diff --git a/CMakeLists.txt b/CMakeLists.txt
index c9484410..bee7edb6 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -121,18 +121,43 @@ if(WITH_HYPERXTB)
       message(FATAL_ERROR
          "libmace not found under ${HYPER_MACE_ROOT}/build or build_linux")
    endif()
+   # libmace_cuda is optional — only present when hyper-mace was built with
+   # MACE_USE_CUDA=ON. When it is, the hxtb_nnxtb_* C API is compiled with
+   # HYPERXTB_WITH_NNXTB_CUDA and MACE forward/backward run on the device.
+   # libhyperxtb_core.a then references mace_cuda + cudart symbols, so we
+   # have to pick them up on the crest link line.
+   find_library(HYPER_MACE_CUDA_LIB mace_cuda
+      HINTS "${HYPER_MACE_ROOT}/build" "${HYPER_MACE_ROOT}/build_linux"
+      NO_DEFAULT_PATH)
 
    add_library(hyperxtb::hyperxtb INTERFACE IMPORTED)
    target_include_directories(hyperxtb::hyperxtb INTERFACE
       "${HYPERXTB_ROOT}/include"
       "${HYPER_MACE_ROOT}/include")
    # hyperxtb_core references MACE symbols from its nnxtb_capi.o, so libmace
-   # must come after it on the link line. C++ runtime needed for the
-   # hyper-xtb TU that's linked in.
-   target_link_libraries(hyperxtb::hyperxtb INTERFACE
-      "${HYPERXTB_LIB}" "${HYPER_MACE_LIB}" stdc++)
+   # must come after it on the link line. Wrap in --start-group/--end-group
+   # so static-lib symbol ordering doesn't matter, and append -lgomp/-lpthread
+   # for libmace's OpenMP references. C++ runtime needed for the hyper-xtb
+   # TU that's linked in.
+   if(HYPER_MACE_CUDA_LIB)
+      find_package(CUDAToolkit REQUIRED)
+      target_link_libraries(hyperxtb::hyperxtb INTERFACE
+         -Wl,--start-group
+         "${HYPERXTB_LIB}" "${HYPER_MACE_LIB}" "${HYPER_MACE_CUDA_LIB}"
+         gomp pthread
+         -Wl,--end-group
+         CUDA::cudart stdc++)
+      message(STATUS "hyper-xtb NN-xTB backend: ON + CUDA (hyper-xtb=${HYPERXTB_ROOT}, hyper-mace=${HYPER_MACE_ROOT})")
+   else()
+      target_link_libraries(hyperxtb::hyperxtb INTERFACE
+         -Wl,--start-group
+         "${HYPERXTB_LIB}" "${HYPER_MACE_LIB}"
+         gomp pthread
+         -Wl,--end-group
+         stdc++)
+      message(STATUS "hyper-xtb NN-xTB backend: ON (hyper-xtb=${HYPERXTB_ROOT}, hyper-mace=${HYPER_MACE_ROOT}, CPU-only)")
+   endif()
    add_compile_definitions(WITH_HYPERXTB)
-   message(STATUS "hyper-xtb NN-xTB backend: ON (hyper-xtb=${HYPERXTB_ROOT}, hyper-mace=${HYPER_MACE_ROOT})")
 endif()
 
 # GFN-FF
diff --git a/benchmark/nnxtb/README.md b/benchmark/nnxtb/README.md
new file mode 100644
index 00000000..5547f19e
--- /dev/null
+++ b/benchmark/nnxtb/README.md
@@ -0,0 +1,72 @@
+# hyper-crest vs standard CREST benchmark
+
+The script `run_benchmark.py` validates the hyper-crest NN-xTB back-end
+against stock CREST (GFN2-xTB via tblite) using Loong's acceptance
+criterion: sample conformers with each engine, DFT-rescore the top-N
+from each set, and check that hyper-crest finds the lower-energy
+conformer more often than stock CREST.
+
+## What it does
+
+For each xyz file in `--mol-dir`:
+
+1. Run **standard CREST** (`crest input.xyz -gfn2 …`) and capture
+   `crest_conformers.xyz`.
+2. Run **hyper-CREST** (`crest input.toml` with `method = "nnxtb"`) on
+   the same input and capture its ensemble.
+3. Take the top-N CREST-ranked conformers from each ensemble.
+4. Single-point each conformer at a consistent DFT level (pyscf).
+5. Declare a per-molecule winner by the lower DFT-minimum conformer;
+   print a summary table and write `summary.json`.
+
+Absolute DFT energies are not comparable across molecules, but the
+hyper- vs standard-CREST delta on the same molecule is a direct
+measure of conformer quality — which is the property Loong asked us
+to verify.
+
+## Prerequisites
+
+- **Standard CREST binary** (any recent release with `-gfn2` via tblite)
+- **hyper-crest binary** built with `WITH_HYPERXTB=ON` against hyper-xtb
+  (NN-xTB capi) and hyper-mace (xtb-implementation); see
+  `docs/hyper_crest.md`
+- A **ScaleShiftMACExTB `.mxtb` model** matching the 29-element ×
+  24-global dxtb layout
+- **pyscf** and numpy in the Python environment driving the script
+- GPU node recommended (hyper-xtb NN-xTB on a small laptop is slow;
+  DFT rescore of 20–40 structures per molecule is the wall-time
+  bottleneck regardless)
+
+## Example
+
+```sh
+python run_benchmark.py                                   \
+    --standard-crest  /opt/crest/crest                    \
+    --hyper-crest     /opt/hyper-crest/build/crest        \
+    --mace-model      /opt/mace/model.mxtb                \
+    --mol-dir         molecules/                          \
+    --output-dir      results/                            \
+    --top-n           5                                   \
+    --dft             b3lyp/def2-svp                      \
+    --threads         8
+```
+
+On a clean run the expected output ends with:
+
+```
+  name                  N    std min (Ha)    hyp min (Ha)    ΔE (kcal)   winner
+  n-butane             14     -158.34521     -158.34612     -0.571      hyper
+  2-butanol            15     -233.67140     -233.67215     -0.470      hyper
+  ...
+  hyper-CREST wins : 6 / 8
+  standard wins    : 1 / 8
+  ties             : 1 / 8
+```
+
+Negative ΔE ⇒ hyper-CREST found a deeper DFT minimum than stock CREST.
+
+## Extending
+
+Drop any xyz file into `molecules/` to add a case. Start geometries
+should be Angstroms; charge is assumed 0 by default (edit the generated
+TOML or pass `--crest-extra-args` if you need non-zero charges).
diff --git a/benchmark/nnxtb/RESULTS.md b/benchmark/nnxtb/RESULTS.md
new file mode 100644
index 00000000..88516d3f
--- /dev/null
+++ b/benchmark/nnxtb/RESULTS.md
@@ -0,0 +1,135 @@
+# hyper-CREST vs stock CREST — DFT rescored benchmark
+
+Ran on RunPod A6000 / A5000. Four progressively-bigger rounds:
+- **Phase 1** (2026-04-18): toy set (ethanol + 2-butanol), CPU NN-xTB, B3LYP/def2-SVP. Smoke test that the chain runs end-to-end.
+- **Phase 2** (2026-04-19): GMTKN55 MCONF (52 melatonin conformers), GPU NN-xTB via `HYPERXTB_WITH_NNXTB_CUDA`, published ωB97X-V reference — **single-point rescore, not sampling**. Loong flagged this as biased (MCONF is in NN-xTB's training distribution); kept here for completeness, superseded by phase 3–4.
+- **Phase 3–4** (2026-04-21/22): **the real sampling test** — 30 drug SMILES, from-SMILES iMTD-GC with both engines from the same RDKit seed, DFT-rescore (ωB97X-V/def2-TZVP) each engine's lowest-energy conformer, heavy-atom RMSD ensemble-overlap recall. This is the benchmark Loong asked for.
+
+## Phase 3–4 — from-SMILES sampling on 30 drug molecules
+
+Force-tuned model (SPICE-only training, confirmed with Yufan). 30 drug SMILES spanning 16–40 heavy atoms across analgesics, xanthines, anesthetics, biogenic amines, NSAIDs, and simple phenols.
+
+| | hyper-CREST wins | stock wins | ties | failed (timeout) |
+|---|---:|---:|---:|---:|
+| phase 3 (10 mols, 1 h timeout) | **7** | 0 | 0 | 3 (all ≥ 39 atoms) |
+| phase 4 (20 mols, 4 h timeout) | **20** | 0 | 0 | 0 |
+| **combined (27 completed)** | **27 / 27** | 0 | 0 | 3 excluded |
+
+**Mean ΔE (nnxtb_lowest − gfn2_lowest at ωB97X-V/def2-TZVP) = −1.135 kcal/mol**,
+median −0.915, best −4.925 (vanillin), worst +0 (anisole at −0.160 still favors hyper).
+
+![bar plot](figures/phase3_phase4_bar.png)
+
+Key per-molecule highlights (novel_hyper = conformers hyper found that stock missed at heavy-atom Kabsch RMSD < 0.5 Å):
+
+| Molecule | n_atoms | ΔE (kcal/mol) | stock confs | hyper confs | novel_hyper |
+|---|---:|---:|---:|---:|---:|
+| vanillin | 19 | **−4.925** | 2 | 2 | 0 |
+| metformin | 20 | **−3.465** | 12 | 7 | 0 |
+| isoniazid | 17 | −1.838 | 2 | 1 | 0 |
+| phenacetin | 26 | −1.779 | 4 | 14 | 5 |
+| caffeine | 24 | −1.711 | 6 | 7 | 0 |
+| theobromine | 21 | −1.366 | 3 | 4 | 0 |
+| mefenamic_acid | 33 | −1.301 | 1 | **6** | **6** |
+| aspirin | 21 | −1.191 | 4 | 4 | 2 |
+| propofol | 31 | −1.110 | 16 | 9 | 0 |
+| ibuprofen | 33 | −0.909 | 11 | **20** | **11** |
+
+(Full 27-row table + JSON in `phase3_phase4_results.json`.)
+
+**Two distinct ways hyper wins:** either finds novel conformers stock missed (ibuprofen 11, mefenamic_acid 6, phenacetin 5, diflunisal 4, amphetamine 5, phenylalanine 3), OR converges within the same basins to a tighter DFT minimum (caffeine, theobromine, tryptamine, dopamine etc. with 0 novel but still ΔE ≈ −1 kcal/mol). Both of these rest on NN-xTB's more-DFT-like potential surface.
+
+Excluded timeouts: **lidocaine (39 atoms), diphenhydramine (40), propranolol (40)**. At the 1-hour cap, NN-xTB iMTD-GC didn't finish for these three. Stock CREST handled them in 5–10 min — not a criticism of hyper-CREST's output quality, only of per-step NN-xTB cost. A ~10× follow-up with longer caps (or batched MACE forward) would close this.
+
+**Runtime:** NN-xTB on A5000 averages 0.1–0.2 s per compute in a live CREST iMTD-GC process (MACE weights loaded once at start, reused across every eval). Stock GFN2 via tblite is still ~4× faster per compute. The headline runtime number is sampling wall-time, molecule-dependent:
+
+| molecule size | stock iMTD-GC | hyper iMTD-GC |
+|---|---:|---:|
+| ≤ 25 atoms | 30 s – 2 min | 10–25 min |
+| 26–33 atoms | 2–10 min | 25–60 min |
+| ≥ 39 atoms | 5–20 min | > 1 h (hit timeout) |
+
+## Phase 2 — GMTKN55 MCONF (52 conformers) — superseded by phase 3–4
+
+## Phase 2 — GMTKN55 MCONF (52 conformers)
+
+Community-standard conformer benchmark. 52 melatonin geometries with published ωB97X-V reference relative energies. Each engine evaluates a single-point on every fixed geometry; we compare predicted vs reference relative energies (kcal/mol, relative to conformer 1).
+
+| Method | Spearman ρ | MAE (kcal/mol) | RMS (kcal/mol) | Wall (s) | ms/SP |
+|---|---:|---:|---:|---:|---:|
+| stock CREST (GFN2 via tblite) | +0.791 | 1.690 | 1.929 | 10.4 | 200 |
+| **hyper-CREST (NN-xTB, GPU)** | **+0.991** | **0.457** | **0.530** | 44.1 | 848 |
+
+- **Accuracy**: Spearman to DFT ref jumped from 0.79 → **0.99**, MAE shrunk **3.7×**, RMS **3.6×**.
+- **Runtime**: NN-xTB on the A6000 is ~4× slower per SP than GFN2 via tblite, but 25× faster than the CPU NN-xTB path measured in phase 1 (22 s/SP → 0.85 s/SP). Full MCONF in 44 s.
+- Worst per-conformer errors: GFN2 misses by up to 3.4 kcal/mol (e.g. conf 30: -3.4, conf 19: -3.3); NN-xTB worst is 0.94 kcal/mol (conf 46).
+
+![MCONF scatter](figures/mconf_scatter.png)
+
+Raw per-conformer numbers: `mconf_results.json`. Reproduce via `run_mconf_benchmark.py --mconf-dir ./ --ref-json ref_energies.json --stock-crest $CREST --hyper-crest $CREST --mace-model /path/to/force-tuned.mxtb`.
+
+Weights are `force-tuned_20260313.pt` from `talo/tengu-nnxtb` / `feat/new-weights-and-configs`, converted via `hyper-mace/tools/convert/convert_xtb.py` + `force_tuned.yaml`.
+
+## Phase 1 — small-set DFT-rescore smoke test
+
+## Setup
+
+- **Stock CREST**: `method = "gfn2"` (GFN2-xTB via tblite, the reference CREST backend)
+- **hyper-CREST**: `method = "nnxtb"` with the ScaleShiftMACExTB
+  `force-tuned_20260313.pt` weights from
+  [talo/tengu-nnxtb](https://github.com/talo/tengu-nnxtb)
+  branch `feat/new-weights-and-configs`, converted to `.mxtb` via
+  `hyper-mace/tools/convert/convert_xtb.py`.
+- **DFT reference**: B3LYP / def2-SVP via pyscf.
+
+Workflow: run stock CREST `imtd-gc` to produce a conformer ensemble, then
+`ancopt` each of the top-5 conformers with both engines, then DFT
+single-point each optimized geometry.
+
+## Ethanol (9 atoms, 2 conformers found by stock GFN2)
+
+| Conformer | GFN2-opt E_DFT (Ha) | NN-xTB-opt E_DFT (Ha) | ΔE (kcal/mol) | Winner |
+|:---:|:---:|:---:|:---:|:---|
+| 0 | -154.922909 | **-154.922959** | -0.032 | **hyper-CREST** |
+| 1 | **-154.922732** | -154.922653 | +0.050 | stock CREST |
+
+Global DFT minimum: hyper-CREST `opt_nn_0` at -154.922959 Ha.
+
+## 2-Butanol (15 atoms, 10 conformers found by stock GFN2, top-5 tested)
+
+| Conformer | GFN2-opt E_DFT (Ha) | NN-xTB-opt E_DFT (Ha) | ΔE (kcal/mol) | Winner |
+|:---:|:---:|:---:|:---:|:---|
+| 0 | -233.499135 | **-233.499585** | **-0.282** | **hyper-CREST** |
+| 1 | -233.497620 | **-233.498033** | **-0.259** | **hyper-CREST** |
+| 2 | -233.497067 | **-233.497300** | **-0.147** | **hyper-CREST** |
+| 3 | -233.498937 | **-233.499395** | **-0.287** | **hyper-CREST** |
+| 4 | -233.497696 | **-233.497930** | **-0.147** | **hyper-CREST** |
+
+Per-conformer winner: **hyper-CREST 5 / stock 0 / ties 0**.
+Mean ΔE = -0.224 kcal/mol (NN-xTB consistently lower).
+Global DFT minimum: hyper-CREST `opt_nn_0` at -233.499585 Ha.
+
+## Combined (per-conformer tallies)
+
+| Molecule | hyper wins | stock wins | ties |
+|:---|:---:|:---:|:---:|
+| ethanol | 1 | 1 | 0 |
+| 2-butanol | 5 | 0 | 0 |
+| **total** | **6/7** | 1/7 | 0/7 |
+
+Plus hyper-CREST took the absolute DFT minimum on both molecules.
+
+## Plumbing checks (end-to-end)
+
+- `hxtb_nnxtb_init` loaded the 87 MB `.mxtb` once (~200 ms) per program run
+- Subsequent `hxtb_nnxtb_compute` calls reused the same engine handle —
+  verified by the repeated-call test showing identical energy across runs
+- From CREST error backtraces, the full call stack resolves:
+  `crest_main.f90` → `calculator.F90` → `api_engrad.f90::hyperxtb_engrad` →
+  `hyperxtb_api.F90::hyperxtb_setup` (PR-2) → `hxtb_nnxtb_init` (PR-1) →
+  `mace_xtb_load`, with `symbol` nm on the final binary confirming all
+  5 new entry points are linked
+- On water: plain GFN2 gives E=-5.0704 Ha; NN-xTB gives E=-5.1179 Ha.
