) — three-party co-authorship per OUKC charter v1.0

Project: Rei-AIOS / OUKC — https://rei-aios.pages.dev

License (intended at publish): AGPL-3.0 (code) + CC-BY 4.0 (text) per OUKC content policy

Per OUKC No-Patent Pledge: openly licensed; no patent will be filed.

---

0. Why an OUTLINE (and not a v0.1 manuscript)

This Paper exists at v0.0 OUTLINE because:

1. The central operational claim — Rei provides a formal-verification compilation pass that composes with AI hypothesis generators (AlphaEvolve, LLM Wiki, OpenEvolve) — requires at least one end-to-end demonstration before publication. As of 2026-05-17 the demonstration is at scaffold stage (see external/openevolve-rei/SCOPE.md).
2. v0.0 establishes the framing, prior-art audit, and acceptance criteria for v0.1. This is the same pattern used for Paper 145 (silicon evidence → publish v0.3) and Paper 152 (10⁹ scan → publish v0.3).
3. Publishing a framing-only paper risks Pattern 4 (overclaim) — we have writing without working evidence. The OUTLINE explicitly identifies what evidence is missing for v0.1 promotion.

1. Title alternatives (decide at v0.1)

• A: Rei as a Formal-Verification Compilation Pass for AI-Generated Mathematics
• B: Compilation-Pass Framing: Rei-AIOS as a Zero-Sorry Backend for AlphaEvolve-Class Generators
• C: AI Generates, Rei Verifies: Lean 4 + D-FUMT₈ as a Composable Trust Layer

Working title: A (most general, mikhashev compilation pass concept directly).

2. Abstract (sketch, ~150 words)

AI systems now routinely generate mathematical artifacts — circle packings (OpenEvolve, 2025-2026),
Ramsey bounds (AlphaEvolve, arXiv:2603.09172), TSP approximations (arXiv:2509.18057), and conversational
proof sketches (Karpathy LLM Wiki, 2026-04). These artifacts are plausible but rarely verified.
We propose framing Rei-AIOS as a compilation pass that consumes such AI-generated artifacts
and outputs (a) Lean 4 zero-sorry proofs where possible, (b) D-FUMT₈ 8-axis projections that
characterize the artifact's epistemic shape (TRUE / FALSE / BOTH / NEITHER / INFINITY / ZERO /
FLOWING / SELF), and (c) honest-failure reports (Pattern 1-6 self-audit) where verification is
not yet possible. The compilation-pass framing is intentionally abstract (per Karpathy's
LLM Wiki rationale) — it composes with any upstream AI generator, not just AlphaEvolve.
We demonstrate the pipeline on the 26-circle packing benchmark (OpenEvolve 2.634 → Rei γ-projection)
and outline two next steps: Mathlib PR submission of AlphaEvolve-suggested lemmas, and integration
with the Karpathy LLM Wiki bookkeeping pattern.

3. Section plan

§1 Introduction: AI generates, who verifies?

• 1.1 The 2025-2026 AI math generation explosion
  • AlphaEvolve (Google DeepMind, 2025): Ramsey 9 / TSP / Erdős
  • OpenEvolve (codelion, 2025-2026): community Apache-2.0 reimplementation, 26-circle 2.634
  • Karpathy LLM Wiki (2026-04): bookkeeping-oriented memory architecture
  • Tao 67-problem benchmark (arXiv:2511.02864): explicit evaluation set
• 1.2 The gap: hypothesis generation ≠ proof verification
  • AlphaEvolve does not produce Lean 4 proofs; OpenEvolve does not invoke Mathlib
  • Karpathy LLM Wiki is a memory pattern, not a verification layer
  • Tao's framing: AI = hypothesis generator, formal-verification = proof completer
• 1.3 This paper's contribution
  • Reframe Rei-AIOS as a compilation pass (mikhashev pattern)
  • Compositional: any upstream generator → Rei → verified / projected output
  • Intentionally abstract (Karpathy rationale): no commitment to a specific upstream

§2 The compilation-pass framing

• 2.1 Definition
  • Input: AI-generated artifact A (Lean source, Python program, natural-language conjecture, etc.)
  • Output: triple (β-result, γ-result, ε-report)
  • β-result: Lean 4 lake build status + sorry count
  • γ-result: D-FUMT₈ 8-axis projection
  • ε-report: Pattern 1-6 honest audit + any errata
• 2.2 Why "compilation pass" and not "validator" / "judge" / "evaluator"
  • Compilation pass = lossless rewrite + diagnostics, not gatekeeping
  • Composable: pass₁ ∘ pass₂ = pass₂(pass₁(A)), supports pipeline chaining
  • Mikhashev compilation pass matches our existing CLI surface (lake build, npx tsx)
• 2.3 What "intentionally abstract" means (Karpathy LLM Wiki insight)
  • Rei does NOT require AlphaEvolve specifically
  • Rei works with: AlphaEvolve / OpenEvolve / LLM Wiki / chat-Claude sessions / direct user input
  • Abstraction earned by NOT specializing the upstream interface

