ericchansen · ericchansen · May 28, 2026 · May 31, 2026 · Jun 1, 2026
diff --git a/README.md b/README.md
@@ -17,16 +17,21 @@ benchmarks/
 │   ├── results/*.json                   # 100 optimizer runs (L-BFGS-B, Nelder-Mead, Optax, …) — full-matrix CLI output
 │   └── forcefields/                     # Optimized force fields per run from the matrix output
 ├── rh-enamide/                          # Rh-enamide TS (Donoghue 2008), 9 molecules
-│   └── convergence/                     # Ratio-gated end-to-end optimization
+│   ├── convergence/                     # Canonical: QFUERZA-start (starting_point="qfuerza", current default)
+│   └── from-published/                  # Opt-out baseline: published-FF-start (starting_point="published")
 ├── heck-relay/                          # Heck relay (Rosales 2020), 23 molecules
-│   ├── convergence/                     # Ratio-gated end-to-end optimization
+│   ├── convergence/                     # Canonical: QFUERZA-start
+│   ├── from-published/                  # Opt-out baseline: published-FF-start
 │   └── diagnostic/                      # Three-baseline diagnostic (q2mm#277 loader bug)
 ├── pd-allyl-amination/                  # Pd allyl amination (Wahlers 2021), 21 molecules
-│   └── convergence/
+│   ├── convergence/                     # Canonical: QFUERZA-start
+│   └── from-published/                  # Opt-out baseline: published-FF-start
 ├── pd-1,4-conjugate-addition/           # Pd 1,4-conjugate addition (Wahlers 2021), 10 molecules
-│   └── convergence/
+│   ├── convergence/                     # Canonical: QFUERZA-start
+│   └── from-published/                  # Opt-out baseline: published-FF-start
 └── rh-1,4-conjugate-addition/           # Rh 1,4-conjugate addition (Wahlers 2022), 10 molecules
-    └── convergence/
+    ├── convergence/                     # Canonical: QFUERZA-start
+    └── from-published/                  # Opt-out baseline: published-FF-start
 
 qfuerza-zenodo/                          # QFUERZA paper validation data (Farrugia 2025)
 ├── README.md
@@ -38,9 +43,17 @@ Two standard directory layouts:
 
 - `convergence/` — output of `scripts/regenerate_convergence_results.py`
   in q2mm, the canonical end-to-end ratio-gated optimization pipeline.
+  As of q2mm PR #290, the canonical default is `starting_point="qfuerza"`
+  (Hessian-derived bond/angle scalars on the chemist-provided OPT
+  topology — see the [QFUERZA-recovery doc](https://github.com/ericchansen/q2mm/blob/master/docs/benchmarks/qfuerza-recovery.md)).
   Every published-FF system has exactly one of these.  Contains
-  `validation_results.json`, `paper_metrics.json`, and the optimized
-  `.fld` force field.
+  `validation_results.json`, `paper_metrics.json`, the optimized
+  `.fld` force field, and `per_param_comparison.md`.
+- `from-published/` — same script, opt-out baseline produced with
+  `--starting-point published`.  These are the legacy
+  literature-start runs preserved for backend-vs-literature
+  comparisons and side-by-side with the canonical QFUERZA-start
+  results.  Same file layout as `convergence/`.
 - `results/` + `forcefields/` — output of the legacy full-matrix
   `q2mm-benchmark` CLI.  Currently kept only for `ch3f/`, which is
   the source of the optimizer-matrix table in

diff --git a/benchmarks/CONVERGENCE_README.md b/benchmarks/CONVERGENCE_README.md
@@ -3,13 +3,29 @@
 This directory and its sibling `<system>/convergence/` directories hold
 the canonical baseline data for the published-FF benchmark systems.
 
-## What's in each `<system>/convergence/` directory
+As of q2mm PR #290, the canonical layout is:
+
+- `<system>/convergence/` — **canonical default**: QFUERZA-start
+  (`starting_point="qfuerza"`). Bond/angle force constants and
+  equilibria are derived from the QM Hessian via QFUERZA (Farrugia
+  2025) on top of the chemist-provided OPT substructure topology.
+- `<system>/from-published/` — opt-out baseline: published-FF-start
+  (`starting_point="published"`). Literature OPT values are used as
+  the initial parameters. Preserved for backend-vs-literature
+  comparison.
+
+Both subdirectories follow the same file layout (see below). See
+[`docs/benchmarks/qfuerza-recovery.md`](https://github.com/ericchansen/q2mm/blob/master/docs/benchmarks/qfuerza-recovery.md)
+for the framing.
+
+## What's in each `<system>/convergence/` (or `from-published/`) directory
 