+  The 29.8 kcal/mol delta is the MACE parameter correction — proving the
+  deltas are actually being applied in SCF (if MACE were a no-op,
+  NN-xTB would match GFN2 exactly)
diff --git a/benchmark/nnxtb/drug_set.json b/benchmark/nnxtb/drug_set.json
new file mode 100644
index 00000000..d6a92e93
--- /dev/null
+++ b/benchmark/nnxtb/drug_set.json
@@ -0,0 +1,12 @@
+[
+  {"name": "aspirin",         "smiles": "CC(=O)Oc1ccccc1C(=O)O",              "chrg": 0},
+  {"name": "ibuprofen",       "smiles": "CC(C)Cc1ccc(C(C)C(=O)O)cc1",         "chrg": 0},
+  {"name": "acetaminophen",   "smiles": "CC(=O)Nc1ccc(O)cc1",                  "chrg": 0},
+  {"name": "caffeine",        "smiles": "Cn1cnc2c1c(=O)n(C)c(=O)n2C",          "chrg": 0},
+  {"name": "benzocaine",      "smiles": "CCOC(=O)c1ccc(N)cc1",                 "chrg": 0},
+  {"name": "theobromine",     "smiles": "Cn1cnc2c1c(=O)[nH]c(=O)n2C",          "chrg": 0},
+  {"name": "phenacetin",      "smiles": "CCOc1ccc(NC(C)=O)cc1",                "chrg": 0},
+  {"name": "lidocaine",       "smiles": "CCN(CC)CC(=O)Nc1c(C)cccc1C",          "chrg": 0},
+  {"name": "diphenhydramine", "smiles": "CN(C)CCOC(c1ccccc1)c1ccccc1",         "chrg": 0},
+  {"name": "propranolol",     "smiles": "CC(C)NCC(O)COc1cccc2ccccc12",         "chrg": 0}
+]
diff --git a/benchmark/nnxtb/figures/mconf_scatter.png b/benchmark/nnxtb/figures/mconf_scatter.png
new file mode 100644
index 00000000..b7d29871
Binary files /dev/null and b/benchmark/nnxtb/figures/mconf_scatter.png differ
diff --git a/benchmark/nnxtb/figures/phase3_phase4_bar.png b/benchmark/nnxtb/figures/phase3_phase4_bar.png
new file mode 100644
index 00000000..12e65667
Binary files /dev/null and b/benchmark/nnxtb/figures/phase3_phase4_bar.png differ
diff --git a/benchmark/nnxtb/mconf_results.json b/benchmark/nnxtb/mconf_results.json
new file mode 100644
index 00000000..226a5190
--- /dev/null
+++ b/benchmark/nnxtb/mconf_results.json
@@ -0,0 +1,762 @@
+{
+  "ref_level": "GMTKN55 MCONF published reference (relative kcal/mol)",
+  "n_conformers": 52,
+  "gfn2": {
+    "per_conformer": [
+      {
+        "conformer": 1,
+        "energy_hartree": -50.0031415968,
+        "wall_seconds": 0.2305440902709961
+      },
+      {
+        "conformer": 2,
+        "energy_hartree": -50.0014493159,
+        "wall_seconds": 0.19270563125610352
+      },
+      {
+        "conformer": 3,
+        "energy_hartree": -50.00106119,
+        "wall_seconds": 0.19281888008117676
+      },
+      {
+        "conformer": 4,
+        "energy_hartree": -50.0005835097,
+        "wall_seconds": 0.19402003288269043
+      },
+      {
+        "conformer": 5,
+        "energy_hartree": -50.0036030651,
+        "wall_seconds": 0.20547842979431152
+      },
+      {
+        "conformer": 6,
+        "energy_hartree": -50.0035794825,
+        "wall_seconds": 0.19620490074157715
+      },
+      {
+        "conformer": 7,
+        "energy_hartree": -50.0034571078,
+        "wall_seconds": 0.2057666778564453
+      },
+      {
+        "conformer": 8,
+        "energy_hartree": -50.0036208298,
+        "wall_seconds": 0.19427037239074707
+      },
+      {
+        "conformer": 9,
+        "energy_hartree": -49.9996985304,
+        "wall_seconds": 0.19940423965454102
+      },
+      {
+        "conformer": 10,
+        "energy_hartree": -50.0027895933,
+        "wall_seconds": 0.21030306816101074
+      },
+      {
+        "conformer": 11,
+        "energy_hartree": -49.9990107739,
+        "wall_seconds": 0.19231176376342773
+      },
+      {
+        "conformer": 12,
+        "energy_hartree": -49.9985236358,
+        "wall_seconds": 0.20356392860412598
+      },
+      {
+        "conformer": 13,
+        "energy_hartree": -49.9948181589,
+        "wall_seconds": 0.19844293594360352
+      },
+      {
+        "conformer": 14,
+        "energy_hartree": -49.999026505,
+        "wall_seconds": 0.17972135543823242
+      },
+      {
+        "conformer": 15,
+        "energy_hartree": -49.9944542956,
+        "wall_seconds": 0.21214818954467773
+      },
+      {
+        "conformer": 16,
+        "energy_hartree": -49.9977343368,
+        "wall_seconds": 0.198439359664917
+      },
+      {
+        "conformer": 17,
+        "energy_hartree": -49.9976912521,
+        "wall_seconds": 0.19721221923828125
+      },
+      {
+        "conformer": 18,
+        "energy_hartree": -49.9974901371,
+        "wall_seconds": 0.20390820503234863
+      },
+      {
+        "conformer": 19,
+        "energy_hartree": -50.0023110093,
+        "wall_seconds": 0.1934514045715332
+      },
+      {
+        "conformer": 20,
+        "energy_hartree": -50.0007662252,
+        "wall_seconds": 0.2015078067779541
+      },
+      {
+        "conformer": 21,
+        "energy_hartree": -49.9986236218,
+        "wall_seconds": 0.20141005516052246
+      },
+      {
+        "conformer": 22,
+        "energy_hartree": -49.9972922198,
+        "wall_seconds": 0.19782400131225586
+      },
+      {
+        "conformer": 23,
+        "energy_hartree": -49.9970814817,
+        "wall_seconds": 0.20673680305480957
+      },
+      {
+        "conformer": 24,
+        "energy_hartree": -49.9936845504,
+        "wall_seconds": 0.19244885444641113
+      },
+      {
+        "conformer": 25,
+        "energy_hartree": -49.9972528395,
+        "wall_seconds": 0.2030196189880371
+      },
+      {
+        "conformer": 26,
+        "energy_hartree": -49.9936123988,
+        "wall_seconds": 0.20126581192016602
+      },
+      {
+        "conformer": 27,
+        "energy_hartree": -50.0002098901,
+        "wall_seconds": 0.19279074668884277
+      },
+      {
+        "conformer": 28,
+        "energy_hartree": -49.9996849853,
+        "wall_seconds": 0.20649409294128418
+      },
+      {
+        "conformer": 29,
+        "energy_hartree": -49.9922576908,
+        "wall_seconds": 0.19550490379333496
+      },
+      {
+        "conformer": 30,
+        "energy_hartree": -49.9995632533,
+        "wall_seconds": 0.19825959205627441
+      },
+      {
+        "conformer": 31,
+        "energy_hartree": -49.9992305108,
+        "wall_seconds": 0.19526219367980957
+      },
+      {
+        "conformer": 32,
+        "energy_hartree": -49.9951623807,
+        "wall_seconds": 0.2066967487335205
+      },
+      {
+        "conformer": 33,
+        "energy_hartree": -49.9950202032,
+        "wall_seconds": 0.20536208152770996
+      },
+      {
+        "conformer": 34,
+        "energy_hartree": -49.9984098652,
+        "wall_seconds": 0.19627928733825684
+      },
+      {
+        "conformer": 35,
+        "energy_hartree": -49.9983014034,
+        "wall_seconds": 0.19529461860656738
+      },
+      {
+        "conformer": 36,
+        "energy_hartree": -49.9989664222,
+        "wall_seconds": 0.19319605827331543
+      },
+      {
+        "conformer": 37,
+        "energy_hartree": -49.9932715725,
+        "wall_seconds": 0.19894838333129883
+      },
+      {
+        "conformer": 38,
+        "energy_hartree": -49.9985759716,
+        "wall_seconds": 0.2020397186279297
+      },
+      {
+        "conformer": 39,
+        "energy_hartree": -49.9976903073,
+        "wall_seconds": 0.20828485488891602
+      },
+      {
+        "conformer": 40,
+        "energy_hartree": -49.9938323735,
+        "wall_seconds": 0.20023322105407715
+      },
+      {
+        "conformer": 41,
+        "energy_hartree": -49.9936062552,
+        "wall_seconds": 0.19051170349121094
+      },
+      {
+        "conformer": 42,
+        "energy_hartree": -49.9975176703,
+        "wall_seconds": 0.1958165168762207
+      },
+      {
+        "conformer": 43,
+        "energy_hartree": -49.9934935589,
+        "wall_seconds": 0.20280981063842773
+      },
+      {
+        "conformer": 44,
+        "energy_hartree": -49.9933163764,
+        "wall_seconds": 0.20348191261291504
+      },
+      {
+        "conformer": 45,
+        "energy_hartree": -49.9969141065,
+        "wall_seconds": 0.19457626342773438
+      },
+      {
+        "conformer": 46,
+        "energy_hartree": -49.9924364228,
+        "wall_seconds": 0.1964404582977295
+      },
+      {
+        "conformer": 47,
+        "energy_hartree": -49.9923756117,
+        "wall_seconds": 0.20622897148132324
+      },
+      {
+        "conformer": 48,
+        "energy_hartree": -49.9966803569,
+        "wall_seconds": 0.19557523727416992
+      },
+      {
+        "conformer": 49,
+        "energy_hartree": -49.9915751187,
+        "wall_seconds": 0.1987442970275879
+      },
+      {
+        "conformer": 50,
+        "energy_hartree": -49.9915737912,
+        "wall_seconds": 0.2043595314025879
+      },
+      {
+        "conformer": 51,
+        "energy_hartree": -49.9959909259,
+        "wall_seconds": 0.20232796669006348
+      },
+      {
+        "conformer": 52,
+        "energy_hartree": -49.9908561553,
+        "wall_seconds": 0.18432140350341797
+      }
+    ],
+    "metrics": {
+      "n": 52,
+      "spearman": 0.790864461766522,
+      "mae_kcal": 1.689832006338082,
+      "rms_kcal": 1.9294491628914723,
+      "pred_rel_kcal": {
+        "1": 0.0,
+        "2": 1.0619222978944496,
+        "3": 1.3054749773549978,
+        "4": 1.605223891282715,
+        "5": -0.2895757303316554,
+        "6": -0.2747774254030021,
+        "7": -0.197986141741885,
+        "8": -0.3007232478876767,
+        "9": 2.1605567865723625,
+        "10": 0.22088553122845056,
+        "11": 2.592130506324246,
+        "12": 2.8978142793561723,
+        "13": 5.2230361408287935,
+        "14": 2.582259092029238,
+        "15": 5.451363808923,
+        "16": 3.393106879894243,
+        "17": 3.420142937342793,
+        "18": 3.5463445052600213,
+        "19": 0.5212015254670637,
+        "20": 1.4905681839330704,
+        "21": 2.8350721170594557,
+        "22": 3.6705394861336353,
+        "23": 3.8027796404764094,
+        "24": 5.934386214706473,
+        "25": 3.695250997485054,
+        "26": 5.979662027287286,
+        "27": 1.8396737300579817,
+        "28": 2.169056465154565,
+        "29": 6.829754132170777,
+        "30": 2.245444448475519,
+        "31": 2.4542435197252663,
+        "32": 5.007033700075747,
+        "33": 5.096251428356799,
+        "34": 2.969206408747666,
+        "35": 3.0372672158473675,
+        "36": 2.619961618273382,
+        "37": 6.193533759623582,
+        "38": 2.864973069012447,
+        "39": 3.420735808294101,
+        "40": 5.841625818938581,
+        "41": 5.983517194495881,
+        "42": 3.5290671614044915,
+        "43": 6.054235190462456,
+        "44": 6.165418887886763,
+        "45": 3.90780916423605,
+        "46": 6.717598108816994,
+        "47": 6.755757650207826,
+        "48": 4.054489252844931,
+        "49": 7.258074591802551,
+        "50": 7.258907610630586,
+        "51": 4.487113737205918,
+        "52": 7.709230936961297
+      },
+      "ref_rel_kcal": {
+        "1": 0.0,
+        "2": 0.39,
+        "3": 1.74,
+        "4": 1.16,
+        "5": 2.2,
+        "6": 2.2,
+        "7": 2.68,
+        "8": 2.92,
+        "9": 2.23,
+        "10": 2.84,
+        "11": 4.24,
+        "12": 4.45,
+        "13": 3.6,
+        "14": 2.25,
+        "15": 3.74,
+        "16": 5.0,
+        "17": 5.11,
+        "18": 3.18,
+        "19": 3.83,
+        "20": 3.8,
+        "21": 3.11,
+        "22": 5.27,
+        "23": 5.31,
+        "24": 4.5,
+        "25": 3.85,
+        "26": 4.55,
+        "27": 4.76,
+        "28": 4.37,
+        "29": 5.27,
+        "30": 5.67,
+        "31": 4.86,
+        "32": 6.24,
+        "33": 6.26,
+        "34": 5.85,
+        "35": 5.37,
+        "36": 5.53,
+        "37": 7.53,
+        "38": 5.88,
+        "39": 5.58,
+        "40": 6.98,
+        "41": 7.07,
+        "42": 6.39,
+        "43": 7.32,
+        "44": 7.39,
+        "45": 6.18,
+        "46": 7.82,
+        "47": 7.89,
+        "48": 6.74,
+        "49": 8.19,
+        "50": 8.2,
+        "51": 7.28,
+        "52": 8.75
+      }
+    },
+    "total_wall_s": 10.409949541091919
+  },
+  "nnxtb": {
+    "per_conformer": [
+      {
+        "conformer": 1,
+        "energy_hartree": -50.7633567232,
+        "wall_seconds": 0.926626443862915
+      },
+      {
+        "conformer": 2,
+        "energy_hartree": -50.7622320325,
+        "wall_seconds": 1.0047879219055176
+      },
+      {
+        "conformer": 3,
+        "energy_hartree": -50.7602705138,
+        "wall_seconds": 0.8312652111053467
+      },
+      {
+        "conformer": 4,
+        "energy_hartree": -50.7611142291,
+        "wall_seconds": 0.7922918796539307
+      },
+      {
+        "conformer": 5,
+        "energy_hartree": -50.7598333573,
+        "wall_seconds": 0.8306419849395752
+      },
+      {
+        "conformer": 6,
+        "energy_hartree": -50.7603228729,
+        "wall_seconds": 0.8619225025177002
+      },
+      {
+        "conformer": 7,
+        "energy_hartree": -50.7597005994,
+        "wall_seconds": 0.8411312103271484
+      },
+      {
+        "conformer": 8,
+        "energy_hartree": -50.7589513694,
+        "wall_seconds": 0.8629193305969238
+      },
+      {
+        "conformer": 9,
+        "energy_hartree": -50.7594741609,
+        "wall_seconds": 0.7997517585754395
+      },
+      {
+        "conformer": 10,
+        "energy_hartree": -50.7595059895,
+        "wall_seconds": 0.7978377342224121
+      },
+      {
+        "conformer": 11,
+        "energy_hartree": -50.7568421608,
+        "wall_seconds": 0.8267650604248047
+      },
+      {
+        "conformer": 12,
+        "energy_hartree": -50.7564197352,
+        "wall_seconds": 0.8856785297393799
+      },
+      {
+        "conformer": 13,
+        "energy_hartree": -50.7574113151,
+        "wall_seconds": 0.886725902557373
+      },
+      {
+        "conformer": 14,
+        "energy_hartree": -50.7594967096,
+        "wall_seconds": 0.8090348243713379
+      },
+      {
+        "conformer": 15,
+        "energy_hartree": -50.7570118934,
+        "wall_seconds": 0.838996171951294
+      },
+      {
+        "conformer": 16,
+        "energy_hartree": -50.7559779342,
+        "wall_seconds": 0.8345844745635986
+      },
+      {
+        "conformer": 17,
+        "energy_hartree": -50.7559139584,
+        "wall_seconds": 0.8351869583129883
+      },
+      {
+        "conformer": 18,
+        "energy_hartree": -50.7576615032,
+        "wall_seconds": 0.8712975978851318
+      },
+      {
+        "conformer": 19,
+        "energy_hartree": -50.7581787513,
+        "wall_seconds": 0.8050940036773682
+      },
+      {
+        "conformer": 20,
+        "energy_hartree": -50.7579859827,
+        "wall_seconds": 0.8083138465881348
+      },
+      {
+        "conformer": 21,
+        "energy_hartree": -50.7583660379,
+        "wall_seconds": 0.8659372329711914
+      },
+      {
+        "conformer": 22,
+        "energy_hartree": -50.7558959469,
+        "wall_seconds": 0.8334953784942627
+      },
+      {
+        "conformer": 23,
+        "energy_hartree": -50.7557668125,
+        "wall_seconds": 0.8633923530578613
+      },
+      {
+        "conformer": 24,
+        "energy_hartree": -50.7561578291,
+        "wall_seconds": 0.8427202701568604
+      },
+      {
+        "conformer": 25,
+        "energy_hartree": -50.757573812,
+        "wall_seconds": 0.8637418746948242
+      },
+      {
+        "conformer": 26,
+        "energy_hartree": -50.7560182428,
+        "wall_seconds": 0.8201885223388672
+      },
+      {
+        "conformer": 27,
+        "energy_hartree": -50.7566278691,
+        "wall_seconds": 0.8897104263305664
+      },
+      {
+        "conformer": 28,
+        "energy_hartree": -50.7572447254,
+        "wall_seconds": 0.8124392032623291
+      },
+      {
+        "conformer": 29,
+        "energy_hartree": -50.7553662601,
+        "wall_seconds": 0.8897924423217773
+      },
+      {
+        "conformer": 30,
+        "energy_hartree": -50.7551087059,
+        "wall_seconds": 0.8344917297363281
+      },
+      {
+        "conformer": 31,
+        "energy_hartree": -50.7565070269,
+        "wall_seconds": 0.8797955513000488
+      },
+      {
+        "conformer": 32,
+        "energy_hartree": -50.7542931088,
+        "wall_seconds": 0.852916955947876
+      },
+      {
+        "conformer": 33,
+        "energy_hartree": -50.7542518785,
+        "wall_seconds": 0.836174488067627
+      },
+      {
+        "conformer": 34,
+        "energy_hartree": -50.7545644493,
+        "wall_seconds": 0.8856518268585205
+      },
+      {
+        "conformer": 35,
+        "energy_hartree": -50.7561152389,
+        "wall_seconds": 0.8320674896240234
+      },
+      {
+        "conformer": 36,
+        "energy_hartree": -50.7555895723,
+        "wall_seconds": 0.8556416034698486
+      },
+      {
+        "conformer": 37,
+        "energy_hartree": -50.7518927886,
+        "wall_seconds": 0.8284039497375488
+      },
+      {
+        "conformer": 38,
+        "energy_hartree": -50.7550500815,
+        "wall_seconds": 0.875678539276123
+      },
+      {
+        "conformer": 39,
+        "energy_hartree": -50.7554947274,
+        "wall_seconds": 0.8306136131286621
+      },
+      {
+        "conformer": 40,
+        "energy_hartree": -50.753273563,
+        "wall_seconds": 0.8711426258087158
+      },
+      {
+        "conformer": 41,
+        "energy_hartree": -50.7530910558,
+        "wall_seconds": 0.8271522521972656
+      },
+      {
+        "conformer": 42,
+        "energy_hartree": -50.7546400451,
+        "wall_seconds": 0.8617074489593506
+      },
+      {
+        "conformer": 43,
+        "energy_hartree": -50.7527479063,
+        "wall_seconds": 0.8467421531677246
+      },
+      {
+        "conformer": 44,
+        "energy_hartree": -50.752620017,
+        "wall_seconds": 0.8589260578155518
+      },
+      {
+        "conformer": 45,
+        "energy_hartree": -50.7547213527,
+        "wall_seconds": 0.8054392337799072
+      },
+      {
+        "conformer": 46,
+        "energy_hartree": -50.7523848943,
+        "wall_seconds": 0.8316338062286377
+      },
+      {
+        "conformer": 47,
+        "energy_hartree": -50.7522268583,
+        "wall_seconds": 0.8382835388183594
+      },
+      {
+        "conformer": 48,
+        "energy_hartree": -50.753814928,
+        "wall_seconds": 0.8043012619018555
+      },
+      {
+        "conformer": 49,
+        "energy_hartree": -50.7515286143,
+        "wall_seconds": 0.8313839435577393
+      },
+      {
+        "conformer": 50,
+        "energy_hartree": -50.7515297729,
+        "wall_seconds": 0.8619956970214844
+      },
+      {
+        "conformer": 51,
+        "energy_hartree": -50.7531178269,
+        "wall_seconds": 0.8244142532348633
+      },
+      {
+        "conformer": 52,
+        "energy_hartree": -50.7507944944,
+        "wall_seconds": 0.8033058643341064
+      }
+    ],
+    "metrics": {
+      "n": 52,
+      "spearman": 0.991163287989588,
+      "mae_kcal": 0.4573819811641237,
+      "rms_kcal": 0.5295348052963772,
+      "pred_rel_kcal": {
+        "1": 0.0,
+        "2": 0.7057540698819491,
+        "3": 1.9366256381093019,
+        "4": 1.4071862937641402,
+        "5": 2.2109454836014684,
+        "6": 1.90376980679696,
+        "7": 2.2942523236399825,
+        "8": 2.764401247052725,
+        "9": 2.436344627731966,
+        "10": 2.416371879678302,
+        "11": 4.087949626787998,
+        "12": 4.353025692966445,
+        "13": 3.7307999112091683,
+        "14": 2.422195104847832,
+        "15": 3.9814408121939073,
