Publish DroneVLM-2 release package and technical report

README.md

@@ -472,7 +472,7 @@ These automated metrics are regression signals, not proof of physical-world safe
 
 The methods report explains how detector annotations became image-question-JSON supervision, why the four response envelopes exist, how observed model failures became preference data, and how the evidence gate prevents partial evaluations from becoming release claims:
 
-**[Miril-DroneVLM-2B-2: Compiling Detector Annotations into a Typed Aerial VLM](./technical_report.pdf)**
+**[Miril-DroneVLM-2B-2: Turning Aerial Detector Labels into a Structured Vision-Language Interface](./technical_report.pdf)**
 
 The report treats data design as the central engineering contribution and clearly separates verified dataset properties from model-quality and operational-safety claims.
 
@@ -480,7 +480,7 @@ The report treats data design as the central engineering contribution and clearl
 
 ```bibtex
 @techreport{sturges2026mirildrone2b2,
-  title       = {Miril-DroneVLM-2B-2: Compiling Detector Annotations into a Typed Aerial VLM},
+  title       = {Miril-DroneVLM-2B-2: Turning Aerial Detector Labels into a Structured Vision-Language Interface},
   author      = {Sturges, Stephan},
   institution = {Miril.ai},
   year        = {2026},
@@ -504,7 +504,15 @@ Treat those measured deltas as authoritative. The CUDA 8-bit and 4-bit
 packages use different bitsandbytes algorithms, so bit width alone does not
 rank their multimodal fidelity.
 
-The MLX 4-bit artifact remains available as a measured compact variant, not as an equivalent replacement for MLX 8-bit. On the complete balanced independent benchmark it preserves valid JSON at 100.0% after rounding but drops to 55.8% exact schema validity and 56.9% route accuracy, predominantly because pointing responses use `intent` where the contract requires `action`. The variant card contains the route-level evidence, operational warning, and the mixed 4/8-bit experiment planned for generation three.
+The MLX 4-bit artifact remains available as a measured compact variant, not as an equivalent replacement for MLX 8-bit. On the complete balanced independent benchmark it preserves valid JSON at 100.0% after rounding but drops to 55.8% exact schema validity and 56.9% route accuracy, predominantly because pointing responses use `intent` where the contract requires `action`.
+
+The export tensor audit passes. All expected key/value-owner and vision `k_norm` tensors are present; later language layers intentionally reuse shared key/value states. The leading hypothesis is therefore precision loss from uniform affine 4-bit conversion of token embeddings and linear modules, not a missing tensor. The variant card contains the full evidence and operational warning.
+
+## Next Experiments
+
+- **Mixed-bit MLX:** keep token embeddings, the language head, and selected projections at 8-bit, then require recovery of the `action` key, exact route, and exact schema on the same benchmark cases.
+- **Gemma 4 E4B:** repeat the four-schema training and evaluation protocol at the larger effective model scale while holding the text corpus, grouped splits, and release gates fixed.
+- **Paired spoken input:** replace a deterministic subset of typed questions with ElevenLabs-generated speech while preserving the image and JSON target. Split-isolated synthesis voices, varied voice settings, local acoustic augmentation, held-out voices, and later human recordings will test whether audio training closes the observed spoken-input gap without degrading typed requests.
 
 ## WALDO Lineage
 

technical_report.pdf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2314e40d060b4386ad1d396a2bf92e313167bbf1250cdd81bd829a2c8d620848
-size 1266065
+oid sha256:e4b3e52dfefae8f76c1301a19d877f13262b7aa1d3db8d5b29fde7be1e0d0d12
+size 1212612