 | File | Purpose |
 |------|---------|
 | `validation_results.json` | Top-level per-system summary: param counts, initial objective score, initial JaxLoss, ratio, ratio status, per-category Seminario fit, optimization result (when run). |
 | `paper_metrics.json` | Per-category Seminario + post-optimization fit (R², RMSD, MAE, n_refs for bond_length, bond_angle, eig_diagonal). |
 | `<system>_optimized.fld` | (Optional) Optimized force field, written only when the ratio gate passes and L-BFGS-B optimization actually ran. |
+| `per_param_comparison.md` | (QFUERZA-start runs) per-parameter comparison of published vs QFUERZA-optimized OPT-row scalars, bucketed by chemical motif. |
 
 Every JSON file is **strict JSON** (no `Infinity`/`NaN` — diverged values
 are encoded as structured `ratio_status: "diverged"` with `ratio: null`),
@@ -32,18 +48,16 @@ From a checkout of `ericchansen/q2mm`, with the supporting-info data
 extracted (or `Q2MM_SUPPORTING_INFO` pointing at it):
 
 ```bash
-# Regenerate baseline only (~60-90 min for all 5 systems on a single GPU)
+# Canonical: QFUERZA-start (default since q2mm PR #290) — writes to convergence/
 python scripts/regenerate_convergence_results.py \
-    --skip-optimization \
     --output-dir ../q2mm-data/benchmarks
 
-# Add optimization to systems that pass the ratio gate
+# Opt-out: published-FF-start baseline — writes to from-published/
 python scripts/regenerate_convergence_results.py \
-    --system rh-enamide \
-    --system pd-allyl \
+    --starting-point published \
     --output-dir ../q2mm-data/benchmarks
 
-# Experimental: bypass the ratio gate for borderline systems
+# Single system, ratio-gate bypass for borderline systems
 python scripts/regenerate_convergence_results.py \
     --system pd-conjugate \
     --ratio-tol none \

diff --git a/benchmarks/CROSS_SYSTEM_R2_RMSD.md b/benchmarks/CROSS_SYSTEM_R2_RMSD.md
@@ -0,0 +1,63 @@
+## R² / RMSD comparison: published paper vs q2mm @ published vs q2mm @ QFUERZA
+
+Per-system, per-category goodness of fit between MM predictions and the QM training data. _Same reference data is used for both q2mm columns_ (the published TSFF papers use the same training set, evaluated through MacroModel/MM3* instead of q2mm/JaxEngine).
+
+### rh-enamide
+
+| Metric | Published paper R²/RMSD | q2mm @ published start | q2mm @ QFUERZA start |
+|---|---:|---:|---:|
+| Bond length | R²=— / RMSD=0.03 Å | R²=0.9892 / RMSD=0.03859 Å | R²=0.9791 / RMSD=0.05371 Å |
+| Bond angle | R²=— / RMSD=2 deg | R²=0.9545 / RMSD=3.206 deg | R²=0.9503 / RMSD=3.349 deg |
+| Hessian eig (diag) | R²=— / RMSD=— mdyn/Å | R²=0.9683 / RMSD=0.06762 mdyn/Å | R²=0.9689 / RMSD=0.067 mdyn/Å |
+| Final OF | — | 2.699e+05 | 2.943e+05 |
+| Optimizer L-BFGS-B iters | — | 13 | 20 |
+
+_Paper note: Donoghue 2008 is paywalled. Bond RMSD ≤ 0.03 Å and angle RMSD < 2° are upper bounds from Tables 5–6. Energy RMSD ~0.4 kcal/mol from Table 7. Eigenvalue R² not reported by the paper. External-selectivity MUE = 0.6 kcal/mol over 18 test cases._
+
+### pd-allyl
+
+| Metric | Published paper R²/RMSD | q2mm @ published start | q2mm @ QFUERZA start |
+|---|---:|---:|---:|
+| Bond length | R²=0.988 / RMSD=— Å | R²=0.04612 / RMSD=0.37 Å | R²=0.4155 / RMSD=0.2896 Å |
+| Bond angle | R²=0.988 / RMSD=— deg | R²=0.3311 / RMSD=14.26 deg | R²=0.4905 / RMSD=12.45 deg |