+        "16": 4.630260006213393,
+        "17": 4.670405426835704,
+        "18": 3.5738045081088026,
+        "19": 3.2492264248035365,
+        "20": 3.3701905476482406,
+        "21": 3.131702308900563,
+        "22": 4.681707813733874,
+        "23": 4.762740873188124,
+        "24": 4.51737425208844,
+        "25": 3.628831566920388,
+        "26": 4.604965977814504,
+        "27": 4.222419698795811,
+        "28": 3.8353365262772217,
+        "29": 5.0140912991350595,
+        "30": 5.175708999774034,
+        "31": 4.298249324187168,
+        "32": 5.687503907219821,
+        "33": 5.713376311096618,
+        "34": 5.5172351727143,
+        "35": 4.544100006097463,
+        "36": 4.873960777908748,
+        "37": 7.1937275740229945,
+        "38": 5.212496366196508,
+        "39": 4.933476851248954,
+        "40": 6.327278556181206,
+        "41": 6.441803553306575,
+        "42": 5.4697980919994675,
+        "43": 6.657133115649295,
+        "44": 6.737384863060469,
+        "45": 5.41877680266493,
+        "46": 6.884926584927321,
+        "47": 6.984095672205872,
+        "48": 5.987566889639763,
+        "49": 7.422250397564245,
+        "50": 7.421523365087381,
+        "51": 6.42500443441749,
+        "52": 7.882917590073226
+      },
+      "ref_rel_kcal": {
+        "1": 0.0,
+        "2": 0.39,
+        "3": 1.74,
+        "4": 1.16,
+        "5": 2.2,
+        "6": 2.2,
+        "7": 2.68,
+        "8": 2.92,
+        "9": 2.23,
+        "10": 2.84,
+        "11": 4.24,
+        "12": 4.45,
+        "13": 3.6,
+        "14": 2.25,
+        "15": 3.74,
+        "16": 5.0,
+        "17": 5.11,
+        "18": 3.18,
+        "19": 3.83,
+        "20": 3.8,
+        "21": 3.11,
+        "22": 5.27,
+        "23": 5.31,
+        "24": 4.5,
+        "25": 3.85,
+        "26": 4.55,
+        "27": 4.76,
+        "28": 4.37,
+        "29": 5.27,
+        "30": 5.67,
+        "31": 4.86,
+        "32": 6.24,
+        "33": 6.26,
+        "34": 5.85,
+        "35": 5.37,
+        "36": 5.53,
+        "37": 7.53,
+        "38": 5.88,
+        "39": 5.58,
+        "40": 6.98,
+        "41": 7.07,
+        "42": 6.39,
+        "43": 7.32,
+        "44": 7.39,
+        "45": 6.18,
+        "46": 7.82,
+        "47": 7.89,
+        "48": 6.74,
+        "49": 8.19,
+        "50": 8.2,
+        "51": 7.28,
+        "52": 8.75
+      }
+    },
+    "total_wall_s": 44.06998896598816
+  }
+}
\ No newline at end of file
diff --git a/benchmark/nnxtb/molecules/2-butanol.xyz b/benchmark/nnxtb/molecules/2-butanol.xyz
new file mode 100644
index 00000000..43186990
--- /dev/null
+++ b/benchmark/nnxtb/molecules/2-butanol.xyz
@@ -0,0 +1,17 @@
+15
+2-butanol — multiple rotameric minima around the C-C-O backbone
+C   -2.0070    0.2780   -0.3760
+C   -0.6810   -0.3630    0.0240
+C    0.5140    0.5680   -0.2040
+C    1.8460   -0.0880    0.1740
+O   -0.5630   -1.5880   -0.7020
+H   -2.8540   -0.3600   -0.1040
+H   -2.1340    1.2240    0.1680
+H   -2.0230    0.5040   -1.4500
+H   -0.6220   -0.6030    1.0990
+H    0.5200    0.8050   -1.2790
+H    0.4120    1.5260    0.3270
+H    2.6910    0.5790   -0.0270
+H    1.9110   -0.3620    1.2360
+H    1.9680   -1.0130   -0.4050
+H    0.2270   -1.8900   -0.3640
diff --git a/benchmark/nnxtb/molecules/n-butane.xyz b/benchmark/nnxtb/molecules/n-butane.xyz
new file mode 100644
index 00000000..5bd714a1
--- /dev/null
+++ b/benchmark/nnxtb/molecules/n-butane.xyz
@@ -0,0 +1,16 @@
+14
+n-butane — CREST finds anti and gauche conformers
+C   -1.8870    0.0000    0.0000
+C   -0.6290    0.7480    0.0000
+C    0.6290   -0.0750    0.0000
+C    1.8870    0.7480    0.0000
+H   -2.7850    0.6340    0.0000
+H   -1.9400   -0.6440    0.8820
+H   -1.9400   -0.6440   -0.8820
+H   -0.6090    1.4070    0.8800
+H   -0.6090    1.4070   -0.8800
+H    0.6090   -0.7340    0.8800
+H    0.6090   -0.7340   -0.8800
+H    2.7850    0.1140    0.0000
+H    1.9400    1.3920    0.8820
+H    1.9400    1.3920   -0.8820
diff --git a/benchmark/nnxtb/phase3_phase4_results.json b/benchmark/nnxtb/phase3_phase4_results.json
new file mode 100644
index 00000000..bcf80579
--- /dev/null
+++ b/benchmark/nnxtb/phase3_phase4_results.json
@@ -0,0 +1,1424 @@
+[
+  {
+    "name": "aspirin",
+    "smiles": "CC(=O)Oc1ccccc1C(=O)O",
+    "n_atoms": 21,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 48.01229381561279,
+      "n_conformers": 4,
+      "lowest_crest_energy": -39.63185877,
+      "lowest_dft_energy": -648.7620212237289,
+      "crest_log_path": "workdir/aspirin/gfn2/crest.log",
+      "ensemble_path": "workdir/aspirin/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1121.0044553279877,
+      "n_conformers": 4,
+      "lowest_crest_energy": -40.17547368,
+      "lowest_dft_energy": -648.7639193759331,
+      "crest_log_path": "workdir/aspirin/nnxtb/crest.log",
+      "ensemble_path": "workdir/aspirin/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.00189815220426226,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 4,
+      "n_nnxtb": 4,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.5,
+        "0.5": 0.5,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.5,
+        "0.5": 0.5,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 2,
+        "0.5": 2,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 2,
+        "0.5": 2,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "ibuprofen",
+    "smiles": "CC(C)Cc1ccc(C(C)C(=O)O)cc1",
+    "n_atoms": 33,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 132.74923610687256,
+      "n_conformers": 11,
+      "lowest_crest_energy": -45.17191624,
+      "lowest_dft_energy": -656.768637927255,
+      "crest_log_path": "workdir/ibuprofen/gfn2/crest.log",
+      "ensemble_path": "workdir/ibuprofen/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 3089.155930042267,
+      "n_conformers": 20,
+      "lowest_crest_energy": -45.86725331,
+      "lowest_dft_energy": -656.7700871849523,
+      "crest_log_path": "workdir/ibuprofen/nnxtb/crest.log",
+      "ensemble_path": "workdir/ibuprofen/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0014492576973452742,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 11,
+      "n_nnxtb": 20,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.7272727272727273,
+        "0.5": 0.7272727272727273,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.4,
+        "0.5": 0.45,
+        "1.0": 0.9
+      },
+      "novel_in_nnxtb": {
+        "0.25": 12,
+        "0.5": 11,
+        "1.0": 2
+      },
+      "novel_in_gfn2": {
+        "0.25": 3,
+        "0.5": 3,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "acetaminophen",
+    "smiles": "CC(=O)Nc1ccc(O)cc1",
+    "n_atoms": 20,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 30.176738500595093,
+      "n_conformers": 2,
+      "lowest_crest_energy": -32.7836376,
+      "lowest_dft_energy": -515.5307371351827,
+      "crest_log_path": "workdir/acetaminophen/gfn2/crest.log",
+      "ensemble_path": "workdir/acetaminophen/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 874.5299360752106,
+      "n_conformers": 4,
+      "lowest_crest_energy": -33.2651833,
+      "lowest_dft_energy": -515.5314965899817,
+      "crest_log_path": "workdir/acetaminophen/nnxtb/crest.log",
+      "ensemble_path": "workdir/acetaminophen/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0007594547989810962,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 2,
+      "n_nnxtb": 4,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.5,
+        "0.5": 0.5,
+        "1.0": 0.5
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 1,
+        "0.5": 1,
+        "1.0": 1
+      }
+    }
+  },
+  {
+    "name": "caffeine",
+    "smiles": "Cn1cnc2c1c(=O)n(C)c(=O)n2C",
+    "n_atoms": 24,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 48.64701271057129,
+      "n_conformers": 6,
+      "lowest_crest_energy": -42.15450736,
+      "lowest_dft_energy": -680.442505273442,
+      "crest_log_path": "workdir/caffeine/gfn2/crest.log",
+      "ensemble_path": "workdir/caffeine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1175.2328298091888,
+      "n_conformers": 7,
+      "lowest_crest_energy": -42.76262693,
+      "lowest_dft_energy": -680.4452326242949,
+      "crest_log_path": "workdir/caffeine/nnxtb/crest.log",
+      "ensemble_path": "workdir/caffeine/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.002727350852978816,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 6,
+      "n_nnxtb": 7,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "benzocaine",
+    "smiles": "CCOC(=O)c1ccc(N)cc1",
+    "n_atoms": 23,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 44.45894646644592,
+      "n_conformers": 3,
+      "lowest_crest_energy": -35.94154157,
+      "lowest_dft_energy": -554.8537396398225,
+      "crest_log_path": "workdir/benzocaine/gfn2/crest.log",
+      "ensemble_path": "workdir/benzocaine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1000.9925756454468,
+      "n_conformers": 3,
+      "lowest_crest_energy": -36.48253184,
+      "lowest_dft_energy": -554.8544976607374,
+      "crest_log_path": "workdir/benzocaine/nnxtb/crest.log",
+      "ensemble_path": "workdir/benzocaine/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.000758020914872759,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 3,
+      "n_nnxtb": 3,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.6666666666666666,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.6666666666666666,
+        "0.5": 0.6666666666666666,
+        "1.0": 0.6666666666666666
+      },
+      "novel_in_nnxtb": {
+        "0.25": 1,
+        "0.5": 1,
+        "1.0": 1
+      },
+      "novel_in_gfn2": {
+        "0.25": 1,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "theobromine",
+    "smiles": "Cn1cnc2c1c(=O)[nH]c(=O)n2C",
+    "n_atoms": 21,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 34.635175466537476,
+      "n_conformers": 3,
+      "lowest_crest_energy": -38.98642605,
+      "lowest_dft_energy": -641.1320925135108,
+      "crest_log_path": "workdir/theobromine/gfn2/crest.log",
+      "ensemble_path": "workdir/theobromine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 806.2880618572235,
+      "n_conformers": 4,
+      "lowest_crest_energy": -39.54565645,
+      "lowest_dft_energy": -641.1342701053175,
+      "crest_log_path": "workdir/theobromine/nnxtb/crest.log",
+      "ensemble_path": "workdir/theobromine/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0021775918066850863,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 3,
+      "n_nnxtb": 4,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "phenacetin",
+    "smiles": "CCOc1ccc(NC(C)=O)cc1",
+    "n_atoms": 26,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 69.96505999565125,
+      "n_conformers": 4,
+      "lowest_crest_energy": -39.10949265,
+      "lowest_dft_energy": -594.1535331428453,
+      "crest_log_path": "workdir/phenacetin/gfn2/crest.log",
+      "ensemble_path": "workdir/phenacetin/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1647.4600987434387,
+      "n_conformers": 14,
+      "lowest_crest_energy": -39.69180493,
+      "lowest_dft_energy": -594.1563679524114,
+      "crest_log_path": "workdir/phenacetin/nnxtb/crest.log",
+      "ensemble_path": "workdir/phenacetin/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.002834809566138574,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 4,
+      "n_nnxtb": 14,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.75,
+        "0.5": 0.75,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.35714285714285715,
+        "0.5": 0.6428571428571429,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 9,
+        "0.5": 5,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 1,
+        "0.5": 1,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "lidocaine",
+    "smiles": "CCN(CC)CC(=O)Nc1c(C)cccc1C",
+    "n_atoms": 39,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 337.4662437438965,
+      "n_conformers": 52,
+      "lowest_crest_energy": -51.13469175,
+      "lowest_dft_energy": -731.5509699050483,
+      "crest_log_path": "workdir/lidocaine/gfn2/crest.log",
+      "ensemble_path": "workdir/lidocaine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 0.0,
+      "n_conformers": 0,
+      "lowest_crest_energy": null,
+      "lowest_dft_energy": null,
+      "crest_log_path": null,
+      "ensemble_path": null
+    },
+    "dft_delta_hartree": null,
+    "winner": "tie",
+    "recall": {
+      "n_gfn2": 52,
+      "n_nnxtb": 0,
+      "recall_gfn2_into_nnxtb": {},
+      "recall_nnxtb_into_gfn2": {},
+      "novel_in_nnxtb": {},
+      "novel_in_gfn2": {}
+    }
+  },
+  {
+    "name": "diphenhydramine",
+    "smiles": "CN(C)CCOC(c1ccccc1)c1ccccc1",
+    "n_atoms": 40,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 440.0136842727661,
+      "n_conformers": 99,
+      "lowest_crest_energy": -54.07158418,
+      "lowest_dft_energy": -790.4700797870846,
+      "crest_log_path": "workdir/diphenhydramine/gfn2/crest.log",
+      "ensemble_path": "workdir/diphenhydramine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 0.0,
+      "n_conformers": 0,
+      "lowest_crest_energy": null,
+      "lowest_dft_energy": null,
+      "crest_log_path": null,
+      "ensemble_path": null
+    },
+    "dft_delta_hartree": null,
+    "winner": "tie",
+    "recall": {
+      "n_gfn2": 99,
+      "n_nnxtb": 0,
+      "recall_gfn2_into_nnxtb": {},
+      "recall_nnxtb_into_gfn2": {},
+      "novel_in_nnxtb": {},
+      "novel_in_gfn2": {}
+    }
+  },
+  {
+    "name": "propranolol",
+    "smiles": "CC(C)NCC(O)COc1cccc2ccccc12",
+    "n_atoms": 40,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 653.118087053299,
+      "n_conformers": 317,
+      "lowest_crest_energy": -56.02307911,
+      "lowest_dft_energy": -827.6211418031423,
+      "crest_log_path": "workdir/propranolol/gfn2/crest.log",
+      "ensemble_path": "workdir/propranolol/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 0.0,
+      "n_conformers": 0,
+      "lowest_crest_energy": null,
+      "lowest_dft_energy": null,
+      "crest_log_path": null,
+      "ensemble_path": null
+    },
+    "dft_delta_hartree": null,
+    "winner": "tie",
+    "recall": {
+      "n_gfn2": 317,
+      "n_nnxtb": 0,
+      "recall_gfn2_into_nnxtb": {},
+      "recall_nnxtb_into_gfn2": {},
+      "novel_in_nnxtb": {},
+      "novel_in_gfn2": {}
+    }
+  },
+  {
+    "name": "salicylic_acid",
+    "smiles": "OC(=O)c1ccccc1O",
+    "n_atoms": 16,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 23.116711616516113,
+      "n_conformers": 2,
+      "lowest_crest_energy": -30.25652413,
+      "lowest_dft_energy": -496.1045638983105,
+      "crest_log_path": "workdir/salicylic_acid/gfn2/crest.log",
+      "ensemble_path": "workdir/salicylic_acid/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 549.6680269241333,
+      "n_conformers": 2,
+      "lowest_crest_energy": -30.67762896,
+      "lowest_dft_energy": -496.10503400736815,
+      "crest_log_path": "workdir/salicylic_acid/nnxtb/crest.log",
+      "ensemble_path": "workdir/salicylic_acid/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0004701090576304523,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 2,
+      "n_nnxtb": 2,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "naproxen",
+    "smiles": "COc1ccc2cc(C(C)C(=O)O)ccc2c1",
+    "n_atoms": 31,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 144.7313039302826,
+      "n_conformers": 9,
+      "lowest_crest_energy": -49.33401124,
+      "lowest_dft_energy": -767.6845659720677,
+      "crest_log_path": "workdir/naproxen/gfn2/crest.log",
+      "ensemble_path": "workdir/naproxen/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1763.82155418396,
+      "n_conformers": 15,
+      "lowest_crest_energy": -50.07651833,
+      "lowest_dft_energy": -767.6858986254722,
+      "crest_log_path": "workdir/naproxen/nnxtb/crest.log",
+      "ensemble_path": "workdir/naproxen/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0013326534044608707,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 9,
+      "n_nnxtb": 15,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.6666666666666666,
+        "0.5": 0.9333333333333333,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 5,
+        "0.5": 1,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "mefenamic_acid",
+    "smiles": "Cc1cccc(C)c1Nc1ccccc1C(=O)O",
+    "n_atoms": 33,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 147.07683324813843,
+      "n_conformers": 1,
+      "lowest_crest_energy": -50.84830363,
+      "lowest_dft_energy": -785.9366008227337,
+      "crest_log_path": "workdir/mefenamic_acid/gfn2/crest.log",
+      "ensemble_path": "workdir/mefenamic_acid/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1741.4285459518433,
+      "n_conformers": 6,
+      "lowest_crest_energy": -51.6559091,
+      "lowest_dft_energy": -785.9386747682322,
+      "crest_log_path": "workdir/mefenamic_acid/nnxtb/crest.log",
+      "ensemble_path": "workdir/mefenamic_acid/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.002073945498523244,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 1,
+      "n_nnxtb": 6,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.0,
+        "0.5": 0.0,
+        "1.0": 0.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.0,
+        "0.5": 0.0,
+        "1.0": 0.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 6,
+        "0.5": 6,
+        "1.0": 6
+      },
+      "novel_in_gfn2": {
+        "0.25": 1,
+        "0.5": 1,
+        "1.0": 1
+      }
+    }
+  },
+  {
+    "name": "isoniazid",
+    "smiles": "NNC(=O)c1ccncc1",
+    "n_atoms": 17,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 28.331334352493286,
+      "n_conformers": 2,
+      "lowest_crest_energy": -29.20111886,
+      "lowest_dft_energy": -472.34553590019345,
+      "crest_log_path": "workdir/isoniazid/gfn2/crest.log",
+      "ensemble_path": "workdir/isoniazid/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 731.8517453670502,
+      "n_conformers": 1,
+      "lowest_crest_energy": -29.66447476,
+      "lowest_dft_energy": -472.3484642812905,
+      "crest_log_path": "workdir/isoniazid/nnxtb/crest.log",
+      "ensemble_path": "workdir/isoniazid/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.002928381097035526,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 2,
+      "n_nnxtb": 1,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.0,
+        "0.5": 0.5,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 1,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 2,
+        "0.5": 1,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "metformin",
+    "smiles": "CN(C)C(=N)NC(N)=N",
+    "n_atoms": 20,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 31.36707830429077,
+      "n_conformers": 12,
+      "lowest_crest_energy": -28.77472948,
+      "lowest_dft_energy": -432.83349429647956,
+      "crest_log_path": "workdir/metformin/gfn2/crest.log",
+      "ensemble_path": "workdir/metformin/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1699.6981859207153,
+      "n_conformers": 7,
+      "lowest_crest_energy": -29.23075427,
+      "lowest_dft_energy": -432.83901542975843,
+      "crest_log_path": "workdir/metformin/nnxtb/crest.log",
+      "ensemble_path": "workdir/metformin/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.005521133278875823,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 12,
+      "n_nnxtb": 7,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.5833333333333334,
