diff --git a/marker_experiments/downstream/cluster/submit_extcaps.sh b/marker_experiments/downstream/cluster/submit_extcaps.sh
new file mode 100755
index 00000000..f0c12ce3
--- /dev/null
+++ b/marker_experiments/downstream/cluster/submit_extcaps.sh
@@ -0,0 +1,210 @@
+#!/usr/bin/env bash
+# The bnd_wpd_extcaps downstream sweep: one arm, seeds 0-2, matched to the BPE runs.
+#
+#     CORE_SAFE_ARMS="" marker_experiments/downstream/cluster/submit_extcaps.sh
+#
+# bnd_wpd_extcaps differs from bnd_wpd_caps only in which side of the boundary markers the
+# caps code sits on: `The` is `<^><|>the<|>` here against `<|><^>the<|>` there. The
+# lowercase span is therefore a suffix of the cased one and BPE may cover both with one
+# piece. Every other setting matches the committed BPE runs so the difference in the
+# downstream numbers is attributable to that one change.
+#
+# ONE job for all three seeds, not three. A node carries 4 GPUs and a node-hour bills as 4
+# GPU-hours whether or not a GPU is touched, so three single-GPU runs on three nodes bill
+# 12 GPU-hours an hour to use 3. They run concurrently on GPUs 0, 1 and 2 of one node.
+# Concurrency is safe because runner.py gives each run its own base dir keyed by
+# tokenizer_id, which carries the seed, and symlinks the shared read-only ClimbMix shards
+# and CORE bundle into it; the bpb byte table that concurrent runs used to race on is
+# per-run. ENCODE_WORKERS is set because run_downstream_eval otherwise sizes its encode
+# pool for a run that owns the whole node.
+
+set -euo pipefail
+
+REPO="$(cd "$(dirname "${BASH_SOURCE[0]}")/../../.." && pwd)"
+cd "${REPO}"
+source marker_experiments/downstream/cluster/env.sh
+
+ARM=${ARM:-bnd_wpd_extcaps}
+# Which arms CORE can score. Required, and MAY be empty: run core_prefix_check.py on the
+# trained tokenizer and pass what it supports. Getting it wrong loses the whole run:
+# base_eval raises on the first violation and reports nothing, bits-per-byte included.
+# No `:` in the test, so an explicit empty value is honoured.
+: "${CORE_SAFE_ARMS?set CORE_SAFE_ARMS explicitly; the empty string is a valid value}"
+# No default. a0229 and infra01 bill different budgets and carry different fairshare.
+: "${ACCOUNT:?set ACCOUNT (a0229 or infra01)}"
+SEEDS=${SEEDS:-0,1,2}
+TRAINER=${TRAINER:-bpe}
+DEPTH=${DEPTH:-12}
+VOCAB=${VOCAB:-34685}
+CORPUS=${CORPUS:-fineweb_en_5gb}
+NUM_SHARDS=${NUM_SHARDS:-8}
+GPUS=${GPUS:-1}
+TRAIN_WORKERS=${TRAIN_WORKERS:-90}
+ENCODE_WORKERS=${ENCODE_WORKERS:-90}
+# Not `${OUT:-...}`: env.sh has already exported OUT=results/marker_downstream, so
+# defaulting from OUT would drop this sweep's logs in with the committed BPE runs and
+# collect_results.py would then rebuild one TSV over both.
+OUT_DIR=${OUT_DIR:-results/extcaps_downstream}
+NANOCHAT_BASE_DIR=/capstor/scratch/cscs/cmeister747/marker_downstream/nanochat_base
+
+# `git status --porcelain`, not `git diff`. Neither `git diff HEAD` nor
+# `git diff --cached` reports untracked files, so the guard passed while a new,
+# untracked file that the job runs was described by no commit.
+DIRT=$(git status --porcelain)
+if [[ -n "${DIRT}" ]]; then
+  echo "refusing to submit: working tree differs from HEAD, so the jobs would run code" >&2
+  echo "no commit describes. Untracked files count." >&2
+  echo "${DIRT}" >&2
+  exit 1
+fi
+
+TOK="marker_experiments/downstream/tokenizers/${CORPUS}_${ARM}_${TRAINER}_v${VOCAB}.json.gz"
+[[ -f "$TOK" ]] || { echo "missing tokenizer: $TOK" >&2; exit 1; }
+
+# Verify the tokenizer is exactly the matched size before spending any GPU time. The
+# adapter reports vocab+1 for the synthetic BOS.
+n=$(PYTHONPATH="${REPO}:${REPO}/eval/py-nanochat" uv run python -c "
+from marker_experiments.downstream.boundary_tokenizer import BoundaryBPETokenizer
+from pynanochat.tokenizer_adapter import ScriptBPETokenizerAdapter
+print(ScriptBPETokenizerAdapter(BoundaryBPETokenizer.load('$TOK')).get_vocab_size())
+")
+if [[ "$n" != "$((VOCAB + 1))" ]]; then
+  echo "refusing to submit: ${ARM} vocabulary is ${n}, expected $((VOCAB + 1))" >&2
+  exit 1
+fi
+echo "-- ${ARM}: vocabulary ${n}, matched"
+
+# The shards must already be there. runner.py refuses to download from inside a run, so a
+# missing shard would fail three runs an hour in rather than here.
+have=$(ls "${NANOCHAT_BASE_DIR}/base_data_climbmix"/shard_*.parquet 2>/dev/null | wc -l)
+if (( have < NUM_SHARDS + 1 )); then
+  echo "refusing to submit: ${NANOCHAT_BASE_DIR}/base_data_climbmix holds ${have} shard(s)," >&2
+  echo "and this sweep declares ${NUM_SHARDS} train shards plus a val shard." >&2
+  exit 1
+fi
+echo "-- shards: ${have} present, ${NUM_SHARDS} train + 1 val required"
+
+mkdir -p "${OUT_DIR}/slurm" "${OUT_DIR}/logs"
+# No `|| true`. A scheduler hiccup would leave IN_FLIGHT empty, every seed would look
+# absent, and a re-run would submit duplicates. Two jobs with one tokenizer_id share
+# NANOCHAT_BASE/runs/<id>, hence the checkpoint directory and token_bytes.pt, and both
+# tee to one log. That is how the first sweep of this project was lost.
+if ! IN_FLIGHT=$(squeue -u "$USER" -h -o "%j" 2>&1); then
+  echo "refusing to submit: squeue failed, so in-flight jobs cannot be detected and" >&2
+  echo "this would submit duplicates sharing a checkpoint directory and a log." >&2
+  echo "${IN_FLIGHT}" >&2
+  exit 1
+fi
+JOB="extcaps_${ARM}_${TRAINER}_d${DEPTH}"
+if grep -qx "${JOB}" <<< "$IN_FLIGHT"; then
+  echo "-- ${JOB}: already queued or running, not submitting"; exit 0
+fi
+
+IFS=',' read -ra SEED_LIST <<< "$SEEDS"
+TODO=()
+for seed in "${SEED_LIST[@]}"; do
+  log="${OUT_DIR}/logs/${ARM}_${TRAINER}_d${DEPTH}_s${seed}.log"
+  if [[ -s "$log" ]] && grep -q "artifact_dir" "$log"; then
+    echo "-- seed ${seed}: already has a result, not submitting"; continue
+  fi
+  TODO+=("$seed")
+done
+if (( ${#TODO[@]} == 0 )); then echo "nothing to do"; exit 0; fi
+# One GPU per seed, one node, so the node's 4 GPUs are the ceiling. Past that the loop
+# below would hand two seeds the same CUDA_VISIBLE_DEVICES and they would share a GPU.
+GPUS_PER_NODE=4
+if (( ${#TODO[@]} > GPUS_PER_NODE )); then
+  echo "refusing to submit: ${#TODO[@]} seeds but only ${GPUS_PER_NODE} GPUs on a node." >&2
+  echo "Run them in batches, or extend this script to allocate more nodes." >&2
+  exit 1
+fi
+echo "-- submitting seeds: ${TODO[*]}"
+
+# Build the concurrent launch block: one run_arms.sh per seed, each pinned to its own GPU.
+LAUNCH=""
+gpu=0
+for seed in "${TODO[@]}"; do
+  LAUNCH+="
+CUDA_VISIBLE_DEVICES=${gpu} ARMS=${ARM} SEEDS=${seed} \\
+  marker_experiments/downstream/run_arms.sh > \"${OUT_DIR}/slurm/\${SLURM_JOB_ID}_s${seed}.out\" 2>&1 &
+pids+=(\$!); seeds+=(${seed})
+echo \"[pack] seed ${seed} on GPU ${gpu}, pid \${pids[-1]}\"
+"
+  gpu=$((gpu + 1))
+done
+
+jid=$(sbatch --parsable --job-name="${JOB}" --account="${ACCOUNT}" --partition=normal \
+    --nodes=1 --ntasks-per-node=1 --time=12:00:00 \
+    --output="${OUT_DIR}/slurm/%x_%j.out" --error="${OUT_DIR}/slurm/%x_%j.err" \
+    --wrap="
+# No -e: a failing seed must reach the wait loop below and be reported, not abort the job.
+set -uo pipefail
+cd ${REPO} || exit 1
+source marker_experiments/downstream/cluster/env.sh
+# The venv installs script_bpe and pynanochat as editable pointers into the MAIN
+# checkout, so without this the job runs that tree's library code no matter which
+# checkout the driver script came from, and the clean-tree guard checks the wrong
+# tree. Both entries are needed: REPO alone still resolves pynanochat to the main one.
