Upload 2 files

README.md

@@ -52,7 +52,7 @@ Run it:
 llama serve -m ./voiced/model.gguf
 ```
 
-> **We tried it on Orion (n=10 heuristic, not a guarantee):** We cast [Gryphe/Gemma-4-26B-A4B-StyleTune-V2](https://huggingface.co/Gryphe/Gemma-4-26B-A4B-StyleTune-V2) voice onto [TheDrummer/Orion-26B-A4B-v1.4](https://huggingface.co/TheDrummer/Orion-26B-A4B-v1.4) — on our 10 prompts, phrasing was ~98% different (1.6% shared at `1.0`, 3.2% at greedy via our 50-phrase heuristic). Same cast on its abliterated sibling [llmfan46/gemma-4-26B-A4B-it-qat-q4_0-unquantized-uncensored-heretic](https://huggingface.co/llmfan46/gemma-4-26B-A4B-it-qat-q4_0-unquantized-uncensored-heretic) looped; re-cast as `delta` stopped looping and kept the voice. We got StyleTune on the QAT version with just `voice cast delta` — one file out, same family/shape required.
+> **We tried it on Orion:** full comparison (314 benchmark items, 32 writing pairs) in [Evaluation](#evaluation). Same cast on its abliterated sibling [llmfan46/gemma-4-26B-A4B-it-qat-q4_0-unquantized-uncensored-heretic](https://huggingface.co/llmfan46/gemma-4-26B-A4B-it-qat-q4_0-unquantized-uncensored-heretic) looped; re-cast as `delta` stopped looping and kept the voice — one file out, same family/shape required.
 
 No adapter at runtime. `voice cast` bakes the voice into a complete `.gguf` — the head at `Q8_0`, everything else untouched. Just `llama serve -m`.
 

voice.py

@@ -3430,7 +3430,7 @@ def _compute_delta(voice_path, base_path, delta_folder, delta_name, args=None):
         header = {
             vname: {"dtype": "F16", "shape": list(delta_f16.shape), "data_offsets": [0, total_elems * 2]},
             "delta.voice.marker": {"dtype": "U8", "shape": [1], "data_offsets": [total_elems * 2, total_elems * 2 + 1]},
-            "__metadata__": {"type": "delta", "base": str(base_path), "voice": str(voice_path)},
+            "__metadata__": {"type": "delta", "base": _public_source(str(base_path)), "voice": _public_source(str(voice_path))},
         }
         hj = json.dumps(header).encode("utf-8")
         with open(tmp_out, "wb") as out:
@@ -3497,7 +3497,7 @@ def _compute_delta(voice_path, base_path, delta_folder, delta_name, args=None):
     header = {
         vname: {"dtype": "F16", "shape": list(vshape), "data_offsets": [0, total_elems * 2]},
         marker_name: {"dtype": "U8", "shape": [1], "data_offsets": [total_elems * 2, total_elems * 2 + 1]},
-        "__metadata__": {"type": "delta", "base": str(base_path), "voice": str(voice_path)},
+        "__metadata__": {"type": "delta", "base": _public_source(str(base_path)), "voice": _public_source(str(voice_path))},
     }
     hj = json.dumps(header).encode("utf-8")
 
@@ -3769,6 +3769,42 @@ def _list_bases():
 #  CMD: info
 # ---------------------------------------------------------------------------
 
+
+def _verify_voice_file(path, meta):
+    """Cross-check voice.json claims against the artifact's own header.
+    Returns warning strings (empty = agrees). Never raises: an unreadable
+    file is a note, not a failure."""
+    out = []
+    if str(path).endswith(".gguf"):
+        return out  # quant type names differ; nothing truthful to compare
+    try:
+        header, tens = read_safetensors(str(path))
+    except Exception:
+        return ["  ? Could not read the artifact header; showing voice.json as-is."]
+    names = [k for k in tens if k != "delta.voice.marker"]
+    tname = meta.get("tensor_name")
+    if tname not in tens and names:
+        tname = names[0]
+    if tname not in tens:
+        return ["  ? No tensors found in the artifact."]
+    info = tens[tname]
+    claimed_shape = meta.get("shape")
+    if claimed_shape is not None and list(claimed_shape) != list(info["shape"]):
+        out.append(f"  ! MISMATCH shape: voice.json says {list(claimed_shape)}, "
+                   f"file holds {list(info['shape'])} — file wins.")
+    claimed_dtype = (meta.get("dtype") or "").upper()
+    if claimed_dtype and claimed_dtype != str(info["dtype"]).upper():
+        out.append(f"  ! MISMATCH dtype: voice.json says {meta.get('dtype')}, "
+                   f"file holds {info['dtype']} — file wins.")
+    has_marker = "delta.voice.marker" in tens
+    if meta.get("type") == "delta" and not has_marker:
+        out.append("  ! voice.json claims delta but the file has no delta marker.")
+    if meta.get("type") != "delta" and has_marker:
+        out.append("  ! File carries a delta marker but voice.json type is "
+                   f"'{meta.get('type', 'voice')}'.")
+    return out
+
+
 def cmd_info(args):
     voicename = args.voicename.strip()
     voice_path, voice_meta = _registry_resolve(voicename)
@@ -3807,6 +3843,8 @@ def cmd_info(args):
         _say(f"  License: {_trunc(meta['license'], 80)}")
     _say(f"  Saved:   {meta.get('downloaded_at', meta.get('fetched_at', '—'))}")
     _say(f"  File:    {voice_path}")
+    for w in _verify_voice_file(str(voice_path), meta):
+        _warn(w)
 
 
 # ---------------------------------------------------------------------------