+        "0.5": 0.6666666666666666,
+        "1.0": 0.8333333333333334
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 5,
+        "0.5": 4,
+        "1.0": 2
+      }
+    }
+  },
+  {
+    "name": "propofol",
+    "smiles": "CC(C)c1cccc(C(C)C)c1O",
+    "n_atoms": 31,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 70.74724316596985,
+      "n_conformers": 16,
+      "lowest_crest_energy": -38.95120625,
+      "lowest_dft_energy": -543.400274118748,
+      "crest_log_path": "workdir/propofol/gfn2/crest.log",
+      "ensemble_path": "workdir/propofol/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1340.336655139923,
+      "n_conformers": 9,
+      "lowest_crest_energy": -39.56918793,
+      "lowest_dft_energy": -543.4020432852391,
+      "crest_log_path": "workdir/propofol/nnxtb/crest.log",
+      "ensemble_path": "workdir/propofol/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.001769166491044416,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 16,
+      "n_nnxtb": 9,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.625,
+        "0.5": 0.875,
+        "1.0": 0.9375
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.7777777777777778,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 2,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 6,
+        "0.5": 2,
+        "1.0": 1
+      }
+    }
+  },
+  {
+    "name": "theophylline",
+    "smiles": "Cn1c(=O)c2[nH]cnc2n(C)c1=O",
+    "n_atoms": 21,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 30.571295261383057,
+      "n_conformers": 3,
+      "lowest_crest_energy": -38.97975216,
+      "lowest_dft_energy": -641.1295842222779,
+      "crest_log_path": "workdir/theophylline/gfn2/crest.log",
+      "ensemble_path": "workdir/theophylline/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 801.1018433570862,
+      "n_conformers": 4,
+      "lowest_crest_energy": -39.5426107,
+      "lowest_dft_energy": -641.1311923208185,
+      "crest_log_path": "workdir/theophylline/nnxtb/crest.log",
+      "ensemble_path": "workdir/theophylline/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.001608098540600622,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 3,
+      "n_nnxtb": 4,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "diflunisal",
+    "smiles": "OC(=O)c1cc(-c2ccc(F)cc2F)ccc1O",
+    "n_atoms": 26,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 66.79989218711853,
+      "n_conformers": 4,
+      "lowest_crest_energy": -53.58776026,
+      "lowest_dft_energy": -925.6611570210091,
+      "crest_log_path": "workdir/diflunisal/gfn2/crest.log",
+      "ensemble_path": "workdir/diflunisal/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1079.118264913559,
+      "n_conformers": 6,
+      "lowest_crest_energy": -54.30655955,
+      "lowest_dft_energy": -925.6626186858812,
+      "crest_log_path": "workdir/diflunisal/nnxtb/crest.log",
+      "ensemble_path": "workdir/diflunisal/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0014616648720675585,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 4,
+      "n_nnxtb": 6,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.5,
+        "0.5": 0.5,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.3333333333333333,
+        "0.5": 0.3333333333333333,
+        "1.0": 0.5
+      },
+      "novel_in_nnxtb": {
+        "0.25": 4,
+        "0.5": 4,
+        "1.0": 3
+      },
+      "novel_in_gfn2": {
+        "0.25": 2,
+        "0.5": 2,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "vanillin",
+    "smiles": "COc1cc(C=O)ccc1O",
+    "n_atoms": 19,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 28.573541402816772,
+      "n_conformers": 2,
+      "lowest_crest_energy": -33.37599475,
+      "lowest_dft_energy": -535.3609855656035,
+      "crest_log_path": "workdir/vanillin/gfn2/crest.log",
+      "ensemble_path": "workdir/vanillin/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 780.5972700119019,
+      "n_conformers": 2,
+      "lowest_crest_energy": -33.84761208,
+      "lowest_dft_energy": -535.3688337930745,
+      "crest_log_path": "workdir/vanillin/nnxtb/crest.log",
+      "ensemble_path": "workdir/vanillin/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.00784822747095859,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 2,
+      "n_nnxtb": 2,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.5,
+        "0.5": 0.5,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 1,
+        "0.5": 1,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "amphetamine",
+    "smiles": "CC(N)Cc1ccccc1",
+    "n_atoms": 23,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 37.29288387298584,
+      "n_conformers": 10,
+      "lowest_crest_energy": -28.80264717,
+      "lowest_dft_energy": -405.57219336443694,
+      "crest_log_path": "workdir/amphetamine/gfn2/crest.log",
+      "ensemble_path": "workdir/amphetamine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1097.9997432231903,
+      "n_conformers": 12,
+      "lowest_crest_energy": -29.29383082,
+      "lowest_dft_energy": -405.5736507353753,
+      "crest_log_path": "workdir/amphetamine/nnxtb/crest.log",
+      "ensemble_path": "workdir/amphetamine/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0014573709383398636,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 10,
+      "n_nnxtb": 12,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.7,
+        "0.5": 0.8,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.5833333333333334,
+        "0.5": 0.5833333333333334,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 5,
+        "0.5": 5,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 3,
+        "0.5": 2,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "guaiacol",
+    "smiles": "COc1ccccc1O",
+    "n_atoms": 17,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 39.411060094833374,
+      "n_conformers": 3,
+      "lowest_crest_energy": -27.18713808,
+      "lowest_dft_energy": -422.02856331246574,
+      "crest_log_path": "workdir/guaiacol/gfn2/crest.log",
+      "ensemble_path": "workdir/guaiacol/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 762.9869558811188,
+      "n_conformers": 2,
+      "lowest_crest_energy": -27.5747573,
+      "lowest_dft_energy": -422.02883769611304,
+      "crest_log_path": "workdir/guaiacol/nnxtb/crest.log",
+      "ensemble_path": "workdir/guaiacol/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0002743836473086958,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 3,
+      "n_nnxtb": 2,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.6666666666666666,
+        "0.5": 0.6666666666666666,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 1,
+        "0.5": 1,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "tryptamine",
+    "smiles": "NCCc1c[nH]c2ccccc12",
+    "n_atoms": 24,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 90.28900742530823,
+      "n_conformers": 16,
+      "lowest_crest_energy": -33.37171619,
+      "lowest_dft_energy": -497.83620706660145,
+      "crest_log_path": "workdir/tryptamine/gfn2/crest.log",
+      "ensemble_path": "workdir/tryptamine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1232.4667434692383,
+      "n_conformers": 15,
+      "lowest_crest_energy": -33.93268248,
+      "lowest_dft_energy": -497.8376391098442,
+      "crest_log_path": "workdir/tryptamine/nnxtb/crest.log",
+      "ensemble_path": "workdir/tryptamine/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.00143204324274393,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 16,
+      "n_nnxtb": 15,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.8125,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.6666666666666666,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 5,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 3,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "tyramine",
+    "smiles": "NCCc1ccc(O)cc1",
+    "n_atoms": 21,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 36.59581732749939,
+      "n_conformers": 18,
+      "lowest_crest_energy": -29.70879768,
+      "lowest_dft_energy": -441.483097265115,
+      "crest_log_path": "workdir/tyramine/gfn2/crest.log",
+      "ensemble_path": "workdir/tyramine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1097.3079569339752,
+      "n_conformers": 16,
+      "lowest_crest_energy": -30.17364805,
+      "lowest_dft_energy": -441.48410761411384,
+      "crest_log_path": "workdir/tyramine/nnxtb/crest.log",
+      "ensemble_path": "workdir/tyramine/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0010103489988182446,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 18,
+      "n_nnxtb": 16,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.7222222222222222,
+        "0.5": 0.9444444444444444,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.8125,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 3,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 5,
+        "0.5": 1,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "dopamine",
+    "smiles": "NCCc1ccc(O)c(O)c1",
+    "n_atoms": 22,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 58.376280307769775,
+      "n_conformers": 19,
+      "lowest_crest_energy": -33.78296422,
+      "lowest_dft_energy": -516.7167678752379,
+      "crest_log_path": "workdir/dopamine/gfn2/crest.log",
+      "ensemble_path": "workdir/dopamine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1062.716608285904,
+      "n_conformers": 23,
+      "lowest_crest_energy": -34.27652884,
+      "lowest_dft_energy": -516.7180059033582,
+      "crest_log_path": "workdir/dopamine/nnxtb/crest.log",
+      "ensemble_path": "workdir/dopamine/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0012380281202695187,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 19,
+      "n_nnxtb": 23,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.7894736842105263,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.9565217391304348,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 1,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 4,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "histamine",
+    "smiles": "NCCc1c[nH]cn1",
+    "n_atoms": 17,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 25.417662382125854,
+      "n_conformers": 24,
+      "lowest_crest_energy": -24.05426496,
+      "lowest_dft_energy": -360.23070174233317,
+      "crest_log_path": "workdir/histamine/gfn2/crest.log",
+      "ensemble_path": "workdir/histamine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 873.7109248638153,
+      "n_conformers": 17,
+      "lowest_crest_energy": -24.43970524,
+      "lowest_dft_energy": -360.2319224322582,
+      "crest_log_path": "workdir/histamine/nnxtb/crest.log",
+      "ensemble_path": "workdir/histamine/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0012206899250486458,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 24,
+      "n_nnxtb": 17,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.5833333333333334,
+        "0.5": 0.9583333333333334,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.7058823529411765,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 5,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 10,
+        "0.5": 1,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "phenylalanine",
+    "smiles": "N[C@@H](Cc1ccccc1)C(=O)O",
+    "n_atoms": 23,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 109.4740538597107,
+      "n_conformers": 19,
+      "lowest_crest_energy": -35.92535601,
+      "lowest_dft_energy": -554.8523131031861,
+      "crest_log_path": "workdir/phenylalanine/gfn2/crest.log",
+      "ensemble_path": "workdir/phenylalanine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 2116.501625061035,
+      "n_conformers": 19,
+      "lowest_crest_energy": -36.4737757,
+      "lowest_dft_energy": -554.8538596792761,
+      "crest_log_path": "workdir/phenylalanine/nnxtb/crest.log",
+      "ensemble_path": "workdir/phenylalanine/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0015465760899360248,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 19,
+      "n_nnxtb": 19,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.7894736842105263,
+        "0.5": 0.8947368421052632,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.7368421052631579,
+        "0.5": 0.8421052631578947,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 5,
+        "0.5": 3,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 4,
+        "0.5": 2,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "serotonin",
+    "smiles": "NCCc1c[nH]c2ccc(O)cc12",
+    "n_atoms": 25,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 60.269389629364014,
+      "n_conformers": 17,
+      "lowest_crest_energy": -37.4447207,
+      "lowest_dft_energy": -573.0667509148659,
+      "crest_log_path": "workdir/serotonin/gfn2/crest.log",
+      "ensemble_path": "workdir/serotonin/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1289.9519147872925,
+      "n_conformers": 36,
+      "lowest_crest_energy": -38.03198218,
+      "lowest_dft_energy": -573.0684095151045,
+      "crest_log_path": "workdir/serotonin/nnxtb/crest.log",
+      "ensemble_path": "workdir/serotonin/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0016586002385565735,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 17,
+      "n_nnxtb": 36,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.9411764705882353,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.6944444444444444,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 11,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 1,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "coumarin",
+    "smiles": "O=c1ccc2ccccc2o1",
+    "n_atoms": 17,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 22.506465196609497,
+      "n_conformers": 1,
+      "lowest_crest_energy": -30.45757697,
+      "lowest_dft_energy": -497.0624483716068,
+      "crest_log_path": "workdir/coumarin/gfn2/crest.log",
+      "ensemble_path": "workdir/coumarin/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 511.5465302467346,
+      "n_conformers": 1,
+      "lowest_crest_energy": -30.92409841,
+      "lowest_dft_energy": -497.06302752853526,
+      "crest_log_path": "workdir/coumarin/nnxtb/crest.log",
+      "ensemble_path": "workdir/coumarin/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0005791569284383513,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 1,
+      "n_nnxtb": 1,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "paracetamol_N",
+    "smiles": "O=C(N)c1ccc(O)cc1",
+    "n_atoms": 17,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 32.846306800842285,
+      "n_conformers": 2,
+      "lowest_crest_energy": -29.60647831,
+      "lowest_dft_energy": -476.22051366713754,
+      "crest_log_path": "workdir/paracetamol_N/gfn2/crest.log",
+      "ensemble_path": "workdir/paracetamol_N/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 678.7731022834778,
+      "n_conformers": 3,
+      "lowest_crest_energy": -30.04803652,
+      "lowest_dft_energy": -476.2210789764954,
+      "crest_log_path": "workdir/paracetamol_N/nnxtb/crest.log",
+      "ensemble_path": "workdir/paracetamol_N/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0005653093578530388,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 2,
+      "n_nnxtb": 3,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.5,
+        "0.5": 0.5,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.6666666666666666,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 1,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 1,
+        "0.5": 1,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "anisole",
+    "smiles": "COc1ccccc1",
+    "n_atoms": 16,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 19.603131532669067,
+      "n_conformers": 1,
+      "lowest_crest_energy": -23.11319507,
+      "lowest_dft_energy": -346.7943890396327,
+      "crest_log_path": "workdir/anisole/gfn2/crest.log",
+      "ensemble_path": "workdir/anisole/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 652.470059633255,
+      "n_conformers": 2,
+      "lowest_crest_energy": -23.47137836,
+      "lowest_dft_energy": -346.79464481000986,
+      "crest_log_path": "workdir/anisole/nnxtb/crest.log",
+      "ensemble_path": "workdir/anisole/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.00025577037717994244,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 1,
+      "n_nnxtb": 2,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.5,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 1,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  }
+]
\ No newline at end of file
diff --git a/benchmark/nnxtb/phase3_results.json b/benchmark/nnxtb/phase3_results.json
new file mode 100644
index 00000000..8f88dfab
--- /dev/null
+++ b/benchmark/nnxtb/phase3_results.json
@@ -0,0 +1,444 @@
+[
+  {
+    "name": "aspirin",
+    "smiles": "CC(=O)Oc1ccccc1C(=O)O",
+    "n_atoms": 21,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 48.01229381561279,
+      "n_conformers": 4,
+      "lowest_crest_energy": -39.63185877,
+      "lowest_dft_energy": -648.7620212237289,
+      "crest_log_path": "workdir/aspirin/gfn2/crest.log",
+      "ensemble_path": "workdir/aspirin/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1121.0044553279877,
+      "n_conformers": 4,
+      "lowest_crest_energy": -40.17547368,
+      "lowest_dft_energy": -648.7639193759331,
+      "crest_log_path": "workdir/aspirin/nnxtb/crest.log",
+      "ensemble_path": "workdir/aspirin/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.00189815220426226,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 4,
+      "n_nnxtb": 4,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.5,
+        "0.5": 0.5,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.5,
+        "0.5": 0.5,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 2,
+        "0.5": 2,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 2,
+        "0.5": 2,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "ibuprofen",
+    "smiles": "CC(C)Cc1ccc(C(C)C(=O)O)cc1",
+    "n_atoms": 33,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 132.74923610687256,
+      "n_conformers": 11,
+      "lowest_crest_energy": -45.17191624,
+      "lowest_dft_energy": -656.768637927255,
+      "crest_log_path": "workdir/ibuprofen/gfn2/crest.log",
+      "ensemble_path": "workdir/ibuprofen/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 3089.155930042267,
+      "n_conformers": 20,
+      "lowest_crest_energy": -45.86725331,
+      "lowest_dft_energy": -656.7700871849523,
+      "crest_log_path": "workdir/ibuprofen/nnxtb/crest.log",
+      "ensemble_path": "workdir/ibuprofen/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0014492576973452742,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 11,
+      "n_nnxtb": 20,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.7272727272727273,
+        "0.5": 0.7272727272727273,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.4,
+        "0.5": 0.45,
+        "1.0": 0.9
+      },
+      "novel_in_nnxtb": {
+        "0.25": 12,
+        "0.5": 11,
+        "1.0": 2
+      },
+      "novel_in_gfn2": {
+        "0.25": 3,
+        "0.5": 3,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "acetaminophen",
+    "smiles": "CC(=O)Nc1ccc(O)cc1",
+    "n_atoms": 20,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 30.176738500595093,
+      "n_conformers": 2,
+      "lowest_crest_energy": -32.7836376,
+      "lowest_dft_energy": -515.5307371351827,