+export PYTHONPATH=\"${REPO}:${REPO}/eval/py-nanochat:\${PYTHONPATH:-}\"
+export NANOCHAT_BASE=${NANOCHAT_BASE_DIR}
+export OUT=${OUT_DIR}
+export TRAINER=${TRAINER} DEPTH=${DEPTH} GPUS=${GPUS}
+export CORE_SAFE_ARMS=\"${CORE_SAFE_ARMS}\"
+# run_arms.sh ends by copying \$OUT/results.tsv over marker_experiments/paper/generated
+# and regenerating the tables. This sweep TSV holds three extcaps rows, and those paper
+# artifacts are tracked and carry the full BPE grid. Keep them out of it.
+export SKIP_PAPER_ARTIFACTS=1
+export VOCAB=${VOCAB} CORPUS=${CORPUS} NUM_SHARDS=${NUM_SHARDS}
+export TRAIN_WORKERS=${TRAIN_WORKERS} ENCODE_WORKERS=${ENCODE_WORKERS}
+# Pinned, not inherited. --wrap runs under the submitting shell environment, so a
+# leftover SMOKE=1 would make every seed train 20 iterations, tag itself _smoke, and
+# still print a result block that collect_results.py parses into a row.
+export SMOKE=0
+export TAG_SUFFIX=
+
+COMMIT=\$(git rev-parse HEAD)
+MAIN_COMMIT=\$(git -C /users/cmeister747/script_tok rev-parse HEAD)
+MAIN_DIRTY=\$(git -C /users/cmeister747/script_tok status --porcelain | wc -l)
+VENDOR=\$(git -C eval/py-nanochat/vendor/nanochat rev-parse HEAD)
+TOK_SHA=\$(sha256sum ${TOK} | cut -d' ' -f1)
+echo \"[job] \${SLURM_JOB_ID} ${JOB} commit=\${COMMIT}\"
+echo \"[job] main_checkout=\${MAIN_COMMIT} dirty=\${MAIN_DIRTY} nanochat=\${VENDOR}\"
+echo \"[job] tokenizer=${TOK}\"
+echo \"[job] tokenizer_sha256=\${TOK_SHA}\"
+echo \"[job] node=\$(hostname) cores=\$(nproc) gpus=\$(nvidia-smi -L | wc -l)\"
+echo \"[job] seeds=${TODO[*]} core_safe_arms='${CORE_SAFE_ARMS}' encode_workers=${ENCODE_WORKERS}\"
+uv pip freeze > \"${OUT_DIR}/slurm/\${SLURM_JOB_ID}.freeze\" 2>/dev/null || true
+
+# Count the GPUs on the node that was actually allocated. The submit-side check used a
+# constant, which is the login node's answer to a question only the compute node can
+# settle; two runs handed the same device would not be clean single-GPU replicates.
+NGPU=\$(nvidia-smi --query-gpu=index --format=csv,noheader | wc -l)
+if (( ${#TODO[@]} > NGPU )); then
+  echo \"[pack] refusing: ${#TODO[@]} seeds on a node with \${NGPU} GPU(s)\" >&2
+  exit 1
+fi
+echo \"[job] allocated GPUs: \${NGPU}\"
+
+pids=(); seeds=()
+${LAUNCH}
+# Report every failure rather than dying on the first, so one bad seed does not hide the
+# other two.
+fail=0
+for i in \"\${!pids[@]}\"; do
+  if wait \"\${pids[\$i]}\"; then
+    echo \"[pack] seed \${seeds[\$i]} ok\"
+  else
+    echo \"[pack] seed \${seeds[\$i]} FAILED\"
+    fail=1
+  fi
+done
+
+# One authoritative collection after the join. Each run_arms.sh also collects when it
+# finishes, and three concurrent writers to one TSV can interleave; this pass runs when
+# every log is complete and overwrites whatever they left.
+uv run python marker_experiments/downstream/collect_results.py \\
+    --logs-dir \"${OUT_DIR}/logs\" --out \"${OUT_DIR}/results.tsv\" || fail=1
+echo \"[pack] done, fail=\${fail}\"
+exit \${fail}
+")
+echo "-- ${JOB}: job ${jid}, ${#TODO[@]} seed(s) on one node"
diff --git a/marker_experiments/downstream/extcaps_preflight.py b/marker_experiments/downstream/extcaps_preflight.py
new file mode 100644
index 00000000..796cfe9a
--- /dev/null
+++ b/marker_experiments/downstream/extcaps_preflight.py
@@ -0,0 +1,131 @@
+#!/usr/bin/env python3
+"""Tokenizer-side checks for the bnd_wpd_extcaps downstream sweep, before any GPU time.
+
+Three sweeps have been discarded for measurement bugs, so every property the sweep relies
+on is measured here rather than assumed:
+
+  * vocabulary is exactly the matched size (the adapter reports vocab+1, for the synthetic
+    BOS it adds on top of the trained vocabulary);
+  * the held-out slice round-trips with zero failures;
+  * chars/token beside the arms this one is meant to be compared against;
+  * only BOS carries zero bytes in the table evaluate_bpb masks on, and every token that
+    decodes to the empty string carries the floored count of 1. A boundary marker decodes
+    to nothing, and without the floor the cost of predicting it would be dropped from the
+    bpb numerator for exactly the arms under test.
+
+The CORE prefix property is measured separately by core_prefix_check.py, which needs the
+CORE bundle.
+
+    uv run python marker_experiments/downstream/extcaps_preflight.py
+"""
+
+import json
+import os
+import sys
+
+import cyclopts
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+REPO = os.path.dirname(os.path.dirname(HERE))
+sys.path.insert(0, REPO)
+
+from marker_experiments.downstream.train_matched import eval_compression  # noqa: E402
+
+app = cyclopts.App()
+
+
+@app.default
+def main(
+    arm: str = "bnd_wpd_extcaps",
+    against: str = "bnd_wpd_caps,plain",
+    corpus: str = "fineweb_en_5gb",
+    trainer: str = "bpe",
+    vocab: int = 34_685,
+    eval_texts: str = os.path.join(REPO, "marker_experiments", "eval_texts", "en.json"),
+) -> None:
+    """Check the trained tokenizer and print the numbers the sweep is gated on.
+
+    Args:
+        arm: The arm under test.
+        against: Comma-separated arms to report chars/token beside it.
+        corpus: Corpus name in the tokenizer filename.
+        trainer: Trainer name in the tokenizer filename.
+        vocab: Matched total vocabulary the arm must land on exactly.
+        eval_texts: Held-out slice for chars/token and round-trip.
+    """
+    from marker_experiments.downstream.boundary_tokenizer import BoundaryBPETokenizer
+    from pynanochat.tokenizer_adapter import ScriptBPETokenizerAdapter
+    from pynanochat.tokenizer import special_aware_token_bytes, special_token_ids
+
+    tok_dir = os.path.join(HERE, "tokenizers")
+    path = os.path.join(tok_dir, f"{corpus}_{arm}_{trainer}_v{vocab}.json.gz")
+    if not os.path.exists(path):
+        raise SystemExit(f"missing tokenizer: {path}")
+
+    failures = []
+    print(f"tokenizer: {path}")
+    tokenizer = BoundaryBPETokenizer.load(path)
+    adapter = ScriptBPETokenizerAdapter(tokenizer)
+
+    # 1. vocabulary
+    n_adapter = adapter.get_vocab_size()
+    n_trained = len(tokenizer.tokens)
+    print(f"\n[1] vocabulary  trained={n_trained:,}  adapter={n_adapter:,} (+1 synthetic BOS)")
+    if n_trained != vocab:
+        failures.append(f"trained vocabulary {n_trained} != {vocab}")
+    if n_adapter != vocab + 1:
+        failures.append(f"adapter vocabulary {n_adapter} != {vocab + 1}")
+
+    # 2 + 3. round-trip and chars/token, beside the comparison arms
+    print(f"\n[2/3] held-out slice {os.path.relpath(eval_texts, REPO)}")
+    stats = eval_compression(tokenizer, eval_texts)
+    if not stats:
+        raise SystemExit(f"no eval slice at {eval_texts}")
+    print(f"    {arm:<20} chars/token {stats['eval_chars_per_token']:.4f}  "
+          f"roundtrip_failures {stats['roundtrip_failures']}  "
+          f"({stats['eval_chars']:,} chars, {stats['eval_tokens']:,} tokens)")
+    if stats["roundtrip_failures"] != 0:
+        failures.append(f"{stats['roundtrip_failures']} round-trip failures on the held-out slice")
+
+    base = stats["eval_chars_per_token"]
+    for other in [a.strip() for a in against.split(",") if a.strip()]:
+        other_path = os.path.join(tok_dir, f"{corpus}_{other}_{trainer}_v{vocab}.json.gz")
+        if not os.path.exists(other_path):