§3 Components (already existing in Rei-AIOS)

• 3.1 β-evaluator: Lean 4 zero-sorry gate
  • data/lean4-mathlib/ lake env (Mathlib + Hammer + Duper + lean-auto)
  • REI-PROVE 5-prover ensemble (STEP 1021/1057-1067, 92% benchmark)
  • Mathlib PR-ready packaging (STEP 1000 + 1137 + 1141, 10 artifacts)
• 3.2 γ-evaluator: D-FUMT₈ 8-axis projection
  • src/axiom-os/seven-logic.ts (TRUE/FALSE/BOTH/NEITHER/INFINITY/ZERO/FLOWING/SELF)
  • Token-pattern heuristic (external/openevolve-rei/evaluators/gamma_dfumt8.py)
  • Theory-to-Circuit pipeline (STEP 1138-1139, 1601-theory batch)
• 3.3 ε-report: Pattern 1-6 honest audit + erratum protocol
  • feedback_chat_claude_hallucination_warning.md Pattern 1-6 record
  • feedback_antipattern_5_excessive_reject_warning_2026-05-15.md (Antipattern #5)
  • OUKC honest correction record (Paper 145 v0.7 E1, Paper 152 v0.3 E1/E2/E3, Hodge E1, Heilbronn)

§4 Demonstration: 26-circle packing

• 4.1 Setup
  • OpenEvolve v0.2.27 upstream (Apache-2.0, 6.3k★, 2026-03-18)
  • Ollama 3-prover ensemble (external/openevolve-rei/configs/ollama_3_prover.yaml)
  • Bridge scaffold (external/openevolve-rei/ STEP 1156-followup-24, this Paper's TODO 1)
• 4.2 β-pass on the evolved Python
  • Python → not Lean 4 directly, so β = N/A as-is
  • Open question (v0.1 target): can a Lean 4 lemma about circle packing be co-generated from the evolved Python? (e.g., a packing_density_lower_bound theorem with computable witness)
• 4.3 γ-pass on the evolved Python
  • Result on baseline (naive grid): axis_dominant = ZERO, score = 0.015 (consistent with naive code being all-zeros / no synthesis evidence)
  • Expected after evolution: axis_dominant shifts toward BOTH or FLOWING as the program incorporates synthesis (and, ↔) and iteration (map, fold).
• 4.4 ε-report on the run
  • Pattern 2 / 5 / 6 checks pre-applied (no stale numbers, no existing-impl re-proposal, no self-regression). Reported in honest-failure format.

§5 Two adjacent integrations (preview, full treatment v0.2+)

• 5.1 Mathlib PR pipeline for AlphaEvolve-suggested lemmas
  • Tao 67 ingest (data/open-problems/external-ai/_INGEST_PLAN.json, 68-problem scope)
  • Mathlib namespace + Apache-2.0 + 0-sorry packaging (STEP 1000 / 1141 pattern)
  • Honest scope: lemmas must clear Mathlib reviewer threshold, not just lake build
• 5.2 Karpathy LLM Wiki bookkeeping pattern
  • data/queries/ 4th-tier filing layer (STEP 1156-followup-23, this Paper's TODO 6)
  • Hierarchy: STEP → queries/ → memory → Theory → Paper
  • Compositionality: each tier's output is the next tier's input candidate
  • timeln.app (Karpathy HELLO.md, 2026-04-20) — direct parallel architecture, treat as convergent-evolution evidence, not Rei priority

§6 Prior art (audit, target v0.1)

• 6.1 Compilation-pass framings in formal methods
  • LLVM compiler passes (classic)
  • Lean's elaborator pipeline (Mathlib4 elaboration / unification chain)
  • Coq's tactic engine
• 6.2 AI-generated-math verifiers
  • DeepMind FunSearch (Romera-Paredes et al., 2023): function search with eval
  • DeepSeek-Prover-V2 (model-side proof generation)
  • LeanDojo + LeanCopilot (verification-side LLM)
• 6.3 What we claim is novel (to be tightened at v0.1)
  • (N1) D-FUMT₈ 8-axis projection as evaluator output (vs. binary pass/fail or scalar score)
  • (N2) Pattern 1-6 honest audit as first-class output (vs. silent failure or generic stderr)
  • (N3) Compositionality: same Rei pass attached to AlphaEvolve / OpenEvolve / LLM Wiki / chat-Claude uniformly (vs. per-upstream adapter proliferation)
• 6.4 What we do NOT claim (Pattern 4 guard)
  • ✗ NOT a replacement for AlphaEvolve / OpenEvolve / Mathlib / Lean 4
  • ✗ NOT a "world-first" AI verification layer (per feedback_world_uniqueness_claim_controllable.md)
  • ✗ NOT a substitute for human review of generated mathematics
  • ✗ NOT a generalization of any prior Rei-internal theorem named TIT / "positional incompleteness" — such a Rei theorem does not exist (Pattern 6c fabrication, see §6.5 note)