+      "crest_log_path": "workdir/acetaminophen/gfn2/crest.log",
+      "ensemble_path": "workdir/acetaminophen/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 874.5299360752106,
+      "n_conformers": 4,
+      "lowest_crest_energy": -33.2651833,
+      "lowest_dft_energy": -515.5314965899817,
+      "crest_log_path": "workdir/acetaminophen/nnxtb/crest.log",
+      "ensemble_path": "workdir/acetaminophen/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0007594547989810962,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 2,
+      "n_nnxtb": 4,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.5,
+        "0.5": 0.5,
+        "1.0": 0.5
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 1,
+        "0.5": 1,
+        "1.0": 1
+      }
+    }
+  },
+  {
+    "name": "caffeine",
+    "smiles": "Cn1cnc2c1c(=O)n(C)c(=O)n2C",
+    "n_atoms": 24,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 48.64701271057129,
+      "n_conformers": 6,
+      "lowest_crest_energy": -42.15450736,
+      "lowest_dft_energy": -680.442505273442,
+      "crest_log_path": "workdir/caffeine/gfn2/crest.log",
+      "ensemble_path": "workdir/caffeine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1175.2328298091888,
+      "n_conformers": 7,
+      "lowest_crest_energy": -42.76262693,
+      "lowest_dft_energy": -680.4452326242949,
+      "crest_log_path": "workdir/caffeine/nnxtb/crest.log",
+      "ensemble_path": "workdir/caffeine/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.002727350852978816,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 6,
+      "n_nnxtb": 7,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "benzocaine",
+    "smiles": "CCOC(=O)c1ccc(N)cc1",
+    "n_atoms": 23,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 44.45894646644592,
+      "n_conformers": 3,
+      "lowest_crest_energy": -35.94154157,
+      "lowest_dft_energy": -554.8537396398225,
+      "crest_log_path": "workdir/benzocaine/gfn2/crest.log",
+      "ensemble_path": "workdir/benzocaine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1000.9925756454468,
+      "n_conformers": 3,
+      "lowest_crest_energy": -36.48253184,
+      "lowest_dft_energy": -554.8544976607374,
+      "crest_log_path": "workdir/benzocaine/nnxtb/crest.log",
+      "ensemble_path": "workdir/benzocaine/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.000758020914872759,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 3,
+      "n_nnxtb": 3,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.6666666666666666,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.6666666666666666,
+        "0.5": 0.6666666666666666,
+        "1.0": 0.6666666666666666
+      },
+      "novel_in_nnxtb": {
+        "0.25": 1,
+        "0.5": 1,
+        "1.0": 1
+      },
+      "novel_in_gfn2": {
+        "0.25": 1,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "theobromine",
+    "smiles": "Cn1cnc2c1c(=O)[nH]c(=O)n2C",
+    "n_atoms": 21,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 34.635175466537476,
+      "n_conformers": 3,
+      "lowest_crest_energy": -38.98642605,
+      "lowest_dft_energy": -641.1320925135108,
+      "crest_log_path": "workdir/theobromine/gfn2/crest.log",
+      "ensemble_path": "workdir/theobromine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 806.2880618572235,
+      "n_conformers": 4,
+      "lowest_crest_energy": -39.54565645,
+      "lowest_dft_energy": -641.1342701053175,
+      "crest_log_path": "workdir/theobromine/nnxtb/crest.log",
+      "ensemble_path": "workdir/theobromine/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0021775918066850863,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 3,
+      "n_nnxtb": 4,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "phenacetin",
+    "smiles": "CCOc1ccc(NC(C)=O)cc1",
+    "n_atoms": 26,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 69.96505999565125,
+      "n_conformers": 4,
+      "lowest_crest_energy": -39.10949265,
+      "lowest_dft_energy": -594.1535331428453,
+      "crest_log_path": "workdir/phenacetin/gfn2/crest.log",
+      "ensemble_path": "workdir/phenacetin/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1647.4600987434387,
+      "n_conformers": 14,
+      "lowest_crest_energy": -39.69180493,
+      "lowest_dft_energy": -594.1563679524114,
+      "crest_log_path": "workdir/phenacetin/nnxtb/crest.log",
+      "ensemble_path": "workdir/phenacetin/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.002834809566138574,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 4,
+      "n_nnxtb": 14,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.75,
+        "0.5": 0.75,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.35714285714285715,
+        "0.5": 0.6428571428571429,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 9,
+        "0.5": 5,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 1,
+        "0.5": 1,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "lidocaine",
+    "smiles": "CCN(CC)CC(=O)Nc1c(C)cccc1C",
+    "n_atoms": 39,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 337.4662437438965,
+      "n_conformers": 52,
+      "lowest_crest_energy": -51.13469175,
+      "lowest_dft_energy": -731.5509699050483,
+      "crest_log_path": "workdir/lidocaine/gfn2/crest.log",
+      "ensemble_path": "workdir/lidocaine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 0.0,
+      "n_conformers": 0,
+      "lowest_crest_energy": null,
+      "lowest_dft_energy": null,
+      "crest_log_path": null,
+      "ensemble_path": null
+    },
+    "dft_delta_hartree": null,
+    "winner": "tie",
+    "recall": {
+      "n_gfn2": 52,
+      "n_nnxtb": 0,
+      "recall_gfn2_into_nnxtb": {},
+      "recall_nnxtb_into_gfn2": {},
+      "novel_in_nnxtb": {},
+      "novel_in_gfn2": {}
+    }
+  },
+  {
+    "name": "diphenhydramine",
+    "smiles": "CN(C)CCOC(c1ccccc1)c1ccccc1",
+    "n_atoms": 40,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 440.0136842727661,
+      "n_conformers": 99,
+      "lowest_crest_energy": -54.07158418,
+      "lowest_dft_energy": -790.4700797870846,
+      "crest_log_path": "workdir/diphenhydramine/gfn2/crest.log",
+      "ensemble_path": "workdir/diphenhydramine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 0.0,
+      "n_conformers": 0,
+      "lowest_crest_energy": null,
+      "lowest_dft_energy": null,
+      "crest_log_path": null,
+      "ensemble_path": null
+    },
+    "dft_delta_hartree": null,
+    "winner": "tie",
+    "recall": {
+      "n_gfn2": 99,
+      "n_nnxtb": 0,
+      "recall_gfn2_into_nnxtb": {},
+      "recall_nnxtb_into_gfn2": {},
+      "novel_in_nnxtb": {},
+      "novel_in_gfn2": {}
+    }
+  },
+  {
+    "name": "propranolol",
+    "smiles": "CC(C)NCC(O)COc1cccc2ccccc12",
+    "n_atoms": 40,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 653.118087053299,
+      "n_conformers": 317,
+      "lowest_crest_energy": -56.02307911,
+      "lowest_dft_energy": -827.6211418031423,
+      "crest_log_path": "workdir/propranolol/gfn2/crest.log",
+      "ensemble_path": "workdir/propranolol/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 0.0,
+      "n_conformers": 0,
+      "lowest_crest_energy": null,
+      "lowest_dft_energy": null,
+      "crest_log_path": null,
+      "ensemble_path": null
+    },
+    "dft_delta_hartree": null,
+    "winner": "tie",
+    "recall": {
+      "n_gfn2": 317,
+      "n_nnxtb": 0,
+      "recall_gfn2_into_nnxtb": {},
+      "recall_nnxtb_into_gfn2": {},
+      "novel_in_nnxtb": {},
+      "novel_in_gfn2": {}
+    }
+  }
+]
\ No newline at end of file
diff --git a/benchmark/nnxtb/phase4_a_results.json b/benchmark/nnxtb/phase4_a_results.json
new file mode 100644
index 00000000..9f375bb1
--- /dev/null
+++ b/benchmark/nnxtb/phase4_a_results.json
@@ -0,0 +1,492 @@
+[
+  {
+    "name": "salicylic_acid",
+    "smiles": "OC(=O)c1ccccc1O",
+    "n_atoms": 16,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 23.116711616516113,
+      "n_conformers": 2,
+      "lowest_crest_energy": -30.25652413,
+      "lowest_dft_energy": -496.1045638983105,
+      "crest_log_path": "workdir/salicylic_acid/gfn2/crest.log",
+      "ensemble_path": "workdir/salicylic_acid/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 549.6680269241333,
+      "n_conformers": 2,
+      "lowest_crest_energy": -30.67762896,
+      "lowest_dft_energy": -496.10503400736815,
+      "crest_log_path": "workdir/salicylic_acid/nnxtb/crest.log",
+      "ensemble_path": "workdir/salicylic_acid/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0004701090576304523,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 2,
+      "n_nnxtb": 2,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "naproxen",
+    "smiles": "COc1ccc2cc(C(C)C(=O)O)ccc2c1",
+    "n_atoms": 31,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 144.7313039302826,
+      "n_conformers": 9,
+      "lowest_crest_energy": -49.33401124,
+      "lowest_dft_energy": -767.6845659720677,
+      "crest_log_path": "workdir/naproxen/gfn2/crest.log",
+      "ensemble_path": "workdir/naproxen/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1763.82155418396,
+      "n_conformers": 15,
+      "lowest_crest_energy": -50.07651833,
+      "lowest_dft_energy": -767.6858986254722,
+      "crest_log_path": "workdir/naproxen/nnxtb/crest.log",
+      "ensemble_path": "workdir/naproxen/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0013326534044608707,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 9,
+      "n_nnxtb": 15,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.6666666666666666,
+        "0.5": 0.9333333333333333,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 5,
+        "0.5": 1,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "mefenamic_acid",
+    "smiles": "Cc1cccc(C)c1Nc1ccccc1C(=O)O",
+    "n_atoms": 33,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 147.07683324813843,
+      "n_conformers": 1,
+      "lowest_crest_energy": -50.84830363,
+      "lowest_dft_energy": -785.9366008227337,
+      "crest_log_path": "workdir/mefenamic_acid/gfn2/crest.log",
+      "ensemble_path": "workdir/mefenamic_acid/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1741.4285459518433,
+      "n_conformers": 6,
+      "lowest_crest_energy": -51.6559091,
+      "lowest_dft_energy": -785.9386747682322,
+      "crest_log_path": "workdir/mefenamic_acid/nnxtb/crest.log",
+      "ensemble_path": "workdir/mefenamic_acid/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.002073945498523244,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 1,
+      "n_nnxtb": 6,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.0,
+        "0.5": 0.0,
+        "1.0": 0.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.0,
+        "0.5": 0.0,
+        "1.0": 0.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 6,
+        "0.5": 6,
+        "1.0": 6
+      },
+      "novel_in_gfn2": {
+        "0.25": 1,
+        "0.5": 1,
+        "1.0": 1
+      }
+    }
+  },
+  {
+    "name": "isoniazid",
+    "smiles": "NNC(=O)c1ccncc1",
+    "n_atoms": 17,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 28.331334352493286,
+      "n_conformers": 2,
+      "lowest_crest_energy": -29.20111886,
+      "lowest_dft_energy": -472.34553590019345,
+      "crest_log_path": "workdir/isoniazid/gfn2/crest.log",
+      "ensemble_path": "workdir/isoniazid/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 731.8517453670502,
+      "n_conformers": 1,
+      "lowest_crest_energy": -29.66447476,
+      "lowest_dft_energy": -472.3484642812905,
+      "crest_log_path": "workdir/isoniazid/nnxtb/crest.log",
+      "ensemble_path": "workdir/isoniazid/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.002928381097035526,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 2,
+      "n_nnxtb": 1,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.0,
+        "0.5": 0.5,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 1,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 2,
+        "0.5": 1,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "metformin",
+    "smiles": "CN(C)C(=N)NC(N)=N",
+    "n_atoms": 20,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 31.36707830429077,
+      "n_conformers": 12,
+      "lowest_crest_energy": -28.77472948,
+      "lowest_dft_energy": -432.83349429647956,
+      "crest_log_path": "workdir/metformin/gfn2/crest.log",
+      "ensemble_path": "workdir/metformin/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1699.6981859207153,
+      "n_conformers": 7,
+      "lowest_crest_energy": -29.23075427,
+      "lowest_dft_energy": -432.83901542975843,
+      "crest_log_path": "workdir/metformin/nnxtb/crest.log",
+      "ensemble_path": "workdir/metformin/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.005521133278875823,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 12,
+      "n_nnxtb": 7,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.5833333333333334,
+        "0.5": 0.6666666666666666,
+        "1.0": 0.8333333333333334
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 5,
+        "0.5": 4,
+        "1.0": 2
+      }
+    }
+  },
+  {
+    "name": "propofol",
+    "smiles": "CC(C)c1cccc(C(C)C)c1O",
+    "n_atoms": 31,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 70.74724316596985,
+      "n_conformers": 16,
+      "lowest_crest_energy": -38.95120625,
+      "lowest_dft_energy": -543.400274118748,
+      "crest_log_path": "workdir/propofol/gfn2/crest.log",
+      "ensemble_path": "workdir/propofol/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1340.336655139923,
+      "n_conformers": 9,
+      "lowest_crest_energy": -39.56918793,
+      "lowest_dft_energy": -543.4020432852391,
+      "crest_log_path": "workdir/propofol/nnxtb/crest.log",
+      "ensemble_path": "workdir/propofol/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.001769166491044416,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 16,
+      "n_nnxtb": 9,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.625,
+        "0.5": 0.875,
+        "1.0": 0.9375
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.7777777777777778,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 2,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 6,
+        "0.5": 2,
+        "1.0": 1
+      }
+    }
+  },
+  {
+    "name": "theophylline",
+    "smiles": "Cn1c(=O)c2[nH]cnc2n(C)c1=O",
+    "n_atoms": 21,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 30.571295261383057,
+      "n_conformers": 3,
+      "lowest_crest_energy": -38.97975216,
+      "lowest_dft_energy": -641.1295842222779,
+      "crest_log_path": "workdir/theophylline/gfn2/crest.log",
+      "ensemble_path": "workdir/theophylline/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 801.1018433570862,
+      "n_conformers": 4,
+      "lowest_crest_energy": -39.5426107,
+      "lowest_dft_energy": -641.1311923208185,
+      "crest_log_path": "workdir/theophylline/nnxtb/crest.log",
+      "ensemble_path": "workdir/theophylline/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.001608098540600622,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 3,
+      "n_nnxtb": 4,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "diflunisal",
+    "smiles": "OC(=O)c1cc(-c2ccc(F)cc2F)ccc1O",
+    "n_atoms": 26,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 66.79989218711853,
+      "n_conformers": 4,
+      "lowest_crest_energy": -53.58776026,
+      "lowest_dft_energy": -925.6611570210091,
+      "crest_log_path": "workdir/diflunisal/gfn2/crest.log",
+      "ensemble_path": "workdir/diflunisal/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1079.118264913559,
+      "n_conformers": 6,
+      "lowest_crest_energy": -54.30655955,
+      "lowest_dft_energy": -925.6626186858812,
+      "crest_log_path": "workdir/diflunisal/nnxtb/crest.log",
+      "ensemble_path": "workdir/diflunisal/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0014616648720675585,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 4,
+      "n_nnxtb": 6,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.5,
+        "0.5": 0.5,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.3333333333333333,
+        "0.5": 0.3333333333333333,
+        "1.0": 0.5
+      },
+      "novel_in_nnxtb": {
+        "0.25": 4,
+        "0.5": 4,
+        "1.0": 3
+      },
+      "novel_in_gfn2": {
+        "0.25": 2,
+        "0.5": 2,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "vanillin",
+    "smiles": "COc1cc(C=O)ccc1O",
+    "n_atoms": 19,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 28.573541402816772,
+      "n_conformers": 2,
+      "lowest_crest_energy": -33.37599475,
+      "lowest_dft_energy": -535.3609855656035,
+      "crest_log_path": "workdir/vanillin/gfn2/crest.log",
+      "ensemble_path": "workdir/vanillin/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 780.5972700119019,
+      "n_conformers": 2,
+      "lowest_crest_energy": -33.84761208,
+      "lowest_dft_energy": -535.3688337930745,
+      "crest_log_path": "workdir/vanillin/nnxtb/crest.log",
+      "ensemble_path": "workdir/vanillin/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.00784822747095859,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 2,
+      "n_nnxtb": 2,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.5,
+        "0.5": 0.5,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 1,
+        "0.5": 1,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "amphetamine",
+    "smiles": "CC(N)Cc1ccccc1",
+    "n_atoms": 23,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 37.29288387298584,
+      "n_conformers": 10,
+      "lowest_crest_energy": -28.80264717,
+      "lowest_dft_energy": -405.57219336443694,
+      "crest_log_path": "workdir/amphetamine/gfn2/crest.log",
+      "ensemble_path": "workdir/amphetamine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1097.9997432231903,
+      "n_conformers": 12,
+      "lowest_crest_energy": -29.29383082,
+      "lowest_dft_energy": -405.5736507353753,
+      "crest_log_path": "workdir/amphetamine/nnxtb/crest.log",
+      "ensemble_path": "workdir/amphetamine/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0014573709383398636,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 10,
+      "n_nnxtb": 12,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.7,
+        "0.5": 0.8,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.5833333333333334,
+        "0.5": 0.5833333333333334,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 5,
+        "0.5": 5,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 3,
+        "0.5": 2,
+        "1.0": 0
+      }
+    }
+  }
+]
\ No newline at end of file
diff --git a/benchmark/nnxtb/phase4_b_results.json b/benchmark/nnxtb/phase4_b_results.json
new file mode 100644
index 00000000..ec7ff69e
--- /dev/null
+++ b/benchmark/nnxtb/phase4_b_results.json
@@ -0,0 +1,492 @@
+[
+  {
+    "name": "guaiacol",
+    "smiles": "COc1ccccc1O",
+    "n_atoms": 17,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 39.411060094833374,
+      "n_conformers": 3,
+      "lowest_crest_energy": -27.18713808,
+      "lowest_dft_energy": -422.02856331246574,
+      "crest_log_path": "workdir/guaiacol/gfn2/crest.log",
+      "ensemble_path": "workdir/guaiacol/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 762.9869558811188,
+      "n_conformers": 2,
+      "lowest_crest_energy": -27.5747573,
+      "lowest_dft_energy": -422.02883769611304,
+      "crest_log_path": "workdir/guaiacol/nnxtb/crest.log",