+            print(f"    {other:<20} MISSING at {other_path}")
+            failures.append(f"comparison tokenizer missing: {other_path}")
+            continue
+        other_tok = BoundaryBPETokenizer.load(other_path)
+        s = eval_compression(other_tok, eval_texts)
+        delta = 100 * (base - s["eval_chars_per_token"]) / s["eval_chars_per_token"]
+        print(f"    {other:<20} chars/token {s['eval_chars_per_token']:.4f}  "
+              f"roundtrip_failures {s['roundtrip_failures']}  "
+              f"({arm} is {delta:+.2f}% against it)")
+
+    # 4. the byte table evaluate_bpb masks on
+    print("\n[4] byte table (0 bytes -> loss masked by evaluate_bpb)")
+    counts = special_aware_token_bytes(adapter)
+    specials = special_token_ids(adapter)
+    bos = adapter.get_bos_token_id()
+    zeros = [i for i, c in enumerate(counts) if c == 0]
+    empty = [i for i in range(n_adapter) if i not in specials and adapter.decode([i]) == ""]
+    print(f"    special ids {sorted(specials)}  bos id {bos}")
+    print(f"    ids with 0 bytes: {zeros}")
+    print(f"    non-special ids decoding to the empty string: {len(empty)}")
+    if empty:
+        bad = [i for i in empty if counts[i] != 1]
+        print(f"    of those, byte count 1: {len(empty) - len(bad)}; not 1: {len(bad)}")
+        if bad:
+            failures.append(f"{len(bad)} empty-decoding tokens are not floored to 1 byte")
+    if zeros != [bos]:
+        failures.append(f"ids with 0 bytes are {zeros}, expected exactly [{bos}] (BOS)")
+
+    print("\n" + "=" * 60)
+    if failures:
+        for f in failures:
+            print(f"FAIL: {f}")
+        raise SystemExit(1)
+    print("all tokenizer-side checks passed")
+
+
+if __name__ == "__main__":
+    app()
diff --git a/marker_experiments/downstream/tokenizers/fineweb_ar_5gb_bnd_w_mingram_v34685.json.gz b/marker_experiments/downstream/tokenizers/fineweb_ar_5gb_bnd_w_mingram_v34685.json.gz
new file mode 100644
index 00000000..1e5b117e
Binary files /dev/null and b/marker_experiments/downstream/tokenizers/fineweb_ar_5gb_bnd_w_mingram_v34685.json.gz differ
diff --git a/marker_experiments/downstream/tokenizers/fineweb_ar_5gb_bnd_wp_mingram_v34685.json.gz b/marker_experiments/downstream/tokenizers/fineweb_ar_5gb_bnd_wp_mingram_v34685.json.gz
new file mode 100644
index 00000000..314e26c8
Binary files /dev/null and b/marker_experiments/downstream/tokenizers/fineweb_ar_5gb_bnd_wp_mingram_v34685.json.gz differ
diff --git a/marker_experiments/downstream/tokenizers/fineweb_ar_5gb_bnd_wpd_mingram_v34685.json.gz b/marker_experiments/downstream/tokenizers/fineweb_ar_5gb_bnd_wpd_mingram_v34685.json.gz
new file mode 100644
index 00000000..b51e5577
Binary files /dev/null and b/marker_experiments/downstream/tokenizers/fineweb_ar_5gb_bnd_wpd_mingram_v34685.json.gz differ
diff --git a/marker_experiments/downstream/tokenizers/fineweb_ar_5gb_plain_mingram_v34685.json.gz b/marker_experiments/downstream/tokenizers/fineweb_ar_5gb_plain_mingram_v34685.json.gz
new file mode 100644
index 00000000..cd41c1ca
Binary files /dev/null and b/marker_experiments/downstream/tokenizers/fineweb_ar_5gb_plain_mingram_v34685.json.gz differ
diff --git a/marker_experiments/downstream/tokenizers/fineweb_de_5gb_bnd_w_mingram_v34685.json.gz b/marker_experiments/downstream/tokenizers/fineweb_de_5gb_bnd_w_mingram_v34685.json.gz
new file mode 100644
index 00000000..21555152
Binary files /dev/null and b/marker_experiments/downstream/tokenizers/fineweb_de_5gb_bnd_w_mingram_v34685.json.gz differ
diff --git a/marker_experiments/downstream/tokenizers/fineweb_de_5gb_bnd_wp_mingram_v34685.json.gz b/marker_experiments/downstream/tokenizers/fineweb_de_5gb_bnd_wp_mingram_v34685.json.gz
new file mode 100644
index 00000000..b1adb1aa
Binary files /dev/null and b/marker_experiments/downstream/tokenizers/fineweb_de_5gb_bnd_wp_mingram_v34685.json.gz differ
diff --git a/marker_experiments/downstream/tokenizers/fineweb_de_5gb_bnd_wpd_mingram_v34685.json.gz b/marker_experiments/downstream/tokenizers/fineweb_de_5gb_bnd_wpd_mingram_v34685.json.gz
new file mode 100644
index 00000000..2bbb18da
Binary files /dev/null and b/marker_experiments/downstream/tokenizers/fineweb_de_5gb_bnd_wpd_mingram_v34685.json.gz differ
diff --git a/marker_experiments/downstream/tokenizers/fineweb_de_5gb_plain_mingram_v34685.json.gz b/marker_experiments/downstream/tokenizers/fineweb_de_5gb_plain_mingram_v34685.json.gz
new file mode 100644
index 00000000..1f50ae3a
Binary files /dev/null and b/marker_experiments/downstream/tokenizers/fineweb_de_5gb_plain_mingram_v34685.json.gz differ
diff --git a/marker_experiments/downstream/tokenizers/fineweb_en_5gb_bnd_w_mingram_v34685.json.gz b/marker_experiments/downstream/tokenizers/fineweb_en_5gb_bnd_w_mingram_v34685.json.gz
new file mode 100644
index 00000000..919d12aa
Binary files /dev/null and b/marker_experiments/downstream/tokenizers/fineweb_en_5gb_bnd_w_mingram_v34685.json.gz differ
diff --git a/marker_experiments/downstream/tokenizers/fineweb_en_5gb_bnd_wp_mingram_v34685.json.gz b/marker_experiments/downstream/tokenizers/fineweb_en_5gb_bnd_wp_mingram_v34685.json.gz
new file mode 100644
index 00000000..4837b8c0
Binary files /dev/null and b/marker_experiments/downstream/tokenizers/fineweb_en_5gb_bnd_wp_mingram_v34685.json.gz differ
diff --git a/marker_experiments/downstream/tokenizers/fineweb_en_5gb_bnd_wpd_extcaps_bpe_v34685.json.gz b/marker_experiments/downstream/tokenizers/fineweb_en_5gb_bnd_wpd_extcaps_bpe_v34685.json.gz
new file mode 100644
index 00000000..e5c94121
Binary files /dev/null and b/marker_experiments/downstream/tokenizers/fineweb_en_5gb_bnd_wpd_extcaps_bpe_v34685.json.gz differ
diff --git a/marker_experiments/downstream/tokenizers/fineweb_fi_5gb_bnd_w_mingram_v34685.json.gz b/marker_experiments/downstream/tokenizers/fineweb_fi_5gb_bnd_w_mingram_v34685.json.gz
new file mode 100644
index 00000000..dc011345
Binary files /dev/null and b/marker_experiments/downstream/tokenizers/fineweb_fi_5gb_bnd_w_mingram_v34685.json.gz differ
diff --git a/marker_experiments/downstream/tokenizers/fineweb_fi_5gb_bnd_wp_mingram_v34685.json.gz b/marker_experiments/downstream/tokenizers/fineweb_fi_5gb_bnd_wp_mingram_v34685.json.gz
new file mode 100644
index 00000000..6d833a57
Binary files /dev/null and b/marker_experiments/downstream/tokenizers/fineweb_fi_5gb_bnd_wp_mingram_v34685.json.gz differ
diff --git a/marker_experiments/downstream/tokenizers/fineweb_fi_5gb_bnd_wpd_mingram_v34685.json.gz b/marker_experiments/downstream/tokenizers/fineweb_fi_5gb_bnd_wpd_mingram_v34685.json.gz
new file mode 100644
index 00000000..c4659bd2
Binary files /dev/null and b/marker_experiments/downstream/tokenizers/fineweb_fi_5gb_bnd_wpd_mingram_v34685.json.gz differ
diff --git a/marker_experiments/downstream/tokenizers/fineweb_fi_5gb_plain_mingram_v34685.json.gz b/marker_experiments/downstream/tokenizers/fineweb_fi_5gb_plain_mingram_v34685.json.gz
new file mode 100644
index 00000000..a79899fc
Binary files /dev/null and b/marker_experiments/downstream/tokenizers/fineweb_fi_5gb_plain_mingram_v34685.json.gz differ
diff --git a/marker_experiments/downstream/tokenizers/fineweb_ru_5gb_bnd_w_mingram_v34685.json.gz b/marker_experiments/downstream/tokenizers/fineweb_ru_5gb_bnd_w_mingram_v34685.json.gz
new file mode 100644
index 00000000..0b46059c