+| Hessian eig (diag) | R²=0.998 / RMSD=— mdyn/Å | R²=-2.822 / RMSD=0.2789 mdyn/Å | R²=-2.594 / RMSD=0.2705 mdyn/Å |
+| Final OF | — | 7.991e+06 | 6.140e+06 |
+| Optimizer L-BFGS-B iters | — | 2 | 3 |
+
+_Paper note: Main text: 'R² values for the internal validation range from 0.988 to 0.998 for geometric and Hessian eigenvalues.' The 0.988 is the AGGREGATE geometric R² (bonds+angles+torsions combined) — not per-category. The 0.998 is Hessian eigenvalue R². Charges R² = 0.822. Energy values are EXTERNAL: 77 selectivity predictions, R² = 0.41, MUE = 4.4 kJ/mol = 1.05 kcal/mol._
+
+### pd-conjugate
+
+| Metric | Published paper R²/RMSD | q2mm @ published start | q2mm @ QFUERZA start |
+|---|---:|---:|---:|
+| Bond length | R²=0.99 / RMSD=— Å | R²=0.9495 / RMSD=0.07567 Å | R²=0.4048 / RMSD=0.2598 Å |
+| Bond angle | R²=0.99 / RMSD=— deg | R²=0.03686 / RMSD=17.9 deg | R²=0.1546 / RMSD=16.77 deg |
+| Hessian eig (diag) | R²=— / RMSD=— mdyn/Å | R²=-9.642 / RMSD=0.4119 mdyn/Å | R²=-4.641 / RMSD=0.2999 mdyn/Å |
+| Final OF | — | 7.235e+06 | 6.220e+06 |
+| Optimizer L-BFGS-B iters | — | 3 | 5 |
+
+_Paper note: Internal R² > 0.99 (aggregate, slopes ~1.01) per docs citing paper. No per-category or eigenvalue R² accessible. DOI correction: 10.1021/acs.joc.1c00136 is correct (the joc.0c02918 in some README files is wrong, 404s). Wahlers dissertation Ch. 5 (not Ch. 4). Energy: 82 external predictions, R² = 0.877, MUE = 1.8 kJ/mol = 0.43 kcal/mol._
+
+### rh-conjugate
+
+| Metric | Published paper R²/RMSD | q2mm @ published start | q2mm @ QFUERZA start |
+|---|---:|---:|---:|
+| Bond length | R²=0.95 / RMSD=— Å | R²=0.8219 / RMSD=0.1651 Å | R²=0.777 / RMSD=0.1847 Å |
+| Bond angle | R²=0.95 / RMSD=— deg | R²=0.5398 / RMSD=15.2 deg | R²=0.3374 / RMSD=18.24 deg |
+| Hessian eig (diag) | R²=— / RMSD=— mdyn/Å | R²=-12.85 / RMSD=0.4639 mdyn/Å | R²=-4.746 / RMSD=0.2988 mdyn/Å |
+| Final OF | — | 5.095e+06 | 7.668e+06 |
+| Optimizer L-BFGS-B iters | — | 4 | 2 |
+
+_Paper note: No published paper. Internal aggregate (all categories pooled): bisphosphine slopes 0.94–1.01, R² 0.91–0.99; diene slopes 1.0–1.07, R² 0.92–0.99. No per-category breakdown. R² 0.95 cells are midpoint of those ranges. Energy values are bisphosphine 67-structure EXTERNAL validation: R² = 0.64, MUE = 4.1 kJ/mol = 0.98 kcal/mol. Diene class: R² = 0.37, MUE = 5.3 kJ/mol. Wahlers dissertation Ch. 6 (not Ch. 5)._
+
+### heck-relay
+
+| Metric | Published paper R²/RMSD | q2mm @ published start | q2mm @ QFUERZA start |
+|---|---:|---:|---:|
+| Bond length | R²=— / RMSD=— Å | R²=0.983 / RMSD=0.04261 Å | R²=-6228 / RMSD=25.77 Å |
+| Bond angle | R²=— / RMSD=— deg | R²=0.9088 / RMSD=5.19 deg | R²=-8.348 / RMSD=52.56 deg |
+| Hessian eig (diag) | R²=— / RMSD=— mdyn/Å | R²=-14.28 / RMSD=0.5056 mdyn/Å | R²=-4.662 / RMSD=0.3078 mdyn/Å |
+| Final OF | — | 1.454e+06 | 1.317e+08 |
+| Optimizer L-BFGS-B iters | — | 7 | 0 |
+
+_Paper note: Main text says slope/R² 'near unity' for bonds/angles/torsions/eigenvalues — numeric values are in SI Figs S3–S6 (not accessible). Training-set geometry mean RMSD = 0.25 Å. Energy values are EXTERNAL selectivity validation against 151 corrected experimental points: R² = 0.89, MUE = 1.8 kJ/mol = 0.43 kcal/mol, RMSD = 2.3 kJ/mol = 0.55 kcal/mol._