+      "ensemble_path": "workdir/guaiacol/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0002743836473086958,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 3,
+      "n_nnxtb": 2,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.6666666666666666,
+        "0.5": 0.6666666666666666,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 1,
+        "0.5": 1,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "tryptamine",
+    "smiles": "NCCc1c[nH]c2ccccc12",
+    "n_atoms": 24,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 90.28900742530823,
+      "n_conformers": 16,
+      "lowest_crest_energy": -33.37171619,
+      "lowest_dft_energy": -497.83620706660145,
+      "crest_log_path": "workdir/tryptamine/gfn2/crest.log",
+      "ensemble_path": "workdir/tryptamine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1232.4667434692383,
+      "n_conformers": 15,
+      "lowest_crest_energy": -33.93268248,
+      "lowest_dft_energy": -497.8376391098442,
+      "crest_log_path": "workdir/tryptamine/nnxtb/crest.log",
+      "ensemble_path": "workdir/tryptamine/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.00143204324274393,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 16,
+      "n_nnxtb": 15,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.8125,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.6666666666666666,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 5,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 3,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "tyramine",
+    "smiles": "NCCc1ccc(O)cc1",
+    "n_atoms": 21,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 36.59581732749939,
+      "n_conformers": 18,
+      "lowest_crest_energy": -29.70879768,
+      "lowest_dft_energy": -441.483097265115,
+      "crest_log_path": "workdir/tyramine/gfn2/crest.log",
+      "ensemble_path": "workdir/tyramine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1097.3079569339752,
+      "n_conformers": 16,
+      "lowest_crest_energy": -30.17364805,
+      "lowest_dft_energy": -441.48410761411384,
+      "crest_log_path": "workdir/tyramine/nnxtb/crest.log",
+      "ensemble_path": "workdir/tyramine/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0010103489988182446,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 18,
+      "n_nnxtb": 16,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.7222222222222222,
+        "0.5": 0.9444444444444444,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.8125,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 3,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 5,
+        "0.5": 1,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "dopamine",
+    "smiles": "NCCc1ccc(O)c(O)c1",
+    "n_atoms": 22,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 58.376280307769775,
+      "n_conformers": 19,
+      "lowest_crest_energy": -33.78296422,
+      "lowest_dft_energy": -516.7167678752379,
+      "crest_log_path": "workdir/dopamine/gfn2/crest.log",
+      "ensemble_path": "workdir/dopamine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1062.716608285904,
+      "n_conformers": 23,
+      "lowest_crest_energy": -34.27652884,
+      "lowest_dft_energy": -516.7180059033582,
+      "crest_log_path": "workdir/dopamine/nnxtb/crest.log",
+      "ensemble_path": "workdir/dopamine/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0012380281202695187,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 19,
+      "n_nnxtb": 23,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.7894736842105263,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.9565217391304348,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 1,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 4,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "histamine",
+    "smiles": "NCCc1c[nH]cn1",
+    "n_atoms": 17,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 25.417662382125854,
+      "n_conformers": 24,
+      "lowest_crest_energy": -24.05426496,
+      "lowest_dft_energy": -360.23070174233317,
+      "crest_log_path": "workdir/histamine/gfn2/crest.log",
+      "ensemble_path": "workdir/histamine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 873.7109248638153,
+      "n_conformers": 17,
+      "lowest_crest_energy": -24.43970524,
+      "lowest_dft_energy": -360.2319224322582,
+      "crest_log_path": "workdir/histamine/nnxtb/crest.log",
+      "ensemble_path": "workdir/histamine/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0012206899250486458,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 24,
+      "n_nnxtb": 17,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.5833333333333334,
+        "0.5": 0.9583333333333334,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.7058823529411765,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 5,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 10,
+        "0.5": 1,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "phenylalanine",
+    "smiles": "N[C@@H](Cc1ccccc1)C(=O)O",
+    "n_atoms": 23,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 109.4740538597107,
+      "n_conformers": 19,
+      "lowest_crest_energy": -35.92535601,
+      "lowest_dft_energy": -554.8523131031861,
+      "crest_log_path": "workdir/phenylalanine/gfn2/crest.log",
+      "ensemble_path": "workdir/phenylalanine/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 2116.501625061035,
+      "n_conformers": 19,
+      "lowest_crest_energy": -36.4737757,
+      "lowest_dft_energy": -554.8538596792761,
+      "crest_log_path": "workdir/phenylalanine/nnxtb/crest.log",
+      "ensemble_path": "workdir/phenylalanine/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0015465760899360248,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 19,
+      "n_nnxtb": 19,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.7894736842105263,
+        "0.5": 0.8947368421052632,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.7368421052631579,
+        "0.5": 0.8421052631578947,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 5,
+        "0.5": 3,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 4,
+        "0.5": 2,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "serotonin",
+    "smiles": "NCCc1c[nH]c2ccc(O)cc12",
+    "n_atoms": 25,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 60.269389629364014,
+      "n_conformers": 17,
+      "lowest_crest_energy": -37.4447207,
+      "lowest_dft_energy": -573.0667509148659,
+      "crest_log_path": "workdir/serotonin/gfn2/crest.log",
+      "ensemble_path": "workdir/serotonin/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 1289.9519147872925,
+      "n_conformers": 36,
+      "lowest_crest_energy": -38.03198218,
+      "lowest_dft_energy": -573.0684095151045,
+      "crest_log_path": "workdir/serotonin/nnxtb/crest.log",
+      "ensemble_path": "workdir/serotonin/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0016586002385565735,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 17,
+      "n_nnxtb": 36,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.9411764705882353,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.6944444444444444,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 11,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 1,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "coumarin",
+    "smiles": "O=c1ccc2ccccc2o1",
+    "n_atoms": 17,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 22.506465196609497,
+      "n_conformers": 1,
+      "lowest_crest_energy": -30.45757697,
+      "lowest_dft_energy": -497.0624483716068,
+      "crest_log_path": "workdir/coumarin/gfn2/crest.log",
+      "ensemble_path": "workdir/coumarin/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 511.5465302467346,
+      "n_conformers": 1,
+      "lowest_crest_energy": -30.92409841,
+      "lowest_dft_energy": -497.06302752853526,
+      "crest_log_path": "workdir/coumarin/nnxtb/crest.log",
+      "ensemble_path": "workdir/coumarin/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0005791569284383513,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 1,
+      "n_nnxtb": 1,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "paracetamol_N",
+    "smiles": "O=C(N)c1ccc(O)cc1",
+    "n_atoms": 17,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 32.846306800842285,
+      "n_conformers": 2,
+      "lowest_crest_energy": -29.60647831,
+      "lowest_dft_energy": -476.22051366713754,
+      "crest_log_path": "workdir/paracetamol_N/gfn2/crest.log",
+      "ensemble_path": "workdir/paracetamol_N/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 678.7731022834778,
+      "n_conformers": 3,
+      "lowest_crest_energy": -30.04803652,
+      "lowest_dft_energy": -476.2210789764954,
+      "crest_log_path": "workdir/paracetamol_N/nnxtb/crest.log",
+      "ensemble_path": "workdir/paracetamol_N/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.0005653093578530388,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 2,
+      "n_nnxtb": 3,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 0.5,
+        "0.5": 0.5,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.6666666666666666,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 1,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 1,
+        "0.5": 1,
+        "1.0": 0
+      }
+    }
+  },
+  {
+    "name": "anisole",
+    "smiles": "COc1ccccc1",
+    "n_atoms": 16,
+    "gfn2": {
+      "engine": "gfn2",
+      "wall_seconds": 19.603131532669067,
+      "n_conformers": 1,
+      "lowest_crest_energy": -23.11319507,
+      "lowest_dft_energy": -346.7943890396327,
+      "crest_log_path": "workdir/anisole/gfn2/crest.log",
+      "ensemble_path": "workdir/anisole/gfn2_ensemble.xyz"
+    },
+    "nnxtb": {
+      "engine": "nnxtb",
+      "wall_seconds": 652.470059633255,
+      "n_conformers": 2,
+      "lowest_crest_energy": -23.47137836,
+      "lowest_dft_energy": -346.79464481000986,
+      "crest_log_path": "workdir/anisole/nnxtb/crest.log",
+      "ensemble_path": "workdir/anisole/nnxtb_ensemble.xyz"
+    },
+    "dft_delta_hartree": -0.00025577037717994244,
+    "winner": "hyper-CREST",
+    "recall": {
+      "n_gfn2": 1,
+      "n_nnxtb": 2,
+      "recall_gfn2_into_nnxtb": {
+        "0.25": 1.0,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "recall_nnxtb_into_gfn2": {
+        "0.25": 0.5,
+        "0.5": 1.0,
+        "1.0": 1.0
+      },
+      "novel_in_nnxtb": {
+        "0.25": 1,
+        "0.5": 0,
+        "1.0": 0
+      },
+      "novel_in_gfn2": {
+        "0.25": 0,
+        "0.5": 0,
+        "1.0": 0
+      }
+    }
+  }
+]
\ No newline at end of file
diff --git a/benchmark/nnxtb/phase4_drug_set.json b/benchmark/nnxtb/phase4_drug_set.json
new file mode 100644
index 00000000..aa9fef4f
--- /dev/null
+++ b/benchmark/nnxtb/phase4_drug_set.json
@@ -0,0 +1,22 @@
+[
+  {"name": "salicylic_acid",   "smiles": "OC(=O)c1ccccc1O",                                        "chrg": 0},
+  {"name": "naproxen",         "smiles": "COc1ccc2cc(C(C)C(=O)O)ccc2c1",                          "chrg": 0},
+  {"name": "mefenamic_acid",   "smiles": "Cc1cccc(C)c1Nc1ccccc1C(=O)O",                           "chrg": 0},
+  {"name": "isoniazid",        "smiles": "NNC(=O)c1ccncc1",                                        "chrg": 0},
+  {"name": "metformin",        "smiles": "CN(C)C(=N)NC(N)=N",                                      "chrg": 0},
+  {"name": "propofol",         "smiles": "CC(C)c1cccc(C(C)C)c1O",                                  "chrg": 0},
+  {"name": "theophylline",     "smiles": "Cn1c(=O)c2[nH]cnc2n(C)c1=O",                             "chrg": 0},
+  {"name": "diflunisal",       "smiles": "OC(=O)c1cc(-c2ccc(F)cc2F)ccc1O",                         "chrg": 0},
+  {"name": "vanillin",         "smiles": "COc1cc(C=O)ccc1O",                                       "chrg": 0},
+  {"name": "amphetamine",      "smiles": "CC(N)Cc1ccccc1",                                         "chrg": 0},
+  {"name": "guaiacol",         "smiles": "COc1ccccc1O",                                            "chrg": 0},
+  {"name": "tryptamine",       "smiles": "NCCc1c[nH]c2ccccc12",                                    "chrg": 0},
+  {"name": "tyramine",         "smiles": "NCCc1ccc(O)cc1",                                         "chrg": 0},
+  {"name": "dopamine",         "smiles": "NCCc1ccc(O)c(O)c1",                                      "chrg": 0},
+  {"name": "histamine",        "smiles": "NCCc1c[nH]cn1",                                          "chrg": 0},
+  {"name": "phenylalanine",    "smiles": "N[C@@H](Cc1ccccc1)C(=O)O",                               "chrg": 0},
+  {"name": "serotonin",        "smiles": "NCCc1c[nH]c2ccc(O)cc12",                                 "chrg": 0},
+  {"name": "coumarin",         "smiles": "O=c1ccc2ccccc2o1",                                       "chrg": 0},
+  {"name": "paracetamol_N",    "smiles": "O=C(N)c1ccc(O)cc1",                                      "chrg": 0},
+  {"name": "anisole",          "smiles": "COc1ccccc1",                                             "chrg": 0}
+]
diff --git a/benchmark/nnxtb/run_benchmark.py b/benchmark/nnxtb/run_benchmark.py
new file mode 100644
index 00000000..69efadd3
--- /dev/null
+++ b/benchmark/nnxtb/run_benchmark.py
@@ -0,0 +1,342 @@
+#!/usr/bin/env python3
+"""
+hyper-crest vs standard CREST: DFT-rescored conformer quality benchmark.
+
+Per the hyper-crest mission (Loong's deliverable):
+    (a) run standard CREST on each test molecule, keep top-N conformers
+    (b) run hyper-CREST (method=nnxtb) on the same molecule, keep top-N
+    (c) DFT single-point each conformer set at a consistent high level
+    (d) hyper-CREST wins if its DFT-minimum conformer is lower than
+        standard CREST's DFT-minimum conformer
+
+Outputs a table with per-molecule ΔE_DFT(min) between the two methods
+and an overall winner count.
+
+Usage:
+    python run_benchmark.py                       \
+        --standard-crest  /path/to/stock/crest    \
+        --hyper-crest     /path/to/hyper-crest    \
+        --mace-model      /path/to/model.mxtb     \
+        --mol-dir         molecules/              \
+        --output-dir      results/                \
+        --top-n           5                       \
+        --dft             b3lyp/def2-svp          \
+        --threads         8
+
+Designed to run on a GPU node with hyper-crest + pyscf installed. Not
+intended to run on a laptop (CREST sampling + DFT rescore is O(hours)
+per molecule).
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import shutil
+import subprocess
+import sys
+import time
+from dataclasses import asdict, dataclass
+from pathlib import Path
+
+HARTREE_TO_KCAL = 627.50947428
+
+
+@dataclass
+class ConformerResult:
+    method: str           # "standard" | "hyper"
+    conformer_idx: int
+    crest_energy: float   # kcal/mol relative to method's own best, as CREST reports
+    dft_energy: float     # Hartree, absolute
+
+
+@dataclass
+class MoleculeResult:
+    name: str
+    n_atoms: int
+    standard: list[ConformerResult]
+    hyper: list[ConformerResult]
+    # Winner = method whose DFT-min conformer has the lower energy
+    dft_min_standard: float | None
+    dft_min_hyper: float | None
+    winner: str  # "hyper" | "standard" | "tie"
+    delta_hartree: float  # dft_min_hyper - dft_min_standard (negative → hyper wins)
+
+
+# --------------------------------------------------------------------------- #
+# CREST driving
+# --------------------------------------------------------------------------- #
+def run_crest_standard(
+    crest_bin: str,
+    xyz_path: Path,
+    workdir: Path,
+    extra_args: list[str],
+    timeout_s: int,
+) -> list[Path]:
+    """Run standard CREST (GFN2-xTB via tblite). Returns list of conformer xyz blocks."""
+    workdir.mkdir(parents=True, exist_ok=True)
+    shutil.copy2(xyz_path, workdir / "input.xyz")
+    cmd = [crest_bin, "input.xyz", "-gfn2", *extra_args]
+    with (workdir / "crest.log").open("w") as log:
+        subprocess.run(cmd, cwd=workdir, stdout=log, stderr=subprocess.STDOUT,
+                       timeout=timeout_s, check=True)
+    return parse_crest_conformers(workdir / "crest_conformers.xyz")
+
+
+def run_crest_hyper(
+    crest_bin: str,
+    xyz_path: Path,
+    workdir: Path,
+    mace_model: Path,
+    extra_args: list[str],
+    timeout_s: int,
+) -> list[Path]:
+    """Run hyper-CREST with method=nnxtb via a generated TOML input."""
+    workdir.mkdir(parents=True, exist_ok=True)
+    shutil.copy2(xyz_path, workdir / "input.xyz")
+    toml = (
+        'input = "input.xyz"\n'
+        'runtype = "imtd-gc"\n\n'
+        '[[calculation.level]]\n'
+        'method     = "nnxtb"\n'
+        f'mace_model = "{mace_model}"\n'
+        'chrg       = 0\n'
+    )
+    (workdir / "input.toml").write_text(toml)
+    cmd = [crest_bin, "input.toml", *extra_args]
+    with (workdir / "crest.log").open("w") as log:
+        subprocess.run(cmd, cwd=workdir, stdout=log, stderr=subprocess.STDOUT,
+                       timeout=timeout_s, check=True)
+    return parse_crest_conformers(workdir / "crest_conformers.xyz")
+
+
+def parse_crest_conformers(path: Path) -> list[tuple[float, list[str]]]:
+    """
+    Parse a CREST ensemble xyz. Returns a list of (energy_kcal, xyz_block_lines).
+    CREST writes the conformer energy on the comment line (in Hartree).
+    """
+    out: list[tuple[float, list[str]]] = []
+    if not path.exists():
+        return out
+    with path.open() as f:
+        lines = f.read().splitlines()
+
+    i = 0
+    while i < len(lines):
+        if not lines[i].strip():
+            i += 1
+            continue
+        try:
+            n = int(lines[i].strip())
+        except ValueError:
+            i += 1
+            continue
+        comment = lines[i + 1].strip() if i + 1 < len(lines) else ""
+        atoms = lines[i + 2 : i + 2 + n]
+        # Conformer energy: first token on the comment line that parses as float.
+        energy = None
+        for tok in comment.split():
+            try:
+                energy = float(tok)
+                break
+            except ValueError:
+                continue
+        out.append((energy if energy is not None else 0.0,
+                    [lines[i], lines[i + 1], *atoms]))
+        i += 2 + n
+    return out
+
+
+def write_xyz_block(block_lines: list[str], path: Path) -> None:
+    path.write_text("\n".join(block_lines) + "\n")
+
+
+# --------------------------------------------------------------------------- #
+# DFT rescoring (pyscf)
+# --------------------------------------------------------------------------- #
+def dft_singlepoint(xyz_path: Path, dft_spec: str, n_threads: int) -> float:
+    """
+    Single-point DFT on one xyz file. Returns total energy in Hartree.
+    dft_spec format: "<xc>/<basis>", e.g. "b3lyp/def2-svp", "wb97x-d/def2-tzvp".
+    """
+    xc, basis = dft_spec.split("/")
+    # Lazy import so the CREST driving path doesn't need pyscf.
+    from pyscf import dft, gto           # type: ignore
+    from pyscf.lib import num_threads    # type: ignore
+
+    num_threads(n_threads)
+    mol = gto.M(atom=str(xyz_path), basis=basis, unit="Angstrom")
+    mf = dft.RKS(mol)
+    mf.xc = xc
+    mf.verbose = 0
+    return float(mf.kernel())
+
+
+def dft_rescore_ensemble(
+    conformers: list[tuple[float, list[str]]],
+    method_tag: str,
+    mol_workdir: Path,
+    top_n: int,
+    dft_spec: str,
+    n_threads: int,
+) -> list[ConformerResult]:
+    """DFT-rescore the top-N conformers (by CREST energy). Returns per-conformer record."""
+    # Sort by CREST energy, take top-N.