Binary files /dev/null and b/marker_experiments/downstream/tokenizers/fineweb_ru_5gb_bnd_w_mingram_v34685.json.gz differ
diff --git a/marker_experiments/downstream/tokenizers/fineweb_ru_5gb_bnd_wp_mingram_v34685.json.gz b/marker_experiments/downstream/tokenizers/fineweb_ru_5gb_bnd_wp_mingram_v34685.json.gz
new file mode 100644
index 00000000..2f8e8573
Binary files /dev/null and b/marker_experiments/downstream/tokenizers/fineweb_ru_5gb_bnd_wp_mingram_v34685.json.gz differ
diff --git a/marker_experiments/downstream/tokenizers/fineweb_ru_5gb_bnd_wpd_mingram_v34685.json.gz b/marker_experiments/downstream/tokenizers/fineweb_ru_5gb_bnd_wpd_mingram_v34685.json.gz
new file mode 100644
index 00000000..f2fe9239
Binary files /dev/null and b/marker_experiments/downstream/tokenizers/fineweb_ru_5gb_bnd_wpd_mingram_v34685.json.gz differ
diff --git a/marker_experiments/downstream/tokenizers/fineweb_ru_5gb_plain_mingram_v34685.json.gz b/marker_experiments/downstream/tokenizers/fineweb_ru_5gb_plain_mingram_v34685.json.gz
new file mode 100644
index 00000000..7a51ac1e
Binary files /dev/null and b/marker_experiments/downstream/tokenizers/fineweb_ru_5gb_plain_mingram_v34685.json.gz differ
diff --git a/marker_experiments/paper/generated/eval_goldfish.json b/marker_experiments/paper/generated/eval_goldfish.json
index 10b099e4..1981131f 100644
--- a/marker_experiments/paper/generated/eval_goldfish.json
+++ b/marker_experiments/paper/generated/eval_goldfish.json
@@ -11,6 +11,18 @@
     "train_corpus": "fineweb_ar_5gb",
     "train_tokens": 1329072646
   },
+  "fineweb_ar_5gb_bnd_w_mingram_v34685": {
+    "eval_chars": 8036366,
+    "eval_chars_per_token": 3.658174710400105,
+    "eval_corpus": "goldfish:arb_arab.txt",
+    "eval_docs": 31081,
+    "eval_tokens": 2196824,
+    "roundtrip_failures": 0,
+    "slice_hash": "aec9fd6c5acc275c",
+    "train_chars": 4997549743,
+    "train_corpus": "fineweb_ar_5gb",
+    "train_tokens": 1321433705
+  },
   "fineweb_ar_5gb_bnd_wp_bpe_v34685": {
     "eval_chars": 8036366,
     "eval_chars_per_token": 3.8977919015372198,
@@ -23,6 +35,18 @@
     "train_corpus": "fineweb_ar_5gb",
     "train_tokens": 1237944737
   },
+  "fineweb_ar_5gb_bnd_wp_mingram_v34685": {
+    "eval_chars": 8036366,
+    "eval_chars_per_token": 3.9236006819606564,
+    "eval_corpus": "goldfish:arb_arab.txt",
+    "eval_docs": 31081,
+    "eval_tokens": 2048212,
+    "roundtrip_failures": 0,
+    "slice_hash": "aec9fd6c5acc275c",
+    "train_chars": 4997549743,
+    "train_corpus": "fineweb_ar_5gb",
+    "train_tokens": 1230294575
+  },
   "fineweb_ar_5gb_bnd_wpd_bpe_v34685": {
     "eval_chars": 8036366,
     "eval_chars_per_token": 3.9653428241961737,
@@ -47,6 +71,18 @@
     "train_corpus": "fineweb_ar_5gb",
     "train_tokens": 1253230657
   },
+  "fineweb_ar_5gb_bnd_wpd_mingram_v34685": {
+    "eval_chars": 8036366,
+    "eval_chars_per_token": 3.9922334823646297,
+    "eval_corpus": "goldfish:arb_arab.txt",
+    "eval_docs": 31081,
+    "eval_tokens": 2013000,
+    "roundtrip_failures": 0,
+    "slice_hash": "aec9fd6c5acc275c",
+    "train_chars": 4997549743,
+    "train_corpus": "fineweb_ar_5gb",
+    "train_tokens": 1210371332
+  },
   "fineweb_ar_5gb_plain_bpe_v34685": {
     "eval_chars": 8036366,
     "eval_chars_per_token": 3.973489193550939,
@@ -59,6 +95,18 @@
     "train_corpus": "fineweb_ar_5gb",
     "train_tokens": 1222664188
   },
+  "fineweb_ar_5gb_plain_mingram_v34685": {
+    "eval_chars": 8036366,
+    "eval_chars_per_token": 4.023890889226368,
+    "eval_corpus": "goldfish:arb_arab.txt",
+    "eval_docs": 31081,
+    "eval_tokens": 1997163,
+    "roundtrip_failures": 0,
+    "slice_hash": "aec9fd6c5acc275c",
+    "train_chars": 4997549743,
+    "train_corpus": "fineweb_ar_5gb",
+    "train_tokens": 1208558599
+  },
   "fineweb_ar_5gb_quick_bnd_w_bpe_v34685": {
     "eval_chars": 8036366,
     "eval_chars_per_token": 3.63050661240731,
@@ -263,6 +311,18 @@
     "train_corpus": "fineweb_de_5gb",
     "train_tokens": 1216560299
   },
+  "fineweb_de_5gb_bnd_w_mingram_v34685": {
+    "eval_chars": 8033174,
+    "eval_chars_per_token": 4.153148760984306,
+    "eval_corpus": "goldfish:deu_latn.txt",
+    "eval_docs": 23953,
+    "eval_tokens": 1934237,
+    "roundtrip_failures": 0,
+    "slice_hash": "eca22877fba9c9eb",
+    "train_chars": 4997457262,
+    "train_corpus": "fineweb_de_5gb",
+    "train_tokens": 1196858906
+  },
   "fineweb_de_5gb_bnd_wp_bpe_v34685": {
     "eval_chars": 8033174,
     "eval_chars_per_token": 4.4384604461791,
@@ -275,6 +335,18 @@
     "train_corpus": "fineweb_de_5gb",
     "train_tokens": 1125504486
   },
+  "fineweb_de_5gb_bnd_wp_mingram_v34685": {
+    "eval_chars": 8033174,
+    "eval_chars_per_token": 4.518179471272527,
+    "eval_corpus": "goldfish:deu_latn.txt",
+    "eval_docs": 23953,
+    "eval_tokens": 1777967,
+    "roundtrip_failures": 0,
+    "slice_hash": "eca22877fba9c9eb",
+    "train_chars": 4997457262,
+    "train_corpus": "fineweb_de_5gb",
+    "train_tokens": 1105770329
+  },
   "fineweb_de_5gb_bnd_wpd_bpe_v34685": {
     "eval_chars": 8033174,
     "eval_chars_per_token": 4.527109092343586,
@@ -299,6 +371,18 @@
     "train_corpus": "fineweb_de_5gb",
     "train_tokens": 1102478035
   },
+  "fineweb_de_5gb_bnd_wpd_mingram_v34685": {
+    "eval_chars": 8033174,
+    "eval_chars_per_token": 4.610899694297027,
+    "eval_corpus": "goldfish:deu_latn.txt",
+    "eval_docs": 23953,
+    "eval_tokens": 1742214,
+    "roundtrip_failures": 0,
+    "slice_hash": "eca22877fba9c9eb",
+    "train_chars": 4997457262,
+    "train_corpus": "fineweb_de_5gb",
+    "train_tokens": 1085260981
+  },
   "fineweb_de_5gb_plain_bpe_v34685": {
     "eval_chars": 8033174,
     "eval_chars_per_token": 4.574380040578023,
@@ -311,6 +395,18 @@
     "train_corpus": "fineweb_de_5gb",
     "train_tokens": 1098186861
   },
+  "fineweb_de_5gb_plain_mingram_v34685": {
+    "eval_chars": 8033174,
+    "eval_chars_per_token": 4.665331697143889,
+    "eval_corpus": "goldfish:deu_latn.txt",
+    "eval_docs": 23953,
+    "eval_tokens": 1721887,
+    "roundtrip_failures": 0,
+    "slice_hash": "eca22877fba9c9eb",
+    "train_chars": 4997457262,
+    "train_corpus": "fineweb_de_5gb",
+    "train_tokens": 1076721751
+  },
   "fineweb_de_5gb_quick_bnd_w_bpe_v34685": {
     "eval_chars": 8033174,
     "eval_chars_per_token": 4.084043140552506,
@@ -515,6 +611,18 @@
     "train_corpus": "fineweb_en_5gb",
     "train_tokens": 1255889200
   },
+  "fineweb_en_5gb_bnd_w_mingram_v34685": {
+    "eval_chars": 8109368,
+    "eval_chars_per_token": 4.086721855990034,
+    "eval_corpus": "goldfish:eng_latn.txt",
+    "eval_docs": 22442,
+    "eval_tokens": 1984321,
+    "roundtrip_failures": 0,
+    "slice_hash": "8bca50f02b720acc",
+    "train_chars": 4994221434,
+    "train_corpus": "fineweb_en_5gb",
+    "train_tokens": 1248287691
+  },
   "fineweb_en_5gb_bnd_wp_bpe_v34685": {
     "eval_chars": 8109368,
     "eval_chars_per_token": 4.399205585853309,
@@ -527,6 +635,18 @@
     "train_corpus": "fineweb_en_5gb",
     "train_tokens": 1162736882
   },
+  "fineweb_en_5gb_bnd_wp_mingram_v34685": {
+    "eval_chars": 8109368,
+    "eval_chars_per_token": 4.4284568273722895,
+    "eval_corpus": "goldfish:eng_latn.txt",
+    "eval_docs": 22442,
+    "eval_tokens": 1831195,
+    "roundtrip_failures": 0,
+    "slice_hash": "8bca50f02b720acc",
+    "train_chars": 4994221434,