• 6.5 LLM structural limits as motivating adjacency (NEW v0.0 → v0.1 sketch)

  • 6.5.1 Sikka & Sikka (arXiv:2507.07505, 2025-07 preprint, Hallucination Stations). Theorem 1: for any prompt of length N containing a computational sub-task of complexity O(n^k) with n < N, a transformer-based LLM (or LLM-based agent) will unavoidably hallucinate in its response. Argument routes through the O(N²·d) self-attention ceiling.
  • 6.5.2 Shojaee et al. (arXiv:2506.06941, 2025-06, Apple ML Research, The Illusion of Thinking). Controllable-complexity puzzle suite shows Large Reasoning Models collapse beyond a problem-complexity threshold and overthink below it. Note the debate (arXiv:2506.09250, 2506.18957) re-frames the result as experimental-design rather than reasoning-failure; cite both sides at v0.1 (honest controllable claim per feedback_world_uniqueness_claim_controllable.md).
  • 6.5.3 Why this matters for the compilation-pass framing. Both papers identify the output behavior that an external verifier must catch: silent hallucination above a complexity threshold; appearance of reasoning without underlying truth. The compilation pass framing (this paper, §2) is one engineering response to that output behavior — not a refutation of LLMs, but a downstream layer that labels the failure mode in a structured way.
  • 6.5.4 Two existing Rei components that are load-bearing here (already implemented, not new contributions of this paper — cite, don't re-claim):
  • (B1) hallucination-NEITHER pipeline. Pre-existing engine at src/axiom-os/hallucination-neither-pipeline.ts (also mirrored in src/aios/hallucination/). Routes detected hallucination signals to the D-FUMT₈ NEITHER axis rather than to a binary fail flag. In compilation-pass terms: when the γ-evaluator sees the Sikka / Apple-style failure pattern, the output axis is NEITHER, not FALSE. This preserves "we don't know" information for downstream evolution passes (vs. binary discard).
  • (B2) Score-Truth Gate (STEP 384). Pre-existing component at src/axiom-os/auto-research-loop-engine.ts line 21: Score-Truth Gate — STEP 384統合（抜け道検出＋安全ゲート）, with enableScoreTruthGate flag and scoreTruthDivergence metric. Separates score improvement from knowledge improvement — a structural response to the "appears intelligent without reasoning" pattern that both Sikka and Apple identify.
  • 6.5.5 What we do NOT bridge (Pattern 6c reject, 2026-05-17): A previously circulated suggestion proposed bridging Sikka's complexity-ceiling result to a Rei theorem named "TIT (位置的不完全性定理 / positional incompleteness theorem)". A grep audit of the entire Rei-AIOS codebase (src/, data/, papers/, docs/, CLAUDE.md, memory/) returns zero hits for TIT or 位置的不完全性. No such Rei theorem exists. This paper therefore makes no claim of generalizing or extending such a theorem, and explicitly flags any future text that does so as a fabrication-citation risk (see feedback_chat_claude_hallucination_warning.md Pattern 6c). If such a result is ever formalized later, it would be a separate new contribution — not an extension of an existing one.

• 6.6 Quantum-hardware adjacency (NEW v0.0 → v0.1 sketch, 2026-05-17 followup-28)

