userdevice-rtus-0.1.0 — first public weights release (stage-D2, best-by-DISTS iter50000)

README.md

@@ -126,10 +126,28 @@ Those terms bind the dataset; this release distributes a model and
 redistributes no images. Full record in
 [`docs/PROVENANCE.md`](https://github.com/ltomes/userdevice-rtus/blob/main/docs/PROVENANCE.md).
 
-## How it was evaluated
+## How it was evaluated — and what this release scored
 
 A checkpoint is accepted by passing four legs on identical frames, not
-because a number went up:
+because a number went up. **This release passed all four**, measured
+2026-09-05 on the project's validation split:
+
+| Leg | Result for 0.1.0 | Verdict |
+|---|---|---|
+| Sweep (DISTS) | iter050000 at 0.141940, best of eleven | selected |
+| Face gate | 0 violations in 60 images | **PASS** |
+| Temporal | 5.9147 vs bicubic 6.0240 (−1.8%) | **PASS** — steadier than bicubic |
+| Invention probe | 6.14% vs teacher 8.50% | **PASS** — invents less than the teacher |
+
+Read the invention rate against bicubic rather than as an absolute: bicubic
+cannot invent detail, so its 2.14% is the detector's false-positive floor.
+The kill condition is exceeding the teacher, and this release sits below it.
+
+The sweep that made the selection ships with the weights as
+`userdevice-rtus-0.1.0-sweep.json`, so the choice of iteration 50000 over
+the other ten is checkable rather than asserted.
+
+What the four legs are:
 
 1. **Sweep** — PSNR/SSIM/DISTS/LPIPS across every checkpoint against
    bicubic and the teacher, selecting **best-by-DISTS** rather than the last

SHA256SUMS

@@ -1,6 +1,8 @@
 cda73933e89138ca243685b3812316adf23ea10ff4dc8def416b11a96d128513  ./userdevice-rtus-0.1.0-1080p.onnx
 e0f6423ba8c0cdd26d950d38de3136a70eb661696a5e70d748174b9dd88ad35a  ./userdevice-rtus-0.1.0-540p.onnx
 66c20ecc1ab290cf2c50a095c8cc03393c2e6bb66ad7d1a4bedbec831d93978a  ./userdevice-rtus-0.1.0-720p.onnx
+74f1815b6f4dce584901c484e8d6303213a7d828d7c3f06403554a6fa4d30222  ./userdevice-rtus-0.1.0-invention.json
 6f6ef08965905c1a14b3ae173f9f3df015c393764fdb5966e774191b5e7ab8ed  ./userdevice-rtus-0.1.0-parity.npz
 05bd9b0d76318f759a3a2b6af04d682801242a7854ef6cf55577bee9fcf97e61  ./userdevice-rtus-0.1.0-sweep.json
+c2db265eaf25e5aecf6196383c89681a32518c35f5b038b4e0a78ce554c1d161  ./userdevice-rtus-0.1.0-temporal.json
 b13c380e91337aceb05ea35a3cb2aae6ec208d1e5fe8455d4400ec330434311f  ./userdevice-rtus-0.1.0.safetensors

userdevice-rtus-0.1.0-invention.json

@@ -0,0 +1,26 @@
+{
+  "bicubic": {
+    "hf_recovery": 0.42613253251124816,
+    "invention_rate": 0.021418362661037922,
+    "corr_p05": 0.3920679330825806,
+    "corr_median": 0.6276682317256927,
+    "tiles": 12652,
+    "frames": 60
+  },
+  "teacher_web": {
+    "hf_recovery": 0.7187647518301425,
+    "invention_rate": 0.08498611722050611,
+    "corr_p05": 0.27736336588859556,
+    "corr_median": 0.6105428338050842,
+    "tiles": 12652,
+    "frames": 60
+  },
+  "d2_release_0.1.0": {
+    "hf_recovery": 0.6850620948788754,
+    "invention_rate": 0.061409031315949915,
+    "corr_p05": 0.3084971964359283,
+    "corr_median": 0.6467152237892151,
+    "tiles": 12652,
+    "frames": 60
+  }
+}

userdevice-rtus-0.1.0-temporal.json

@@ -0,0 +1,5 @@
+{
+ "bicubic": 6.024002206307871,
+ "gt_control": 0.0,
+ "d2_release_0.1.0": 5.91470356949881
+}