+    "train_corpus": "fineweb_en_5gb",
+    "train_tokens": 1155078497
+  },
   "fineweb_en_5gb_bnd_wpd_bpe_v34685": {
     "eval_chars": 8109368,
     "eval_chars_per_token": 4.484629958778006,
@@ -551,6 +671,18 @@
     "train_corpus": "fineweb_en_5gb",
     "train_tokens": 1140849054
   },
+  "fineweb_en_5gb_bnd_wpd_extcaps_bpe_v34685": {
+    "eval_chars": 8109368,
+    "eval_chars_per_token": 4.490657209893949,
+    "eval_corpus": "goldfish:eng_latn.txt",
+    "eval_docs": 22442,
+    "eval_tokens": 1805831,
+    "roundtrip_failures": 0,
+    "slice_hash": "8bca50f02b720acc",
+    "train_chars": 4994221434,
+    "train_corpus": "fineweb_en_5gb",
+    "train_tokens": 1140257359
+  },
   "fineweb_en_5gb_bnd_wpd_mingram_v34685": {
     "eval_chars": 8109368,
     "eval_chars_per_token": 4.51586439322601,
@@ -791,6 +923,18 @@
     "train_corpus": "fineweb_fi_5gb",
     "train_tokens": 1161948277
   },
+  "fineweb_fi_5gb_bnd_w_mingram_v34685": {
+    "eval_chars": 8013043,
+    "eval_chars_per_token": 4.6646689750241,
+    "eval_corpus": "goldfish:fin_latn.txt",
+    "eval_docs": 29307,
+    "eval_tokens": 1717816,
+    "roundtrip_failures": 0,
+    "slice_hash": "1df047431ded0222",
+    "train_chars": 4994960579,
+    "train_corpus": "fineweb_fi_5gb",
+    "train_tokens": 1143138540
+  },
   "fineweb_fi_5gb_bnd_wp_bpe_v34685": {
     "eval_chars": 8013043,
     "eval_chars_per_token": 4.928285921463975,
@@ -803,6 +947,18 @@
     "train_corpus": "fineweb_fi_5gb",
     "train_tokens": 1090389249
   },
+  "fineweb_fi_5gb_bnd_wp_mingram_v34685": {
+    "eval_chars": 8013043,
+    "eval_chars_per_token": 4.974218378995788,
+    "eval_corpus": "goldfish:fin_latn.txt",
+    "eval_docs": 29307,
+    "eval_tokens": 1610915,
+    "roundtrip_failures": 0,
+    "slice_hash": "1df047431ded0222",
+    "train_chars": 4994960579,
+    "train_corpus": "fineweb_fi_5gb",
+    "train_tokens": 1071616061
+  },
   "fineweb_fi_5gb_bnd_wpd_bpe_v34685": {
     "eval_chars": 8013043,
     "eval_chars_per_token": 4.988341922141695,
@@ -827,6 +983,18 @@
     "train_corpus": "fineweb_fi_5gb",
     "train_tokens": 1071141065
   },
+  "fineweb_fi_5gb_bnd_wpd_mingram_v34685": {
+    "eval_chars": 8013043,
+    "eval_chars_per_token": 5.035618024059982,
+    "eval_corpus": "goldfish:fin_latn.txt",
+    "eval_docs": 29307,
+    "eval_tokens": 1591273,
+    "roundtrip_failures": 0,
+    "slice_hash": "1df047431ded0222",
+    "train_chars": 4994960579,
+    "train_corpus": "fineweb_fi_5gb",
+    "train_tokens": 1055162976
+  },
   "fineweb_fi_5gb_plain_bpe_v34685": {
     "eval_chars": 8013043,
     "eval_chars_per_token": 5.011788554069547,
@@ -839,6 +1007,18 @@
     "train_corpus": "fineweb_fi_5gb",
     "train_tokens": 1071826379
   },
+  "fineweb_fi_5gb_plain_mingram_v34685": {
+    "eval_chars": 8013043,
+    "eval_chars_per_token": 5.152810591338592,
+    "eval_corpus": "goldfish:fin_latn.txt",
+    "eval_docs": 29307,
+    "eval_tokens": 1555082,
+    "roundtrip_failures": 0,
+    "slice_hash": "1df047431ded0222",
+    "train_chars": 4994960579,
+    "train_corpus": "fineweb_fi_5gb",
+    "train_tokens": 1038682926
+  },
   "fineweb_fi_5gb_quick_bnd_w_bpe_v34685": {
     "eval_chars": 8013043,
     "eval_chars_per_token": 4.622411624901933,
@@ -1355,6 +1535,18 @@
     "train_corpus": "fineweb_ru_5gb",
     "train_tokens": 1280518773
   },
+  "fineweb_ru_5gb_bnd_w_mingram_v34685": {
+    "eval_chars": 8015461,
+    "eval_chars_per_token": 3.890271531645041,
+    "eval_corpus": "goldfish:rus_cyrl.txt",
+    "eval_docs": 27615,
+    "eval_tokens": 2060386,
+    "roundtrip_failures": 0,
+    "slice_hash": "0896372a26fffac9",
+    "train_chars": 4999671485,
+    "train_corpus": "fineweb_ru_5gb",
+    "train_tokens": 1268156020
+  },
   "fineweb_ru_5gb_bnd_wp_bpe_v34685": {
     "eval_chars": 8015461,
     "eval_chars_per_token": 4.268870090016584,
@@ -1367,6 +1559,18 @@
     "train_corpus": "fineweb_ru_5gb",
     "train_tokens": 1162777299
   },
+  "fineweb_ru_5gb_bnd_wp_mingram_v34685": {
+    "eval_chars": 8015461,
+    "eval_chars_per_token": 4.3162588400135915,
+    "eval_corpus": "goldfish:rus_cyrl.txt",
+    "eval_docs": 27615,
+    "eval_tokens": 1857039,
+    "roundtrip_failures": 0,
+    "slice_hash": "0896372a26fffac9",
+    "train_chars": 4999671485,
+    "train_corpus": "fineweb_ru_5gb",
+    "train_tokens": 1150370586
+  },
   "fineweb_ru_5gb_bnd_wpd_bpe_v34685": {
     "eval_chars": 8015461,
     "eval_chars_per_token": 4.343928875304166,
@@ -1391,6 +1595,18 @@
     "train_corpus": "fineweb_ru_5gb",
     "train_tokens": 1139746207
   },
+  "fineweb_ru_5gb_bnd_wpd_mingram_v34685": {
+    "eval_chars": 8015461,
+    "eval_chars_per_token": 4.394340152255266,
+    "eval_corpus": "goldfish:rus_cyrl.txt",
+    "eval_docs": 27615,
+    "eval_tokens": 1824042,
+    "roundtrip_failures": 0,
+    "slice_hash": "0896372a26fffac9",
+    "train_chars": 4999671485,
+    "train_corpus": "fineweb_ru_5gb",
+    "train_tokens": 1130568761
+  },
   "fineweb_ru_5gb_plain_bpe_v34685": {
     "eval_chars": 8015461,
     "eval_chars_per_token": 4.32163225914284,
@@ -1403,6 +1619,18 @@
     "train_corpus": "fineweb_ru_5gb",
     "train_tokens": 1152140966
   },
+  "fineweb_ru_5gb_plain_mingram_v34685": {
+    "eval_chars": 8015461,
+    "eval_chars_per_token": 4.373656876204528,
+    "eval_corpus": "goldfish:rus_cyrl.txt",
+    "eval_docs": 27615,
+    "eval_tokens": 1832668,
+    "roundtrip_failures": 0,
+    "slice_hash": "0896372a26fffac9",
+    "train_chars": 4999671485,
+    "train_corpus": "fineweb_ru_5gb",
+    "train_tokens": 1139033861
+  },
   "fineweb_ru_5gb_quick_bnd_w_bpe_v34685": {
     "eval_chars": 8015461,
     "eval_chars_per_token": 3.845729159169965,
diff --git a/marker_experiments/paper/generated/manifest_parts/fineweb_ar_5gb_bnd_w_mingram_v34685.json b/marker_experiments/paper/generated/manifest_parts/fineweb_ar_5gb_bnd_w_mingram_v34685.json
new file mode 100644
index 00000000..d78dd629
--- /dev/null
+++ b/marker_experiments/paper/generated/manifest_parts/fineweb_ar_5gb_bnd_w_mingram_v34685.json
@@ -0,0 +1,19 @@
+{
+  "fineweb_ar_5gb_bnd_w_mingram_v34685": {
+    "additional_vocab_size": 32974,
+    "arm": "bnd_w",
+    "atomic_vocab": 1711,
+    "corpus": "fineweb_ar_5gb",
+    "eval_chars": 745245,
+    "eval_chars_per_token": 3.3515999190483687,
+    "eval_slice": "marker_experiments/eval_texts/ar.json",
+    "eval_tokens": 222355,
+    "lang": "ar",
+    "overshoot_factor": 1.15,
+    "path": "downstream/tokenizers/fineweb_ar_5gb_bnd_w_mingram_v34685.json.gz",
+    "roundtrip_failures": 0,
+    "total_vocab": 34685,
+    "train_seconds": 782,
+    "trainer": "mingram"
+  }
+}
\ No newline at end of file
diff --git a/marker_experiments/paper/generated/manifest_parts/fineweb_ar_5gb_bnd_wp_mingram_v34685.json b/marker_experiments/paper/generated/manifest_parts/fineweb_ar_5gb_bnd_wp_mingram_v34685.json
new file mode 100644
index 00000000..4023cc44
--- /dev/null
+++ b/marker_experiments/paper/generated/manifest_parts/fineweb_ar_5gb_bnd_wp_mingram_v34685.json
@@ -0,0 +1,19 @@
+{
+  "fineweb_ar_5gb_bnd_wp_mingram_v34685": {
+    "additional_vocab_size": 32974,
+    "arm": "bnd_wp",
+    "atomic_vocab": 1711,
+    "corpus": "fineweb_ar_5gb",
+    "eval_chars": 745245,
+    "eval_chars_per_token": 3.683532854219595,