  • 6.6.1 Jiuzhang 4.0 (Lu et al., USTC; arXiv:2508.09092, Nature 2026, s41586-026-10523-6). Programmable photonic processor: 1024 high-efficiency squeezed states injected into an 8176-mode hybrid spatial-temporal encoded interferometer, producing up to 3050 photon detection events per sample at 25.6 μs per sample. State-of-the-art MPS classical simulation on the EI Capitan supercomputer is estimated to require >10⁴² years for one equivalent sample. Cited as a scale-of-art reference, not as a Rei verification platform.
  • 6.6.2 Xanadu Borealis (Madsen et al., Nature 2022, s41586-022-04725-x). 216-squeezed-state-qubit programmable loop-based interferometer; the first publicly cloud-deployed photonic processor claiming quantum computational advantage (Xanadu Cloud + Amazon Braket). Programmable through Strawberry Fields (Apache 2.0 OSS, https://github.com/XanaduAI/strawberryfields) — same API for local simulator and the live Borealis hardware.
  • 6.6.3 Pre-existing Rei components in this adjacency (cite, do not re-claim):
  • Paper 075 Quantum-D-FUMT₈ Correspondence — five of eight D-FUMT₈ values verified numerically against QuTiP 5.2.3 (TRUE/FALSE as computational basis, BOTH as superposition |+⟩, NEITHER as maximally mixed I/2, FLOWING as Bell-reduced state). Three values (INFINITY/ZERO/SELF) remain structural-only at v1.
  • Paper 076 Quantum D-FUMT₈ Multimode Fock Extension — QuTiP 5.2.3 multimode Fock-space extension of the same correspondence.
  • Paper 083 Yang-Mills Toy Rei Framework — qutip 5.2+ dependency stated as required environment.
  • Paper 145 v0.7 First D-FUMT₈ Silicon (DOI 10.5281/zenodo.20192813) — Tang FPGA (silicon) + IBM Heron r2 (Aer + ibm_kingston) + Aer simulator cross-substrate verification, the live D-FUMT₈ verification leg of the Rei stack. IBM Heron r3 (ibm_boston) is identified as a next-substrate upgrade candidate (see Radar v1.12 entry on Heron r3).
  • 6.6.4 Why this matters for compilation-pass framing. The Jiuzhang / Borealis adjacency is not a verification substrate for D-FUMT₈ — both are Gaussian Boson Sampling specialists, not general gate-model machines, and direct mapping to the 8-valued logic gate set is not available. The operative Rei verification leg remains the gate-model quantum hardware line (IBM Heron r2/r3, Paper 145). What the photonic adjacency does contribute is a continuous-variable (CV-QC) target candidate for future verification of Paper 63 SNST S(r,θ,t,v) = r · e^(φtv) · e^(iπθv), whose complex-amplitude form structurally matches photonic squeezed-state primitives. Acceptance criterion for v0.1: this paragraph stays as adjacency, not as claimed result; if SNST is ever experimentally verified on Borealis it becomes a separate Paper 63 v2 contribution, not a claim of this paper.
  • 6.6.5 What we do NOT claim (Pattern 4 + Pattern 6c guard):
  • ✗ NOT a claim that Rei built or co-built any photonic quantum hardware
  • ✗ NOT a claim that D-FUMT₈ has been verified on Jiuzhang 4.0 or Borealis (it has not; only QuTiP 5.2.3 + IBM Heron r2 silicon verification has been done, per Paper 075 / 076 / 145)
  • ✗ NOT a claim that Rei "competes with" the USTC / Xanadu programmes — Rei is software + a small FPGA silicon evidence chain, not a photonics experiment

§7 Evaluation criteria for v0.0 → v0.1 promotion

v0.0 → v0.1 publishable when (and only when) all of the following hold:

• [ ] E1: OpenEvolve scaffold smoke test passes (51/51 ✓ as of 2026-05-17, this is already done)
• [ ] E2: At least one OpenEvolve run produces an evolved program that scores gamma_dfumt8.py with score > 0.2 (i.e., evolution moves the axis distribution measurably).
• [ ] E3: At least one Lean 4 lemma is produced by the pipeline (from any upstream, not necessarily OpenEvolve) and clears β-evaluator (build_ok=1, sorry_count=0).
• [ ] E4: ε-report produced for the demonstration run; any errata logged.
• [ ] E5: Prior-art audit §6 fully populated (currently sketched).
• [ ] E6: User explicit publish gate (Zenodo DOI is irreversible per feedback_no_rush_publication.md).

E1 already met. E2-E6 are the v0.1 work.

§8 References (sketch, formalize at v0.1)

