>[!warning]
>This content has not been peer reviewed.
# Methodology protocol — Graph rule search
A priori protocol for scientifically rigorous search. Defined before execution.
---
## 1. Success criteria
### Primary (full robustness)
- **d_B** in [1.19, 1.25] (1.22 ± 0.03)
- **d_s** in [1.31, 1.35] (4/3 ± 0.02)
- **Multi-seed:** N ≥ 3 seeds, all converged, all d_s in [1.29, 1.36], mean d_s in band, variance < 0.15
- **Success claim:** Only when all above are met
### Secondary (documented single-seed)
- If no config achieves primary within resource budget: report best config with explicit caveat
- **Caveat text:** "Single-seed result (seed X); not robust across seeds. See robustness results."
- Do not claim "found" without this qualification
---
## 2. Search procedure
### Phases (order fixed)
1. baseline
2. dw_diagnostic
3. ledger
4. ledger_balanced
5. ledger_d_s
6. ledger_p_init
7. ledger_hybrid_I
8. ledger_anneal
9. ledger_fine
10. ledger_search
11. ledger_quasicrystal
12. extended_search, biconnected, two_phase, spectral_fit, alt_rules, relaxed_long
### Rules
- No post-hoc addition of phases to "rescue" a run
- Log all phase results (positive and negative)
- Multi-seed robustness is final validation before any unqualified "found" claim
---
## 3. Measurement settings
| Param | Value | Notes |
|-------|-------|-------|
| L | 24 | Lattice side (finite-size effects apply) |
| T | 2000 | Evolution steps |
| measure_interval | 150 | Steps between dimension measurement |
| n_walks | 800 | Spectral dimension random walks |
| max_steps | 80 | Max steps per walk |
| converged_n | 2 | Consecutive measurements within threshold |
| eta_mode | ledger | μ from solve_fidelity_equilibrium |
---
## 4. Theoretical baseline
**3D bond percolation at p_c:**
- **d_B** ≈ 1.22 (Stauffer & Aharony)
- **d_s** = 4/3 (Alexander-Orbach; proven for d≥6; ≈1.33 for d=3)
Our target bands match this theoretical baseline. See [[expanded theory/knowledge/The backbone dimension at the percolation threshold]].
---
## 5. Train / validation
- **Search seeds:** 42, 123, 456 (tuning / phase selection)
- **Holdout seeds:** 789, 2024, 3141 (final validation, never used for tuning)
- **Report:** "Config selected on seeds {42, 123, 456}; validated on holdout {789, 2024, 3141}"
---
## 6. Scientific verification workflow (FSS)
For a scientifically relevant result, use finite-size scaling with a **scale-invariant** rule:
1. **Rule:** `ledger_scale_invariant` — intensive $I_{ij}$ and $N$ only (no $L$, $n$, $|E|$). See [[Graph rule verification results]] for the extensive-noise bug diagnosis.
2. **FSS:** `python rst_graph_rule_fss.py --rule ledger_scale_invariant --d-method chemical`
3. **Evaluate:** $d_B(L)$ should flatten with $L$; PASS if $d_B(\infty) \in [1.17, 1.27]$.
Outputs: `fss_extrapolation.png`, `fss_summary.txt`, `fss_data.csv`.
---
## 7. Limitations (stated explicitly)
1. **Finite L:** L=24 is far below literature (L~100–1000); dimensions may shift at larger L
2. **Seed dependence:** Some seeds land in bad attractors; no config yet achieves full robustness across all tested seeds
3. **Estimator bias:** Box-counting and spectral fits depend on ℓ-range and t-range; recursive μ can couple to estimator
4. **Initial condition:** Bond percolation at p_c; real universe may have different init
5. **No holdout (current):** If resource-limited, all seeds used for both search and report; document this