+    "eval_slice": "marker_experiments/eval_texts/ar.json",
+    "eval_tokens": 202318,
+    "lang": "ar",
+    "overshoot_factor": 1.15,
+    "path": "downstream/tokenizers/fineweb_ar_5gb_bnd_wp_mingram_v34685.json.gz",
+    "roundtrip_failures": 0,
+    "total_vocab": 34685,
+    "train_seconds": 784,
+    "trainer": "mingram"
+  }
+}
\ No newline at end of file
diff --git a/marker_experiments/paper/generated/manifest_parts/fineweb_ar_5gb_bnd_wpd_mingram_v34685.json b/marker_experiments/paper/generated/manifest_parts/fineweb_ar_5gb_bnd_wpd_mingram_v34685.json
new file mode 100644
index 00000000..fec0a6a4
--- /dev/null
+++ b/marker_experiments/paper/generated/manifest_parts/fineweb_ar_5gb_bnd_wpd_mingram_v34685.json
@@ -0,0 +1,19 @@
+{
+  "fineweb_ar_5gb_bnd_wpd_mingram_v34685": {
+    "additional_vocab_size": 32974,
+    "arm": "bnd_wpd",
+    "atomic_vocab": 1711,
+    "corpus": "fineweb_ar_5gb",
+    "eval_chars": 745245,
+    "eval_chars_per_token": 3.843311484371374,
+    "eval_slice": "marker_experiments/eval_texts/ar.json",
+    "eval_tokens": 193907,
+    "lang": "ar",
+    "overshoot_factor": 1.15,
+    "path": "downstream/tokenizers/fineweb_ar_5gb_bnd_wpd_mingram_v34685.json.gz",
+    "roundtrip_failures": 0,
+    "total_vocab": 34685,
+    "train_seconds": 787,
+    "trainer": "mingram"
+  }
+}
\ No newline at end of file
diff --git a/marker_experiments/paper/generated/manifest_parts/fineweb_ar_5gb_plain_mingram_v34685.json b/marker_experiments/paper/generated/manifest_parts/fineweb_ar_5gb_plain_mingram_v34685.json
new file mode 100644
index 00000000..a7211465
--- /dev/null
+++ b/marker_experiments/paper/generated/manifest_parts/fineweb_ar_5gb_plain_mingram_v34685.json
@@ -0,0 +1,19 @@
+{
+  "fineweb_ar_5gb_plain_mingram_v34685": {
+    "additional_vocab_size": 32975,
+    "arm": "plain",
+    "atomic_vocab": 1710,
+    "corpus": "fineweb_ar_5gb",
+    "eval_chars": 745245,
+    "eval_chars_per_token": 3.817051746303287,
+    "eval_slice": "marker_experiments/eval_texts/ar.json",
+    "eval_tokens": 195241,
+    "lang": "ar",
+    "overshoot_factor": 1.15,
+    "path": "downstream/tokenizers/fineweb_ar_5gb_plain_mingram_v34685.json.gz",
+    "roundtrip_failures": 0,
+    "total_vocab": 34685,
+    "train_seconds": 825,
+    "trainer": "mingram"
+  }
+}
\ No newline at end of file
diff --git a/marker_experiments/paper/generated/manifest_parts/fineweb_de_5gb_bnd_w_mingram_v34685.json b/marker_experiments/paper/generated/manifest_parts/fineweb_de_5gb_bnd_w_mingram_v34685.json
new file mode 100644
index 00000000..48a91c81
--- /dev/null
+++ b/marker_experiments/paper/generated/manifest_parts/fineweb_de_5gb_bnd_w_mingram_v34685.json
@@ -0,0 +1,19 @@
+{
+  "fineweb_de_5gb_bnd_w_mingram_v34685": {
+    "additional_vocab_size": 32974,
+    "arm": "bnd_w",
+    "atomic_vocab": 1711,
+    "corpus": "fineweb_de_5gb",
+    "eval_chars": 1404827,
+    "eval_chars_per_token": 3.4056329560413188,
+    "eval_slice": "marker_experiments/eval_texts/de.json",
+    "eval_tokens": 412501,
+    "lang": "de",
+    "overshoot_factor": 1.15,
+    "path": "downstream/tokenizers/fineweb_de_5gb_bnd_w_mingram_v34685.json.gz",
+    "roundtrip_failures": 0,
+    "total_vocab": 34685,
+    "train_seconds": 1262,
+    "trainer": "mingram"
+  }
+}
\ No newline at end of file
diff --git a/marker_experiments/paper/generated/manifest_parts/fineweb_de_5gb_bnd_wp_mingram_v34685.json b/marker_experiments/paper/generated/manifest_parts/fineweb_de_5gb_bnd_wp_mingram_v34685.json
new file mode 100644
index 00000000..49177422
--- /dev/null
+++ b/marker_experiments/paper/generated/manifest_parts/fineweb_de_5gb_bnd_wp_mingram_v34685.json
@@ -0,0 +1,19 @@
+{
+  "fineweb_de_5gb_bnd_wp_mingram_v34685": {
+    "additional_vocab_size": 32974,
+    "arm": "bnd_wp",
+    "atomic_vocab": 1711,
+    "corpus": "fineweb_de_5gb",
+    "eval_chars": 1404827,
+    "eval_chars_per_token": 3.725037188669189,
+    "eval_slice": "marker_experiments/eval_texts/de.json",
+    "eval_tokens": 377131,
+    "lang": "de",
+    "overshoot_factor": 1.15,
+    "path": "downstream/tokenizers/fineweb_de_5gb_bnd_wp_mingram_v34685.json.gz",
+    "roundtrip_failures": 0,
+    "total_vocab": 34685,
+    "train_seconds": 1250,
+    "trainer": "mingram"
+  }
+}
\ No newline at end of file
diff --git a/marker_experiments/paper/generated/manifest_parts/fineweb_de_5gb_bnd_wpd_mingram_v34685.json b/marker_experiments/paper/generated/manifest_parts/fineweb_de_5gb_bnd_wpd_mingram_v34685.json
new file mode 100644
index 00000000..a69f5437
--- /dev/null
+++ b/marker_experiments/paper/generated/manifest_parts/fineweb_de_5gb_bnd_wpd_mingram_v34685.json
@@ -0,0 +1,19 @@
+{
+  "fineweb_de_5gb_bnd_wpd_mingram_v34685": {
+    "additional_vocab_size": 32974,
+    "arm": "bnd_wpd",
+    "atomic_vocab": 1711,
+    "corpus": "fineweb_de_5gb",
+    "eval_chars": 1404827,
+    "eval_chars_per_token": 3.90280675532639,
+    "eval_slice": "marker_experiments/eval_texts/de.json",
+    "eval_tokens": 359953,
+    "lang": "de",
+    "overshoot_factor": 1.15,
+    "path": "downstream/tokenizers/fineweb_de_5gb_bnd_wpd_mingram_v34685.json.gz",
+    "roundtrip_failures": 0,
+    "total_vocab": 34685,
+    "train_seconds": 1270,
+    "trainer": "mingram"
+  }
+}
\ No newline at end of file
diff --git a/marker_experiments/paper/generated/manifest_parts/fineweb_de_5gb_plain_mingram_v34685.json b/marker_experiments/paper/generated/manifest_parts/fineweb_de_5gb_plain_mingram_v34685.json
new file mode 100644
index 00000000..2a58b0de
--- /dev/null
+++ b/marker_experiments/paper/generated/manifest_parts/fineweb_de_5gb_plain_mingram_v34685.json
@@ -0,0 +1,19 @@
+{
+  "fineweb_de_5gb_plain_mingram_v34685": {
+    "additional_vocab_size": 32975,
+    "arm": "plain",
+    "atomic_vocab": 1710,
+    "corpus": "fineweb_de_5gb",
+    "eval_chars": 1404827,
+    "eval_chars_per_token": 3.8938173533194376,
+    "eval_slice": "marker_experiments/eval_texts/de.json",
+    "eval_tokens": 360784,
+    "lang": "de",
+    "overshoot_factor": 1.15,
+    "path": "downstream/tokenizers/fineweb_de_5gb_plain_mingram_v34685.json.gz",
+    "roundtrip_failures": 0,
+    "total_vocab": 34685,
+    "train_seconds": 1363,
+    "trainer": "mingram"
+  }
+}
\ No newline at end of file
diff --git a/marker_experiments/paper/generated/manifest_parts/fineweb_en_5gb_bnd_w_mingram_v34685.json b/marker_experiments/paper/generated/manifest_parts/fineweb_en_5gb_bnd_w_mingram_v34685.json
new file mode 100644
index 00000000..c7cb64bd
--- /dev/null
+++ b/marker_experiments/paper/generated/manifest_parts/fineweb_en_5gb_bnd_w_mingram_v34685.json
@@ -0,0 +1,19 @@
+{
+  "fineweb_en_5gb_bnd_w_mingram_v34685": {
+    "additional_vocab_size": 32974,
+    "arm": "bnd_w",
+    "atomic_vocab": 1711,
+    "corpus": "fineweb_en_5gb",
+    "eval_chars": 3602925,
+    "eval_chars_per_token": 3.1113530047245535,
+    "eval_slice": "marker_experiments/eval_texts/en.json",
+    "eval_tokens": 1157993,
+    "lang": "en",
+    "overshoot_factor": 1.15,
+    "path": "downstream/tokenizers/fineweb_en_5gb_bnd_w_mingram_v34685.json.gz",
+    "roundtrip_failures": 0,
+    "total_vocab": 34685,
+    "train_seconds": 850,
+    "trainer": "mingram"
+  }
+}
\ No newline at end of file
diff --git a/marker_experiments/paper/generated/manifest_parts/fineweb_en_5gb_bnd_wp_mingram_v34685.json b/marker_experiments/paper/generated/manifest_parts/fineweb_en_5gb_bnd_wp_mingram_v34685.json