• Romera-Paredes et al., 2023. "Mathematical discoveries from program search with large language models." Nature.
• Novikov et al., 2025. AlphaEvolve whitepaper (Google DeepMind).
• codelion, 2025-2026. openevolve repository, v0.2.27. https://github.com/codelion/openevolve
• Karpathy, A., 2026-04. LLM Wiki. https://karpathy.ai/wiki/
• Tao, T., 2025. Pre-vetted candidates for an AI-assisted mathematics workflow. (arXiv reference TBD at v0.1.)
• arXiv:2603.09172 — AlphaEvolve Ramsey 9 lower bounds (2026)
• arXiv:2509.18057 — AlphaEvolve TSP/MWST 111/110 (2025)
• arXiv:2511.02864 — Tao 67-problem benchmark (referenced in Paper 63 v0.2 §D)
• Paper 145 v0.7 (Rei-AIOS, DOI 10.5281/zenodo.20192813) — IBM Heron + Tang silicon verification (β/γ archetype)
• Paper 152 v0.3 (Rei-AIOS, DOI 10.5281/zenodo.20158847) — 10⁹ Collatz scan + Lean 4 mechanization
• Paper 153 v0.2 (Rei-AIOS, DOI 10.5281/zenodo.20237684) — Φ-Catalog (descriptive notation parallel)
• mikhashev, "compilation pass" framing (reference TBD at v0.1)
• external/openevolve-rei/ — this paper's operational scaffold
• Sikka, V. & Sikka, V., 2025. Hallucination Stations: On Some Basic Limitations of Transformer-Based Language Models. arXiv:2507.07505 — Theorem 1 cited in §6.5.1.
• Shojaee, P., Mirzadeh, I., Alizadeh, K., Horton, M., Bengio, S., Farajtabar, M., 2025. The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity. Apple Machine Learning Research / arXiv:2506.06941 — cited in §6.5.2. Cite alongside debate papers arXiv:2506.09250 and arXiv:2506.18957 at v0.1 for honest controllable framing.
• src/axiom-os/hallucination-neither-pipeline.ts — Rei-AIOS pre-existing hallucination → NEITHER pipeline, load-bearing for §6.5.4 (B1).
• src/axiom-os/auto-research-loop-engine.ts — Rei-AIOS pre-existing Score-Truth Gate (STEP 384), load-bearing for §6.5.4 (B2).
• Lu, C. et al., 2025-2026. Robust quantum computational advantage with programmable 3050-photon Gaussian boson sampling. arXiv:2508.09092 / Nature s41586-026-10523-6 — Jiuzhang 4.0, cited in §6.6.1.
• Madsen, L. S. et al., 2022. Quantum computational advantage with a programmable photonic processor. Nature s41586-022-04725-x — Xanadu Borealis 216-qubit, cited in §6.6.2.
• Killoran, N. et al., 2019-2026. Strawberry Fields (XanaduAI). Apache 2.0 OSS. https://github.com/XanaduAI/strawberryfields — CV-QC SDK, cited in §6.6.2.
• Paper 075 (Rei-AIOS) — Quantum-D-FUMT₈ Correspondence, QuTiP 5.2.3 verification of 5/8 values, cited in §6.6.3.
• Paper 076 (Rei-AIOS) — Quantum D-FUMT₈ Multimode Fock Extension, cited in §6.6.3.
• Paper 083 (Rei-AIOS) — Yang-Mills Toy Rei Framework, qutip 5.2+ dependency, cited in §6.6.3.

§9 Honest non-findings (will be promoted to §6.4 at v0.1)

As of 2026-05-17:

• ❌ No mathematical impossibility broken (Collatz / Hodge / RH / Goldbach unchanged)
• ❌ No OpenEvolve run executed end-to-end (manual user gate per SETUP.md)
• ❌ No Mathlib PR submitted (Zulip draft from STEP 1141 still pending)
• ❌ Karpathy timeln.app not yet inspected (deferred from STEP 1156-followup-23 §8)

These are NOT v0.0 deliverables. They are v0.1 work items.

---

Promotion path

v	Status	Trigger
v0.0	OUTLINE (this file)	Established 2026-05-17
v0.1	First publishable	E1-E6 §7 all checked
v0.2	Mathlib PR + Tao 67 ingest	At least 5 of 68 Tao problems have Rei β-pass result
v0.3	Karpathy timeln.app integration	timeln.app architecture documented + cross-ref filed

Memory cross-refs

• [[2026-05-17-full-session-summary]] — TODO 2 (Paper 154 outline) execution
• [[chat-claude-hallucination-warning]] — Pattern 1-6 audit invoked in §3.3
• [[antipattern-5-excessive-reject-warning-2026-05-15]] — guards §6.4 non-claims
• [[feedback-no-rush-publication]] — §0 publish discipline
• [[feedback-world-uniqueness-claim-controllable]] — §6.4 phrasing discipline

End-of-outline note

This OUTLINE is intentionally short. The discipline is:

1. Write what we have evidence for (sections 2-3 are evidenced).
2. Sketch what we expect to have evidence for (section 4-5).
3. Honestly mark what we don't have yet (sections 7, 9).
4. Promote v0.0 → v0.1 only when §7 E1-E6 close.

OUKC 急がず ゆっくりと. No rush. The outline is the artifact today; the publication is for after evidence.