model card: code ships in the Docker image (no public git repo); link labs page

README.md

@@ -40,11 +40,13 @@ curl -X POST localhost:8088/classify_multi_criteria -H 'content-type: applicatio
                    "someone other than the speaker is anxious"]}'
 ```
 
-**From source** — code at <https://github.com/nope-net/predicate> (Apache-2.0): download the
-artifact zip + the LoRA from this repo, then `Scorer.from_path("...zip")`. The artifact's
-canary verifies the hidden-state frame at load and **refuses to serve on a numerical
-mismatch** (e.g. a different GPU architecture — the canary here was generated on an RTX 4000
-Ada; pack/serve on the same GPU family). Built under `torch 2.6.0` / `transformers 5.9.0`.
+**From source** — the inference code is **Apache-2.0 and ships inside the Docker image**
+(`/app/predicate` + `/app/service`). To run the probe yourself, use these weights with that
+code: `Scorer.from_path("base_probe_granite_L07_armA_L28_amortized_kvpop_calibrated.zip")`
+(point the manifest's `lora_path` at the LoRA in this repo). The artifact's **canary** verifies
+the hidden-state frame at load and **refuses to serve on a numerical mismatch** — e.g. a
+different GPU architecture (the canary here was generated on an RTX 4000 Ada; pack and serve on
+the same GPU family). Built under `torch 2.6.0` / `transformers 5.9.0`.
 
 ## What it reads well
 - Topic, sentiment, intent, stance
@@ -83,4 +85,4 @@ Predicate wins universality + calibration + cost.
 Predicate is a content classifier — **not predictive, diagnostic, or therapeutic**, and **not a
 substitute for human judgment**. Scores are signals for routing, ranking, and human review.
 
-Recipe, benchmarks, and the full investigation trail: <https://github.com/nope-net/predicate>
+Project page, recipe, and benchmarks: <https://labs.nope.net/predicate>