new file mode 100644
index 00000000..d26aaabf
--- /dev/null
+++ b/marker_experiments/paper/generated/manifest_parts/fineweb_en_5gb_bnd_wp_mingram_v34685.json
@@ -0,0 +1,19 @@
+{
+  "fineweb_en_5gb_bnd_wp_mingram_v34685": {
+    "additional_vocab_size": 32974,
+    "arm": "bnd_wp",
+    "atomic_vocab": 1711,
+    "corpus": "fineweb_en_5gb",
+    "eval_chars": 3602925,
+    "eval_chars_per_token": 3.5446230402342276,
+    "eval_slice": "marker_experiments/eval_texts/en.json",
+    "eval_tokens": 1016448,
+    "lang": "en",
+    "overshoot_factor": 1.15,
+    "path": "downstream/tokenizers/fineweb_en_5gb_bnd_wp_mingram_v34685.json.gz",
+    "roundtrip_failures": 0,
+    "total_vocab": 34685,
+    "train_seconds": 855,
+    "trainer": "mingram"
+  }
+}
\ No newline at end of file
diff --git a/marker_experiments/paper/generated/manifest_parts/fineweb_en_5gb_bnd_wpd_extcaps_bpe_v34685.json b/marker_experiments/paper/generated/manifest_parts/fineweb_en_5gb_bnd_wpd_extcaps_bpe_v34685.json
new file mode 100644
index 00000000..bff32272
--- /dev/null
+++ b/marker_experiments/paper/generated/manifest_parts/fineweb_en_5gb_bnd_wpd_extcaps_bpe_v34685.json
@@ -0,0 +1,19 @@
+{
+  "fineweb_en_5gb_bnd_wpd_extcaps_bpe_v34685": {
+    "additional_vocab_size": 32972,
+    "arm": "bnd_wpd_extcaps",
+    "atomic_vocab": 1713,
+    "corpus": "fineweb_en_5gb",
+    "eval_chars": 3602925,
+    "eval_chars_per_token": 3.74635027388543,
+    "eval_slice": "marker_experiments/eval_texts/en.json",
+    "eval_tokens": 961716,
+    "lang": "en",
+    "overshoot_factor": null,
+    "path": "downstream/tokenizers/fineweb_en_5gb_bnd_wpd_extcaps_bpe_v34685.json.gz",
+    "roundtrip_failures": 0,
+    "total_vocab": 34685,
+    "train_seconds": 153,
+    "trainer": "bpe"
+  }
+}
\ No newline at end of file
diff --git a/marker_experiments/paper/generated/manifest_parts/fineweb_fi_5gb_bnd_w_mingram_v34685.json b/marker_experiments/paper/generated/manifest_parts/fineweb_fi_5gb_bnd_w_mingram_v34685.json
new file mode 100644
index 00000000..b8b6ec18
--- /dev/null
+++ b/marker_experiments/paper/generated/manifest_parts/fineweb_fi_5gb_bnd_w_mingram_v34685.json
@@ -0,0 +1,19 @@
+{
+  "fineweb_fi_5gb_bnd_w_mingram_v34685": {
+    "additional_vocab_size": 32974,
+    "arm": "bnd_w",
+    "atomic_vocab": 1711,
+    "corpus": "fineweb_fi_5gb",
+    "eval_chars": 983997,
+    "eval_chars_per_token": 3.0845914157816203,
+    "eval_slice": "marker_experiments/eval_texts/fi.json",
+    "eval_tokens": 319004,
+    "lang": "fi",
+    "overshoot_factor": 1.15,
+    "path": "downstream/tokenizers/fineweb_fi_5gb_bnd_w_mingram_v34685.json.gz",
+    "roundtrip_failures": 0,
+    "total_vocab": 34685,
+    "train_seconds": 1805,
+    "trainer": "mingram"
+  }
+}
\ No newline at end of file
diff --git a/marker_experiments/paper/generated/manifest_parts/fineweb_fi_5gb_bnd_wp_mingram_v34685.json b/marker_experiments/paper/generated/manifest_parts/fineweb_fi_5gb_bnd_wp_mingram_v34685.json
new file mode 100644
index 00000000..bbd6d543
--- /dev/null
+++ b/marker_experiments/paper/generated/manifest_parts/fineweb_fi_5gb_bnd_wp_mingram_v34685.json
@@ -0,0 +1,19 @@
+{
+  "fineweb_fi_5gb_bnd_wp_mingram_v34685": {
+    "additional_vocab_size": 32974,
+    "arm": "bnd_wp",
+    "atomic_vocab": 1711,
+    "corpus": "fineweb_fi_5gb",
+    "eval_chars": 983997,
+    "eval_chars_per_token": 3.402385134574424,
+    "eval_slice": "marker_experiments/eval_texts/fi.json",
+    "eval_tokens": 289208,
+    "lang": "fi",
+    "overshoot_factor": 1.15,
+    "path": "downstream/tokenizers/fineweb_fi_5gb_bnd_wp_mingram_v34685.json.gz",
+    "roundtrip_failures": 0,
+    "total_vocab": 34685,
+    "train_seconds": 1801,
+    "trainer": "mingram"
+  }
+}
\ No newline at end of file
diff --git a/marker_experiments/paper/generated/manifest_parts/fineweb_fi_5gb_bnd_wpd_mingram_v34685.json b/marker_experiments/paper/generated/manifest_parts/fineweb_fi_5gb_bnd_wpd_mingram_v34685.json
new file mode 100644
index 00000000..e6972419
--- /dev/null
+++ b/marker_experiments/paper/generated/manifest_parts/fineweb_fi_5gb_bnd_wpd_mingram_v34685.json
@@ -0,0 +1,19 @@
+{
+  "fineweb_fi_5gb_bnd_wpd_mingram_v34685": {
+    "additional_vocab_size": 32974,
+    "arm": "bnd_wpd",
+    "atomic_vocab": 1711,
+    "corpus": "fineweb_fi_5gb",
+    "eval_chars": 983997,
+    "eval_chars_per_token": 3.584453478460429,
+    "eval_slice": "marker_experiments/eval_texts/fi.json",
+    "eval_tokens": 274518,
+    "lang": "fi",
+    "overshoot_factor": 1.15,
+    "path": "downstream/tokenizers/fineweb_fi_5gb_bnd_wpd_mingram_v34685.json.gz",
+    "roundtrip_failures": 0,
+    "total_vocab": 34685,
+    "train_seconds": 1795,
+    "trainer": "mingram"
+  }
+}
\ No newline at end of file
diff --git a/marker_experiments/paper/generated/manifest_parts/fineweb_fi_5gb_plain_mingram_v34685.json b/marker_experiments/paper/generated/manifest_parts/fineweb_fi_5gb_plain_mingram_v34685.json
new file mode 100644
index 00000000..b771e761
--- /dev/null
+++ b/marker_experiments/paper/generated/manifest_parts/fineweb_fi_5gb_plain_mingram_v34685.json
@@ -0,0 +1,19 @@
+{
+  "fineweb_fi_5gb_plain_mingram_v34685": {
+    "additional_vocab_size": 32975,
+    "arm": "plain",
+    "atomic_vocab": 1710,
+    "corpus": "fineweb_fi_5gb",
+    "eval_chars": 983997,
+    "eval_chars_per_token": 3.622163734079364,
+    "eval_slice": "marker_experiments/eval_texts/fi.json",
+    "eval_tokens": 271660,
+    "lang": "fi",
+    "overshoot_factor": 1.15,
+    "path": "downstream/tokenizers/fineweb_fi_5gb_plain_mingram_v34685.json.gz",
+    "roundtrip_failures": 0,
+    "total_vocab": 34685,
+    "train_seconds": 1944,
+    "trainer": "mingram"
+  }
+}
\ No newline at end of file
diff --git a/marker_experiments/paper/generated/manifest_parts/fineweb_ru_5gb_bnd_w_mingram_v34685.json b/marker_experiments/paper/generated/manifest_parts/fineweb_ru_5gb_bnd_w_mingram_v34685.json
new file mode 100644
index 00000000..a429595f
--- /dev/null
+++ b/marker_experiments/paper/generated/manifest_parts/fineweb_ru_5gb_bnd_w_mingram_v34685.json
@@ -0,0 +1,19 @@
+{
+  "fineweb_ru_5gb_bnd_w_mingram_v34685": {
+    "additional_vocab_size": 32974,
+    "arm": "bnd_w",
+    "atomic_vocab": 1711,
+    "corpus": "fineweb_ru_5gb",
+    "eval_chars": 1894911,
+    "eval_chars_per_token": 3.00272397094431,
+    "eval_slice": "marker_experiments/eval_texts/ru.json",
+    "eval_tokens": 631064,
+    "lang": "ru",
+    "overshoot_factor": 1.15,
+    "path": "downstream/tokenizers/fineweb_ru_5gb_bnd_w_mingram_v34685.json.gz",
+    "roundtrip_failures": 0,
+    "total_vocab": 34685,
+    "train_seconds": 771,
+    "trainer": "mingram"
+  }
+}
\ No newline at end of file
diff --git a/marker_experiments/paper/generated/manifest_parts/fineweb_ru_5gb_bnd_wp_mingram_v34685.json b/marker_experiments/paper/generated/manifest_parts/fineweb_ru_5gb_bnd_wp_mingram_v34685.json
new file mode 100644
index 00000000..c8c74ec9
--- /dev/null
+++ b/marker_experiments/paper/generated/manifest_parts/fineweb_ru_5gb_bnd_wp_mingram_v34685.json
@@ -0,0 +1,19 @@
+{
+  "fineweb_ru_5gb_bnd_wp_mingram_v34685": {
+    "additional_vocab_size": 32974,
+    "arm": "bnd_wp",
+    "atomic_vocab": 1711,
+    "corpus": "fineweb_ru_5gb",