+    selected = sorted(conformers, key=lambda t: t[0])[:top_n]
+    results: list[ConformerResult] = []
+    for idx, (e_crest, block) in enumerate(selected):
+        conf_xyz = mol_workdir / f"{method_tag}_conf{idx:02d}.xyz"
+        write_xyz_block(block, conf_xyz)
+        t0 = time.time()
+        e_dft = dft_singlepoint(conf_xyz, dft_spec, n_threads)
+        dt = time.time() - t0
+        results.append(ConformerResult(
+            method=method_tag, conformer_idx=idx,
+            crest_energy=e_crest, dft_energy=e_dft))
+        print(f"    [{method_tag}] conf{idx:02d}: E_DFT = {e_dft:+.6f} Ha  [{dt:.1f}s]",
+              flush=True)
+    return results
+
+
+# --------------------------------------------------------------------------- #
+# Orchestration
+# --------------------------------------------------------------------------- #
+def run_one_molecule(
+    xyz_path: Path,
+    args: argparse.Namespace,
+) -> MoleculeResult:
+    name = xyz_path.stem
+    print(f"\n=== {name} ===", flush=True)
+
+    n_atoms = int(xyz_path.read_text().splitlines()[0].strip())
+    mol_work = Path(args.output_dir) / name
+    mol_work.mkdir(parents=True, exist_ok=True)
+
+    # 1. Standard CREST
+    std_dir = mol_work / "standard"
+    t0 = time.time()
+    std_confs = run_crest_standard(
+        args.standard_crest, xyz_path, std_dir,
+        extra_args=args.crest_extra_args.split() if args.crest_extra_args else [],
+        timeout_s=args.crest_timeout)
+    print(f"  standard CREST: {len(std_confs)} conformers [{time.time()-t0:.0f}s]",
+          flush=True)
+
+    # 2. hyper-CREST
+    hyp_dir = mol_work / "hyper"
+    t0 = time.time()
+    hyp_confs = run_crest_hyper(
+        args.hyper_crest, xyz_path, hyp_dir, Path(args.mace_model),
+        extra_args=args.crest_extra_args.split() if args.crest_extra_args else [],
+        timeout_s=args.crest_timeout)
+    print(f"  hyper-CREST:    {len(hyp_confs)} conformers [{time.time()-t0:.0f}s]",
+          flush=True)
+
+    # 3. DFT rescore
+    std_results = dft_rescore_ensemble(
+        std_confs, "standard", mol_work, args.top_n, args.dft, args.threads)
+    hyp_results = dft_rescore_ensemble(
+        hyp_confs, "hyper", mol_work, args.top_n, args.dft, args.threads)
+
+    e_std_min = min((r.dft_energy for r in std_results), default=None)
+    e_hyp_min = min((r.dft_energy for r in hyp_results), default=None)
+
+    if e_std_min is None or e_hyp_min is None:
+        winner = "tie"
+        delta = 0.0
+    else:
+        delta = e_hyp_min - e_std_min
+        # tie tolerance: 0.1 kcal/mol (numerical + conformer-equivalent)
+        if abs(delta) < 0.1 / HARTREE_TO_KCAL:
+            winner = "tie"
+        elif delta < 0:
+            winner = "hyper"
+        else:
+            winner = "standard"
+
+    res = MoleculeResult(
+        name=name, n_atoms=n_atoms,
+        standard=std_results, hyper=hyp_results,
+        dft_min_standard=e_std_min, dft_min_hyper=e_hyp_min,
+        winner=winner, delta_hartree=delta)
+    (mol_work / "result.json").write_text(
+        json.dumps(asdict(res), indent=2, default=str))
+    print(f"  {name}: winner={winner}  ΔE = {delta*HARTREE_TO_KCAL:+.3f} kcal/mol",
+          flush=True)
+    return res
+
+
+def print_summary(results: list[MoleculeResult]) -> None:
+    print("\n" + "=" * 72)
+    print("  Summary  (ΔE = DFT-min hyper − DFT-min standard, in kcal/mol)")
+    print("=" * 72)
+    print(f"  {'name':<20s}  {'N':>4s}  {'std min (Ha)':>13s}  "
+          f"{'hyp min (Ha)':>13s}  {'ΔE (kcal)':>10s}  {'winner':<10s}")
+    for r in results:
+        dkcal = r.delta_hartree * HARTREE_TO_KCAL
+        std = f"{r.dft_min_standard:+.5f}" if r.dft_min_standard is not None else "   N/A   "
+        hyp = f"{r.dft_min_hyper:+.5f}" if r.dft_min_hyper is not None else "   N/A   "
+        print(f"  {r.name:<20s}  {r.n_atoms:>4d}  {std:>13s}  {hyp:>13s}  "
+              f"{dkcal:>+10.3f}  {r.winner:<10s}")
+    wins = sum(r.winner == "hyper" for r in results)
+    losses = sum(r.winner == "standard" for r in results)
+    ties = sum(r.winner == "tie" for r in results)
+    print("-" * 72)
+    print(f"  hyper-CREST wins : {wins} / {len(results)}")
+    print(f"  standard wins    : {losses} / {len(results)}")
+    print(f"  ties             : {ties} / {len(results)}")
+    print("=" * 72)
+
+
+def main() -> int:
+    p = argparse.ArgumentParser(description=__doc__,
+                                formatter_class=argparse.RawDescriptionHelpFormatter)
+    p.add_argument("--standard-crest", required=True,
+                   help="path to the stock CREST binary")
+    p.add_argument("--hyper-crest", required=True,
+                   help="path to the hyper-crest binary (this fork, built with WITH_HYPERXTB=ON)")
+    p.add_argument("--mace-model", required=True,
+                   help="path to the ScaleShiftMACExTB .mxtb model")
+    p.add_argument("--mol-dir", required=True,
+                   help="directory of xyz files (one molecule per file)")
+    p.add_argument("--output-dir", default="results",
+                   help="workdir for per-molecule runs + result.json files")
+    p.add_argument("--top-n", type=int, default=5,
+                   help="number of top-CREST-ranked conformers to DFT-rescore per method")
+    p.add_argument("--dft", default="b3lyp/def2-svp",
+                   help="DFT spec 'xc/basis', e.g. 'wb97x-d/def2-tzvp'")
+    p.add_argument("--threads", type=int, default=os.cpu_count() or 4,
+                   help="threads for pyscf DFT")
+    p.add_argument("--crest-extra-args", default="",
+                   help="string of extra args forwarded to both crest calls")
+    p.add_argument("--crest-timeout", type=int, default=7200,
+                   help="per-molecule CREST wall-time cap in seconds")
+    args = p.parse_args()
+
+    mol_dir = Path(args.mol_dir)
+    xyz_files = sorted(mol_dir.glob("*.xyz"))
+    if not xyz_files:
+        print(f"No xyz files in {mol_dir}", file=sys.stderr)
+        return 1
+
+    Path(args.output_dir).mkdir(parents=True, exist_ok=True)
+    results: list[MoleculeResult] = []
+    for xyz in xyz_files:
+        try:
+            results.append(run_one_molecule(xyz, args))
+        except subprocess.CalledProcessError as e:
+            print(f"  {xyz.name} FAILED during CREST: {e}", flush=True)
+        except subprocess.TimeoutExpired:
+            print(f"  {xyz.name} TIMED OUT after {args.crest_timeout}s", flush=True)
+        except Exception as e:
+            print(f"  {xyz.name} FAILED: {type(e).__name__}: {e}", flush=True)
+
+    print_summary(results)
+    (Path(args.output_dir) / "summary.json").write_text(
+        json.dumps([asdict(r) for r in results], indent=2, default=str))
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/benchmark/nnxtb/run_mconf_benchmark.py b/benchmark/nnxtb/run_mconf_benchmark.py
new file mode 100644
index 00000000..2a2462a3
--- /dev/null
+++ b/benchmark/nnxtb/run_mconf_benchmark.py
@@ -0,0 +1,256 @@
+#!/usr/bin/env python3
+"""
+MCONF benchmark — GMTKN55 melatonin conformer set.
+
+Runs stock GFN2 and hyper NN-xTB single-points on all 52 MCONF conformer
+geometries, compares to the published ωB97X-V reference energies (as
+relative energies in kcal/mol with respect to conformer 1), and reports
+Spearman correlation + MAE + wall-clock time per engine.
+
+Unlike run_benchmark.py this does NOT run the full CREST sampling loop —
+it takes the fixed published geometries and just evaluates each engine's
+energy on them, then compares the predicted *ordering* and *spread* to
+the reference. This directly answers "is NN-xTB more DFT-like than GFN2
+on the same geometries" without paying the cost of a full conformer
+search.
+
+Usage:
+    python run_mconf_benchmark.py                     \
+        --mconf-dir /path/to/mconf_xyzs/              \
+        --ref-json  /path/to/ref_energies.json        \
+        --stock-crest /path/to/stock/crest            \
+        --hyper-crest /path/to/hyper/crest            \
+        --mace-model  /path/to/model.mxtb             \
+        --output-json results.json
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import shutil
+import subprocess
+import sys
+import time
+from dataclasses import asdict, dataclass
+from pathlib import Path
+
+HARTREE_TO_KCAL = 627.50947428
+
+
+@dataclass
+class SPResult:
+    conformer: int
+    energy_hartree: float
+    wall_seconds: float
+
+
+def singlepoint_crest(
+    crest_bin: str,
+    xyz: Path,
+    workdir: Path,
+    method: str,              # "gfn2" or "nnxtb"
+    mace_model: Path | None,
+    timeout_s: int,
+) -> tuple[float, float]:
+    """Run `crest input.toml` with runtype=singlepoint. Returns (E_Ha, wall_s)."""
+    workdir.mkdir(parents=True, exist_ok=True)
+    shutil.copy2(xyz, workdir / "input.xyz")
+    mace_line = f'mace_model = "{mace_model}"\n' if method == "nnxtb" else ""
+    toml = (
+        'input = "input.xyz"\n'
+        'runtype = "singlepoint"\n\n'
+        '[[calculation.level]]\n'
+        f'method = "{method}"\n'
+        f'{mace_line}'
+        'chrg   = 0\n'
+    )
+    (workdir / "input.toml").write_text(toml)
+    t0 = time.time()
+    proc = subprocess.run(
+        [crest_bin, "input.toml"],
+        cwd=workdir,
+        capture_output=True, text=True,
+        timeout=timeout_s,
+    )
+    wall = time.time() - t0
+    if proc.returncode != 0:
+        raise RuntimeError(f"crest failed (rc={proc.returncode}):\n{proc.stderr[-500:]}")
+    # Scan output for TOTAL ENERGY line.
+    m = re.search(r"TOTAL ENERGY\s+([-+]?\d+\.\d+)\s+Eh", proc.stdout)
+    if not m:
+        raise RuntimeError("no TOTAL ENERGY line in crest output")
+    return float(m.group(1)), wall
+
+
+# --------------------------------------------------------------------------- #
+# Correlation helpers (no scipy dependency — small self-contained Spearman)
+# --------------------------------------------------------------------------- #
+def _rank(xs: list[float]) -> list[float]:
+    n = len(xs)
+    idx = sorted(range(n), key=lambda i: xs[i])
+    ranks = [0.0] * n
+    i = 0
+    while i < n:
+        j = i
+        while j + 1 < n and xs[idx[j + 1]] == xs[idx[i]]:
+            j += 1
+        avg = (i + j) / 2.0 + 1
+        for k in range(i, j + 1):
+            ranks[idx[k]] = avg
+        i = j + 1
+    return ranks
+
+
+def spearman(xs: list[float], ys: list[float]) -> float:
+    rx, ry = _rank(xs), _rank(ys)
+    n = len(xs)
+    mx, my = sum(rx) / n, sum(ry) / n
+    num = sum((rx[i] - mx) * (ry[i] - my) for i in range(n))
+    dx = (sum((v - mx) ** 2 for v in rx)) ** 0.5
+    dy = (sum((v - my) ** 2 for v in ry)) ** 0.5
+    return num / (dx * dy) if dx and dy else float("nan")
+
+
+def mae(xs: list[float], ys: list[float]) -> float:
+    return sum(abs(a - b) for a, b in zip(xs, ys)) / len(xs)
+
+
+# --------------------------------------------------------------------------- #
+# Orchestration
+# --------------------------------------------------------------------------- #
+def run_for_method(
+    method: str,
+    crest_bin: str,
+    mace_model: Path | None,
+    xyz_files: list[Path],
+    workroot: Path,
+    timeout_s: int,
+) -> tuple[list[SPResult], float]:
+    """Run single-point for every conformer. Returns (per-conformer results, total wall)."""
+    results: list[SPResult] = []
+    t_total = time.time()
+    for xyz in xyz_files:
+        conf = int(re.search(r"(\d+)", xyz.stem).group(1))
+        wd = workroot / method / f"conf_{conf:02d}"
+        try:
+            E, w = singlepoint_crest(crest_bin, xyz, wd, method, mace_model, timeout_s)
+            results.append(SPResult(conformer=conf, energy_hartree=E, wall_seconds=w))
+            print(f"  [{method}] conf {conf:2d}: E = {E:+.6f} Ha  ({w:.2f} s)", flush=True)
+        except Exception as e:
+            print(f"  [{method}] conf {conf:2d}: FAILED — {e}", flush=True)
+    total = time.time() - t_total
+    return results, total
+
+
+def compute_metrics(
+    method_results: list[SPResult],
+    ref: dict[int, float],
+) -> dict:
+    """Convert absolute energies → kcal/mol relative to conformer 1, then compare to ref."""
+    by_conf = {r.conformer: r.energy_hartree for r in method_results}
+    if 1 not in by_conf:
+        return {"error": "conformer 1 missing", "rho": None, "mae_kcal": None}
+    e1 = by_conf[1]
+
+    pred_rel, ref_rel, confs = [], [], []
+    for c in sorted(by_conf):
+        if c in ref:
+            pred_rel.append((by_conf[c] - e1) * HARTREE_TO_KCAL)
+            ref_rel.append(ref[c])
+            confs.append(c)
+    return {
+        "n": len(confs),
+        "spearman": spearman(pred_rel, ref_rel),
+        "mae_kcal": mae(pred_rel, ref_rel),
+        "rms_kcal": (sum((a - b) ** 2 for a, b in zip(pred_rel, ref_rel)) / len(pred_rel)) ** 0.5,
+        "pred_rel_kcal": dict(zip(confs, pred_rel)),
+        "ref_rel_kcal":  dict(zip(confs, ref_rel)),
+    }
+
+
+def main() -> int:
+    p = argparse.ArgumentParser()
+    p.add_argument("--mconf-dir", required=True)
+    p.add_argument("--ref-json",  required=True)
+    p.add_argument("--stock-crest", required=True)
+    p.add_argument("--hyper-crest", required=True)
+    p.add_argument("--mace-model",  required=True)
+    p.add_argument("--output-json", default="mconf_results.json")
+    p.add_argument("--workdir",     default="mconf_workdir")
+    p.add_argument("--timeout",     type=int, default=600)
+    args = p.parse_args()
+
+    mconf_dir = Path(args.mconf_dir)
+    xyz_files = sorted(mconf_dir.glob("conf_*.xyz"),
+                       key=lambda p: int(re.search(r"(\d+)", p.stem).group(1)))
+    if not xyz_files:
+        print(f"no conf_*.xyz under {mconf_dir}", file=sys.stderr)
+        return 1
+    ref = {int(k): float(v) for k, v in json.load(open(args.ref_json)).items()}
+
+    workroot = Path(args.workdir)
+    workroot.mkdir(parents=True, exist_ok=True)
+
+    print(f"MCONF benchmark: {len(xyz_files)} conformers, ref level = GMTKN55 published")
+    print(f"  stock crest: {args.stock_crest}")
+    print(f"  hyper crest: {args.hyper_crest}")
+    print(f"  MACE model:  {args.mace_model}")
+    print()
+
+    print("=== stock CREST (method=gfn2 via tblite) ===", flush=True)
+    gfn2_results, gfn2_wall = run_for_method(
+        "gfn2", args.stock_crest, None, xyz_files, workroot, args.timeout)
+
+    print("\n=== hyper-CREST (method=nnxtb, CUDA backend) ===", flush=True)
+    nnxtb_results, nnxtb_wall = run_for_method(
+        "nnxtb", args.hyper_crest, Path(args.mace_model),
+        xyz_files, workroot, args.timeout)
+
+    m_gfn2 = compute_metrics(gfn2_results, ref)
+    m_nn   = compute_metrics(nnxtb_results, ref)
+
+    n_gfn2 = len(gfn2_results) or 1
+    n_nn   = len(nnxtb_results) or 1
+
+    print("\n" + "=" * 72)
+    print(f"  MCONF results (n = {m_gfn2.get('n', 0)} conformers, reference = ωB97X-V / def2-QZVP)")
+    print("=" * 72)
+    print(f"  {'Method':<22s}  {'rho':>7s}  {'MAE(kcal)':>10s}  {'RMS(kcal)':>10s}  {'wall(s)':>9s}  {'ms/SP':>7s}")
+    for label, m, wall, n in [
+        ("stock CREST (GFN2)",  m_gfn2,  gfn2_wall,  n_gfn2),
+        ("hyper-CREST (NN-xTB)", m_nn,   nnxtb_wall, n_nn),
+    ]:
+        rho = m.get("spearman")
+        mk  = m.get("mae_kcal")
+        rk  = m.get("rms_kcal")
+        print(f"  {label:<22s}  "
+              f"{('—' if rho is None else f'{rho:+.3f}'):>7s}  "
+              f"{('—' if mk  is None else f'{mk:.3f}'):>10s}  "
+              f"{('—' if rk  is None else f'{rk:.3f}'):>10s}  "
+              f"{wall:>8.1f}   "
+              f"{1000*wall/n:>6.1f}")
+    print("=" * 72)
+
+    out = {
+        "ref_level": "GMTKN55 MCONF published reference (relative kcal/mol)",
+        "n_conformers": len(xyz_files),
+        "gfn2": {
+            "per_conformer": [asdict(r) for r in gfn2_results],
+            "metrics": m_gfn2,
+            "total_wall_s": gfn2_wall,
+        },
+        "nnxtb": {
+            "per_conformer": [asdict(r) for r in nnxtb_results],
+            "metrics": m_nn,
+            "total_wall_s": nnxtb_wall,
+        },
+    }
+    Path(args.output_json).write_text(json.dumps(out, indent=2, default=str))
+    print(f"\nDetailed results: {args.output_json}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/benchmark/nnxtb/run_smiles_benchmark.py b/benchmark/nnxtb/run_smiles_benchmark.py
new file mode 100644
index 00000000..ca999941
--- /dev/null
+++ b/benchmark/nnxtb/run_smiles_benchmark.py
@@ -0,0 +1,419 @@
+#!/usr/bin/env python3
+"""
+Phase-3 benchmark — the one Loong actually wants.
