# Prop-01 — 7-Day Symbolic Regression Watch

**Date:** 2026-09-05 ~00:20 America/New_York  
**Window:** new arXiv abs ~2026-08-29 → 2026-09-05  
**Track:** Prop-01 units SR / PhySO / PySR / SRBench / InsightSR follow-ups  
**Context:** PhySO capped ladder done (`prop01_support_claim` false); E1 live pilot running; InsightSR parked until measured rates (`lit-insightSR-vs-physo.md`).  
**Rule:** Real abs only; HTTP 200 verified. No fabrication. No stack-change recommendation.

---

## Executive take

**In-window SR traffic is sparse.** One clear GP/SR methods paper (`2609.03352` GPU constant optimization) is relevant to the **PySR/Operon-style** evolutionary path, not to PhySO’s units-RL core. **No new PhySO, SRBench, or InsightSR follow-up abs** in the 7d window. InsightSR (`2608.25291`, 2026-08-26) sits just **outside** 7d and remains **parked** pending E1 numbers.

**Stack unchanged:** PhySO + PySR + SRBench. Finish E1 before any InsightSR or constant-opt ablation.

---

## Ranked watch (in-window)

### 1. Efficient Constant Optimization for Symbolic Regression with GPU-Accelerated Tree-Based Genetic Programming
- **arXiv:** [2609.03352](https://arxiv.org/abs/2609.03352) · 2026-09-03 · Mao, Liu, Lu, Zhao, Jiang et al. · HTTP 200  
- **Claim (from abs):** GPU-resident batched Levenberg–Marquardt constant optimization for tree GP SR; addresses per-generation cost that leads GPU frameworks to omit / weaken constant opt.  
- **Prop-01 link:** Directly adjacent to **PySR / Operon** constant-fitting bottlenecks. If E1 shows PySR recovery lagging on constant-heavy targets, this is the first lit follow-up to evaluate — **not** a PhySO substitute.  
- **Does it change E1?** **No.** Do not insert mid-pilot. Log as post-E1 optional ablation vs bare PySR.  
- **Gaps:** Confirm whether code is released; compare against PySR’s existing constant optimization; units still absent.

### 2–N. Other keyword hits
Broader “symbolic / regression / uncertainty” queries returned mostly **off-topic** (soil Proctor corpus ML baselines, MIMO HW, etc.). **Not ranked** for Prop-01.

---

## Near-window / stack (not 7d news)

| Item | Role | Status |
|------|------|--------|
| [2608.25291](https://arxiv.org/abs/2608.25291) InsightSR | LLM outer loop on PySR; soft dimensional seeds | **Parked** — compare notes `lit-insightSR-vs-physo.md`; wait for E1 measured rates |
| [2303.03192](https://arxiv.org/abs/2303.03192) PhySO | Hard in-situ units SR | **Primary engine** · [github.com/WassimTenachi/PhySO](https://github.com/WassimTenachi/PhySO) HTTP 200 |
| [2305.01582](https://arxiv.org/abs/2305.01582) PySR | Fast evolutionary SR | **Primary engine** · [github.com/MilesCranmer/PySR](https://github.com/MilesCranmer/PySR) HTTP 200 |
| [2107.14351](https://arxiv.org/abs/2107.14351) SRBench | Eval harness | **Eval** · [github.com/cavalab/srbench](https://github.com/cavalab/srbench) HTTP 200 |
| [2503.19043](https://arxiv.org/abs/2503.19043) QDSR | QD + physics constraints (2025) | Competitor reference only |

---

## Implications for FLUX ENGINE / E1

1. **Do not** change `e1_live.yaml` or budgets based on this watch.  
2. After E1 posts measured linear/harmonic rates:  
   - If headroom vs support claim remains and PySR underperforms on constants → consider `2609.03352`-style constant-opt note (still no InsightSR).  
   - If PhySO units path is the bottleneck → InsightSR still optional outer loop only (prior CoS/HG agreement).  
3. Writing Bot: no “SOTA SR” language until E1 artifacts exist.

---

## Sources
- arXiv API queries: `symbolic regression`, PhySO/PySR, dimensional+SR, equation discovery+physics, SRBench/Feynman (2026-09-05)  
- Abs HTTP 200: `2609.03352`, stack ids above  
- Companions: `lit-insightSR-vs-physo.md`, `lit-interim-continuous.md`, `lit-tracks-top3.md`