+    "eval_chars": 1894911,
+    "eval_chars_per_token": 3.345446360001059,
+    "eval_slice": "marker_experiments/eval_texts/ru.json",
+    "eval_tokens": 566415,
+    "lang": "ru",
+    "overshoot_factor": 1.15,
+    "path": "downstream/tokenizers/fineweb_ru_5gb_bnd_wp_mingram_v34685.json.gz",
+    "roundtrip_failures": 0,
+    "total_vocab": 34685,
+    "train_seconds": 772,
+    "trainer": "mingram"
+  }
+}
\ No newline at end of file
diff --git a/marker_experiments/paper/generated/manifest_parts/fineweb_ru_5gb_bnd_wpd_mingram_v34685.json b/marker_experiments/paper/generated/manifest_parts/fineweb_ru_5gb_bnd_wpd_mingram_v34685.json
new file mode 100644
index 00000000..02c0f4a0
--- /dev/null
+++ b/marker_experiments/paper/generated/manifest_parts/fineweb_ru_5gb_bnd_wpd_mingram_v34685.json
@@ -0,0 +1,19 @@
+{
+  "fineweb_ru_5gb_bnd_wpd_mingram_v34685": {
+    "additional_vocab_size": 32974,
+    "arm": "bnd_wpd",
+    "atomic_vocab": 1711,
+    "corpus": "fineweb_ru_5gb",
+    "eval_chars": 1894911,
+    "eval_chars_per_token": 3.5068864048562016,
+    "eval_slice": "marker_experiments/eval_texts/ru.json",
+    "eval_tokens": 540340,
+    "lang": "ru",
+    "overshoot_factor": 1.15,
+    "path": "downstream/tokenizers/fineweb_ru_5gb_bnd_wpd_mingram_v34685.json.gz",
+    "roundtrip_failures": 0,
+    "total_vocab": 34685,
+    "train_seconds": 780,
+    "trainer": "mingram"
+  }
+}
\ No newline at end of file
diff --git a/marker_experiments/paper/generated/manifest_parts/fineweb_ru_5gb_plain_mingram_v34685.json b/marker_experiments/paper/generated/manifest_parts/fineweb_ru_5gb_plain_mingram_v34685.json
new file mode 100644
index 00000000..8a9bc7a5
--- /dev/null
+++ b/marker_experiments/paper/generated/manifest_parts/fineweb_ru_5gb_plain_mingram_v34685.json
@@ -0,0 +1,19 @@
+{
+  "fineweb_ru_5gb_plain_mingram_v34685": {
+    "additional_vocab_size": 32975,
+    "arm": "plain",
+    "atomic_vocab": 1710,
+    "corpus": "fineweb_ru_5gb",
+    "eval_chars": 1894911,
+    "eval_chars_per_token": 3.4422135108230454,
+    "eval_slice": "marker_experiments/eval_texts/ru.json",
+    "eval_tokens": 550492,
+    "lang": "ru",
+    "overshoot_factor": 1.15,
+    "path": "downstream/tokenizers/fineweb_ru_5gb_plain_mingram_v34685.json.gz",
+    "roundtrip_failures": 0,
+    "total_vocab": 34684,
+    "train_seconds": 836,
+    "trainer": "mingram"
+  }
+}
\ No newline at end of file
diff --git a/marker_experiments/paper/generated/results_extcaps.tsv b/marker_experiments/paper/generated/results_extcaps.tsv
new file mode 100644
index 00000000..febccd8d
--- /dev/null
+++ b/marker_experiments/paper/generated/results_extcaps.tsv
@@ -0,0 +1,4 @@
+method	arm	variant	trainer	seed	val_bpb_true	train_bpb_true	val_bpb	train_bpb	byte_factor	depth	vocab_size	tokenizer_id	log
+bnd_wpd_extcaps_bpe	bnd_wpd_extcaps		bpe	0	0.879587190979435	0.7576745998851696	1.04396	0.899265	0.8425487480166243	12	34686	bnd_wpd_extcaps_bpe_d12_s0	bnd_wpd_extcaps_bpe_d12_s0.log
+bnd_wpd_extcaps_bpe	bnd_wpd_extcaps		bpe	1	0.8787623357551267	0.7547753896432444	1.042981	0.895824	0.8425487480166243	12	34686	bnd_wpd_extcaps_bpe_d12_s1	bnd_wpd_extcaps_bpe_d12_s1.log
+bnd_wpd_extcaps_bpe	bnd_wpd_extcaps		bpe	2	0.8795197870795937	0.7582289969613646	1.04388	0.899923	0.8425487480166243	12	34686	bnd_wpd_extcaps_bpe_d12_s2	bnd_wpd_extcaps_bpe_d12_s2.log
diff --git a/marker_experiments/paper/generated/table_intrinsic.tex b/marker_experiments/paper/generated/table_intrinsic.tex
index ed2f231f..288ec25a 100644
--- a/marker_experiments/paper/generated/table_intrinsic.tex
+++ b/marker_experiments/paper/generated/table_intrinsic.tex
@@ -11,17 +11,19 @@
  & train & eval & train & eval & train & eval & train & eval & train & eval & train & eval & train & eval \\
 \midrule
 \multicolumn{15}{l}{\emph{BPE}} \\
-\plainscheme & -- & 4.457 & -- & \textbf{4.574} & -- & \textbf{5.012} & -- & 4.322 & -- & \textbf{3.973} & \textbf{2.288} & \textbf{2.299} & -- & \textbf{4.106} \\
+\plainscheme & -- & 4.457 & -- & \textbf{4.574} & -- & \textbf{5.012} & -- & 4.322 & -- & \textbf{3.973} & \textbf{2.288} & \textbf{2.299} & -- & \underline{4.106} \\
 \bnds{w} & $-8.29$ & $-8.87$ & $-9.73$ & $-10.68$ & $-7.76$ & $-7.74$ & $-10.03$ & $-10.87$ & $-8.01$ & $-8.49$ & $-8.15$ & $-9.12$ & $-8.66$ & $-9.30$ \\
 \bnds{wp} & $-0.94$ & $-1.30$ & $-2.43$ & $-2.97$ & $-1.70$ & $-1.67$ & $-0.91$ & $-1.22$ & $\underline{-1.23}$ & $-1.91$ & $-3.39$ & $-4.00$ & $-1.77$ & $-2.18$ \\
-\bnds{wpd} & $\underline{+0.80}$ & $\underline{+0.62}$ & $\underline{-0.63}$ & $-1.03$ & $\underline{-0.21}$ & $-0.47$ & $\underline{+0.79}$ & $\underline{+0.52}$ & $\mathbf{+0.37}$ & $\underline{-0.21}$ & $\underline{-1.63}$ & $\underline{-2.18}$ & $\mathbf{-0.08}$ & $\underline{-0.46}$ \\
-\bnds{wpdcapsin} & $\mathbf{+0.96}$ & $\mathbf{+0.72}$ & $\mathbf{-0.39}$ & $\underline{-0.78}$ & $\mathbf{+0.06}$ & $\underline{-0.35}$ & $\mathbf{+1.09}$ & $\mathbf{+0.92}$ & $-2.44$ & $-2.52$ & $\underline{-1.63}$ & $-2.22$ & $\underline{-0.39}$ & $-0.71$ \\
+\bnds{wpd} & $+0.80$ & $+0.62$ & $\underline{-0.63}$ & $-1.03$ & $\underline{-0.21}$ & $-0.47$ & $\underline{+0.79}$ & $\underline{+0.52}$ & $\mathbf{+0.37}$ & $\underline{-0.21}$ & $\underline{-1.63}$ & $\underline{-2.18}$ & $\underline{-0.08}$ & $-0.46$ \\
+\bnds{wpdcaps} & $\mathbf{+1.01}$ & $\mathbf{+0.75}$ & -- & -- & -- & -- & -- & -- & -- & -- & -- & -- & $\mathbf{+1.01}$ & $\mathbf{+0.75}$ \\
+\bnds{wpdcapsin} & $\underline{+0.96}$ & $\underline{+0.72}$ & $\mathbf{-0.39}$ & $\underline{-0.78}$ & $\mathbf{+0.06}$ & $\underline{-0.35}$ & $\mathbf{+1.09}$ & $\mathbf{+0.92}$ & $-2.44$ & $-2.52$ & $\underline{-1.63}$ & $-2.22$ & $-0.39$ & $-0.71$ \\
 \midrule
 \multicolumn{15}{l}{\emph{MinGram}} \\
-\plainscheme & -- & \underline{4.500} & -- & -- & -- & -- & -- & -- & -- & -- & \textbf{2.300} & \textbf{2.310} & -- & \textbf{3.405} \\
-\bnds{w} & -- & -- & -- & -- & -- & -- & -- & -- & -- & -- & $-8.36$ & $-9.33$ & $-8.36$ & $-9.33$ \\
-\bnds{wp} & -- & -- & -- & -- & -- & -- & -- & -- & -- & -- & $-3.59$ & $-4.17$ & $-3.59$ & $-4.17$ \\
-\bnds{wpd} & $\mathbf{+0.60}$ & $\mathbf{+0.35}$ & -- & -- & -- & -- & -- & -- & -- & -- & $-1.83$ & $\underline{-2.35}$ & $\mathbf{-0.61}$ & $\underline{-1.00}$ \\
+\plainscheme & -- & \underline{4.500} & \textbf{4.641} & \textbf{4.665} & \textbf{4.809} & \textbf{5.153} & \underline{4.389} & \underline{4.374} & \textbf{4.135} & \textbf{4.024} & \textbf{2.300} & \textbf{2.310} & -- & \textbf{4.171} \\
+\bnds{w} & $-8.55$ & $-9.19$ & $-10.04$ & $-10.98$ & $-9.14$ & $-9.47$ & $-10.18$ & $-11.05$ & $-8.54$ & $-9.09$ & $-8.36$ & $-9.33$ & $-9.13$ & $-9.85$ \\
+\bnds{wp} & $\underline{-1.17}$ & $-1.59$ & $-2.63$ & $-3.15$ & $-3.07$ & $-3.47$ & $-0.99$ & $-1.31$ & $-1.77$ & $-2.49$ & $-3.59$ & $-4.17$ & $-2.20$ & $-2.70$ \\
+\bnds{wpd} & $\mathbf{+0.60}$ & $\mathbf{+0.35}$ & $\underline{-0.79}$ & $\underline{-1.17}$ & $\underline{-1.56}$ & $\underline{-2.27}$ & $\mathbf{+0.75}$ & $\mathbf{+0.47}$ & $\underline{-0.15}$ & $\underline{-0.79}$ & $-1.83$ & $\underline{-2.35}$ & $\mathbf{-0.50}$ & $\underline{-0.96}$ \\
+\bnds{wpdcaps} & -- & -- & -- & -- & -- & -- & -- & -- & -- & -- & -- & -- & -- & -- \\
 \bnds{wpdcapsin} & -- & -- & -- & -- & -- & -- & -- & -- & -- & -- & $\underline{-1.82}$ & $-2.38$ & $\underline{-1.82}$ & $-2.38$ \\
 \bottomrule
 \end{tabular}