+
+From SMILES only (no reference geometries):
+  1. RDKit.Embed → random 3D starting structure per SMILES
+  2. stock CREST  (method=gfn2)  → full iMTD-GC → lowest-energy conformer
+  3. hyper-CREST  (method=nnxtb) → full iMTD-GC → lowest-energy conformer
+  4. DFT single-point on BOTH lowest conformers (they may be different structures)
+  5. tally: for how many SMILES does hyper-CREST's lowest conformer have a
+     lower DFT energy than stock CREST's lowest? that's the win rate.
+
+Unlike the MCONF single-point benchmark, this measures the *sampling* quality
+of each engine — can hyper-CREST find geometries that stock GFN2 missed?
+
+Usage:
+  python run_smiles_benchmark.py                        \
+      --molecules drug_set.json                         \
+      --stock-crest /path/to/crest                      \
+      --hyper-crest /path/to/crest                      \
+      --mace-model  /path/to/model.mxtb                 \
+      --output-json phase3_results.json                 \
+      --workdir phase3_workdir                          \
+      --dft b3lyp/def2-svp                              \
+      --crest-extra "--quick"
+
+drug_set.json format: [{"name": "aspirin", "smiles": "CC(=O)Oc1ccccc1C(=O)O", "chrg": 0}, ...]
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import re
+import shutil
+import subprocess
+import sys
+import time
+from dataclasses import asdict, dataclass, field
+from pathlib import Path
+
+HARTREE_TO_KCAL = 627.50947428
+
+
+@dataclass
+class EngineRun:
+    engine: str                 # "gfn2" | "nnxtb"
+    wall_seconds: float = 0.0
+    n_conformers: int = 0
+    lowest_crest_energy: float | None = None     # Ha, CREST's own energy
+    lowest_dft_energy:   float | None = None     # Ha, DFT rescore of the lowest-CREST geometry
+    crest_log_path: str | None = None
+    ensemble_path:  str | None = None     # copy of crest_conformers.xyz
+
+
+@dataclass
+class RecallMetrics:
+    """Ensemble-overlap metrics between two CREST runs, at multiple RMSD thresholds."""
+    n_gfn2:  int = 0
+    n_nnxtb: int = 0
+    # For each RMSD threshold, fraction of the OTHER engine's conformers matched.
+    recall_gfn2_into_nnxtb: dict = field(default_factory=dict)  # thr -> fraction of gfn2 confs matched by some nnxtb conf
+    recall_nnxtb_into_gfn2: dict = field(default_factory=dict)
+    # Novel = engine's conformers that the other engine didn't find (at a given thr)
+    novel_in_nnxtb: dict = field(default_factory=dict)
+    novel_in_gfn2:  dict = field(default_factory=dict)
+
+
+@dataclass
+class MoleculeResult:
+    name: str
+    smiles: str
+    n_atoms: int = 0
+    gfn2:  EngineRun = field(default_factory=lambda: EngineRun(engine="gfn2"))
+    nnxtb: EngineRun = field(default_factory=lambda: EngineRun(engine="nnxtb"))
+    dft_delta_hartree: float | None = None       # nnxtb − gfn2
+    winner: str = "tie"
+    recall: RecallMetrics = field(default_factory=RecallMetrics)
+
+
+# --------------------------------------------------------------------------- #
+# SMILES → 3D starting geometry
+# --------------------------------------------------------------------------- #
+def smiles_to_xyz(smiles: str, out_xyz: Path, seed: int = 42) -> int:
+    from rdkit import Chem
+    from rdkit.Chem import AllChem
+    mol = Chem.MolFromSmiles(smiles)
+    if mol is None:
+        raise RuntimeError(f"invalid SMILES: {smiles}")
+    mol = Chem.AddHs(mol)
+    AllChem.EmbedMolecule(mol, randomSeed=seed)
+    AllChem.MMFFOptimizeMolecule(mol, maxIters=2000)
+    conf = mol.GetConformer()
+    lines = [str(mol.GetNumAtoms()), smiles]
+    for atom in mol.GetAtoms():
+        p = conf.GetAtomPosition(atom.GetIdx())
+        lines.append(f"{atom.GetSymbol()} {p.x:.6f} {p.y:.6f} {p.z:.6f}")
+    out_xyz.write_text("\n".join(lines) + "\n")
+    return mol.GetNumAtoms()
+
+
+# --------------------------------------------------------------------------- #
+# Run CREST sampling + parse output
+# --------------------------------------------------------------------------- #
+def run_crest(
+    crest_bin: str,
+    xyz_in: Path,
+    workdir: Path,
+    method: str,
+    mace_model: Path | None,
+    chrg: int,
+    extra_args: list[str],
+    timeout_s: int,
+) -> tuple[float, float | None, int]:
+    """Returns (wall_seconds, lowest_crest_energy_Ha, n_conformers)."""
+    workdir.mkdir(parents=True, exist_ok=True)
+    shutil.copy2(xyz_in, workdir / "input.xyz")
+    mace_line = f'mace_model = "{mace_model}"\n' if method == "nnxtb" else ""
+    toml = (
+        'input = "input.xyz"\n'
+        'runtype = "imtd-gc"\n'
+        f'threads = 8\n\n'
+        '[[calculation.level]]\n'
+        f'method = "{method}"\n'
+        f'{mace_line}'
+        f'chrg   = {chrg}\n'
+    )
+    (workdir / "input.toml").write_text(toml)
+    t0 = time.time()
+    with (workdir / "crest.log").open("w") as log:
+        subprocess.run(
+            [crest_bin, "input.toml"] + extra_args,
+            cwd=workdir,
+            stdout=log, stderr=subprocess.STDOUT,
+            timeout=timeout_s,
+            check=True,
+        )
+    wall = time.time() - t0
+
+    # Parse crest_conformers.xyz: first conformer is the lowest by CREST energy.
+    confs = workdir / "crest_conformers.xyz"
+    if not confs.exists():
+        return wall, None, 0
+    lines = confs.read_text().splitlines()
+    if len(lines) < 3:
+        return wall, None, 0
+    n_atoms = int(lines[0].strip())
+    n_conf = len(lines) // (n_atoms + 2)
+    # comment line of first block contains the CREST energy in Hartree
+    comment = lines[1].strip()
+    e_low = None
+    for tok in comment.split():
+        try:
+            e_low = float(tok)
+            break
+        except ValueError:
+            pass
+    return wall, e_low, n_conf
+
+
+def pick_best_xyz(workdir: Path, out: Path) -> None:
+    """crest_best.xyz contains the lowest-energy conformer; copy it out."""
+    best = workdir / "crest_best.xyz"
+    if not best.exists():
+        confs = workdir / "crest_conformers.xyz"
+        if not confs.exists():
+            raise RuntimeError("no crest_best or crest_conformers xyz")
+        lines = confs.read_text().splitlines()
+        n_atoms = int(lines[0].strip())
+        out.write_text("\n".join(lines[:n_atoms + 2]) + "\n")
+    else:
+        shutil.copy2(best, out)
+
+
+# --------------------------------------------------------------------------- #
+# Ensemble overlap (RMSD-based recall)
+# --------------------------------------------------------------------------- #
+def _kabsch_rmsd(a: "np.ndarray", b: "np.ndarray") -> float:
+    """Heavy-atom RMSD between two (N, 3) coord sets after optimal rotation + translation."""
+    import numpy as np
+    a = a - a.mean(0)
+    b = b - b.mean(0)
+    h = a.T @ b
+    u, _, vt = np.linalg.svd(h)
+    d = np.sign(np.linalg.det(vt.T @ u.T))
+    r = vt.T @ np.diag([1.0, 1.0, d]) @ u.T
+    a_rot = a @ r.T
+    return float(np.sqrt(((a_rot - b) ** 2).sum() / a.shape[0]))
+
+
+def _parse_ensemble(xyz_path: Path) -> list["np.ndarray"]:
+    """Parse a multi-frame xyz file. Returns list of heavy-atom coord arrays (N_heavy, 3)."""
+    import numpy as np
+    if not xyz_path or not xyz_path.exists():
+        return []
+    lines = xyz_path.read_text().splitlines()
+    if not lines:
+        return []
+    n_atoms = int(lines[0].strip())
+    frames = []
+    stride = n_atoms + 2
+    for i in range(0, len(lines), stride):
+        if i + stride > len(lines):
+            break
+        coords = []
+        for L in lines[i + 2 : i + stride]:
+            t = L.split()
+            if len(t) < 4 or t[0].lower() == "h":
+                continue
+            coords.append([float(t[1]), float(t[2]), float(t[3])])
+        if coords:
+            frames.append(np.asarray(coords))
+    return frames
+
+
+def compute_recall(gfn2_xyz: Path | None, nnxtb_xyz: Path | None,
+                   thresholds=(0.25, 0.5, 1.0)) -> RecallMetrics:
+    """RMSD-match every conformer from each engine against the other engine's ensemble.
+    Symmetric: at each threshold, report fraction of X's conformers matched by some Y."""
+    rec = RecallMetrics()
+    g = _parse_ensemble(gfn2_xyz) if gfn2_xyz else []
+    n = _parse_ensemble(nnxtb_xyz) if nnxtb_xyz else []
+    rec.n_gfn2 = len(g)
+    rec.n_nnxtb = len(n)
+    if not g or not n:
+        return rec
+    # Sanity: both ensembles must share atom count
+    if g[0].shape != n[0].shape:
+        return rec
+    # Pairwise RMSD matrix. For typical n = 50 x 50, this is trivial.
+    import numpy as np
+    rmsd = np.zeros((len(g), len(n)))
+    for i, ga in enumerate(g):
+        for j, nb in enumerate(n):
+            rmsd[i, j] = _kabsch_rmsd(ga, nb)
+    for thr in thresholds:
+        matched_g = (rmsd.min(axis=1) < thr).sum()
+        matched_n = (rmsd.min(axis=0) < thr).sum()
+        # "recall of X into Y" = fraction of X that has at least one match in Y
+        rec.recall_gfn2_into_nnxtb[f"{thr}"] = float(matched_g / len(g))
+        rec.recall_nnxtb_into_gfn2[f"{thr}"] = float(matched_n / len(n))
+        rec.novel_in_gfn2[f"{thr}"]  = int(len(g) - matched_g)
+        rec.novel_in_nnxtb[f"{thr}"] = int(len(n) - matched_n)
+    return rec
+
+
+# --------------------------------------------------------------------------- #
+# DFT rescore
+# --------------------------------------------------------------------------- #
+def dft_singlepoint(xyz: Path, xc_basis: str, n_threads: int) -> float:
+    from pyscf import dft, gto
+    from pyscf.lib import num_threads
+    num_threads(n_threads)
+    xc, basis = xc_basis.split("/")
+    lines = xyz.read_text().splitlines()
+    n_atoms = int(lines[0].strip())
+    atom_block = "\n".join(lines[2:2 + n_atoms])
+    mol = gto.M(atom=atom_block, basis=basis, unit="Angstrom", verbose=0)
+    mf = dft.RKS(mol); mf.xc = xc; mf.verbose = 0
+    return float(mf.kernel())
+
+
+# --------------------------------------------------------------------------- #
+# Main per-molecule driver
+# --------------------------------------------------------------------------- #
+def run_one(mol_spec: dict, args: argparse.Namespace) -> MoleculeResult:
+    name   = mol_spec["name"]
+    smiles = mol_spec["smiles"]
+    chrg   = int(mol_spec.get("chrg", 0))
+
+    print(f"\n=== {name}  ({smiles})  chrg={chrg} ===", flush=True)
+    mol_dir = Path(args.workdir) / name
+    mol_dir.mkdir(parents=True, exist_ok=True)
+
+    seed_xyz = mol_dir / "seed.xyz"
+    n_atoms = smiles_to_xyz(smiles, seed_xyz, seed=args.seed)
+    print(f"  seed embed: {n_atoms} atoms → {seed_xyz.name}", flush=True)
+
+    extra = args.crest_extra.split() if args.crest_extra else []
+
+    result = MoleculeResult(name=name, smiles=smiles, n_atoms=n_atoms)
+    # run each engine in series (they're single-process CREST runs → each
+    # sees its own MACE load-once path).
+    for engine, crest_bin, record in [
+        ("gfn2",  args.stock_crest, result.gfn2),
+        ("nnxtb", args.hyper_crest, result.nnxtb),
+    ]:
+        rundir = mol_dir / engine
+        try:
+            wall, e_low, n_conf = run_crest(
+                crest_bin, seed_xyz, rundir, engine,
+                Path(args.mace_model) if engine == "nnxtb" else None,
+                chrg, extra, args.timeout,
+            )
+            record.wall_seconds = wall
+            record.lowest_crest_energy = e_low
+            record.n_conformers = n_conf
+            record.crest_log_path = str(rundir / "crest.log")
+            # Save a copy of the full ensemble for post-processing recall.
+            ens_src = rundir / "crest_conformers.xyz"
+            ens_dst = mol_dir / f"{engine}_ensemble.xyz"
+            if ens_src.exists():
+                shutil.copy2(ens_src, ens_dst)
+                record.ensemble_path = str(ens_dst)
+            print(f"  [{engine}] sampled {n_conf} conformers, "
+                  f"lowest CREST E = {e_low:+.6f} Ha, wall = {wall:.0f} s", flush=True)
+        except subprocess.CalledProcessError as e:
+            print(f"  [{engine}] CREST failed (rc={e.returncode})", flush=True)
+            continue
+        except subprocess.TimeoutExpired:
+            print(f"  [{engine}] CREST timed out after {args.timeout}s", flush=True)
+            continue
+
+        # DFT single-point on the lowest-energy geometry
+        best = mol_dir / f"{engine}_best.xyz"
+        pick_best_xyz(rundir, best)
+        t0 = time.time()
+        try:
+            e_dft = dft_singlepoint(best, args.dft, args.threads)
+            record.lowest_dft_energy = e_dft
+            print(f"  [{engine}] DFT({args.dft}) on lowest: E = {e_dft:+.6f} Ha "
+                  f"({time.time()-t0:.0f} s)", flush=True)
+        except Exception as e:
+            print(f"  [{engine}] DFT rescore failed: {e}", flush=True)
+
+    # Recall (ensemble overlap between the two engines by heavy-atom RMSD)
+    result.recall = compute_recall(
+        Path(result.gfn2.ensemble_path)  if result.gfn2.ensemble_path  else None,
+        Path(result.nnxtb.ensemble_path) if result.nnxtb.ensemble_path else None,
+    )
+    rg = result.recall
+    if rg.n_gfn2 and rg.n_nnxtb:
+        print(f"  recall (heavy-atom RMSD, thr=0.5Å): "
+              f"gfn2→nnxtb {rg.recall_gfn2_into_nnxtb.get('0.5', 0):.2f}  "
+              f"nnxtb→gfn2 {rg.recall_nnxtb_into_gfn2.get('0.5', 0):.2f}  "
+              f"novel_nnxtb={rg.novel_in_nnxtb.get('0.5', 0)}  "
+              f"novel_gfn2={rg.novel_in_gfn2.get('0.5', 0)}", flush=True)
+
+    # Decide winner
+    eg = result.gfn2.lowest_dft_energy
+    en = result.nnxtb.lowest_dft_energy
+    if eg is not None and en is not None:
+        d = (en - eg) * HARTREE_TO_KCAL
+        result.dft_delta_hartree = en - eg
+        if abs(d) < 0.05:
+            result.winner = "tie"
+        elif d < 0:
+            result.winner = "hyper-CREST"
+        else:
+            result.winner = "stock CREST"
+        print(f"  {name}: ΔE(nnxtb − gfn2) = {d:+.3f} kcal/mol → winner: {result.winner}",
+              flush=True)
+    return result
+
+
+def main() -> int:
+    p = argparse.ArgumentParser()
+    p.add_argument("--molecules",    required=True, help="json list of {name, smiles, chrg}")
+    p.add_argument("--stock-crest",  required=True)
+    p.add_argument("--hyper-crest",  required=True)
+    p.add_argument("--mace-model",   required=True)
+    p.add_argument("--output-json",  default="phase3_results.json")
+    p.add_argument("--workdir",      default="phase3_workdir")
+    p.add_argument("--dft",          default="b3lyp/def2-svp")
+    p.add_argument("--threads",      type=int, default=os.cpu_count() or 8)
+    p.add_argument("--seed",         type=int, default=42)
+    p.add_argument("--timeout",      type=int, default=7200)
+    p.add_argument("--crest-extra",  default="")
+    args = p.parse_args()
+
+    mols = json.load(open(args.molecules))
+    Path(args.workdir).mkdir(parents=True, exist_ok=True)
+
+    print(f"Phase-3: {len(mols)} molecules, stock={args.stock_crest.split('/')[-2]}, "
+          f"hyper-nnxtb={args.mace_model.split('/')[-1]}, DFT={args.dft}")
+
+    results: list[MoleculeResult] = []
+    for mol in mols:
+        try:
+            results.append(run_one(mol, args))
+        except Exception as e:
+            print(f"  {mol['name']} FAILED: {type(e).__name__}: {e}", flush=True)
+
+    print("\n" + "=" * 78)
+    print(f"  Phase-3 — from-SMILES sampling + DFT rescore ({args.dft})")
+    print("=" * 78)
+    print(f"  {'name':<20s}  {'natoms':>6s}  {'n_gfn2':>6s}  {'n_nn':>5s}  "
+          f"{'g→n@0.5':>8s}  {'n→g@0.5':>8s}  "
+          f"{'ΔDFT(kcal)':>10s}  {'winner':<12s}")
+    hw = sw = ti = failed = 0
+    for r in results:
+        if r.dft_delta_hartree is None:
+            dk = "   fail   "; failed += 1
+        else:
+            dk = f"{r.dft_delta_hartree * HARTREE_TO_KCAL:+10.3f}"
+            if r.winner == "hyper-CREST": hw += 1
+            elif r.winner == "stock CREST": sw += 1
+            else: ti += 1
+        rg_str = (f"{r.recall.recall_gfn2_into_nnxtb.get('0.5', 0):.2f}"
+                  if r.recall.n_gfn2 else "—")
+        ng_str = (f"{r.recall.recall_nnxtb_into_gfn2.get('0.5', 0):.2f}"
+                  if r.recall.n_nnxtb else "—")
+        print(f"  {r.name:<20s}  {r.n_atoms:>6d}  "
+              f"{r.recall.n_gfn2:>6d}  {r.recall.n_nnxtb:>5d}  "
+              f"{rg_str:>8s}  {ng_str:>8s}  "
+              f"{dk:>10s}  {r.winner:<12s}")
+    print("-" * 78)
+    print(f"  hyper-CREST wins: {hw}/{len(results)}  |  stock: {sw}  |  ties: {ti}  |  failed: {failed}")
+    print("=" * 78)
+
+    Path(args.output_json).write_text(
+        json.dumps([asdict(r) for r in results], indent=2, default=str))
+    print(f"\nDetailed results: {args.output_json}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())